Equity market fragmentation and liquidity: the impact of MiFID

Transcription

1 Equity market fragmentation and liquidity: the impact of MiFID (Preliminary) Hans Degryse Frank de Jong y Vincent van Kervel z January 2011 Abstract Changes in European nancial regulation (the Markets in Financial Instruments Directive) in November 2007 allow new trading venues to emerge and compete with traditional stock exchanges (regulated markets). As a result, currently, many stocks are traded on a multitude of trading platforms, i.e. the market has become fragmented. This paper evaluates the impact of fragmentation on liquidity for a sample of 52 Dutch stocks. We consider global liquidity by aggregating order book information over all European trading venues and incorporating the depth of the entire order books, and local liquidity by considering the regulated market only. Controlling for stock characteristics and time e ects, we nd that global liquidity increases with fragmentation, but these bene ts are not enjoyed by all investors as local liquidity reduces. JEL Codes: G10; G14; G15; Keywords: Market microstructure, Market fragmentation, Liquidity, MiFID z CentER, EBC, TILEC, Tilburg University. TILEC-AFM Chair on nancial market regulation. z CentER, Tilburg University z Corresponding author, TILEC, CentER, Tilburg University, [email protected]. The authors thank Mark Van Achter and Gunther Wuyts, and participants at the 2010 Erasmus liquidity conference (Rotterdam), the K.U. Leuven, the 2010 AFM workshop on MiFID (Amsterdam) and Tilburg University for helpful comments and suggestions. 1

2 I Introduction Following the developments in the U.S., European equity markets have seen a proliferation of new trading venues. While traditional stock exchanges had a near monopoly on trading until the beginning of 2007, recent changes in nancial regulation, in particular the Markets in Financial Instruments Directive (MiFID), allow new trading venues to compete for order ow. Consequently, trading has become dispersed over many trading venues, creating a fragmented market place. How fragmentation of trading a ects market quality, and liquidity in particular, has since long interested researchers, regulators, investors and trading institutions. This question needs further reconsideration as technological progress nowadays creates possibilities to connect with di erent trading venues. In this paper, we shed new light on the matter and improve upon previous research by employing a dataset that covers the relevant universe of trading platforms, provides stronger identi cation and allows for improved liquidity metrics. In Europe, the fragmentation of trading is largely a consequence of the implementation of MiFID in November MiFID s main goal is to improve market quality through two channels; by introducing competition between trading venues and by imposing similar degrees of transparency across most trading venues. First, di erent types of trading venues are allowed to compete for order ow with the traditional market. This is expected to reduce the monopoly power of the traditional market, to lower transaction costs (Biais, Martimort, and Rochet, 2000) and to foster technological innovation (Stoll, 2003). However, there is also evidence that fragmentation might reduce liquidity. When order ow becomes fragmented, the probability of nding a counterparty diminishes. Consequently, execution probabilities are lowered, which might cause some investors to leave the market or informed investors to leverage their informational advantage (e.g. Chowdhry and Nanda (1991)). Moreover, a single, consolidated market may enjoy economies of scale resulting in lower processing costs. Second, in order to create a fair level playing eld, most trading venues have to comply with similar enhanced transparency requirements. In particular, trading venues are requested to report continuously on executed transactions and on the quotes they o er. 1 When transparency enhances, information can be accessed 1 An exception of the pre-trade transparency regulation is granted to so-called dark pools, as such rules would con ict with their business model. 2

3 faster and cheaper, hence improving e ciency (e.g. Boehmer, Saar, and Yu (2005)). This paper adds to the literature by investigating the impact of fragmentation on liquidity. Foucault and Menkveld (2008) study competition between the LSE and Euronext, using data from 2004, and nd that fragmentation over these two traditional stock markets improves liquidity. O Hara and Ye (2009) nd that fragmentation lowers transaction costs and increases execution speeds for NYSE and NASDAQ stocks. In this paper, we investigate the impact of competition by new trading venues on market liquidity for a sample of European stocks in the period after the introduction of MiFID. In contrast to the U.S., the impact in Europe may be a ected by di erent regulatory requirements on executed trades. Speci cally, European brokers are required to execute trades against the best available conditions with respect to price, liquidity, transaction costs and likelihood and speed of execution. 2 A consequence of this broad de nition of execution quality is that European brokers can de ne the best-execution benchmark themselves, and may even choose to trade on one trading venue only, typically the traditional exchange. As such, price violations can occur, i.e. a trade is executed at a price worse than the best price available in the market. In the U.S. however, trading venues are obliged to execute trades against the best price available in the market (the National Best Bid or O er, NBBO). Stoll (2006) argues that this rule is controversial as forcing rms to compete primarily on prices undermines the other dimensions of execution quality. An advantage of the trade-through rule is that new entrants are guaranteed to attract order ow when they have the best price in the market, which is not necessarily the case in Europe. This seemingly small deviation in regulation can have substantial e ects on the competitive nature of new entrants (Gomber and Gsell, 2006). The current work con rms that fragmentation has substantially di erent consequences for global traders (with access to all trading venues) and local traders (with access to the regulated market only). We address the impact of fragmentation on market liquidity by creating per rm daily proxies of fragmentation and liquidity, employing information from all relevant trading venues. Speci cally, we study 52 Dutch stocks in a period before the start that fragmentation has set in, until the end of For each stock, we construct a consolidated limit order book (i.e., the limit order books of all trading venues 2 Directive 2004/39 on markets in nancial instruments [2004] O.J. L145/1. 3

4 combined) to get a complete picture of the global liquidity available in the market. Based on the consolidated order book, we analyze liquidity at the best price levels, but also deeper in the order book. This is important, as the e ect of fragmentation on liquidity may not be identical at di erent price levels in the limit order book. Next to global liquidity, we also address the impact of fragmentation on liquidity at the traditional market, which represents local liquidity. Global liquidity is relevant for global traders, i.e. traders that have access to all markets, whereas local liquidity is relevant for local traders, i.e. traders that tap the traditional market only. The panel dataset aims to identify an exogenous relation between liquidity and fragmentation by means of rm*quarter xed e ects and instrumental variables regressions. That is, the rm*quarter dummies only allow for variation in fragmentation and liquidity within a rm-quarter. Hence, they provide robustness to self selection issues, for example that competition might be higher for high volume and more liquid stocks. In addition, the rm*quarter dummies, at least partially, control for dynamic interactions between market structure, competition in the market, the degree of fragmentation and liquidity. Finally, the results based on the rm*quarter dummies are more robust to time e ects such as the nancial crisis. In the instrumental variables regressions we use the daily average order size of the venues that trade the stock as instruments for the degree of fragmentation. This approach solves self selection issues if we assume that the variation in fragmentation generated by the instruments is not related to rm size and initial level of liquidity, given the control variables already in the regression. Our main nding is that there exists an inverted U-shape in the e ect of fragmentation on global liquidity, where the optimal level of fragmentation employing our most conservative estimates improves depth with approximately 32% compared with a completely concentrated market. Interestingly, this improvement holds for liquidity close to the midpoint, i.e. at relatively good price levels, but to a much lesser extent for liquidity deeper in the order book, which improves by only 12%. This result suggests that new entrants mainly improve liquidity close to the midpoint, but do not provide much liquidity deeper in the order book. While global liquidity bene ts from fragmentation, the traditional stock exchange is worse o as local liquidity at good price levels reduces by approximately 10%. This is in contrast to the result of Weston (2002), who nds that the introduction of ECNs improves liquidity on the primary exchange. Finally, competition between trading venues is ercer for large stocks then 4

5 for small stocks, as large stocks are more fragmented and bene t twice as much from fragmentation. In addition, the negative e ect of fragmentation on local liquidity is only experienced by the small stocks. These results suggest that local traders who only trade at the traditional stock market, i.e. do not use Smart Order Routing Technology, can be worse o in a fragmented market, especially for relatively small orders. This conclusion adds to the policy debate on the bene ts and drawbacks of stock market fragmentation, where fragmentation appears to have a bene cial e ect on total market quality, but is not equally enjoyed by all stock market participants. The remainder of this paper is structured as follows. Section II describes the European nancial market after the introduction of MiFID. Section III discusses related literature and challenges in estimating fragmentation and liquidity. The dataset and liquidity measures are described in sections IV and V, while section VI explains the methodology and results. Finally, the conclusion is presented in section VII. II Background on European nancial markets after MiFID This section gives a brief discussion on the contents of the Markets in Financial Instruments Directive (MiFID), e ective November 1, By implementing a single legislation for the European Economic Area, MiFID aims to create a level playing eld for trading venues and investors, which would ultimately improve market quality. The regulation entails three major changes to achieve this goal. First, competition between trading venues is introduced by abolishing the concentration rule 3 and allowing three types of trading systems to compete for order ow. These are regulated markets (RMs), Multilateral Trading Facilities (MTFs) and Systematic Internalisers (SIs). RMs are the traditional exchanges, matching buyers and sellers through an order book or through dealers. A rm chooses on which RM 3 The concentration rule, adopted by some EU members, obliges transactions to be executed at the primary market as opposed to internal settlement. This creates a single and fair market on which all investors post their trades, according to a time and price priority. The repeal of the rule however allows markets to become fragmented and increases competition between di erent trading venues (Ferrarini and Recine, 2006). 5

6 to list, and once listed, MTFs may decide to organize trading in that rm as well. MTFs are similar in matching third party investors, but have di erent regulatory requirements and rules of the game. For example, MTFs and RMs can decide upon the type of orders that can be placed, and the structure of fees, i.e. variable fees as well as make or take fees. 4 xed fees, Typically, MTFs receive order ow from their owners or founders, generating some liquidity. But in order to survive, they should attract order ow from outside investors as well. In terms of traded volume, the most important MTFs with visible liquidity are Chi-X, Bats Europe, NASDAQ OMX and Turquoise. While SIX Swiss exchange is a regulated market, it took over the MTF Virt-X in March 2008, which traded Eurostoxx 50 constituents. Lastly, SIs are organized by investment banks where customers trade against the inventory of the SI or with other clients (Davies, Dufour, and Scott-Quinn, 2006). MiFID allows for new trading platforms and MiFID has at least partially succeeded here. For example, market shares of the primary exchanges in Eurostoxx 50 stocks have reduced from 100% in January 2008 to 78% in December 2009 (Gomber and Pierron, 2010). MiFIDs second keystone refers to transparency and guarantees the ow of information in the market. As the number of trading venues increases, information about available prices and quantities in the order books becomes dispersed. However, in order for investors to decide on the optimal venue and to evaluate order execution, a su cient degree of pre-trade and post-trade transparency is required. In particular, to allow for comparisons between trading venues, pre-trade transparency rules are set in, where trading venues are required to make (part of) their order books public and to continuously update this information. A number of waivers exist regarding pre-trade transparency. In particular, there is the large-in-scale orders waiver, the reference price waiver, the negiotated-trade waiver, and the order management facility waiver. These waivers are used for example by MTFs with completely hidden liquidity, known as dark pools, who only have to report executed trades. Also, SIs do not have to report their quotes for illiquid stocks. 5 4 Make and take fees are costs charged to investors supplying and removing liquidity, respectively. Make fees can be negative, such that providers of liquidity receive a rebate for o ering liquidity. 5 A liquid stock is de ned by the European Commission as having a free oat market capitalization exceeding e500 million and an average daily trading activity of more than 500 trades Ferrarini and Recine (2006). The stock has to be listed on a regulated markets as well. 6

7 The third and nal pillar of MiFID is the introduction of the best-execution rule, which obliges investment rms to execute orders against the best available conditions with respect to price, liquidity, transaction costs and likelihood and speed of execution (Aubry and McKee, 2007). However, such a broad de nition of best-execution policy allows investment rms to decide themselves where to route their orders to. For example, an investment rm may stipulate an execution policy that compares prices with one market only. In absence of a clear benchmark, it becomes di cult for an investor to evaluate the quality of executed trades and the overall performance of an investment rm (Gomber and Gsell, 2006). This is the main di erence between MiFID and its U.S. counterparty, Reg NMS, which only focusses on execution prices. 6 For an extensive summary of the implementation process of MiFID we refer the interested reader to Ferrarini and Recine (2006). III Fragmentation and market quality A Related literature The following subsections summarize the literature on order ow fragmentation and competition. There is a trade-o between a single consolidated market and a fragmented market. A single market bene ts from lower costs, compared with a fragmented market. These consist of the xed costs to set up a new trading venue; xed costs for clearing and settlement; costs of monitoring several trading venues simultaneously; and advanced technological infrastructure to aggregate dispersed information in the market and connect to several trading venues. Also, a single market that is already liquid will attract even more liquidity due to positive network externalities (e.g. Pagano (1989a), Pagano (1989b) and Admati, Amihud, and P eiderer (1991)). Each additional trader reduces the stock s execution risk for other potential traders, attracting more traders. This positive feedback should cause all trades to be executed at a single market, obtaining the highest degree of liquidity. However, while network externalities are still relevant, nowadays they may be realized even when several trading venues coexist. This happens to the extent that the 6 In the U.S., the price of every trade is reported to the consolidated tape, such that the performance of a broker can clearly be evaluated. 7

