The diversity of high frequency traders

Transcription

1 The diversity of high frequency traders Björn Hagströmer & Lars Nordén Stockholm University School of Business September 27, 2012 Abstract The regulatory debate concerning high frequency trading (HFT) emphasizes the importance of distinguishing different HFT strategies and their influence on market quality. Using unique data from NASDAQ OMX Stockholm, we are the first to empirically provide such a distinction for equity markets. Comparing the behavior of market making HFTs to opportunistic HFTs (arbitrage and momentum HFT strategies), we find that market makers constitute the lion share of HFT trading volume (63 72%) and limit order traffic (81 86%). Furthermore, market makers have higher order to trade ratios, lower latency, lower inventory, and supply liquidity more often than opportunistic HFTs. In a natural experiment based on tick size changes, we find that both market making and opportunistic HFT strategies mitigate intraday price volatility. The findings indicate that, e.g., the financial transaction tax proposed by the European Commission, which would render most HFT strategies unprofitable, would primarily hit market makers and increase market volatility. Key words: High frequency trading; Market making; Market quality; Liquidity; Volatility JEL codes: G14, G18 Please send correspondence to Björn Hagströmer, School of Business, Stockholm University, S Stockholm, Sweden. Phone: ; E mail: bjh@fek.su.se. We would like to thank NASDAQ OMX for providing the data, and Petter Dahlström, Mattias Hammarquist, Frank Hatheway, and Björn Hertzberg for useful discussions. Remaining errors are our own. Both authors are grateful to the Jan Wallander and Tom Hedelius foundation and the Tore Browaldh foundation for research support. 1 Electronic copy available at:

2 1 INTRODUCTION Recent advances in information technology employed in equity markets allow traders to process information and submit orders at lightning speed. Firms who utilize the new technology for intraday trading for their own accounts are generally called high frequency traders (HFTs). 1 With typical holding periods measured in seconds or minutes, resulting in large trading volumes, HFTs are now major players in equity markets. Since the arrival of HFTs has coincided with massively increased limit order submissions and cancellations, high intraday price volatility (including flash crashes), and fragmentation of volumes across marketplaces, many voices have been raised for HFT regulation. In the regulatory debate, it is important to recognize that HFT constitutes several different trading strategies. Both the International Organization of Securities Commissions (IOSCO) and the US Securities and Exchange Commission (SEC) have emphasized that distinguishing such strategies is pivotal in the regulatory design. HFT is not a single strategy but it is rather a set of technological arrangements and tools employed in a wide number of strategies, each one having a different market impact and hence raising different regulatory issues. (IOSCO, 2011, p.23) Indeed, any particular proprietary firm may simultaneously be employing many different strategies, some of which generate a large number of trades and some that do not. Conceivably, some of these strategies may benefit market quality and longterm investors and others could be harmful. (SEC, 2010, p.46) 1 We use HFT as abbreviation for both high frequency trader and high frequency trading. 2 Electronic copy available at:

3 In the academic literature, a key empirical issue is the identification of HFT. So far, HFTs have been identified either through classifications made by the exchanges or data driven definitions imposing priors on what HFTs do. 2 None of these methods has been able to distinguish different HFT strategies. Thus, the current empirical literature treats HFTs as a homogenous trader group, forming a gap between the regulatory and academic discussion. The main contribution of this article is to bridge that gap by characterizing HFT subgroups and investigating their respective influence on market quality. Specifically, we use a data set that includes all messages (executions, submissions, and cancellations of limit orders) at the NASDAQ OMX Stockholm equity market. Our data set is unique in that we are able to associate each message with a trader identity, enabling us to track down the strategies of different member firms. With the aid of in house expertise at the NASDAQ OMX, we classify all member firms (about 100) into three categories: HFTs, non HFTs, and hybrid firms that engage both in HFT and trading for clients. Though our classification is similar to that used for US stocks by Brogaard (2011a; 2011b; 2012) and Hendershott and Riordan (2011b), thanks to our access to trader identities, we are able to take the HFT classification a step further than previous literature. Using a metric of how often a member has a limit order posted at the inside quotes, we distinguish HFT market makers from opportunistic HFTs, such as arbitrageurs and directional traders. We base our investigation on the thirty constituent stocks of the OMXS 30 index (a largecap Swedish stock index), which we follow during one month of high market volatility (August, 2011) and one month of relatively calm markets (February, 2012). We find that 2 Brogaard (2011a; 2011b; 2012) and Hendershott and Riordan (2011) utilize a HFT data set with classifications provided by the NASDAQ. Kirilenko et al. (2011) use a data driven definition, classifying HFT as the 7% of intermediaries with the highest trading volume. Menkveld (2011) observes the activities of one particular HFT that dominates trading in Dutch stocks at Chi X. 3

