INTERNATIONAL EVIDENCE ON ALGORITHMIC TRADING

Transcription

1 INTERNATIONAL EVIDENCE ON ALGORITHMIC TRADING Ekkehart Boehmer Kingsley Fong Julie Wu March 14, 2012 Abstract We use a large sample from that incorporates 39 exchanges and an average of 12,800 different common stocks to assess the effect of algorithmic trading (AT) intensity on liquidity in the equity market, short-term volatility, and the informational efficiency of stock prices. We exploit the first availability of co-location facilities to identify the direction of causality. We find that, on average, greater AT intensity improves liquidity and informational efficiency, but increases volatility. The volatility increase is robust to a range of different volatility measures and it is not due to more good volatility that would arise from faster price discovery. These patterns are widespread and are not limited to a few markets, but they vary in the cross-section of stocks. In contrast to the average effect, more AT reduces liquidity in small stocks; has little effect on the liquidity of low-priced or highvolatility stocks; and leads to greater increases in volatility in these stocks. Finally, during days when market making is difficult, AT provide less liquidity, improve efficiency more, and increase volatility more than on other days. Boehmer is from EDHEC Business School, 393 Promenade des Anglais, Nice, France (ekkehart.boehmer@edhec.edu). Fong is from Australian School of Business, UNSW, Sydney, NSW 2052 (k.fong@unsw.edu.au). Wu is from the Terry College of Business, University of Georgia, Athens, GA (juliewu@uga.edu). We thank seminar participants at the Bank of England, Bocconi University, EDHEC Business School, EDHEC Risk Institute London, ESSEC, Monetary Authority of Singapore, National University of Singapore, University of Houston, TCU, and SMU for their helpful comments.

2 Abstract We use a large sample from that incorporates 39 exchanges and an average of 12,800 different common stocks to assess the effect of algorithmic trading (AT) intensity on liquidity in the equity market, short-term volatility, and the informational efficiency of stock prices. We exploit the first availability of co-location facilities to identify the direction of causality. We find that, on average, greater AT intensity improves liquidity and informational efficiency, but increases volatility. The volatility increase is robust to a range of different volatility measures and it is not due to more good volatility that would arise from faster price discovery. These patterns are widespread and are not limited to a few markets, but they vary in the cross-section of stocks. In contrast to the average effect, more AT reduces liquidity in small stocks; has little effect on the liquidity of low-priced or highvolatility stocks; and leads to greater increases in volatility in these stocks. Finally, during days when market making is difficult, AT provide less liquidity, improve efficiency more, and increase volatility more than on other days.

3 1. Introduction By most accounts, high frequency trading (HFT) represents the majority of trading volume in today s markets. HFT refers to orders submitted by algorithms that emit orders or order cancellations in reaction to market updates or other events within milliseconds. Mainly because of their overall importance, but also because HFT strategies are neither transparent nor well understood, there is substantial public policy interest in this issue. Market regulators around the world debate whether HFT should be regulated, and place increasing scrutiny on order submission strategies and their effects in markets that are associated with HFT. Despite this debate and a recent flurry of theoretical and empirical work in this area, we still face many open questions. In this paper, we take a very basic but comprehensive approach that contributes new evidence to this debate. We follow Hendershott, Jones, and Menkveld (2010) and infer proxies for algorithmic trading (AT), a precondition for HFT, from measures that are derived from the intensity of order-related message traffic. We use nine years of intraday security-level quote and trade data for 39 markets around the world. This sample covers an average of 12,800 firms, excluding the U.S. This new and comprehensive sample allows us to exploit variation in algorithmic trading intensity in the cross-section of stocks and the cross-section of markets. We have several objectives. First, we describe the relationship between algorithmic trading and market quality, measured in terms of liquidity, informational efficiency, and short-term volatility. While some studies of HFT have looked outside the U.S. (see Hendershott and Riordan, 2009, or Menkveld, 2010), they are based on relatively small samples. Even the most comprehensive study so far, Hendershott, Jones, and Menkveld (2010), does not use data beyond 2006 and its main analysis is based on a change in trading protocol that happened in These dates arguably precede the real growth of HFT. Overall, we have little understanding of how the relationship

4 between AT and market quality evolves over time, especially outside of the U.S. Our within-country analysis sheds new light on these issues. Second, we exploit the presence of several separate cross-sections of firms. We investigate whether features that are known to affect order submission strategies, such as market cap, share price, and idiosyncratic volatility, also impact the effects of AT on market quality. Third, existing evidence suggests that HFT provides liquidity to other traders, and that fast traders act as informal market makers (Brogaard, 2011b; Jovanovic and Menkveld, 2011). In contrast to exchange-regulated market makers, however, informal market makers are not subject to affirmative obligations, such as a requirement to always provide liquidity on both sides of the market. Therefore, it is likely that informal market makers withdraw from the market when conditions get too difficult. Kirilenko et al. (2011) show that this happened during the Flash Crash of 2010; in this paper, we pose the more general question whether algorithmic traders reduce the intensity of their market making strategies when they become more costly to implement. We find that greater AT intensity is, on average, associated with more liquidity, whether measured at the transaction level or at the daily level, faster price discovery, and greater volatility. These results are remarkably consistent across different markets. They are robust to using different econometric models for estimation and to different measures of volatility. To link AT causally to market quality, we use co-location events as instruments. Co-location events allow fast traders to physically locate their computer hardware next to the exchanges computer to minimize data turnaround times. These events are essential in facilitating AT and represent exogenous shocks to AT that do not directly affect market quality. We use these events as instruments for AT and find evidence that supports causality from AT to market quality more AT improves liquidity and efficiency, but increases volatility. 2