8 technological infrastructure seamlessly links the individual trading venues, creating e ectively one market. From a broker s point of view, the market is then virtually not fragmented, which alleviates the drawbacks of fragmentation (Stoll, 2006). In addition, fragmentation might also enhance market quality, as increased competition among liquidity suppliers forces them to improve their prices, narrowing the bid-ask spreads (e.g. Biais, Martimort, and Rochet (2000) and Battalio (1997)). Con rming a competition e ect, Conrad, Johnson, and Wahal (2003) nd that Alternative Trading Systems in general have lower execution costs compared with traditional brokers. Furthermore, Biais, Bisiere, and Spatt (2010) investigate the competition induced by ECN activity on NASDAQ stocks. They nd that ECNs with smaller tick sizes, tend to undercut the NASDAQ quotes and reduce overall quoted spreads. Di erences between trading venues may arise to cater to the needs of heterogeneous clientele, as investors di er in their preferences for trading speed, order sizes, anonymity and likelihood of execution (Petrella, 2009; Harris, 1993). In the U.S., Boehmer (2005) stresses the trade-o between speed of execution and execution costs on NASDAQ and NYSE, where NASDAQ is more expensive but also faster. In order to attract more investors, new trading venues may apply aggressive pricing schedules, such as make and take fees. Foucault, Kadan, and Kandel (2009) model that asymmetric make and take fees (i.e. liquidity takers pay a higher take fee than liquidity makers receive as make fee) enhance trading activity. The fact that some investors prefer a particular trading venue can also lead to varying degrees of informed trading at each exchange. For instance, the NYSE has been found to attract more informed order ow than the regional dealers (Easley, Kiefer, and O Hara, 1996) and NAS- DAQ market makers (Bessembinder and Kaufman (1997) and A eck-graves, Hedge, and Miller (1994)). Furthermore, Barclay, Hendershott, and McCormick (2003) nd that ECNs attract more informed order ow than NASDAQ market makers, as ECN trades have a larger price impact. Stoll (2003) argues that competition fosters innovation and e ciency, but priority rules may not be maintained. Speci cally, time priority is often violated in fragmented markets, and sometimes also price priority. 7 Foucault and Menkveld (2008) study the competition between an order book run by the LSE (EuroSETS) and Euronext Ams- 7 Time priority is violated when two limit orders with the same price are placed on two venues and the order placed last is executed rst. Price priority is violated, i.e. a trade-through, when an order gets executed against a price worse than the best quoted price in the market. A partial trade-through means that only part of the order could have been executed against a better price. 8

9 terdam for AEX rms in 2004, and nd a trade-through rate of 73%. They call for a prohibition of trade-throughs as it discourages liquidity provision. More recently, this has been studied by Ende, Gomber, and Lutat (2010) as well, who combine the order books of ten trading venues for Eurostoxx 50 stocks in December 2007 and January They nd that 6.7% of all transactions are full trade-throughs and an additional 6.5% are partial trade-throughs. The authors provide two explanations for this result. First, brokers might economize on monitoring costs and discard information on available prices in some trading venues, i.e. do not use Smart Order Routing Technology. Second, trading charges and other fees such as clearing and settlement can either make it more expensive to trade at the market o ering higher visible liquidity, or make it expensive to split up an order among several venues (Degryse, van Achter, and Wuyts, 2010). These explicit transaction costs as well as make and take fees are not taken into account in the combined order book, while they do contribute to the investors execution costs. If the role of explicit transaction costs is large enough, investor could not have done better and in a broad sense, there is no trade-through. Finally, this paper is related to the literature on algorithmic trading, 8 i.e. the use of computer programs to manage and execute trades in electronic limit order books. Algorithmic trading has strongly increased over time, and has drastically a ected the trading environment (Hendershott and Riordan, 2009). In particular, it a ects the level of market fragmentation analyzed in our sample, as computer programs and Smart Order Routing Technology (SORT) allow investors to nd the best liquidity in the market by comparing the order books of individual venues. 9 Moreover, algorithmic trading is related to liquidity as it reduces implicit transaction costs by splitting up a large order into many smaller ones (Hendershott, Jones, and Menkveld, 2010). Programs are also used to identify deviations from the e cient share price, by quickly trading on information in the market or from other securities. Furthermore, programs may provide liquidity when quoted spreads are large, e.g. when it is pro table to do so (Hendershott and Riordan, 2009). Hasbrouck and Saar (2009) describe eeting orders, a relatively new phenomenon in Europe and the U.S., where limit orders are placed and cancelled within two seconds if they are not executed. The authors argue that eeting orders are part of an active search for liquidity and a consequence of improved technology, more hidden liquidity and fragmented markets; three elements 8 Algorithmic trading is also know as High Frequency Trading. 9 See e.g. Gomber and Gsell (2006) for a discussion on SORT and algorithmic trading in Europe. 9

10 that are strongly developing in the European trading landscape. B Challenges in analyzing the impact of fragmentation In estimating the impact of fragmentation on liquidity, several data and methodological issues have to be solved. New solutions to these problems will place the current paper within the closest related literature. First, it is crucial to correctly measure the extent of fragmentation on a stock-bystock basis. We employ a Her ndahl-hirsmann Index (HHI; the sum of the squared market shares) to measure how concentrated trading is over the di erent markets. Our measure of fragmentation then equals 1-HHI which is zero in the absence of fragmentation. We compute the HHI based on executed trades on all visible trading venues. Furthermore, we control for the aggregate trading activity on all non-visible order books such as dark pools, SIs and OTC. 10 Our measure of fragmentation is more accurate than that of O Hara and Ye (2009), where the origin of trades are classi ed as either NASDAQ, NYSE or external. The main bene ts of competition in their paper arise from the external venues, but the degree of fragmentation among those venues is unknown. We create a panel dataset containing daily indicators of fragmentation and liquidity on a stock-by-stock basis over a time period of four years. A panel dataset allows us to control for xed rm, time and rm*time e ects, and provides an accurate identi cation of the impact of fragmentation and liquidity. Such a setup improves on O Hara and Ye (2009), who do a cross sectional regression, ignoring the time series variation in liquidity and fragmentation. Moreover, this setup allows for more exibility than related papers such as Foucault and Menkveld (2008), Chlistalla and Lutat (2009) and Hengelbroc (2010), who study the introduction of a new trading venue (EuroSETS, Chi-X and Turquoise respectively). Speci cally, these articles use a binary variable, i.e. a dummy that equals one after the introduction of the new venue, to estimate the e ect of fragmentation on liquidity. In our setting however, this approach would su er from three problems. First, it assumes that the level of fragmentation on other trading venues has no confounding e ects on the liquidity of 10 This aggregate non-visible orderbook activity does not provide information on which platform the trade took place. We therefore do not include it in our fragmentation indicator as it may represent trades on one venue only or on a panoply of platforms. 10

11 the venues in consideration. Second, this approach ignores the cross sectional variation in fragmentation, where some rms are more heavily traded on the new venue than others. And third, the binary variable does not take into account that a new trading venue might need time to grow and attract order ow. Furthermore, the market as a whole might need time to adjust to a new trading equilibrium, so time series variation in fragmentation is also ignored. Fortunately, the setup of our panel dataset allows for cross sectional and time series variation and analyzes all visible trading venues. A second challenge is to use an appropriate measure of liquidity. We employ two di erent views on liquidity representing the view from a global trader or local trader, respectively. First consider the view from a global trader. We capture global liquidity by aggregating the limit order books of all individual trading venues to obtain a consolidated order book. This is important, as a global trader can use Smart Order Routing Technology (SORT) to access the liquidity of all trading venues simultaneously, so the consolidated order book re ects her real trading opportunities. Based on the consolidated order book, we can construct the traditional liquidity measures, such as the quoted, realized and e ective spreads. In addition, we improve on the work of e.g. O Hara and Ye (2009), Gresse (2010) and Hengelbroc (2010) by estimating a liquidity measure based on the entire depth of the consolidated order book. That is, our data set also includes limit orders further away from the best quotes, 11 which are used to construct a new depth measure. Analyzing depth beyond the best quotes is important, as the e ect of fragmentation might be di erent for liquidity close to the midpoint, compared with liquidity deeper in the order book. Second, we also consider liquidity for a local trader. This represents the view of a trader without access to SORT who trades on the traditional stock market. In that case, we measure liquidity only based upon the order book of Euronext. A third challenge is identi cation: is the observed relation between fragmentation and liquidity causal, or is fragmentation endogenous? Over the years, there may exist dynamic interactions between market structure, competition in the market, the degree of fragmentation and liquidity. For example, changes in market structure will a ect competition and fragmentation over time, which in turn will a ect market structure. However, long-term correlations between these forces will be absorbed by 11 The data contains the ten highest bid and lowest ask prices and their associated quantities, for each trading venue. 11

12 the aforementioned rm*quarter dummies. In addition, the rm*quarter dummies provide additional robustness to time e ects such as the nancial crisis, which a ected some stocks more than others. Furthermore, we execute an instrumental variables regression, as large and liquid stocks may endogenously self select to becoming more fragmented. As instruments we use the average daily order size of the seven visible trading venues that trade the stock. This approach assumes that, after controlling for stock characteristics and rm*quarter dummies, the average daily order size is uncorrelated to any self selection e ects. For exactly this issue, O Hara and Ye (2009) execute a Heckman correction model to account for a potential bias. However, self selection issues are arguably less severe in Europe than in the United States. That is, a new trading venue in Europe typically starts with a test phase in which only a few liquid rms are traded; but will allow trading in all stocks of a certain index simultaneously when it goes live, i.e. there is no selection bias. For example, Chi- X had a two week test phase where the ve largest AEX25 and DAX30 rms were traded, while Turquoise had a ve week test phase where ve large European stocks where traded. This is in strong contrast to the U.S., where some rms may self select to be traded on many di erent platforms. IV Market description, dataset and descriptive statistics A Market description Our dataset contains 52 Dutch stocks forming the constituents of the AEX Large and Mid cap indices. Over time, all of them are traded on several trading platforms. In broad terms, we can summarize the most important trading venues for these stocks into three groups, which we describe shortly. First, there are regulated markets (RMs), such as NYSE Euronext, LSE and Deutsche Boerse. All these markets have an opening and closing auction, and in between there is continuous and anonymous trading through the limit order book. Since Euronext merged with NYSE in April 2007, the order books in Amsterdam, Paris, Brussels and Lisbon act as a fully integrated and single market (Davies, 2008). In our sample, the LSE and Deutsche Boerse are not very important as they attract 12

13 less than 1% of total order ow. Second, there are the new MTFs with visible liquidity, such as Chi-X, Bats Europe, NASDAQ OMX and Turquoise. Chi-X started trading AEX rms in April 2007; Turquoise in August 2008 and Nasdaq OMX and Bats Europe in October Whether these MTFs will survive depends on the current level of liquidity, but also on the quality of the trading technology (e.g. the speed of execution), the number of securities traded, make and take fees and clearing and settlement costs. The third group contains MTFs with completely hidden liquidity (e.g. dark pools), SIs and the Over The Counter market. Dark pools are waived from the pre-trade transparency rules set out by the MiFID due to the nature of their business model. Most dark pools employ a limit order book with similar rules as those at Euronext for example. Other MTFs act as crossing networks, where trades are executed against the midpoint on the primary market, and do not contribute to price discovery. Gomber and Pierron (2010) report that the activity on dark pools and OTC has been fairly constant for European equities in , where they execute approximately 40% of total traded volume. We refer the interested reader to Davies (2008) for a description of some of the individual trading venues. B Dataset Our dataset covers the AEX Large and Midcap constituents from 2006 to 2009, which currently have 25 and 23 stocks respectively. An advantage of using both indices is that we are able to follow stocks that switch. We remove stocks that are in the sample for less than six months or do not have observations in 2008 and 2009, such as Getronics, Pharming, Stork, Numico and ABN Amro. Due to some leavers and joiners, our nal sample has 52 stocks. The data for the 52 AEX Large and Midcap constituents stem from the Thomson Reuters Tick History Data base. This data source covers the seven most relevant European trading venues in Dutch stocks, which have executed more than 99% of the visible order ow: Euronext, Chi-X, Deutsche Boerse, Turquoise, Bats trading, NASDAQ OMX and SIX Swiss exchange (formerly known as Virt-X). 12 We employ 12 The order books of the LSE are discarded as those stocks are denoted in pennies instead of Euros; which in essence are di erent assets. The remaining trading venues with visible liquidity 13