4 within the group of HFTs, market makers represent around 71.5% of the trading volume in August, 2011, and 62.8% in February, During both months, more than 80% of the HFT limit order submissions originate from the market makers. The implication of our findings is that any regulatory policy directed at HFTs as a group would primarily affect market makers. Currently, the European Union (EU) is considering a financial transaction tax of 0.1% on all stock transactions. According to EU projections such a tax would render low margin HFT strategies unprofitable. 3 As market making is generally considered to be good for market quality (see, e.g., Jovanovic and Menkveld, 2011), our results indicate that a financial transaction tax would be negative for equity markets. HFT activity and market quality are two intimately related concepts. As HFTs have their competitive advantage in low margin trades, where they utilize their speed of information processing and order submission, they require high trading volumes to cover their investments in technology. Thus, HFT activity tends to concentrate to liquid stocks. In order to establish whether HFT is good or bad for market quality, exogenous events that influence HFT activity but not market quality directly are needed. Brogaard (2012) uses the short sale ban of 2008 in US equity markets as an exogenous event that removed HFT activity. He finds that the removal of HFT activity caused increased intraday volatility. Studying algorithmic trading (AT), which is a more general concept than HFT, Hendershott et al. (2011) use the automation of quotes on the New York Stock Exchange as exogenous events, and Boehmer et al. (2011) use the availability of colocation services in a cross country investigation. 4 Both find that AT has a positive 3 "Automated Trading in financial markets could be affected by a taxinduced increase in transaction costs, so that these costs would erode the marginal profit. This would especially hold for the business model of highfrequency trading physically closely linked to the trading platforms on which financial institutions undertake numerous high volume but lowmargin transactions." (European Commission, 2011, p.5) 4 Algorithmic trading (AT) is a term that may span all sorts of trading strategies that can be computerized, including trading services provided to clients. 4

5 influence on liquidity. Boehmer et al (2011), however, also observe that short term volatility is amplified by AT. We use tick size changes as exogenous instruments for HFT activity. At European stock exchanges, the tick size (minimum price increment) depends on the stock price level. For example, when the price of a stock increases from SEK 49 to SEK 51, the tick size increases fivefold from SEK 0.01 to SEK 0.05 (SEK is the abbreviation for the Swedish currency krona). We hypothesize that an increased tick size makes market making more profitable and other strategies, such as arbitrage trading, more costly. Thus, we predict that a tick size increase will increase market making HFTs activities and decrease opportunistic HFTs activities, while a tick size decrease would have the opposite effect. Our results confirm that this is indeed the case. We find that in the absence of market making HFT, an exogenous increase in opportunistic HFT activity mitigates intraday volatility. When both opportunistic HFTs and market making HFTs are active, they respond in opposite ways to tick size changes. As the volatility effect is then reversed, we conclude that market makers mitigate volatility as well. Based on our event study results, market making is a HFT activity that mitigates volatility. Thus, as market making constitutes the majority of total HFT activity, the proposed EU financial transaction tax is likely to increase volatility. In general, our findings imply that policy makers, both regulators and exchanges, should encourage HFT market making. Opportunistic HFTs as a group mitigates intraday volatility, but this is a diverse group of strategies. Future research should disaggregate that group to determine the prevalence of malicious strategies. This paper is closely related to Brogaard (2011b). He investigates HFT trading activity in a set of 120 US stocks trading at NASDAQ. Though the categorization between HFTs and 5

6 other trading group is similar in that paper, the ability to distinguish different strategies and form HFT subcategories among the HFTs is a distinct feature of our paper. Our finding that HFTs are indeed a heterogeneous group of traders shows the importance of distinguishing HFT strategies. The only previous paper on HFT that can observe the trades of individual firms is, to our knowledge, Kirilenko et al. (2011). They investigate the trading and quoting surrounding the market turbulence of May 6, 2010, called the flash crash. They define HFT as the 7% most active intermediaries in the market and find that HFTs did not cause but may have amplified the volatility in the flash crash. Their data spans only one security, S&P 500 E mini futures, in three days of extraordinary volatility. Our investigation can be seen as a complement to their paper, as we cover thirty stocks over two months of different volatility levels. Finally, our paper relates to the work of Jovanovic and Menkveld (2011). They develop a model of HFT market making and find that such traders contribute to social welfare under all reasonable parameter values. In an empirical application, they show that one HFT who dominated the trading at Chi X in behaved as the middlemen in the model, with low net inventories, predominantly passive trades, and fast trading. The HFT market makers investigated in this paper are aligned to the same properties. In the next section we present our empirical setting and data, as well as our HFT categorization methodology. Next, we present estimates of various metrics frequently associated with HFT. We also run panel regressions on HFT activity and market quality measures. In Section 4 we divide the HFT group into subcategories. Specifically, we study how market makers differ from other HFTs. Furthermore, we investigate how market maker and opportunistic trader behavior differ across segments of stocks, and 6