5 Volatility is important to traders and issuers. Greater volatility makes limit orders more costly, and may discourage some traders from supplying liquidity. As a result, future liquidity may decrease or the price attached to liquidity risk may increases. Issuers dislike volatility because higher volatility may lower share prices or make subsequent equity issue more difficult. Certain types of volatility could be desirable. For example, prices change faster in response to new information, and volatility could be higher, when markets are more efficient. It is thus conceivable that the greater efficiency that is associated with more AT also produces higher volatility. In this case, the elevated volatility could be desirable because it would be associated with faster price discovery. To address this issue empirically, we hold constant each stock s level of informational efficiency, but we still find that AT increases volatility. Therefore, it is unlikely that the AT-induced change in volatility is due to faster price discovery. Our second objective is to assess the cross-sectional determinants of AT s effect on market quality. While the average effect of AT on market quality is positive, we also find substantial skewness in AT. This suggests that many firms either do not experience AT or are subject to negative consequences. We believe that it is important to understand the cross-sectional determinants of the benefits and costs of greater AT intensity. Specifically, stocks that are larger in terms of market capitalization, have higher share prices, or have low volatility are typically also easier to trade. It is easier to provide liquidity in these stocks, and the trading intensity is high enough to allow for significant algorithmic activity. In particular, high-frequency market making strategies are probably easier to implement in these stocks than in small, low-priced, high volatility stocks. To address these issues, we divide stocks into terciles based on market cap, price, and volatility within each market, and allow the effects of AT to differ according to these characteristics. We find that indeed much of the benefits of high AT intensity accrue to large, high-priced, and low-volatility stock. When AT increases, liquidity actually declines for the smallest terciles of stocks and remains unchanged for low-volatility and low-price stocks. The main costs associated with 3

6 AT, an elevated level of volatility when AT intensity is high, are significantly higher in stocks that are small, low-priced, or have high volatility. Our third objective is to take a closer look at the market-making strategies that are common among algorithmic traders. We cannot observe actual strategies, nor can we identify specific traders that might employ them. We only observe the aggregate effect of AT and can measure how it changes over time at the stock level. We exploit this advantage of our data by designing a timeseries proxy for days when market making strategies are likely to be more costly. Then we examine whether the effects of AT differ on those difficult market-making days. We find that when market making is more costly, AT provides less liquidity; increases the fraction of informed trading, which also enhances efficiency; and their activity increases volatility more than it does on easy marketmaking days. Overall, our results show that algorithmic trading often improves liquidity, but this effect is smaller when market making is difficult and for low-priced or high-volatility stocks. It reverses for small cap stocks, where AT is associated with a decrease in liquidity. AT usually improves efficiency. The main costs associated with AT appear to be elevated levels of volatility. This effect prevails even for large market cap, high price, or low volatility stocks, but it is more pronounced in smaller, low price, or high volatility stocks. Finally, on days when market making is more costly, AT induces less liquidity improvement and increases volatility more than on regular trading days. In future research, we will use characteristics of markets, such as details of the trading protocol, to identify whether the effects of AT are subject to variation across markets. At the country level, we will relate metrics of financial market development, economic growth, stringency of securities regulation, and market activity to costs and benefits of AT. Finally, we will examine whether links exist between the intensity of algorithmic trading and the development of securities markets in a general sense (see Menkveld, 2010). These analyses will help guide policy decisions by identifying which firms or markets, if any, may benefit from regulatory restrictions on fast trading. 4