14 data from all these venues but collect them only during the trading hours of the continuous auction of Euronext Amsterdam, i.e. between to 17.30, Amsterdam time. Therefore, data of the opening and closing auctions at these venues are not included. 13 Each stock-venue combination is reported in a separate le and represents a single order book. Every order book contains the ten best quotes at both sides of the market, i.e. the ten highest bid and lowest ask prices and their associated quantities, summing to 40 variables per observation. 14 A new state of a stock-venue limit order book is created when a limit order arrives, gets cancelled or when a trade takes place. A trade is immediately reported and we observe its associated price and quantity, as well as an update of the order book. Price and time priority rules apply within each stock-venue order book, but not between venues. Furthermore, visible orders have time priority over hidden orders. Hidden orders are not directly observed in the dataset but are detected upon execution. Therefore, we have the same information set as traders have, i.e. the visible part of the order book in a continuous fashion. Our dataset also provides information on dark trades, i.e. trades at dark pools, SIs and Over The Counter (including trades executed over telephone). These dark trades are part of the Thomson Reuters dataset and reported by Markit Boat, a MiFID-compliant trade reporting company. Although these dark data have several issues (e.g. double reporting and missing or double corrections), we use it only as a control variable in our empirical analysis. 15 In addition, we also add the OTC and SI trades reported separately in the MiFID post trade les from Euronext, Xetra, LSE, Chi-X and Stockholm. While we have information regarding price, quantity and time of execution, we do not observe the identity of the underlying trading venue. We nevertheless can include aggregate information on dark trading in our analysis. attract extremely little order ow for the rms in our sample (e.g. NYSE, Milan stock exchange, PLUS group and some smaller MTFs). 13 Unscheduled intra-day auctions are not identi ed in our dataset. These auctions, triggered by transactions that would cause extreme price movements, act as a safety measure and typically last for a few minutes. Given that we will work with daily averages of quote-by-quote liquidity measures, we assume that these auctions will not a ect our results. 14 Part of the sample only has the best ve price levels: Euronext before January Only liquidity deep in the order book is a ected. In section B.5 we execute the analysis on a yearly basis: the results are una ected. 15 To solve these data issues, the Committee of European Securities Regulators (CESR) proposes to standardise the trade reporting rules in their technical report of October 2010 on post trade transparency. 14

15 C Descriptive statistics Figure 1 shows the evolution of the daily traded volume, aggregated over all AEX Large and Mid cap constituents. The graph shows a steady increase in total trading activity, which peaks around the beginning of Moreover, the dominance of Euronext over its challengers is strong, but slowly decreasing over time. This pattern is representative for all regulated markets trading European blue chip stocks, as analyzed in Gomber and Pierron (2010). Finally, while Chi-X started trading AEX rms in April 2007, all MTFs together have attracted signi cant order ow only as of August 2008 (4,5%). The slow start up shows that these venues need time to generate trading activity. In Table I, the characteristics of the di erent stocks and some descriptive statistics are presented. There is considerable variation in rm size (market capitalization), price and trading volume. In the sample, 40 stocks have a market capitalization exceeding one billion Euro, while the 12 remaining stocks have values above 100 million Euro. In order to estimate a stock s volatility, we divide the trading day into 34 fteen-minute periods, and calculate the stock return of each period, based on the spread midpoint at the beginning and end of that period. The standard deviation of these stock returns are daily estimates of realized volatility. 16 The table also shows the average market share of Euronext and dark trades, which covers all trades reported by Markit Boat and OTC from the regulated markets. The market shares are percentages of executed trades as of November 2007 onwards, the period for which Markit Boat data have become available in the dataset. According to our data, 25% of all trades are dark; which can be split up into 29.4% for AEX large cap rms and 18.9% for mid cap rms. V Liquidity and fragmentation A The consolidated order book The goal of this paper is to analyze the impact of equity market fragmentation on liquidity. We follow the approach of Gresse (2010) and distinguish between global 16 The use of realized volatility is well established, see e.g. Andersen, Bollerslev, Diebold, and Ebens (2001). 15

16 traders and local traders. For global traders, global liquidity applies as they employ Smart Order Routing Technology (SORT) to access all trading venues simultaneously; so liquidity should be aggregated over these venues. However, for local traders, representing small investors, SORT might be too expensive because of xed trading charges and costs of adopting this trading technology. We therefore also compute local liquidity and assume that those local traders only have access to the primary market. Empirically, SORT is not used by all investors (e.g. Ende, Gomber, and Lutat (2010) and Foucault and Menkveld (2008)). To analyze the impact of fragmentation for local and global investors, we relate fragmentation to liquidity of the traditional and aggregated market, respectively. Here, Euronext Amsterdam is the primary market and the consolidated order book the aggregated market. To construct the consolidated order book, we follow the methodology of Chlistalla and Lutat (2009) and Foucault and Menkveld (2008), based on snapshots of the limit order book. A snapshot contains the ten best bid and ask prices and associated quantities, for each stock-venue combination. Every minute we take snapshots of all venues and sum the liquidity to obtain a stock s consolidated order book. Therefore, each stock has 510 daily observations (8.5 hours times 60 minutes), containing the order books of the separate trading venues and the consolidated one. B Depth(X) liquidity measure Our rich dataset allows to construct a liquidity measure that incorporates the limit orders beyond the best price levels; which we will refer to as the Depth(X). The measure aggregates the Euro value of the number of shares o ered within a xed interval around the midpoint. Speci cally, the midpoint is the average of the best bid and ask price of the consolidated order book and the interval is an amount X = {10, 20,..., 60} basis points relative to the midpoint. 17 The measure is expressed in Euros and calculated every minute. Equation 1 shows the calculation for the bid and ask side separately, which are summed to obtain Depth(X). Depth(X) is constructed for the consolidated order book and the traditional stock market (i.e., Euronext Amsterdam) separately. De ne price level j = f1; 2; :::; Jg on the pricing grid and the midpoint of 17 Foucault and Menkveld (2008) aggregate liquidity from one up to four ticks away from the best quotes. This measure only looks at the depth dimension, but does not incorporate the price dimension, like ours. Furthermore, we use basis points as tick sizes vary over time. 16

17 the consolidated order book as M, then Depth Ask(X) = Depth Bid(X) = JX j=1 JX j=1 P Ask j P Bid j Q Ask j Q Bid j j P Ask j < M (1 + X) ; (1a) j P Bid j > M (1 X) ; (1b) Depth (X) = Depth Bid(X) + Depth Ask(X): (1c) Figure 2 gives a graphical representation of the depth measure, where liquidity between the horizontal dashed lines is aggregated to obtain Depth(20 bps) and Depth(40 bps). The Depth(X) is calculated every minute and then averaged over the trading day. This gives a proxy for a stock s liquidity on a certain day, where Depth(10 bps) represents liquidity close to the midpoint and Depth(60 bps) also includes liquidity deeper in the order book. Comparing di erent price levels (X) reveals the shape of the order book. For example, if the depth measure increases rapidly in X, the order book is deep while if it increases only slowly, the order book is relatively thin. The Depth(X) is very similar to the Exchange Liquidity Measure, XLM(V), which also analyzes liquidity deeper in the order book (used by e.g. Gomber, Schweickert, and Theissen (2004)). More speci cally, the XLM(V) xes the quantity V of a potential trade, i.e. V equals e , and analyzes the impact on price; while the Depth(X) xes the price, i.e. X equals ten basis points around the midpoint, and analyzes the available quantity. Although both measures estimate the depth and slope of the order book, our approach solves two rather technical issues. First, the impact on price cannot be calculated when a stocks order book has insu cient liquidity to trade e , such that the XLM(V) becomes missing. In contrast, if no additional shares are o ered within the range of X and X + " basis points from the midpoint, then the Depth(X) has a zero increment and Depth(X) = Depth(X + "). Second, the XLM(V) may become negative when the consolidated spread is negative, i.e. when the best ask price of Chi-X is lower than the best bid price of Euronext or vice versa. 18 While negative transaction costs cannot be interpreted meaningfully, 18 Technically, a negative consolidated spread is an arbitrage opportunity, which might not be exploited because of explicit trading costs for example. 17

18 the midpoint and Depth(X) are perfectly identi ed and re ect the available liquidity in a meaningful fashion. Figure 3 shows monthly averages of the Depth measure for 10 and 60 basis points of Euronext Amsterdam and the consolidated order book, over the period 2006 to As expected, these measures covary strongly over time. However, some shocks a ect liquidity close to the midpoint more than liquidity deep in the order book. For example, when comparing the peak of liquidity in October 2007 to the low of February 2008, the ratio of Depth(60 bps) to Depth(10 bps) doubles from 2.0 to 3.9. This variation is present at the aggregate level and is also observed for most individual rms. Finally, the upper panel of Figure 4 shows the shape of the order book by plotting the median of the depth measure against the number of basis points around the midpoint. The di erences between the medians reported here and the means of Figure 3 are very large, a factor 3 to 5, which is in line with high levels of skewness and kurtosis (not reported). Next to substantial variation over time (see Figure 3), there also appear to be large di erences between rms. For example, the 90 th percentile of Depth(10 bps) is e , while the 10 th percentile of Depth(60 bps) is e In the regression analysis we will work with the logarithm of the depth measures, where the 10, 50 and 90 th percentiles are shown in the lower panel of Figure 4. Table II contains the 25, 50 and 75 th percentile of the Depth(X) measure on a yearly basis, along with other liquidity measures discussed in the next section. C Other liquidity measures This section compares our Depth(X) liquidity measure to the more traditional liquidity measures. These are the price impact, e ective and realized spread, based on executed transactions, and the quoted spread and quoted depth, based on quotes in the local and global order books. The quoted depth sums the Euro amount of shares o ered at the best bid and ask price, whereas the quoted spread looks at the associated prices. The appendix (section VIII) contains a formal description of these measures. The 25, 50 and 75 th percentiles of the liquidity measures are reported in Table II, based on daily observations and calculated yearly, for the consolidated order 18

19 book (upper panel) and Euronext Amsterdam (lower panel). The table shows several interesting results. Liquidity close to the midpoint has reduced strongly over time, while liquidity deeper in the order book only marginally. That is, the median of Depth(10 bps) has decreased by 35% from 2006 to 2009, while Depth(60 bps) by only 12%. In addition, the yearly standard deviation of the depth measures has decreased by approximately 50% over the years (not reported). Strikingly, between 2006 and 2009 the median quoted spread has improved by 10%, while the quoted depth has worsened by 68%. This result is in strong contrast with the decrease of 35% in the Depth(10 bps) measure, and shows a shortcoming of the quoted depth and spread. That is, based on the quoted depth and spread alone, one cannot state whether an investor is better o in 2006 or 2009, as this depends on the size of his trade. For a small investor mainly the quoted spread matters, while for a larger one the depth dimension becomes more important. To interpret these numbers, one can argue that a trader who places an order smaller than the median quoted depth in 2009, e30.000, is very likely to be better o with on average 10% (an argument similarly to that made in Hendershott, Jones, and Menkveld (2010)). An additional advantage of the Depth(X) measure is that it is not sensitive to small, price improving orders. For example, the strong reduction in quoted depth over time might be explained by lowered tick sizes, 19 or by increased activity of algorithmic traders, who evidently place many small and price improving limit orders. The impact of these phenomena on the quoted spread and depth also hinges on the degree of fragmentation and may therefore vary over time, which makes comparisons between periods troublesome. Finally, the Depth(X) measure can easily be aggregated over several trading venues, while the quoted depth and spread are only aggregated if the best bid and ask prices of the trading venues in consideration are identical. Turning to the liquidity measures based on executed trades, we observe that the median realized spread has reduced from 2.46 bps in 2006 to 0.02 bps in In this period, the price impact went up with a 2.9 bps while the e ective spread reduced with 0.9 bps. Note that the price impact and realized spread together equal the e ective spread. However, when calculating percentiles, each measure is ranked separately 19 The e ect of the tick size on quoted depth and spread have been subject of analysis in several papers, e.g. Huang and Stoll (2001), Goldstein and Kavajecz (2000). 19

