Equity market fragmentation and liquidity: the
impact of MiFID
(Preliminary)
Hans Degryse
Frank de Jong
yVincent van Kervel
zJanuary 2011
Abstract
Changes in European …nancial regulation (the Markets in Financial Instru-ments 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 fragmenta-tion 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 consid-ering 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
zCentER, EBC, TILEC, Tilburg University. TILEC-AFM Chair on …nancial market regulation. zCentER, Tilburg University
zCorresponding author, TILEC, CentER, Tilburg University, Email: v.l.vankervel@uvt.nl.
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.
I
Introduction
Following the developments in the U.S., European equity markets have seen a prolifer-ation 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 trad-ing venues to compete for order ‡ow. Consequently, tradtrad-ing 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 re-searchers, 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 uni-verse 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 imple-mentation of MiFID in November 2007. 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 tra-ditional market. This is expected to reduce the monopoly power of the tratra-ditional market, to lower transaction costs (Biais, Martimort, and Rochet, 2000) and to foster technological innovation (Stoll, 2003). However, there is also evidence that fragmen-tation might reduce liquidity. When order ‡ow becomes fragmented, the probability of …nding a counterparty diminishes. Consequently, execution probabilities are low-ered, which might cause some investors to leave the market or informed investors to leverage their informational advantage (e.g. Chowdhry and Nanda (1991)). More-over, 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 particu-lar, 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
1An exception of the pre-trade transparency regulation is granted to so-called dark pools, as such rules would con‡ict with their business model.
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 Eu-ronext, using data from 2004, and …nd that fragmentation over these two traditional stock markets improves liquidity. O’Hara and Ye (2009) …nd that fragmentation low-ers 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 regula-tory 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 de-viation 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 2009. For each stock, we construct a consolidated limit order book (i.e., the limit order books of all trading venues
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 re-gressions. 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 frag-mentation 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 frag-mentation on global liquidity, where the optimal level of fragfrag-mentation 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 sug-gests 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
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 Eu-ropean …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
af-ter MiFID
This section gives a brief discussion on the contents of the Markets in Financial Instruments Directive (MiFID), e¤ective November 1, 2007. 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 “con-centration 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
3The “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).
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. …xed fees, variable fees as well as make or take fees.4 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 infor-mation in the market. As the number of trading venues increases, inforinfor-mation 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
4Make 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.
5A liquid stock is de…ned by the European Commission as having a free ‡oat market capitaliza-tion exceedinge500 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.
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 mented market. A single market bene…ts from lower costs, compared with a frag-mented 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 simulta-neously; 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
6In the U.S., the price of every trade is reported to the consolidated tape, such that the perfor-mance of a broker can clearly be evaluated.
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 heteroge-neous 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 asym-metric 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-7Time 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.
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 2008. They …nd that 6.7% of all transactions are full trade-throughs and an addi-tional 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 Tech-nology. 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
8Algorithmic trading is also know as High Frequency Trading.
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 methodolog-ical 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-by-stock 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 liquid-ity 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 se-ries 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
10This 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.
the venues in consideration. Second, this approach ignores the cross sectional varia-tion in fragmentavaria-tion, 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
11The data contains the ten highest bid and lowest ask prices and their associated quantities, for each trading venue.
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 AEChi-X25 and DAChi-X30 …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
sta-tistics
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
less than 1% of total order ‡ow.
Second, there are the new MTFs with visible liquidity, such as Chi-X, Bats Eu-rope, 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 2008. 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 busi-ness 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 2008 - 2009, where they ex-ecute 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
12The 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
data from all these venues but collect them only during the trading hours of the continuous auction of Euronext Amsterdam, i.e. between 09.00 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).
13Unscheduled 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.
14Part of the sample only has the best …ve price levels: Euronext before January 2008. 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.