7 Our paper is organized as follows. We review the theoretical and empirical literature in the next section. In section 3, we discuss our data and define the key variables we use. We discuss our empirical design in section 4 and present our results in section 5. The final section concludes. 2. Literature on algorithmic and high frequency trading HFT is a quite recent phenomenon and has experienced precipitous growth over the past decade. A precondition for HFT is that trading algorithms are available and can be efficiently implemented. While algorithmic trading (AT) can come from either agency algorithms or from proprietary HFT, more AT will most likely imply more HFT. This is because HF traders compete for interacting with agency algorithms, either to supply liquidity or to create short-term alpha (see Hasbrouck and Saar 2011). Only in the 2000s have information technology and market structure developed into an arena that facilitates fast, automated trading. In the U.S., this is mainly a consequence of limit order display rules that were implemented during the 1990s and, in particular, Reg NMS in Other factors also play their roles, including the NYSE s 2003 change to autoquote (mandatory automatic quote updates, as opposed to manual updates initiated by specialists), the development of fast markets that compete with the traditional venues, and the increase in capital available for proprietary trading. Other markets, including Europe, have also adapted trading protocols to facilitate HFT, mostly in the second half of the decade. Here, regulation also played a key role MIFID, for example, provides a framework for off-exchange trading that set the stage for more AT (see Menkveld 2010). Despite being a young literature, analyses of HFT and algorithmic trading reveal an interesting dichotomy. Several theoretical and empirical models analyze HFT s effects on market quality measures, including execution costs, volatility, and informational efficiency. While theoretical models mostly predict negative (or mixed) consequences of having fast traders in the market, the average effects estimated in empirical results tend to be positive. 5

8 A. Theory Hoffman (2010) extends Foucault s (1999) limit order market and allows algorithmic and human traders to compete. Their ability to react faster to new events allows algorithmic traders to evade the adverse selection that is associated with stale limit orders. In this model, the effect of introducing algorithmic traders has ambiguous effects on trading volume and the price impact of human traders, but it decreases the profits of human traders. Considering the overall effect, Hoffman shows that in most cases human traders are strictly worse off when algorithmic trading is widespread. Cartea and Penalva (2011) design a model with liquidity traders, market makers, and HFT. They find that HFT increase overall trading volume, but also volatility and the price impact of liquidity traders. Market makers come out even they lose market share (and thus revenues) for liquidity provision to the HFT, but are compensated with higher rewards for their remaining liquidity supply. The cost for the higher rewards to market making, and for the greater revenues to HFT, are all born by the liquidity traders. McInish and Upson (2011) arrive at similar conclusions using a different mechanism. In their model, strategic fast traders are the first to learn about quote updates and use this privileged information to trade at stale prices with slow traders. Here, too, does HFT activity increase trading costs for (slow) liquidity traders. In Jarrow and Protter s (2011) model, HFT also observe order-flow information faster than other traders. They show that when demand curves are downward sloping, HFT s activity affects price and creates a temporary mispricing that HFT can profitably exploit. In this case, the detrimental effect lies in less efficient pricing in addition to a transfer from slow to fast traders. A similar wealth transfer arises in an earlier model by Brunnermeier and Pederson (2005). They allow traders to follow order anticipation strategies ( predatory trading in their model), a strategy that requires the ability to predict order flow in real time at high frequency and is easily implemented as a trading algorithm. Order anticipators attempt to predict large uninformed orders and then trade ahead of these orders, in the same direction. This increases the costs for the large 6

9 liquidity trader, who will end up trading at relatively inferior prices, perhaps even with the order anticipator. Brunnermeier and Pederson show that this leads to price overshooting and that it withdraws liquidity from the market when it is most needed (by the large trader). As a result, a wealth transfer occurs from the large liquidity trader to the order anticipator. Moreover, they show that the low-liquidity event can trigger systemic liquidity shocks for other traders and markets, thereby multiplying the negative consequences the order anticipator imposes on the market. The models discussed so far generally predict higher costs to uninformed and/or slow liquidity traders in the form of a greater price impact and pre-trade information leakage. Greater execution costs essentially involve a wealth transfer from slow to fast traders, but this does not necessarily have welfare implications. Biais, Foucault, and Moinas (2010) make an elegant argument in this regard. They show that HFT can generate gains either from trade or from adverse selection, which would arise from their faster access to information. But a social planner would only consider gains from trade, not from adverse selection. As a result, HFT overinvest in technology, which leads to socially undesirable outcomes. Overall, existing theoretical models agree that HFT has undesirable consequences for liquidity traders, informational efficiency, and volatility, and these effects may well result in lower social welfare. Jovanovic and Menkveld (2011) also study welfare implications of high frequency trading. In their model, middlemen intermediate between fast limit order and slow market order traders. Depending on parameter values, their entry may increase or decrease trading volume, and also has a mixed effect on welfare. B. Empirical studies The recent spread of HFT has spurred a number of empirical studies that examine its consequences. Their inferences are easiest to synthesize by first categorizing the type of data that each study uses. The ideal data to analyze the consequences of HF and algorithmic trading would allow identification of trader (account) identifiers, which in turn allow the researcher to observe each 7