20 and as such they do not add up exactly. In addition, we have an unbalanced panel because of rms joining and leaving the sample. The lower panel in Table II shows the same liquidity measures for the order book of the traditional stock exchange, which represents the liquidity to a local investor. While in 2006 and 2007 the numbers are highly similar to those of the consolidated order book, in 2009 the Depth(X) on Euronext represents only about 50% of the consolidated depth. In the same year, the quoted depth and spread on Euronext Amsterdam are 20% and 9% worse compared to the consolidated order book. Despite the reduction in Depth(X), the price impact, realized and e ective spread are almost identical to that of the consolidated order book. This nding might be in line with market tipping, where Euronext switches between periods of relatively high liquidity, in which it attracts all trading, and periods of low liquidity, in which trading takes place at competing trading venues. As the price impact, e ective and realized spread are based on trades, relatively liquid periods receive a larger weight in the calculation. D Equity market fragmentation To proxy for the level of fragmentation in each stock, we construct a daily Her ndahl- Hirschman Index (HHI) based on the number of shares traded on each venue, similar to Bennett and Wei (2006) and Weston (2002). HHI it = P N v=1 MS2 v;it, or the squared market share of venue v, summed over all N venues for rm i on day t. A single dominant market has an HHI close to one whereas HHI goes to zero in case of complete fragmentation. In 2009, the average HHI of the sample rms is 0.70, which is in line with other European stocks analyzed by Gomber and Pierron (2010). The U.S. is more fragmented, as NASDAQ and NYSE combined have approximately 65% of market share in 2008, resulting in an HHI lower than 0.50 (O Hara and Ye, 2009). Table III shows the yearly mean, quartiles and standard deviation of HHI, based on the sample rms. As expected, the HHI declines over time, i.e. fragmentation increases. Moreover, the standard deviation of HHI is fairly small in 2006 and 2007, as only few sample rms where traded on Virt-X, LSE and Deutsche Boerse. The variation in HHI mainly stems from changes over time and to a lesser extent of changes between rms; i.e. the time series variation is larger than the cross sectional 20

21 variation. This result is shown in the lower part of Table III, where the standard deviation of HHI is demeaned by rm, by quarter and by rm and quarter. That is, from each observation we subtract o the rm average, or the quarterly average or both, and report the standard deviation. As the table shows, demeaning by rm 0:135 0:149 reduces the standard deviation with only 9% ( ), while demeaning per quarter 0:149 reduces variation with 30%, so cross sectional variation is smaller than variation in the time series. 20 Finally, 43% of the total variation in fragmentation is captured by adding both rm and quarter dummies. Figure 5 shows the 10, 50 and 90 th percentile of HHI over time, calculated on a monthly basis and covering all rms. The sharp increase in fragmentation refers to the period where Chi-X and Turquoise started to attract substantial order ow, September In the next section, we estimate the e ect of fragmentation on various liquidity measures in a regression framework. VI The impact of fragmentation on global and local liquidity A Methodology We employ multivariate regression analysis to study the impact of fragmentation on liquidity. We have a panel data set with N = 52 rms and T = 1022 days ( ), which contains the liquidity and fragmentation measures discussed in section V. The Depth(X) is calculated for X = f10; 20; :::; 60g basis points. As the e ect of fragmentation on liquidity might not be linear, we add a quadratic term. ease the interpretation of the quadratic term, we will use F rag it = 1 To HHI it and F rag 2 it to measure fragmentation, where F rag it = 0 if trading in a rm is completely concentrated. Control variables commonly used in this literature are the rm characteristics described in Table I. 21 Furthermore, we control for algorithmic activity as this has been 20 Indeed, as we have daily observations, demeaning per quarter still leaves some variation in the time series. However, demeaning per day reduces the standard deviation with only 1% compared to demeaning per quarter. 21 Weston (2000) and O Hara and Ye (2009), among others, use similar controls. 21

22 found to improve liquidity (e.g. Hendershott and Riordan (2009)), where we construct a measure similar to Hendershott, Jones, and Menkveld (2010). On average, algorithmic traders place and cancel many limit orders, so the daily number of electronic messages proxies for their activity, i.e. placement and cancellations of limit orders and market orders. However, over time there is an increase in trading volume which would lead to more electronic messages even in the absence of algorithmic traders. To account for increasing trading volumes, the algorithmic trading proxy, Algo it ; is de ned as the number of electronic messages divided by trading volume (in e100), for rm i on day t. Some descriptives of Algo it, and the additional control variables Ln(V olatility) it ; Dark it, Ln(P rice) it ; Ln(Size) it and Ln(V olume) it are presented in Table IV. The regression equation becomes: Liq Measure it = F irm i + 1 F rag it + 2 F rag 2 it + 3 Ln(V olatility) it + 4 Dark it + 5 Ln(P rice) it + 6 Ln(Size) it + (2) 7 Ln(V olume) it + 8 Algo it + " it : In our base speci cation, rm xed e ects and quarter dummies are included 22 and in all regressions we use heteroskedasticity and autocorrelation robust standard errors (Newey-West for panel datasets), based on ve lags. B Results This subsection rst presents our regression results for the base case (i.e., where we include rm xed e ects and quarter dummies) for both the consolidated order book and the traditional stock market. Later, we control for potential endogeneity issues by introducing rm*quarter e ects and two instrumental variables (IV) regressions. The rm*quarter dummies are aimed to control for the simultaneous interactions between market structure, the degree of fragmentation, liquidity and competition in the market. In the rst instrumental variables regression we instrument the degree of fragmentation of day t with that of t 1, such that we remove the unexpected daily shocks from the variation in fragmentation. In the second IV regression we instrument the degree of fragmentation by the average daily order size of each visible venue that trades the stock. This approach may control for potential self selection 22 The results are almost identical when using day or month dummies instead of quarter dummies. 22

23 issues, e.g. that competition tends to be higher for high volume and more liquid stocks (Cantillon and Yin, 2010). We conclude by analyzing large and small rms separately, along with some additional robustness checks. B.1 Regression analysis: base speci cation The regression results for the liquidity measures employing the consolidated order book are reported in Table V. In Table VI, the same regressions are executed for liquidity measures of the traditional stock market, Euronext Amsterdam. The latter represents liquidity available to local investors, while the former liquidity to global investors using SORT. The coe cients of F rag it and F ragit 2 (models (1) to (6), Table V) show that liquidity rst strongly increases with fragmentation and then starts to decrease as the linear term has a positive coe cient and the quadratic a negative one. The results are easier to interpret when shown in a gure where we display the implied results for the six models representing di erent levels of depth. These are displayed in Figure 6, with depth on the vertical axis and the level of fragmentation on the horizontal axis. The gure clearly reveals a trade-o in the bene ts and downsides of fragmentation, where maximum liquidity is obtained at F rag = 0:35. This level of fragmentation is fairly close to the actual level observed in 2009 (where F rag is around 0:30). The patterns are highly similar for all depth levels, although liquidity close to the midpoint bene ts more from fragmentation. The economic magnitudes of the variables are large, where the maximum e ect on Ln Depth(10 bps) is 0:50, meaning that observations here have exp(0:50) = 1:65 or 65% more liquidity than observations in a completely concentrated market. For Depth(60 bps), liquidity improves with 50% at the maximum. The standard deviation of fragmentation is 0.15 in the entire sample (Table III), so variation in fragmentation has a large impact on liquidity throughout the entire order book. We now investigate the impact of fragmentation on the other liquidity indicators as reported in the remaining models from Table V. We observe the following in models (7) to (9). The optimal degree of fragmentation (F rag = 0:35) reduces the price impact and the e ective spread with 6:4 and 6:8 basis points in comparison with a 23

24 completely concentrated market. This is large, considering that the median e ective spread in 2009 is 13:3 basis points. The economic impact of the optimal degree of fragmentation on the e ective spread in our analysis is larger than the one estimated in O Hara and Ye (2009), where the bene t is approximately three basis points for NYSE and NASDAQ rms. 23 The e ect on the realized spread is much smaller with a reduction of only 0.5 basis points. In line with the other liquidity measures, the quoted spread is minimized at F rag = 0:37 and is eight basis points lower compared with a completely concentrated market (model (10)). This result is in stark contrast to the Ln quoted depth (model (11)), which reduces with 27% at F rag = 0:37. The results on the quoted depth point in the opposite direction of those of all other liquidity measures. Moreover, considering the low correlation between the quoted depth and Depth(X) in Table II, it appears that the quoted depth does not perform well in estimating liquidity over longer periods of time. Turning to the control variables of the regressions, we nd that the economic magnitude of Algo it is fairly small and negative. For example, a one standard deviation increase (s = 0:36), lowers the Depth(X) measures with 4%. However, as Algo it might be indirectly related to fragmentation, we want to be careful in interpreting this result. The remaining control variables in the regressions have the expected signs. Larger rms tend to be more liquid, so the coe cient on Ln size is positive and fairly large. The e ect of price is marginally positive and economically very small. As expected, increased trading volumes are associated to better liquidity. Also, volatility has a negative impact on liquidity; especially for liquidity close to the midpoint. Not surprisingly, the price impact strongly increases in volatility, which proxies for the amount of information in the market. The coe cients on Dark it are strongly negative, where a one standard deviation increase (s = 0:18) reduces Depth(10 bps) with 17%. Possibly, dark trading venues gain liquidity at the expense of lit markets, as they are substitutes. In addition, dark trades seem to be correlated to more informed trading, as the coe cient on the price impact is positive (4.13 bps). The impact of fragmentation on liquidity for an investor not using SORT is shown in Table VI. This table contains the regression results where our liquidity measures 23 O Hara and Ye (2009) nd a linear coe cient on marketshare outside the primary markets of 9 basis points, while the average level is 0.35, resulting in a bene t of approximately 3 basis points. 24

25 are based on the order book of Euronext Amsterdam only. To interpret the results, we plot depth against fragmentation in Figure 7. Interestingly, Depth(10 bps) rst slightly improves with fragmentation, where the maximum lies at +10% at F rag = 0:17, but afterwards quickly reduces to -10% at F rag = 0:4. This reduction is in line with the theory of Foucault and Menkveld (2008), where the execution probability on Euronext diminishes as competing venues take away order ow. This side e ect of competition makes Euronext less attractive to liquidity providers, resulting in lower depth. Consequently, small investors, who mainly care for Depth(10 bps) and are limited to trading on Euronext only, are worse o. This result is in contrast to Weston (2002) for instance, who nds that the liquidity on NASDAQ improves when ECNs enter the market and compete for order ow. Likely, the di erence is due to the market structure in the U.S., where NASDAQ market makers lost their oligopolistic rents after the entry of ECNs. The di erence between the global and local order book represents the liquidity available at the competing trading venues, who mainly o er liquidity at Depth(10 bps) and to a much lesser extent at Depth(60). That is, at F rag = 0:40, the MTFs o er 44% of Depth(10 bps) and only 26% of Depth(60 bps). 24 It appears that competition mainly takes place at liquidity close to the midpoint. We now turn to the regressions of the remaining liquidity measures in Table VI, columns (7) - (11). Surprisingly, these are not adversely a ected by fragmentation. That is, despite the reduction in Depth(10 bps), the price impact and e ective and realized spread still improve with fragmentation. It might be the case that Euronext is very liquid on some parts of the day, while relatively illiquid during other parts. As the e ective spread is based on trades, more liquid periods with many trades receive a larger weight in the calculation. In addition, order splitting behavior and smaller average order sizes may also generate lower average e ective spreads. Finally, the quoted spread on Euronext improves with fragmentation, while the quoted depth reduces with 30% at F rag = 0:35. This means that the gains of improved prices are more than o set by the lower quantities o ered, such that the 24 At F rag = 0:40, the consolidated Depth(10 bps) is 63% higher than that of a completely concentrated market, while the local Depth(10 bps) is 8% lower. Therefore, the MTFs o er 163 = 44% of the global Depth(10 bps). 25