15To 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. http://www.cesr-eu.org/data/document/10_882.pdf
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 2008. Moreover, the dominance of Euronext over its challengers is strong, but slowly decreasing over time. This pat-tern 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
16The use of realized volatility is well established, see e.g. Andersen, Bollerslev, Diebold, and Ebens (2001).
traders” and “local traders”. For global traders, global liquidity applies as they employ Smart Order Routing Technology (SORT) to access all trading venues simul-taneously; 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 levelj =f1;2; :::; Jgon the pricing grid and the midpoint of
17Foucault 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.
the consolidated order book asM, then
Depth Ask(X) =
J
X
j=1
PjAsk QAskj j PjAsk < M (1 +X) ; (1a)
Depth Bid(X) =
J
X
j=1
PjBid QBidj j PjBid > 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 be-tween 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 po-tential trade, i.e. V equals e100.000, 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 liq-uidity to trade e100.000, 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,
18Technically, a negative consolidated spread is an arbitrage opportunity, which might not be exploited because of explicit trading costs for example.
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 2009. 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 90th percentile of Depth(10 bps) ise920.000, while the 10th percentile of Depth(60 bps) ise70.000. In the regression analysis we will work with thelogarithm of the depth measures, where the 10, 50 and 90th percentiles are shown in the lower panel of Figure 4. Table II contains the 25, 50 and 75th 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 liq-uidity 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 75th percentiles of the liquidity measures are reported in Ta-ble II, based on daily observations and calculated yearly, for the consolidated order
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 2009. 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
19The 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).
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). HHIit =PNv=1M Sv;it2 , 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
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 reduces the standard deviation with only 9% (0:135 00:149:149), while demeaning per quarter 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 90th 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 2008. 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
lo-cal 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 (2006 – 2009), 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. To ease the interpretation of the quadratic term, we will use F ragit = 1 HHIitand
F rag2
it to measure fragmentation, whereF ragit = 0 if trading in a …rm is completely
concentrated.
Control variables commonly used in this literature are the …rm characteristics de-scribed in Table I.21 Furthermore, we control for algorithmic activity as this has been
20Indeed, 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.
found to improve liquidity (e.g. Hendershott and Riordan (2009)), where we construct a measure similar to Hendershott, Jones, and Menkveld (2010). On average, algo-rithmic 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, Algoit; is
de…ned as the number of electronic messages divided by trading volume (in e100), for …rmi on day t. Some descriptives of Algoit, and the additional control variables
Ln(V olatility)it; Darkit,Ln(P rice)it; Ln(Size)it and Ln(V olume)it are presented in
Table IV. The regression equation becomes:
Liq M easureit = F irmi+ 1F ragit+ 2F ragit2 + 3Ln(V olatility)it+
4Darkit+ 5Ln(P rice)it+ 6Ln(Size)it+ (2)
7Ln(V olume)it+ 8Algoit+"it:
In our base speci…cation, …rm …xed e¤ects and quarter dummies are included22 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
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 ragit and F ragit2 (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 fragmenta-tion, where maximum liquidity is obtained atF rag = 0:35. This level of fragmenta-tion 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 vari-ables are large, where the maximum e¤ect on Ln Depth(10 bps) is0:50, meaning that observations here haveexp(0:50) = 1:65or 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
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= 0:37and 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 ofAlgoit is fairly small and negative. For example, a one standard
devia-tion increase (s= 0:36), lowers the Depth(X) measures with 4%. However, as Algoit
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 ex-pected, 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 Darkit are strongly
nega-tive, 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
23O’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.
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
24At 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¤er163 92163 = 44%of the global Depth(10 bps).
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 prob-lems. 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 in-teractions between market structure, competition in the market, the degree of frag-mentation 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,
but are available upon request.
In the …rst regression, we observe that Depth(10 bps) almost monotonically in-creases 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 frag-mentation 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 com-pared 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,24at 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
dimension of liquidity.
B.3 An instrumental variables approach
The main focus of interest are the coe¢ cients ofF ragi;t and F ragi;t2 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 toF ragi;t (a su¢ ciently strong
instru-ment).
In the …rst speci…cation, we instrument F ragi;t and F rag2i;t with their lagged
valuesF ragi;t 1 and F rag2i;t 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 ragi;t 1 and F ragi;t2 1 as instruments removes
the unanticipated daily ‡uctuations in fragmentation.
In the second speci…cation, we instrumentF ragi;t and F ragi;t2 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 frag-mentation.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) im-proves by 65% compared with a fully concentrated market. This is a consequence
25In all IV regressions the weak identi…cation test (or Paap-Kleibergen Wald test), is strongly rejected.
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 ragi;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.
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 atF 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 monoton-ically 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 im-proves 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 only0:21.
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