10 trader s strategy across stocks and over time. To date, only one study (Kirilenko et al. 2011) has access to this type of data, and it is limited to the trading in index e-minis around the flash crash of May 6, To date, there is no academic study of equity trading that uses data where the researcher can directly identify trader-level order submission strategies and their consequences for algorithmic or HF traders, either over time or across stocks. Researchers follow one of two approaches: infer the portion of algo/hf trading from intraday data; or use data where HF traders are identified as a group. We discuss advantages and disadvantage of both approaches below. The most basic approach uses standard intraday transactions data and either develops proxies for HFT, or infers their actions from the speed with which traders react to market events. On the downside, these approaches do not exactly measure HFT or AT instead, they infer it from the data with relatively unknown consequences for the quality of inference. But the advantage to these approaches is that they permit construction of broad and long panels that allow fairly general inferences. We adhere to this approach in this article, and closely follow Hendershott, Jones, and Menkveld (2010) in using message counts as a proxy for AT activity. Hasbrouck and Saar (2011) and Egginton, VanNess, and VanNess (2010) instead infer HFT activity from periods of apparent highfrequency activity. The former identify episodes of orders that react within milliseconds to market updates. The latter examine high-activity intervals, defined as one-minute periods where the quotesper-minute count exceeds a historical average by 20 standard deviations (and the trading day as a whole is not too different, defined as being less than two standard deviations away from the mean). These samples lose breadth relative to the message-count sample, but are able to study periods where HF activity takes actually place. The second category of data provides summary information about the type of trader. For example, Brogaard (2010, 2011a) and Hendershott and Riordan (2011) use a Nasdaq sample that reveals the aggregate order flow generated by 26 HFT firms that account for about three quarters of sample trading volume. Here, the advantage is that actual HFT can be observed for 8

11 a random sample of 120 stocks. Potential drawbacks include the selection of HFT firms, which have been picked by the exchange that provided the data and, presumably, have been willing to have their order flows disclosed to academics. Because HF strategies are typically considered sensitive both from a legal and competitive perspective, this selection process could conceivably result in orders that are more benign than a random sample of HFT orders. There are also other potential issues that complicate inference from this dataset. First, the sample of 26 HFT firms does not include any of the large proprietary trading desks that, allegedly, are responsible for a substantial portion of HFT. Second, we do observe orders that the sample traders have submitted to other markets. Third, the high percentage of trading volume of those 26 firms is of some concern. High trading volume is not necessarily representative of HF traders instead, they are typically characterized as traders with relatively low volume, but a very high ratio of order messages to trades (see Kirilenko et al and Hasbrouck and Saar 2011). 1 Overall, while these data are currently the most informative about HFT strategies in equities, they also have significant shortcomings that complicate inference. In summary, the broadest data, which in principle would allow the strongest inferences, makes the least clear distinction between HF, algorithmic, and slow trading. In the other extreme, data sets that identify actual HF activity tend to be either small or not necessarily representative for other reasons. Moreover, some of these available data sets are subject to endogeneity concerns, because it is generally not easy to identify whether causality goes from market quality to HFT activity, of from HFT activity to market quality. Against these basic concerns about the data, most but not all results document positive effects of HFT. Hendershott, Jones, and Menkveld (2010) show that algorithmic trading is associated 1 The sample used by Hendershott and Riordan (2009) also falls into this category and it is not subject to a selection concern. They use a short sample of exchange-classified algorithmic trades at Deutsche Boerse. Also similar is Menkveld s (2010) sample, who uses brokerage identities to infer the trades by a single HFT in the European market. These samples allow inferences about algos and HFT, respectively, but are limited to relatively narrow samples. 9

12 with better liquidity and faster price discovery. They use the 2003 introduction of autoquote at the NYSE as an instrument to argue that algorithmic trading causes these market quality improvements. Brogaard (2010, 2011a, 2011b) uses the Nasdaq-selected data on 26 HFTs and shows ambiguous effects on volatility, but improvements in liquidity. Based on HFT activity inferred from millisecond-level responses, Hasbrouck and Saar (2011) find improvements in volatility, spreads, and depth when these fast traders are active. Using the same data as Brogaard, Hendershott and Riordan (2011) document that HFT play an important role in price discovery. Additionally, for a much smaller Deutsche Boerse sample that is not subject to selection concerns, Hendershott and Riordan (2009) find that algorithmic trading makes prices more informative. On the negative side, Kirilenko et al. (2011) argue that HFT worsened (but did not cause) the May 6, 2010 Flash Crash. Because this is the only study that can see exactly what HFT do, it carries significant weight among the empirical work we have so far. Dichev, Huang, and Zhou (2011) find that trading per se generates excess volatility, suggesting that HFT can lead to undesirable levels of volatility. Hasbrouck and Saar (2009) are the first to document the fleeting nature of many limit orders in electronic markets, and question the traditional view that limit orders provide liquidity to the market. This argument raises questions about the quality or usefulness of often short-lived liquidity that HFT supply. Consistent with this concern, Egginton, VanNess, and VanNess (2011) show that periods of extremely active quoting behavior are associated with degraded liquidity and elevated volatility. Importantly, they show that such episodes are surprisingly frequent. While there are good economic reasons for such quote-bunching to occur as a benign by-product of HF liquidity provision, as Hasbrouck and Saar (2011) argue, it is also possible that it arises as a consequence of intentional quote stuffing. McInish and Upson (2011) examine trading around quote changes and compare fast and slow responses. They find that fast traders strategically leave stale orders on the book and that slow traders often interact with these at prices that are inferior to those available elsewhere. Finally, Chaboud et al. (2009) look at HFT in the foreign exchange market and document that the correlation among algorithmic machine orders is much higher than the correlation among 10