26 Depth(10 bps) has reduced. B.2 Regression analysis: rm*time e ects In this section we do the regressions speci ed in (2), but add rm*quarter dummies. Instead of a single dummy for a period of four years, we add 16 dummies per rm. This is similar to Chaboud, Chiquoine, Hjalmarsson, and Vega (2009), who analyze the e ect of algorithmic trading on volatility for currencies, and add separate quarter dummies for each currency pair. This procedure is aimed to solve the following issues. First, the rm*quarter dummies provide additional robustness to the impact of the nancial crisis and industry speci c shocks. For example, if the nancial crisis speci cally a ects certain rms or industries (e.g. the nancial sector), and a ects both liquidity and fragmentation, then the previous analysis would su er from an omitted variables problem, leading to a bias in the coe cients on fragmentation. However, the rm*quarter dummies will capture industry shocks and time-varying rm speci c shocks. Second, the rm*quarter dummies can control for potential self selection problems. For example, Cantillon and Yin (2010) raise the issue that competition might be higher for high volume and more liquid stocks; an e ect that will be absorbed by the rm*quarter dummies. Third, the rm*quarter dummies can, at least partially, control for dynamic interactions between market structure, competition in the market, the degree of fragmentation and liquidity. Speci cally, such interactions are dynamic as, for example, a change in the current market structure will a ect the level of competition in the future, which, in turn, will a ect the market structure in the future. Our approach controls for the long-term interactions of these forces by only allowing for variation in liquidity and fragmentation within a rm-quarter. Accordingly, the dummy variables absorb the variation between quarters, which is likely to be more prone to endogeneity issues. The results for global liquidity reveal a similar pattern as those presented in the original regressions, as shown in panel A of Table VII and displayed in Figure 8. For the sake of brevity, the table does not report the coe cients of the control variables as these have the same signs and magnitudes as those reported in Tables V and VI, 26

27 but are available upon request. In the rst regression, we observe that Depth(10 bps) almost monotonically increases with fragmentation. That is, the maximum of the parabola lies outside the relevant domain, at F rag = 0:67, so there appears to be no harmful e ect of fragmentation on liquidity close to the midpoint. This is not the case for the other depth levels, as the maximum lies around F rag = 0:40, implying a trade-o in the bene ts and drawbacks of fragmentation. In addition, two other ndings emerge from the gure. First, the magnitudes of the coe cients are about 45% to 75% lower, as Depth(10 bps) equals 0:28 at F rag = 0:40; compared with 0:50 in Table V, while Depth(60 bps) is 0:10 compared with the original 0:40 (but still highly signi cant). This is likely because the rm*quarter dummies absorb long-term trends in the measure of fragmentation, while the more noisy day-to-day uctuations remain. From the regression results, it appears that removing the long-term variation dampens the estimated daily e ects. Second, liquidity deeper in the order book bene ts less from fragmentation than liquidity close to the midpoint does. This nding was also con rmed in Figure 6, but becomes more pronounced. The fact that Depth(10 bps) still improves strongly in fragmentation is in line with the earlier nding that competition of new trading venues mainly takes place at liquidity close to the midpoint. Turning to the other liquidity measures in Table VII, columns (7) - (11), we observe that the magnitudes of the coe cients on fragmentation have diminished, but in unreported plots all show a pattern similar to that of the base case. The impact of fragmentation on local liquidity, including rm*quarter e ects, is shown in panel B of Table VII and Figure 9. The gure shows that all the depth measures reduce with fragmentation, while in the base speci cation this only held for Depth(10). However, not all coe cients are signi cant because the magnitude is fairly close to zero while the 750 rm*quarter dummies greatly reduce the statistical power of the regressions. The fact that the coe cients are close to zero can again be attributed to the dummies, which absorb the long-term trends in fragmentation. Similar to the previous subsection, we con rm that the MTFs mainly o er liquidity at Depth(10 bps), and to a much lesser extent at Depth(60). That is, following previous calculations, 24 at F rag = 0:40, the MTFs o er 30% of Depth(10 bps) and only 15% of Depth(60 bps). This is in line with MTFs who speci cally compete on the price 27

28 dimension of liquidity. B.3 An instrumental variables approach The main focus of interest are the coe cients of F rag i;t and F ragi;t 2 in equation (2), but these are potentially biased, as discussed earlier. In order to further alleviate potential endogeneity issues, we employ two instrumental variables speci cations. Formally, the set of instruments needs to be uncorrelated with the error term " it (exogeneity of the instrument), and correlated to F rag i;t (a su ciently strong instrument). In the rst speci cation, we instrument F rag i;t and F ragi;t 2 with their lagged values F rag i;t 1 and F ragi;t 2 1. The intuition is that potentially, an unexpected daily shock can a ect liquidity and fragmentation simultaneously, resulting in a biased coe cient of fragmentation. Using F rag i;t 1 and F ragi;t 2 1 as instruments removes the unanticipated daily uctuations in fragmentation. In the second speci cation, we instrument F rag i;t and F ragi;t 2 with the logarithm of the average daily order size of the seven venues that trade the stock, resulting in seven instruments. This speci cation is aimed to control for a selection bias, where for example large or highly traded rms are more likely to become fragmented. For this reason, O Hara and Ye (2009) estimate a Heckman correction model. Here, the seven instruments are only weakly related to liquidity, but can strongly predict fragmentation. 25 As such, the instruments generate su cient variation in fragmentation, which, we argue, is unrelated to any self selection e ect. In both speci cations, we again add rm*quarter dummies. The results of the rst speci cation are shown in panel A (global liquidity) and B (local liquidity) of Table VIII. Again, for the sake of brevity we do not report the coe cients of the control variables, as these have the same signs and magnitudes as those reported in the base speci cations. First we notice that the coe cients are highly similar to those reported in Table V. At the optimal level of fragmentation, F rag = 0:35, global Depth(10 bps) improves by 65% compared with a fully concentrated market. This is a consequence 25 In all IV regressions the weak identi cation test (or Paap-Kleibergen Wald test), is strongly rejected. 28

29 of two o setting forces that are at play here. On the one hand, the addition of the rm*quarter dummies removes long-term trends in fragmentation, which are more strongly related to liquidity and would lower the estimated coe cients (as in the previous section). On the other hand, by using F rag i;t 1 as instrument, we remove the noisy day-to-day uctuations in fragmentation, which are only weakly related to liquidity. The two forces combined are o setting, as plots of the results are almost identical to those of the base speci cation regressions (not reported). The coe cients in panel B are also very similar to the original ones, but not statistically di erent from zero as the standard errors are strongly increased by the IV procedure. The fact that the IV coe cients are of the same sign and magnitude as the original ones, shows that the part of fragmentation that is known in advance (i.e. the day before) can explain the bene cial e ect of fragmentation on liquidity. The coe cients of the remaining regressions of panel A and B point in the same direction as the original ones, but are lower in terns of magnitude and signi cance. Turning to the second set of IV regressions, reported in panel C and D of Table VIII, we observe the following. In line with earlier results, at F rag = 0:35 global Depth(10 bps) and Depth(60 bps) improve with 110% and 6% respectively, although the level of signi cance has substantially reduced because of the instruments. In addition, the e ect on local liquidity is slightly positive, if anything, but plots do not show a consistent trend. This hints at a relatively poor estimation of the coe cients for local liquidity, and we want to be careful in interpreting these. The second IV approach uses the variation in fragmentation that can be predicted by the average order sizes of the seven venues that trade the stock, after partialling out the e ects of the control variables and rm*quarter dummies. Given these control variables, intuitively this variation in fragmentation is likely not related to rm size or initial level of liquidity, so the instruments seem valid. However, the overidentifying restrictions test (Hansen J statistics) is rejected, meaning that the instruments are not only related to liquidity via fragmentation, but also directly. One reason for rejection is the large number of observations (N = 47:000), as closer inspection of the results reveals that the order sizes are only marginally related to global Depth(10 bps). That is, after we add the logarithm of the average daily order sizes to the main regression, a one standard deviation increase in the order size of Chi-X improves global Depth(10 bps) with only 2%. This is small, and from an economic point of view we argue that the average order sizes are still valid instruments. 29

30 B.4 Small versus large stocks The bene ts and drawbacks of competition on liquidity might hinge on certain stock characteristics, such as rm size. We pursue the point in question by executing the base speci cation regressions for large stocks, with an average market cap exceeding ten billion Euro, and small stocks, with an average market cap below 100 million Euro (as shown in Table I). The results for the global and local order books of 15 large and 14 small sample stocks are reported in Table IX, panel A to D. The coe cients for the consolidated order book are plotted in Figure 10, and show two interesting results. First, the bene ts of competition are higher for large stocks than for small stocks. For large rms, the Depth(10 bps) is 64% higher at F rag = 0:35, while for small rms the maximum, at F rag = 0:18; has 30% more liquidity compared with a completely concentrated market. Second, the gure shows that the bene t of competition for large stocks is monotonically positive, meaning there are no harmful e ects of fragmentation. By contrast, the liquidity of small stocks is negatively a ected for levels of fragmentation exceeding 0:36. This suggests that the bene ts of fragmentation strongly depend on rm size. Both ndings are also con rmed by the other liquidity measures, column (7) to (11) of panel A and B. Turning to the regressions in panel C and D of Table IX, we nd that the local liquidity of large stocks also increases with fragmentation, while that of small stocks strongly decreases. That is, at F rag = 0:35 the Depth(10 bps) of large stocks improves by 12%, while that of small stocks reduces with 38%. Again, this con rms that the drawbacks of a fragmented market place mainly hold for relatively small stocks. The fact that large stocks bene t more from fragmentation is in line with their actual levels of fragmentation, which is 0:41 in 2009, while for small stocks only 0:21. B.5 Robustness checks To investigate the sensitivity of our results, we perform a number of robustness checks. First, we execute the regressions with rm*quarter dummies based on observations in 2008 and 2009 only. The results do not change (not reported), likely because 30

31 fragmentation especially took place in 2008 and This provides an additional robustness to potential time e ects (e.g. the nancial crisis), as the coe cients on fragmentation are estimated within a smaller time window. In addition, this covers for the fact that our dataset contains the ten best price levels on Euronext Amsterdam as of January 2008, while before only the best ve price levels (as mentioned in footnote 14). Finally, this solves the issue that the data by Markit Boat on dark trades is available only as of November Second, we execute the regressions in rst di erences, i.e. use the daily changes instead of the daily levels. By analyzing the day-to-day changes, we remove the longterm trends in the data. The results (currently not reported), are very similar to using rm*quarter dummies. Third, instead of using F rag to measure fragmentation, we use the market share of the traditional market (Euronext Amsterdam), and the qualitative results do not change. Finally, we have plotted higher order polynomials of F rag, and the inverted U-shapes remain: there is indeed on optimal level of fragmentation. VII Conclusion Changes in nancial regulation (the Markets in Financial Instruments Directive) implemented in November 2007 allow for the proliferation of new trading platforms in European equity markets. Currently, many stocks are traded on a variety of trading venues, i.e. the market has become fragmented. How fragmentation a ects liquidity is an important question to investors, nancial institutions and the regulator. This topic is very relevant as technological advances have drastically altered the nancial playing eld. In addition, regulatory di erences between the U.S. and Europe may cause unanticipated side e ects of competition. An example of such a di erence is the best execution rule, which in the U.S. requires brokers to execute trades against the best price available in the market, while European brokers also have to take other dimensions of execution quality into account. As a consequence, European trading platforms can compete on transaction costs, speed of execution and anonymity for example, but price violations may occur. This paper analyzes the e ect of fragmentation on liquidity for a large set of 31

32 Dutch stocks which are representative for the European arena in terms of degree of fragmentation. We explicitly di erentiate between the impact of fragmentation on global and local liquidity, where global liquidity takes all relevant trading venues into account while local liquidity only the traditional stock market. Global liquidity is appropriate for an investor using Smart Order Routing Technology (SORT) whereas local liquidity is relevant when trading fees and costs of technological infrastructure prevent investors from using SORT. The main result is that the e ect of fragmentation on global liquidity shows an inverted U-shape. Maximum liquidity is obtained when the market is partially fragmented, at HHI = 0:65. Based on our most conservative measure, global liquidity is then improved by approximately 35% compared to the the case of a fully concentrated market. However, too much fragmentation lowers global liquidity, as there is a trade-o between the bene ts and downsides. Interestingly, while global liquidity generally improves in fragmentation, local liquidity does not. That is, the competing trading platforms take away liquidity, such that an investor who only has access to the traditional market is worse o. The reduction in liquidity close to the midpoint, i.e. at relatively good prices, can be more than 10% compared to the case of no fragmentation. This suggests that the overall bene ts of fragmentation are not shared by all investors. These results follow from an approach that is aimed to identify an exogenous relation between liquidity and fragmentation by means of rm*quarter xed e ects and instrumental variables regressions. That is, the rm*quarter dummies can control for potential self selection problems, e.g. that competition might be higher for high volume and more liquid stocks. In addition, the rm*quarter dummies provide robustness to dynamic interactions between market structure, competition in the market, the degree of fragmentation and liquidity. Furthermore, the results based on the rm*quarter dummies are more robust to time e ects such as the nancial crisis. In the instrumental variables regressions we use the daily average order size of the venues that trade the stock as instruments for the degree of fragmentation. This approach solves self selection issues if we assume that the variation in fragmentation generated by the instruments is not related to rm size and initial level of liquidity, given the control variables and rm*quarter dummies already in the regression. As additional results, we nd that competition between trading venues is ercer 32

33 for larger stocks, as these are more fragmented and have a higher marginal bene t of fragmentation. Moreover, large stocks do not face the drawbacks of fragmentation like small stocks do. These results add to the policy discussion on competition in nancial markets, and suggest that overall market quality has improved, although some types of stocks and investors are worse o. We argue that the improved liquidity is due to competition between trading venues and suppliers and demanders of liquidity. However, it might be the case that competition makes the market place more e cient, which forces investors to generate improved trading strategies and algorithms, or trading venues to update technological infrastructure. In our view, these would be indirect consequences of a more competitive market place. The current analysis su ers from an obvious shortcoming which is not incorporating the liquidity at dark pools, which could over or underestimate our results. A potential avenue for future research is to incorporate explicit transaction costs in the analysis, which would truly represent the cost of trading to an investor. VIII Appendix: Liquidity measures The liquidity measures other than the Depth(X) are explained in this section. We calculate the price impact and the e ective and realized spreads based on trades and weighted over all trades per day. In contrast, the Depth(X), quoted spread and quoted depth, are liquidity measures based on quotes o ered in the limit order book and time weighted over the trading day. The e ective spread measures direct execution costs while the realized spread takes the order s price impact into account. The realized spread is often considered to be the compensation for the liquidity supplier. Denote MQ o as the quoted midpoint before an order takes place and MQ o+5 the quoted midpoint, but ve minutes later and D = [1,-1] for a buy and a sell order respectively, then Effective half spread = P rice MQ o D 10:000; (3) MQ o Realized half spread = P rice MQ o+5 MQ o D 10:000; (4) 33

34 P rice impact = MQ o+5 MQ o D 10:000: (5) MQ o The price impact, realized and e ective spread are rst calculated per trade, based on the midpoint of that trading venue. Then, all calculations are averaged over the trading day, weighted by traded volume. Next, we average over trading venues, again weighted by trading venue. This approach gives the average spread in the whole market. Limited computer power is the reason we use the midpoint of the trading venue where the trade took place instead of the consolidated midpoint. That is, creating a consolidated midpoint quote-by-qoute, as is required for the e ective and realized spreads, is computationally much more burdensome than creating a consolidated order book using one-minute snapshots. 26 The price impact and realized spread are calculated between , while the e ective spread on Therefore, Effective spread Realized spread + P rice impact. The global quoted spread is based on the best price in the consolidated order book (based on the oneminute snapshot data, see Section A) and expressed in basis points, while the local quoted spread is based on the order book of Euronext. In a similar fashion, the quoted depth aggregates the number of shares times their prices, expressed in Euros, or Quoted spread = P ASK P BID Midpoint 10:000; (6) Quoted depth = P ASK Q ASK + P BID P BID : (7) Note that the quoted depth on Euronext can be larger than that of the consolidated order book, for example when Chi-X o ers a better price but with a lower quantity. The quoted spread of the consolidated order book is always equal or better than that of Euronext. Finally, the quoted depth is identical to the Depth(10 bps) when the quoted spread equals 20 basis points. References Admati, A., Y. Amihud, and P. Pfleiderer (1991): Sunshine trading and nancial market equilibrium, The Review of Financial Studies, 438, Our dataset also has a consolidated tape constructed by Thomson Reuters, containing best prices, quantities and all visible trades in the market. However, extensive checking shows that the time stamp of these trades may di er up to three seconds from the time stamp of the same trades in the original le. 34

35 Affleck-Graves, J., S. Hedge, and R. Miller (1994): Trading Mechanisms and the Components of the Bid-Ask Spread, Journal of Finance, 49(4), Andersen, T., T. Bollerslev, F. Diebold, and H. Ebens (2001): The distribution of realized stockreturn volatility, Journal of Financial Economics, 61, Aubry, N., and M. McKee (2007): MiFID: Where did It Come From, Where is It Taking Us?, Journal of International Banking Law and Regulation, 4, Barclay, M., T. Hendershott, and D. McCormick (2003): Competition among trading venues: Information and trading on electronic communications networks, Journal of Finance, 58, Battalio, R. (1997): Third Market Broker-Dealers: Cost Competitors or Cream Skimmers?, Journal of Finance, 52(1), Bennett, P., and L. Wei (2006): Market structure, fragmentation, and market quality, Journal of Financial Markets, 9(1), Bessembinder, H., and H. Kaufman (1997): A Comparison of Trade Execution Costs for NYSE and NASDAQ-Listed Stocks, The Journal of Financial and Quantitative Analysis, 32, Biais, B., C. Bisiere, and C. Spatt (2010): Imperfect Competition in Financial Markets: An Empirical Study of Island and Nasdaq, Management Science, 56(12), Biais, B., D. Martimort, and J. Rochet (2000): Competing Mechanisms in a Common Value Environment, Econometrica, 68(4), Boehmer, E. (2005): Dimensions of execution quality: Recent evidence for US equity markets, Journal of Financial Economics, 78, Boehmer, E., G. Saar, and L. Yu (2005): Lifting the Veil: An Analysis of Pre-Trade Transparency at the NYSE, Journal of Finance, 60, Cantillon, E., and P. Yin (2010): Competition between Exchanges: A Research Agenda, forthcoming International Journal of Industrial Organization. 35

36 Chaboud, A., B. Chiquoine, E. Hjalmarsson, and C. Vega (2009): Rise of the machines: Algorithmic trading in the foreign exchange market, Discussion paper Federal Reserve System. Chlistalla, M., and M. Lutat (2009): The impact of new execution venues on European equity markets liquidity: The case of Chi-X, Proceedings of the Fifteenth Americas Conference on Information Systems (AMCIS), San Francisco. Chowdhry, B., and V. Nanda (1991): Multimarket Trading and Market Liquidity, The Review of Financial Studies, 4(3), Conrad, J., K. Johnson, and S. Wahal (2003): Institutional Trading and Alternative Trading Systems, Journal of Financial Economics, 70, Davies, R. (2008): MiFID and a changing competitive landscape, Working Paper Series. Davies, R., A. Dufour, and B. Scott-Quinn (2006): The MiFID: Competition in a New European Equity Market Regulatory Structure, Oxford Scholarship Online Monographs, pp Degryse, H., M. van Achter, and G. Wuyts (2010): Internalization, Clearing and Settlement and Market Liquidity, EFA Frankfurt Paper. Easley, D., N. Kiefer, and M. O Hara (1996): Cream-Skimming or Pro t- Sharing? The Curious Role of Purchased Order Flow, The Journal of Finance, 51, Ende, B., P. Gomber, and M. Lutat (2010): Smart order routing technology in the New European Equity Trading Landscape, Software Services for E-Business and E-Society: 9th IFIP WG 6.1 Conference. Ferrarini, G. A., and F. Recine (2006): Investor Protection In Europe: Corporate Law Making, the MiFID and Beyond. Oxford University Press. Foucault, T., O. Kadan, and E. Kandel (2009): Liquidity Cycles and Make/Take Fees in Electronic Markets, EFA 2009 Bergen Meetings Paper. Foucault, T., and A. Menkveld (2008): Competition for order ow and smart order routing systems, Journal of Finance, 63(1),

37 Goldstein, M., and K. Kavajecz (2000): Eights, sixteenths, and market depth: changes in the tick size and liquidity provision on the NYSE, Journal of Financial Economics, 56, Gomber, P., and M. Gsell (2006): Catching Up with Technology - The Impact of Regulatory Changes on ECNs/MTFs and the Trading Venue Landscape in Europe, Competition and Regulation in Network Industries,, 1(4), Gomber, P., and A. Pierron (2010): MiFID Spirit and Reality of a European Financial Markets Directive, Report Published by Celent. Gomber, P., U. Schweickert, and E. Theissen (2004): Zooming in on Liquidity, Working Paper Series. Gresse, C. (2010): Multi-Market Trading and Market Quality, Working Paper. Harris, L. (1993): Consolidation, Fragmentation, Segmentation and Regulation., Financial Markets, Institutions & Instruments, 5, Hasbrouck, J., and G. Saar (2009): Technology and Liquidity Provision: The Blurring of Traditional De nitions, Journal of Financial Markets, 12, Hendershott, T., C. Jones, and A. Menkveld (2010): Does Algorithmic Trading Improve Liquidity?, Journal of Finance, forthcoming. Hendershott, T., and R. Riordan (2009): Algorithmic Trading and Information, Net Institute Working Paper Hengelbroc (2010): Fourteen at one blow: The market entry of Turquoise, Phd dissertation - Bonn University. Huang, R., and H. Stoll (2001): Tick Size, Bid-Ask Spreads and Market Structure, Journal of Financial and Quantitative Analysis, 36, O Hara, M., and M. Ye (2009): Is Market Fragmentation Harming Market Quality?, Journal of Financial Economics, forthcoming. Pagano, M. (1989a): Endogenous Market Thinness and Stock Price Volatility, The Review of Economic Studies, 56,

38 (1989b): Trading volume and asset liquidity, Quarterly Journal of Economics, 104, Petrella, G. (2009): MiFID, Reg NMS and Competition Across Trading Venues in Europe and United States, Working Paper. Stoll, H. (2003): Market Microstructure. Handbook of the Economics of Finance. (2006): Electronic Trading in Stock Markets, Journal of Economic Perspectives, 20(1), Weston, J. (2000): Competition on the Nasdaq and the Impact of Recent Market Reforms, Journal of Finance, 55(6), (2002): Electronic Communication Networks and Liquidity on the Nasdaq, Journal of Financial Services Research, 22,