13 human orders. This raises questions regarding the contribution of algorithms to the transmission of systemic risk. Overall, we make two observations. First, the empirical evidence does not seem consistent with the negative theoretical predictions regarding the consequences of HFT. Instead, on average, algorithmic and HF traders appear to increase liquidity and price discovery. But other empirical work raises concerns about the quality of liquidity, effects on volatility, and about disparities between traders response times that suggest a wealth transfer from slow to fast traders. We believe that these observations demand additional analysis of the broader issues related to algorithmic and HF trading. In this paper, we examine how algorithmic trading is related to market quality and contribute new large-sample, cross-country evidence to this literature. 3. Data We obtain intraday quotes and trades from the Thomson Reuters Tick History (TRTH) database. Our initial sample includes all non-u.s. common stocks covered in the database. We identify the primary listing market for each of these stocks and drop trading in stock that takes place on all other markets. This filter produces stocks trading on 40 primary equity exchanges in 37 countries. 2 We obtain data on daily returns, daily high and low prices, trading volume, security details and financial statement data from Datastream and WorldScope. TRTH, supplied by the Securities Industry Research Centre of Asia-Pacific (SIRCA), provides access to the data feeds from various stock and derivatives exchanges transmitted through the Reuters Integrated Data Network (IDN). The trades, which include odd lots, and quotes are time stamped to the millisecond and the system covers more than 5 million equity and derivatives 2 China has three exchanges (Hong Kong, Shenzhen, and Shanghai), Japan has two (Osaka and Tokyo), all other countries have exactly one exchange included in the sample. 11

14 instruments around the world. TRTH organizes data by the Reuters Instrument Code (RIC). Each RIC code is unique worldwide and is associated with a list of RIC characteristics such asset class (e.g. equity), market, currency denomination, the first and the last data date, and the ISIN and SEDOL where applicable. Each RIC is associated with a data history that describes its changes over time, including changes in currency denomination and company name. Each RIC-market combination is associated with exactly one ISIN. 3 Datastream identifies securities by DSCODE, which is unique to a security- trading venue combination. We retain only the trading location identified as primary quote. In Datastream, this refers to the primary listing location. We are interested in the primary trading location, but in all markets except Germany the two coincide. In Germany, we use XETRA rather than Frankfurt, the primary listing location, because XETRA handles roughly 90% of volume for most stocks. Each DSCODE is associated with a comprehensive list of static securities information including the ISIN. Together with the primary quote designation, each DSCODE is associated with exactly one ISIN. To merge the two data sources we proceed as follows. For each exchange, we obtain the high, low and last trade price for each RIC from TRTH. We find the corresponding trading venue on Datastream and identify the unadjusted daily price, market capitalization, and the adjustment factor (for dilution) for each primary quote DSCODE. Then we match each ISIN associated with a DSCODE with the corresponding ISIN associated with a RIC. The resulting sample consists of all common stocks in TRTH that have an ISIN assigned, are denominated in the primary-market home currency, 3 The RIC for equity has the structure of company code (often, but not always, the same as the local ticker) plus a non-common stock security class modifier called the brokerage character, and followed by. and the exchange code. The brokerage character varies by market and we obtain the brokerage character information from TR s date sensitive market and securities reference system Speedguide. 12

15 and trade in the primary market. We validate the match by comparing the Datastream prices with TRTH prices on the first common data date after adjusting for currency differences 4. TRTH trade and quote data have qualifiers that contain market specific codes denoting the first trade of the day, auction trades, and irregular trades (such as off-market trades or option exercises). In computing intraday bid-ask spreads, effective spreads, returns, and related measures, we only use regular trades and quotes during the continuous trading period and exclude auction trades and irregular trades. Fong, Holden and Trzcinka (2011) provide further TRTH data validation against other international intraday data bases and find TRTH data to be highly accurate. Brockman, Chung and Perignon (2009) compile an intraday trade and quote database from Bloomberg over the 21 month period from October 1, 2002 to June 30, 2004 (455 trading days). The average difference in effective spreads between the two databases over the same period is 3 basis points, again speaking to the accuracy of TRTH. Official trading hours differ across exchanges and over time. We determine each exchange s historical trading hour regime by examining the trade frequency across all stocks in the exchange at 5 minute intervals. We identify the opening and closing of regular trading from spikes and drops in trading activity across all stocks at each exchange. We cross-check our algorithm against the trading hour regime and the trading mechanism entries listed in Reuter s Speedguide and the Handbook of World Stock, Derivative and Commodity Exchanges. A. Variables Our objective in this analysis is to measure the effect of algorithmic trading on market quality. We use variables that describe several dimensions of market quality, focusing on liquidity, volatility, 4 TRTH prices are historical prices in the original currency. Datastream unadjusted prices are historical prices in the current currency unit, e.g. French stocks prior to 1999 were traded in French Franc. We convert Datastream prices to Euro equivalents. 13