39 Table (I) Descriptive statistics of sample rms: cross section The dataset covers daily observations for 52 AEX large and mid cap constituents, from 2006 to All variables in the table are averages. Firm size and traded volume are expressed in millions of Euros. Stdev re ects the daily standard deviation of 15 minute returns on the midpoint and is multiplied by 100. Euronext represents the market share of executed trades on Euronext Amsterdam. O exchange is the market share of Over The Counter trades, Systematic Internalisers and dark pools; this number is available as of November Firm Size Price Volume Stdev O Exchange Euronext Aalberts :01 7:4 0:39 7:95 89:98 Adv. Metal. Group :22 8:6 0:78 17:94 78:89 Aegon :25 161:0 0:46 15:21 76:92 Ahold :52 120:0 0:28 18:61 74:93 Air France :94 63:9 0:40 15:06 78:04 Akzo nobel :93 147:0 0:30 19:59 73:42 Arcadis :39 3:3 0:41 10:78 87:51 Arcellor Mittal :68 388:0 0:50 24:17 70:14 Asm Int :47 7:0 0:44 10:23 86:31 ASML :78 144:0 0:39 16:75 75:52 Bamn Group :83 14:8 0:41 11:81 83:46 Binckbank :60 4:2 0:36 10:53 88:51 Boskalis :49 13:5 0:42 12:73 83:65 Corio :12 26:9 0:36 14:99 79:03 Crucell :49 9:2 0:36 8:49 89:80 CSMN :92 8:0 0:29 11:99 86:03 Draka Hold :01 3:6 0:55 16:18 77:54 DSM :02 73:0 0:29 16:89 76:86 Eurocomm. Prop :47 5:8 0:37 11:14 86:99 Fortis :83 437:0 0:38 13:37 83:51 Fugro :99 25:0 0:34 10:30 84:83 Hagemeyer 2.0 3:76 43:4 0:31 0:00 99:28 Heijmans :07 3:8 0:40 8:30 90:56 Heineken :06 100:0 0:28 18:86 74:15 Imtech :04 9:2 0:40 18:02 77:45 ING :75 904:0 0:44 14:23 81:24 Nutreco :41 15:1 0:28 12:35 85:27 Oce 0.9 9:94 8:5 0:40 10:54 86:81 Ordina :95 2:8 0:41 7:30 89:97 Philips :00 301:0 0:32 21:05 71:25 R. Dutch Shell :22 529:0 0:27 21:57 69:54 R. KPN :93 220:0 0:26 23:25 69:79 R. ten cate :68 2:3 0:40 10:11 87:63 R. Wessanen 0.6 8:48 4:3 0:32 9:10 87:49 Randstad :17 38:3 0:39 14:21 79:11 Reed Elsevier :31 74:1 0:27 20:51 71:97 SBM O shore :38 35:6 0:36 13:18 80:12 Smit Int :09 4:6 0:38 15:30 80:02 Sns Reaal :65 11:5 0:42 12:94 85:50 Tele Atlas :06 37:0 0:33 8:24 68:20 Tnt :95 99:3 0:33 19:65 73:59 Tomtom :32 47:7 0:54 10:44 83:14 Unibail Rodamco :64 172:0 0:36 34:67 57:59 Unilever :62 327:0 0:26 17:58 73:85 Usg People 0.6 9:27 8:5 0:71 19:01 68:01 Vastned :99 5:0 0:32 10:50 87:41 Vdr Moolen 0.2 4:74 2:2 0:35 2:12 97:45 Vedior :70 67:7 0:27 2:99 96:48 Vopak Int :52 8:9 0:30 12:24 84:75 Wavin 0.7 7:94 5:8 0:49 11:83 86:88 Wereldhave :77 16:0 0:28 13:56 81:87 Wolters Kluwers :10 42:1 0:30 13:15 78:09 39

40 Table (II) Descriptive statistics of the various liquidity measures for global and local traders. The table shows the 25, 50 and 75 th percentiles of the liquidity measures on a yearly basis, for the consolidated order book (upper panel) and Euronext Amsterdam (lower panel). The quantiles are based on 52 rms and 250 trading days per year ( obs). The Depth(X) is expressed in e1000s and represents the o ered liquidity within (X) basis points around the midpoint. The quoted depth is the amount of shares, in e1000s, o ered at the best bid and ask price of the global and local order book. The price impact and the e ective, realized and quoted spreads are measured in basis points. The price impact and realized spread are based on a 5 minute timewindow. 40 Year Percentile Global Depth Depth Depth Depth Depth Depth Quoted Spread Quoted Depth Realized Price Impact E ective Local Depth Depth Depth Depth Depth Depth Quoted Spread Quoted Depth Realized Price Impact E ective

41 Table (III) Descriptive statistics of fragmentation. The yearly standard deviation, mean and quartiles of market concentration, HHI, are reported, where HHI is based on the market share of visible traded volume. The lower panel shows the standard deviation of HHI demeaned per rm, per quarter and per rm and quarter. This is done by subtracting the rm average, the quarterly average, and rst rm, and then quarterly average from HHI. Year Stdev Mean 25 th 50 th 75 th Total Stdev HHI demeaned per Firm HHI demeaned per Quarter HHI demeaned per Firm - Quarter Table (IV) Descriptive statistics: time series. The yearly mean, median and standard deviation are reported for the following variables: Ln rm size, Ln traded volume, algorithmic trading, Ln stdev and market share of dark trades. Algo represents the number of electronic messages in the market divided by total traded volume (per e10.000). An electronic message occurs when a limit order in the order book is executed, changed or cancelled. Volatility, Ln Stdev here, is de ned as the natural logarithm of the daily standard deviation of 15 minute returns on the midpoint. Typically, the standard deviation is lower than one, so the natural logarithm becomes negative. Dark share is the percentage of order ow executed at dark pools, SIs and Over The Counter Mean Median Stdev Mean Median Stdev Ln Size Ln Volume Algo (10.000) Ln Stdev Dark share Mean Median Stdev Mean Median Stdev Ln Size Ln Volume Algo (10.000) Ln Stdev Dark share

42 Table (V) The e ect of fragmentation on global liquidity. The dependent variable in models (1) - (6) is the logarithm of the Depth(X) measure based on the consolidated order book. The Depth(X) is expressed in Euros and represents the o ered liquidity within (X) basis points around the midpoint. The quoted, e ective, and realized spread and price impact, (7) - (10), are measured in basis points. Ln quoted depth is the logarithm of the quoted depth in Euros (11). Frag is the degree of market fragmentation, de ned as 1 - HHI. Algo represents the number of electronic messages divided by traded volume in the market (per e100); the other variables are explained in the descriptive statistics and Table I. The regressions are based on 1022 trading days for 52 stocks, and have rm xed e ects and quarter dummies. T-stats are shown below the coe cients, calculated using Newey-West (HAC) standard errors (based on 5 day lags). ***, ** and * denote signi cance at the 1, 5 and 10 percent levels, respectively. 42 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Ln Ln Ln Ln Ln Ln E ective Realized Price Impact Quoted Ln Depth Depth10 Depth20 Depth30 Depth40 Depth50 Depth60 Spread Quoted Frag 2.844*** 2.080*** 2.188*** 2.334*** 2.420*** 2.468*** *** *** *** *** (15.9) (21.0) (26.4) (28.5) (29.7) (30.5) (-16.0) (0.5) (-17.2) (-24.8) (-11.6) Frag *** *** *** *** *** *** 33.94*** ** 40.61*** 55.72*** 0.305** (-13.4) (-15.7) (-19.2) (-21.1) (-22.7) (-24.2) (9.8) (-2.0) (12.8) (19.0) (2.0) Ln Size 1.008*** 0.623*** 0.491*** 0.427*** 0.387*** 0.359*** *** *** *** *** 0.279*** (24.2) (24.8) (24.6) (22.5) (20.9) (19.8) (-15.8) (-7.6) (-8.6) (-9.4) (17.7) Ln Price *** *** *** *** *** *** *** (-0.5) (3.8) (4.7) (4.8) (4.5) (4.0) (-0.4) (-0.7) (0.3) (4.3) (-4.2) Ln Vol 0.576*** 0.429*** 0.385*** 0.353*** 0.327*** 0.309*** *** 0.380** *** *** 0.233*** (40.9) (45.7) (47.0) (47.4) (46.9) (46.5) (-11.3) (2.0) (-17.0) (-29.4) (43.2) Ln Stdev *** *** *** *** *** *** 7.312*** *** 12.28*** 5.733*** *** (-40.0) (-52.2) (-54.0) (-52.5) (-50.7) (-49.5) (31.9) (-23.6) (53.0) (33.5) (-33.7) Dark *** *** *** *** *** *** 2.960*** *** 4.476*** *** (-20.2) (-23.7) (-24.0) (-22.9) (-21.8) (-21.3) (3.7) (-1.4) (7.8) (9.8) (-26.8) Algo *** *** *** *** *** *** 4.565*** *** 4.514*** (-5.4) (-7.1) (-7.6) (-8.4) (-8.8) (-8.6) (14.6) (0.1) (13.7) (15.5) (-0.8) Obs R

43 Table (VI) The e ect of fragmentation on local liquidity. The dependent variable in models (1) - (6) is the logarithm of the Depth(X) measure based on the order book of Euronext Amsterdam. The Depth(X) is expressed in Euros and represents the o ered liquidity within (X) basis points around the midpoint. The quoted, e ective, and realized spread and price impact, (7) - (10), are measured in basis points. Ln quoted depth is the logarithm of the quoted depth in Euros (11). Frag is the degree of market fragmentation, de ned as 1 - HHI. Algo represents the number of electronic messages divided by traded volume in the market (per e100); the other variables are explained in the descriptive statistics and Table I. The regressions are based on 1022 trading days for 52 stocks, and have rm xed e ects and quarter dummies. T-stats are shown below the coe cients, calculated using Newey-West (HAC) standard errors (based on 5 day lags). ***, ** and * denote signi cance at the 1, 5 and 10 percent levels, respectively. 43 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Ln Ln Ln Ln Ln Ln E ective Realized Price Impact Quoted Ln Depth Depth10 Depth20 Depth30 Depth40 Depth50 Depth60 Spread Spread Spread Quoted Frag 1.025*** * 0.411*** 0.589*** 0.687*** *** *** *** *** (5.7) (0.1) (1.8) (4.5) (6.4) (7.4) (-14.7) (0.3) (-17.7) (-18.3) (-5.8) Frag *** ** *** *** *** 36.40*** * 42.82*** 48.65*** *** (-9.6) (-2.2) (-1.6) (-3.3) (-5.0) (-6.3) (9.9) (-1.7) (14.2) (17.0) (-9.1) Ln Size 0.958*** 0.542*** 0.407*** 0.344*** 0.307*** 0.284*** *** *** *** ** 0.224*** (22.5) (20.6) (18.9) (16.1) (14.5) (13.5) (-16.5) (-7.7) (-9.3) (-2.6) (14.0) Ln Price ** ** *** *** *** ** *** *** (-2.1) (2.6) (3.7) (3.6) (3.2) (2.4) (0.2) (-0.5) (0.8) (4.7) (-3.5) Ln Vol 0.578*** 0.426*** 0.382*** 0.351*** 0.326*** 0.310*** *** 0.598*** *** *** 0.244*** (40.1) (43.1) (43.1) (42.6) (41.7) (41.0) (-9.4) (2.9) (-17.4) (-6.6) (45.4) Ln Stdev *** *** *** *** *** *** 7.312*** *** 12.37*** 8.337*** *** (-38.4) (-48.0) (-47.7) (-45.3) (-43.2) (-41.6) (30.8) (-22.9) (53.6) (6.5) (-34.4) Dark *** *** *** *** *** *** 2.348*** ** 4.127*** 4.043*** *** (-20.8) (-23.5) (-23.3) (-22.3) (-21.6) (-21.2) (2.9) (-2.1) (8.0) (8.3) (-27.5) Algo *** *** *** *** *** *** 3.869*** *** 6.010*** *** (-5.9) (-13.0) (-16.2) (-17.4) (-17.5) (-16.8) (12.4) (0.4) (11.9) (18.4) (6.7) Obs R

44 Table (VII) The e ect of fragmentation on liquidity: rm*quarter dummies. The dependent variable in models (1) - (6) is the logarithm of the Depth(X) measure based on the consolidated order book (panel A) and order book of Euronext Amsterdam (panel B). The Depth(X) is expressed in Euros and represents the o ered liquidity within (X) basis points around the midpoint. The quoted, e ective, and realized spread and price impact, (7) - (10), are measured in basis points. Ln quoted depth is the logarithm of the quoted depth in Euros (11). Frag is the degree of market fragmentation, de ned as 1 - HHI. For the sake of brevity, the coe cients on the control variables are not reported, as they are very similar to those of Tables V and VI. The control variables are Ln size, Ln price, Ln volume, Ln volatility, dark and algo, as explained in Table I. The regressions are based on 1022 trading days for 52 stocks. T-stats are shown below the coe cients, calculated using Newey-West (HAC) standard errors (based on 5 day lags). ***, ** and * denote signi cance at the 1, 5 and 10 percent levels, respectively. 44 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Ln Ln Ln Ln Ln Ln E ective Realized Price Impact Quoted Ln Depth Depth10 Depth20 Depth30 Depth40 Depth50 Depth60 Spread Spread Spread Quoted Panel A: Global, Firm * Quarter dummies Frag 0.970*** 0.744*** 0.690*** 0.649*** 0.585*** 0.534*** *** ** * *** (6.3) (9.4) (10.2) (10.4) (9.9) (9.4) (-4.4) (-2.5) (-1.7) (-2.6) (0.2) Frag *** *** *** *** *** *** 10.05** (-2.7) (-4.6) (-6.7) (-7.7) (-7.6) (-7.3) (2.6) (1.0) (1.5) (0.3) (-1.6) R Panel B: Local, Firm * Quarter dummies Frag *** ** ** (1.5) (-0.8) (-1.3) (-1.4) (-1.5) (-1.5) (-4.5) (-2.5) (-2.3) (-1.5) (0.5) Frag *** ** ** ** ** ** 13.97*** *** 7.060** *** (-3.7) (-2.1) (-2.2) (-2.3) (-2.1) (-2.1) (3.3) (0.6) (3.2) (2.6) (-6.3) R