16 and informational efficiency. We describe these variables in this section, along with our proxies for algorithmic trading activity. Execution costs We compute several standard measures of liquidity and execution costs. For each stock, we have the best quoted spread throughout the trading day. For a given time interval s, the relative quoted spread, standardized by the quote midpoint, is defined as RQS s = (Ask s - Bid s ) / ((Ask s + Bid s )/2), (1) where Ask s is the best ask quote and Bid s is the best bid quote in that time interval. When aggregating over a trading day we use time-weighted averages of RQS. To take into account possible price improvement, potentially arising due to hidden liquidity, we compute the relative effective spread, standardized by the quote midpoint at the time of the trade. The RES on the th k trade is defined as RES k = 2D k (ln(p k ) - ln(m k )), (2) where D k is an indicator variable that equals +1 if the sell, P k is the price of the offer (BBO) prevailing at the time of the th k trade is a buy and -1 if the th k trade is a th k trade and M k is the midpoint of the national consolidated best bid and th k trade. RES measures the total price impact of a trade. We decompose this price impact into a permanent (information-related) component, RPI, and a transitory component, the relative realized spread, RRS. We follow standard practice and base both components on the quote midpoint that prevails five minutes after the trade. RRS on the defined as th k trade is RRS k = 2D k (ln(p k ) - ln(m k+5 )), (3) 14

17 where M (k+5) is the midpoint five-minutes after the for providing liquidity. The permanent component, RPI, is defined as th k trade. RRS can be interpreted as the reward RPI k = 0.5 (RES k -RRS k ) = 2D k (ln(m k+5 ) - ln(m k )), (4) and measures the change in quote midpoints that is attributable to the information content of the trade. When aggregating over all trades during a particular day, we produce either trade or (localcurrency) volume-weighted averages of RES, RRS, and RPI. When aggregating across stocks, we equally weight trade-weighted daily averages and volume-weight volume-weighted daily averages. This produces two alternative measures for each market quality dimension that are either consistently equal-weighted or consistently volume-weighted. The latter may give a better picture of profit opportunities in the whole market, but the former may provide a better picture if the costs of doing the largest trades differ fundamentally from the costs of completing small trades. Our last liquidity measure is Amihud illiquidity measure, estimated as the absolute value of daily return divided by the contemporaneous daily dollar trading volume. A larger Amihud ratio indicates that a given volume moves prices by a larger magnitude, and implies lower liquidity. Volatility Our primary measure of volatility is the intra-day range between the highest and lowest prices of a day, standardized by the daily closing price. This measure is useful because it reflects intraday fluctuations in share prices that may trigger or result from algorithmic trading. In addition, we compute four different measures or realized volatility. As lower-frequency measures we employ the absolute value of daily returns, R, and return squared, R 2. Similarly, we compute analogous measures for daily market-adjusted returns. Finally, we use the log of intra-day return variances computed for 10-minute and 30-minute returns. We aggregate within days (if applicable) and across stocks analogously to the execution cost measures above. 15

18 Informational efficiency Following Boehmer and Kelley (2010), we compute intraday measures of informational efficiency. For most of our analysis we rely on intraday measures of quote midpoint autocorrelation, which should be zero at all horizons if prices follow a random walk. Deviations from zero in either direction indicate partial predictability. We first estimate quote midpoint return autocorrelations for each stock-day, based on 30-minute intervals (results are qualitatively identical for 5, 10, 20, and 60- minute return intervals) based on the results from Chordia, Roll, and Subrahmanyam (2005). We use AR30 to denote the absolute value of this autocorrelation. Algorithmic / high frequency trading High frequency algorithmic activity is generally associated with fast order submissions and cancellations (see Hasbrouck and Saar, 2010). The proxy used by Hendershott, Jones, and Menkveld (2010) reflects this concept, and we follow them and use AT, the negative of trading volume in USD100 divided by the number of messages, as our proxy for algorithmic trading activity. It represents the negative of the dollar volume associated with each message (defined as either a trade or a quote update). An increase in this measure reflects an increase in algorithmic activity. Our measure of AT differs in an important way from the one used by HJM, who have access to order-level messages. For our world-wide sample, we only have access to a subset of these messages. We only observe each exchange s best quotes and trades, rather than all order-related messages. This means that we cannot directly capture one important dimension of HFT activity, the high ratio of order (submission and cancel) messages to trades that is so characteristic of many popular HFT strategies. But the HFT strategies that are mentioned, for example, in the SEC 2010 concept release involve activity at the BBO, rather than behind it. Therefore, we believe that HFT activity in our BBO-trade data set is highly correlated with HFT activity in an order-trade data set. We address this issue explicitly by repeating HJM s time-series and panel results using the U.S. portion of our sample constructed from the TAQ database. The time series of AT during the 16

19 overlapping period is very similar to the one presented in HJM. Our replication of the panel estimated using only NYSE activity yield qualitatively identical results, despite the differences between the two tests. 5 Given this result, we have little reason to expect our AT proxy to deliver substantially different results than the HJM proxy. B. Descriptive statistics For inclusion in our final sample we impose a few additional data requirements. We drop Ireland, where fewer than 30 stocks are listed prior to We exclude stocks that have data for fewer than 21 trading days during the sample period. To illustrate the breadth of our sample, Table 1 lists the number of stocks for each market. For the average year, our sample includes about 12,800 firms and we have substantial variation across markets. Over the sample period, the number of listed firms increases by 140% and, in 2009, represents an aggregate market capitalization of USD 16.7 trillion. We first describe the time-series distribution of message traffic (the number of all quote changes and all trades), which is the main component of our algorithmic trading proxy. For each country, Table 2 lists the median number of messages in 2009 (the most recent year in our sample), the mean number of messages, and the change in both between 2001 and We make several important observations. First, message traffic grows steeply over time in all but three markets. This is consistent with AT/HFT playing an increasingly important role around the world. Second, the mean exceeds the median in all markets, and often by an enormous margin. This indicates that a few firms are responsible for a disproportionate chunk of the message traffic. For example, Shanghai has the most messages per stock-day in 2009, and the mean of 13,045 is about 30% larger than the median. At Euronext Amsterdam, however, the mean of 8717 is about 14 times larger than the median. 5 Our analysis differs from theirs in the following ways. First, our AT measures are based on BBO messages rather than all order-level messages. Second, our time series covers rather than Third, we use the full panel rather than only the period surrounding the implementation of the NYSE s 2003 implementation of autoquote. 17

20 Third, growth from 2001 to 2009 differs substantially across markets. The increase in daily message traffic ranges from -47% in Athens to 3707% in Brussels. These two observations have important implications for our test design. They suggest that there is substantial variation of algorithmic trading intensity both across firms and over time. Our empirical approach will account for both. Especially useful is the cross-firm variation, because it allows us to conduct cross-sectional analysis of how AT affects market quality. In Figures 2 and 3, we present the average time-series development in our measures of liquidity, efficiency, and volatility. For each market-day, we first compute an equally weighted mean across firms, and then compute equally weighted means for each market-year. In the figures, we plot the annual time series of averages across markets. The quoted and effective spreads, TWQS and RES, in Figure 2 show similar patterns. For example, RES begins at 25bp in 2001 and declines to 14bp by Afterwards, it increases again to about 23bp, probably as a result of the financial crises around the world. For TWQS, the range is from 60bp in 2001 to 30bp in 2007, and then an increase to 52bp. When we decompose RES into its transient (RRS) and permanent (RPI) components, we again find very similar patterns. Both components decrease until 2007, and then increase again. We also observe that RRS exceeds RPI in every year by more than 50%, and the difference appears to grow over time. Figure 3 show the daily relative price range, a proxy for intraday volatility. It is almost flat from 2001 to 2006 around 0.025, and then begins to increase above 0.04 in This provides an interesting contrast to the liquidity measures that have a pronounced U shape over time. Finally, the lower graph in Figure 3 shows AR, the quote midpoint autocorrelation of returns measured over 10 and 30 minute intervals. Both measures decline over time, indicating an improvement of informational efficiency over time. 18

21 4. Methodology A. Country-specific analysis Our first objective is to identify, for each of the 39 countries, the effects of algorithmic trading intensity on market quality, represented by measures of liquidity, volatility, and informational efficiency. We will document this relationship in panel regressions that control for firm and day fixed effects. These fixed effects prevent us from interpreting systematic patterns in market quality across firms or secular patterns over time as the result of algorithmic trading. To estimate these relationships at the country level, we use two approaches. We employ a panel regression of the form MQit = αi + γt + βatit + δ_xit + εit, (5) where the α i are firm fixed effects, the γ t are day fixed effects, AT is our proxy for algorithmic trading, and X is a vector of control variables. This vector includes share turnover, 1/price, market value of equity, the lagged dependent variable, and the daily price range standardized by the daily closing price (a proxy for volatility). All explanatory variables are lagged by one period to ensure that they are predetermined. Finally, all continuous variables are standardized in the cross-section to make coefficients comparable across countries. For inference within countries we use standard errors that are robust to cross-sectional and time-series heteroskedasticity and within-group autocorrelation based on Arellano and Bond (1991). Across-market inference is based on an equal-weighted means of the 39 market-specific coefficients and simple cross-sectional t-statistics also based on these 39 19