45 Table (VIII) Instrumental Variables regressions of fragmentation on liquidity. Frag i;t is the degree of market fragmentation based on the market shares of all trading venues, de ned as 1 - HHI. Panel A and B show the regressions for global and local depth respectively, where Frag i;t and Frag 2 i;t are instrumented by Frag i;t 1 and Frag 2 i;t 1. Panel C and D show the regressions for global and local depth, where the average daily order size of each of the seven venues that trade the stock are used as instruments. The dependent variable in models (1) - (6) is the logarithm of the Depth(X) measure, based on the global order book (panel A and C) or the local order book (panel B and D). The Depth(X) represents the o ered liquidity within (X) basis points around the midpoint, expressed in Euros. The quoted, e ective, and realized spread and price impact, (7) - (10), are measured in basis points. Ln quoted depth is the logarithm of the quoted depth in Euros (11). The regressions have rm*quarter dummies, t-stats are shown below the coe cients, calculated using Newey-West (HAC) standard errors (based on 5 day lags). For the sake of brevity, we only report the coe cients and t-stats of Frag and Frag 2 ; the signs and magnitudes of the control variables are very similar to those reported in Table V and VI. ***, ** and * denote signi cance at the 1, 5 and 10 percent levels, respectively. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Ln Ln Ln Ln Ln Ln E ective Realized Price Impact Quoted Ln Depth Depth10 Depth20 Depth30 Depth40 Depth50 Depth60 Spread Spread Spread Quoted 45 Panel A: Global, IV Frag i;t 1 and Frag 2 i;t 1 Frag 2.407*** 1.861*** 1.775*** 1.719*** 1.601*** 1.453*** (5.0) (0.251) (0.212) (0.197) (0.187) (0.181) (-1.6) (-1.3) (-0.1) (-0.7) (1.6) Frag *** *** *** *** *** *** *** (-3.2) (0.466) (0.397) (0.370) (0.353) (0.342) (1.6) (1.3) (0.2) (-0.3) (-3.6) R Panel B: Local, IV Frag i;t 1 and Frag 2 i;t 1 Frag *** (0.9) (-0.8) (-1.1) (-1.0) (-0.9) (-0.9) (-2.8) (-1.4) (-1.3) (-0.1) (-0.4) Frag * *** *** (-1.8) (-0.7) (-0.8) (-1.1) (-1.2) (-1.2) (2.8) (1.4) (1.3) (0.4) (-2.9) R Panel C: Global, IV average order sizes Frag 1.257** 1.037*** 1.145*** 1.384*** 1.374*** 1.440*** 80.41*** 48.99*** 31.53*** 37.95*** 3.315*** (2.0) (3.3) (4.9) (6.3) (6.6) (7.2) (7.3) (4.3) (4.6) (7.0) (15.2) Frag * *** *** *** *** *** *** *** *** (1.4) (0.2) (-1.8) (-3.7) (-4.5) (-5.3) (-9.4) (-6.0) (-5.7) (-11.7) (-16.6) R Panel D: Local, IV average order sizes Frag * * *** 67.44*** 30.58*** 65.08*** 3.401*** (-0.7) (-1.8) (-1.7) (-1.1) (-1.1) (-0.9) (8.1) (5.5) (4.7) (3.8) (15.0) Frag ** 4.365** *** *** *** *** *** (2.2) (2.1) (1.5) (0.9) (0.8) (0.6) (-10.0) (-7.1) (-5.4) (-3.3) (-18.1) R

46 Table (IX) The e ect of fragmentation on liquidity: large and small rms. The regressions are executed separately for the 15 smallest stocks (avg market cap < 100 million) and the 14 largest stocks (avg market cap > 10 billion); for the global and local order books. The dependent variable in models (1) - (6) is the logarithm of the Depth(X) measure. The Depth(X) is expressed in Euros and represents the o ered liquidity within (X) basis points around the midpoint. The quoted, e ective, and realized spread and price impact, (7) - (10), are measured in basis points. Ln quoted depth is the logarithm of the quoted depth in Euros (11). Frag is the degree of market fragmentation, de ned as 1 - HHI. For the sake of brevity, the coe cients on the control variables are not reported, as they are very similar to those of Tables V and VI. The control variables are Ln size, Ln price, Ln volume, Ln volatility, dark and algo, as explained in Table I. T-stats are shown below the coe cients and calculated using Newey-West (HAC) standard errors (based on 5 day lags). ***, ** and * denote signi cance at the 1, 5 and 10 percent levels, respectively. 46 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Ln Ln Ln Ln Ln Ln E ective Realized Price Impact Quoted Ln Depth Depth10 Depth20 Depth30 Depth40 Depth50 Depth60 Spread Spread Spread Quoted Panel A: Global, large rms Frag 1.458*** 1.150*** 1.072*** 1.052*** 1.058*** 1.071*** *** ** *** *** *** (10.1) (9.1) (8.1) (7.4) (7.1) (7.1) (-6.0) (-2.2) (-6.5) (-3.3) (-4.5) Frag * * *** *** 10.41** 2.472*** (-2.0) (-1.8) (-1.2) (-1.3) (-1.4) (-1.5) (4.5) (1.6) (4.3) (2.4) (7.7) R Panel B: Global, small rms Frag 2.992*** 2.086*** 1.805*** 1.649*** 1.443*** 1.299*** *** *** (-2.7) (-1.2) (-1.3) (-9.8) (-0.5) Frag *** *** *** *** *** *** 59.00*** 29.72** 29.33* 127.5*** (-6.9) (-8.4) (-8.9) (-9.1) (-8.7) (-8.3) (-0.3) R Panel C: Local, large rms Frag 0.640*** 0.478*** 0.355** 0.352** 0.393** 0.405** *** ** *** *** 0.392*** (5.1) (3.9) (2.6) (2.3) (2.4) (2.5) (-4.6) (-2.4) (-3.9) (-5.4) (2.9) Frag *** * *** *** 14.99*** * (-3.5) (-1.9) (0.3) (0.6) (0.4) (0.5) (3.6) (1.6) (3.2) (5.3) (-1.9) R Panel D: Local, small rms Frag 1.330** *** *** *** (2.1) (1.2) (0.6) (0.5) (0.2) (-0.0) (-5.3) (-0.5) (-4.6) (-6.5) (-0.9) Frag *** *** *** *** *** *** 98.94*** *** 105.5*** (-5.0) (-4.5) (-4.0) (-3.7) (-3.2) (-2.9) (7.3) (1.0) (5.9) (9.2) (-1.1) R

47 Millions of Euros Euronext Germany Other Offexchange ChiX Figure (1) Traded Volume in millions of Euros. The graph displays the daily traded volume in millions, aggregated over the 52 AEX Large and Mid cap constituents. Euronext consists of Amsterdam, Brussels, Paris and Lisbon. Germany combines all the German cities while Other represents Bats Europe, NASDAQ OMX Europe, Virt-x and Turquoise combined. Finally, O exchange represents the order- ow executed Over The Counter, at Systematic Inernalisers and dark pools; however, these numbers are not available prior to November

48 Figure (2) Snapshot of a hypothetical limit order book. The Depth(20) aggregates liquidity o ered within the interval of (M - 20 bps, M + 20 bps), which are 2500 shares on the ask side and 800 on the bidside. Depth(40) contains 4100 and 2800 shares on the ask and bid side respectively. The number of shares o ered are converted to a Euro amount. 48

49 Depth Euronext and Consolidated Millions of Euros m1 2007m1 2008m1 2009m1 2010m1 Month Con_10 Euro_10 Con_60 Euro_60 Figure (3) Local and global depth over time. Monthly averages of Depth(10 bps) and Depth(60 bps) are shown, for Euronext Amsterdam (local) and the consolidated order book (global). The consolidated order book aggregates the order books of Euronext Amsterdam, Deutsche Boerse, Chi-X, Virt-X, Turquoise, Nasdaq OMX Europe and Bats Europe. Depth(10 bps) represents liquidity o ered within 10 basis points around the midpoint, i.e. at good price levels. Depth(60 bps) includes liquidity deeper in the order books as well, o ered up to 60 basis points around the midpoint. The values are averaged over the 52 AEX large and mid cap constituents on a monthly basis and expressed in Euros. 49

50 500 Liquidity Consolidated orderbook Euro in 1000s Log Depth Basis points around midpoint log_con_10 log_con_90 log_con_50 Figure (4) Depth in the consolidated order book. The upper panel shows the median of the Depth(X) measure, expressed in e1000s. The measure aggregates the Euro value of the shares o ered within a xed amount of basis points around the midpoint, shown on the horizontal axes. The consolidated order book represents liquidity to a global investor, where the order books of Euronext Amsterdam, Deutsche Boerse, Chi-X, Virt-X, Turquoise, Nasdaq OMX Europe and Bats Europe are aggregated. The median is based on the 52 AEX large and mid cap constituents between The lower panel displays the 10, 50 and 90 th percentiles of the logarithm of the depth measure. 50

51 HHI: 10, 50 and 90 percentile HHI m1 2007m1 2008m1 2009m1 2010m1 Months hhi_10 hhi_90 hhi_50 Figure (5) Fragmentation of AEX large and Mid cap rms. The monthly 10, 50 and 90 th percentiles of HHI are shown, for the 52 AEX large and mid cap stocks between HHI is based on the number of shares traded at the following trading venues: Euronext (Amsterdam, Brussels, Paris and Lisbon together), Deutsche Boerse, Chi-X, Virt-X, Turquoise, Nasdaq OMX Europe and Bats Europe. Trades executed OTC, on SIs or dark pools are not taken into account, as we analyze the degree of market fragmentation of visible liquidity. 51

52 Ln Depth Fragmentation Ln Depth10 Ln Depth30 Ln Depth50 Ln Depth20 Ln Depth40 Ln Depth60 Figure (6) The e ect of fragmentation on global liquidity. The coe cients of regressions (1) - (6) of Table V are plotted. The vertical axis displays the logarithm of the depth(x), while the horizontal axis shows the level of fragmentation, de ned as (1 - HHI). 52

53 Ln Depth Fragmentation Depth10 Depth 20 Depth30 Depth40 Depth50 Depth60 Figure (7) The e ect of fragmentation on local liquidity. The fragmentation coe cients for Euronext Amsterdam are plotted (regressions (1) - (6) of Table VI). The vertical axis shows the logarithm of Depth(X), while the horizontal axis shows the level of fragmentation, de ned as (1 - HHI). 53

54 0.3 Ln Depth Fragmentation Ln Depth10 Ln Depth30 Ln Depth50 Ln Depth20 Ln Depth40 Ln Depth60 Figure (8) Fragmentation and global liquidity: rm*quarter dummies. The fragmentation coe cients for the consolidated order book are plotted, where the regressions have rm*quarter dummies (regressions (1) - (6) of Table VII). The vertical axis displays the logarithm of the depth(x), while the horizontal axis shows the level of fragmentation, de ned as (1 - HHI). 54

55 Ln Depth Fragmentation Depth10 Depth 20 Depth30 Depth40 Depth50 Depth60 Figure (9) Fragmentation and local liquidity: rm*quarter dummies. The fragmentation coe cients for the order book of Euronext Amsterdam are plotted, where the regressions have rm*quarter dummies (regressions (1) - (6) of Table VII). The vertical axis displays the logarithm of the depth(x), while the horizontal axis shows the level of fragmentation, de ned as (1 - HHI). 55

56 Ln Depth Large stocks Ln Depth Small stocks Fragmentation Depth10 Ln Depth20 Ln Depth30 Ln Depth40 Ln Depth50 Ln Depth60 Figure (10) Fragmentation and global liquidity: small versus large stocks. The fragmentation coe cients for the consolidated order book are plotted, for large and small stocks (regressions (1) - (6) in panel A and B, Table IX). The 14 large stocks have an average market cap exceeding ten billion Euro, while the 15 small caps Large stocks consist of the 14 stocks with an average market cap exceeding ten billion Euro, while the 15 small stocks have a market cap smaller than 100 million Euro. The regressions include rm*quarter dummies. The vertical axis displays the logarithm of the depth(x), while the horizontal axis shows the level of fragmentation, de ned as (1 - HHI). 56