22 observations. This approach should be conservative, because all inference is based only on the 39 market-specific observations. 6 In some of our analysis, we differentiate observations according to cross-sectional firm characteristics including market cap, volatility, and share price. Unless stated otherwise, we determine each day, separately for each market, the lowest and highest tercile of firms based on the most recent 20 trading days. We assign LOW and HIGH dummies to firms in these terciles, respectively. We augment our regression model (5) with the two interactions between AT and each dummy. The interaction coefficients capture the potentially differential effect of AT on market quality in the LOW or HIGH terciles relative to the middle tercile. The total effect of AT for LOW firms is given by the sum of the coefficient on AT and the coefficient on AT*LOW. We interpret results for the HIGH dummy analogously. B. Instruments for algorithmic trading Our second objective is to establish that AT and market quality are not spuriously related to each other. To identify an exogenous change in AT we rely on the first instance of co-location in each country. 7 Co-location refers to locating a trader s order submission algorithm physically close to a trading center s computer. This minimizes response times for data feeds from the exchange and also for order messages from the trader. The introduction of co-location happens at different dates across markets. Because co-location applies to all firms within a market simultaneously, we use a between estimator at the market level. Specifically, we compute the market-value weighted averages for all variables within each market, standardize the resulting time series within each 6 As a robustness check, we use a Fama-McBeth model. For each day, we estimate a cross-sectional regression analogously to equation (5) within each market, but without the time effects. For each market, we then compute the time-series average of each coefficient. Inferences within countries use Newey-West standard errors. Across-country inference uses cross-sectional t-statistics as in the main analysis. This approach produces qualitatively identical results that are not tabulated. 7 Other possible instruments include the introduction of direct market access for traders, DMA, or other updates to the trading protocol that imply a structural change in how AT / HFT is done. 20

23 market, and then perform a two-stage generalized instrumental variable estimation. In the first stage, we regress our AT proxy on time-fixed effects and a co-location dummy. 8 In the second stage, we estimate MQct = αc + γt + βat*ct + δ_xct + εct, (6) where the α c are market fixed effects, the γ t are day fixed effects, AT* is the vector of predicted values from the first-stage regression, and the other variables are market-specific weighted averages of the control variables as described above. For inference we use standard errors that are robust to cross-sectional and time-series heteroskedasticity and within-group autocorrelation based on Arellano and Bond (1991). 5. Regression results A. Within market effects of algorithmic trading As described in section 4, we conduct a two-dimensional analysis within each market using daily two-way fixed-effects regressions. We conservatively conduct global inferences, but to illustrate our estimation we present market-specific results in the appendix. The table shows the 39 countries, whether they are G7 or OECD members, and coefficient estimates and p-values for each market. The p-values are based on dynamic Arellano-Bond standard errors. For example, the effect of AT on RES is for Xetra stocks. Since all variables are standardized each day, this means that a one-standard deviation increase in AT is associated with a 0.01 standard deviation decline in RES. But this effect is heterogeneous across markets. For example, the effect is standard deviations in Shanghai, but only in Singapore. 8 Adding the remaining explanatory variables to the first stage leaves inferences unchanged but results in less precise estimates. 21

24 Panel A of Table 3 presents a summary of the liquidity-related coefficients that we estimate for each market. We omit coefficients on firm fixed effects, daily fixed effects, and control variables. These are estimated as described in equation (5) but not tabulated. The mean effect of AT on RES is , which means that a one-standard deviation increase in AT implies a standard deviation decrease in relative effective spreads. The associated t-statistic is -7.4, using the cross-sectional standard error across the 39 markets. AT reduces (improves) RES in 92% of the markets. 9 We find consistent results for each of the liquidity measures an increase in AT is associated with decreases in RQS, RES, and the Amihud measure. Finally, we document that more AT reduces both the information content of trades (RPI) and the transitory price impact (RRS). This suggests that AT does not necessarily increase the intensity of informed trading. These results suggest that, on average, greater AT intensity is associated with improved liquidity. In Panels B, C, and D, we assess whether the effect of AT varies with market cap, share price, or the 20-day return volatility. For each measure, we contrast the effect for the lowest tercile ( LOW ) with the effect of the largest tercile ( HIGH ). Within each market, we first determine the lowest and highest terciles and accordingly define the LOW and HIGH dummies as described in the data section. In Table 3, we report the mean coefficient of AT, which now represents the effect of AT on liquidity for the middle tercile of the sort variable, and the two interactions. We also report the total effects for LOW and HIGH stocks. In Panel B, we sort by market cap so the LOW and HIGH dummies represent firm size within each market. Looking at Amihud illiquidity, the largest firms have a marginal coefficient of that is significant at the 5% level. This implies that the largest firms experience a reduction in 9 While not tabulated, the corresponding coefficient averages using the Fama-MacBeth approach yield identical inferences regarding the coefficient signs. We note, however, that the magnitude of the AT effect is much larger for the FMB models than for the panel estimation that we present in Table 3. For example, the mean effect of AT on RES is standard deviations with FMB, which is six times larger than the estimate from the panel regressions. The differences arise from different treatment of time effects. FMB allows slope coefficients to vary across days, while the two-way panel nets out an aggregate time trend. 22