7 how market maker activity is related to market quality measures. In Section 5, we present our event study where we analyze the causal effects of HFT on market quality. Section 6 offers a discussion of policy implications as well as concluding remarks. 2 INSTITUTIONAL DETAIL, DATA, AND HFT CLASSIFICATION We study stock trading at the NASDAQ OMX Stockholm exchange (henceforth NOMX St). NOMX St is an electronic limit order book market that is open from 9 am to 5.30 pm every weekday except Swedish bank holidays. If the day before a Swedish bank holiday is a weekday, trading at that day closes at 1 pm. Opening and closing prices are determined in call auctions. During the intermediary continuous trading session traders may post limit orders or market orders. Limit orders are executed by the order of price, time, and visibility. A feature of NOMX St that is distinct to most other stock exchanges is that only large orders may be (partially or completely) hidden. Depending on the average daily turnover of a stock, orders have to be worth at least EUR to be eligible for non visibility (for stocks with average daily turnover exceeding EUR one million, including all stocks in our sample, the order size threshold for hidden liquidity is EUR , and for some of the stocks the threshold is even higher). Accordingly, hidden orders constitute no more than 0.7% of all limit orders in our sample. 5 All messages are entered through the INET trading system, which has capacity to handle more than a million messages per second at less than 0.25 milliseconds average processing time. To cut latency (order processing time) further, NOMX St offers co 5 As a point of comparison, Bessembinder et al. (2009) report that in a sample of 100 stocks at Euronext Paris in April 2003, 44% of the order volume is hidden. In a sample of 99 NASDAQ stocks in October 2010, Hautsch (2012) finds that 14.6% of all trading volume is executed against hidden liquidity. He also reports that hidden liquidity is associated with enormous order activities related to liquidity detection strategies (p.2). The lack of hidden liquidity at NOMX St is likely to induce less liquidity detection strategies, which in in general should lead to lower order to trade ratios. 7

8 location services, where clients can pay a premium to place their servers at the premises of the exchange. NOMX St has roughly one hundred member firms that have the right to submit orders to trade at the stock market. Each member firm may sell the service of trading to clients. The provision of trading services from exchange members to traders may be done via traditional brokers, through direct market access (DMA), or through sponsored access (SA). DMA gives customers access to the market through the infrastructure of the member firm. A typical example is retail investors who get DMA through internet brokers. In the case of SA, the customer uses its own infrastructure but trade under the member identity (MPID) of the sponsor. SA is increasingly popular among algorithmic trading firms, in particular HFTs, as it allows for lower latency (order processing time) than DMA. 2.1 Data, sample selection, and summary statistics We access all messages that are entered into INET. Data used in this paper span the messages of stocks included in OMXS 30, a Swedish stock index including the thirty most traded stocks at NOMX St. Our limitation to the OMXS 30 constituents is due to that HFTs tend to concentrate their activity to the most traded stocks. NOMX St hosts 55 65% of the trading volume in OMXS 30 stocks. The largest competitor is BATS Chi X Europe (25 30%), followed by Burgundy and Turquoise (less than 5% each). 6 The time period studied includes August 2011 and February August 2011 was a highly volatile month where HFT attracted extensive media attention. 7 Being a much less volatile month, February 2012 is included for comparison. As a final limitation, only 6 Market share data is taken from 7 Reasons for the August, 2011, volatility include worries about credit rating downgrades of the United States and France, as well as concerns about the sovereign debt crisis spreading to Italy and Spain. 8

9 messages from the continuous trading session are included, representing roughly 97% of all trades and 99% of all limit order submissions. The data includes all information contained in the messages entered into INET. Most importantly, except for price, volume, time and display properties, all limit orders and executions are associated with MPIDs as well as user identification numbers (USERID). The former identifies the firm that is a member of the exchange through which the message is being entered. The latter identifies who (what broker, trader, or client) at the member firm is responsible for the message. Messages are time stamped to the nanosecond (10 9 second). Table 1 presents summary statistics. 8 The first three rows include means, medians, and standard deviations estimated for August, 2011, and the middle three rows hold the same statistics for February, The bottom three rows contain p values of tests for differences, where each null hypothesis states that there is no difference in each statistic between the two months. The average market capitalization of stocks in our sample is SEK 79 billion in August, 2011, and SEK 91 billion in February, 2012; corresponding to USD 12 billion and USD 14 billion at an exchange rate of SEK/USD 6.6. Thus, the stocks analyzed here are slightly smaller on average than those analyzed in the HFT studies by Brogaard (2011a,b; 2012) and Hendershott and Riordan (2011b), averaging USD 18 billion. The average trading activity is slightly higher in our sample, executions per stock and day on average, as compared to 3090 in their sample. The average relative bid ask spread in our sample ( %) lies in the range between their medium cap and large cap stocks ( %, see Hendershott and Riordan, 2011b). 8 Summary statistics for individual stocks are given in the Appendix. 9

10 It is clear from Table 1 that our two sample months differ substantially. Realized oneminute volatility is in August, 2011, but only in February, 2012 (the difference being statistically significant both in terms of means and medians). August, 2011, also records significantly less liquidity in terms of depth, nominal bid ask spreads, as well as relative bid ask spreads (the liquidity measures are based on limit order book snapshots taken every minute in each stock). The number of trades is 52% higher in August, 2011, than in February, 2012, and the number of limit order submission is 318% higher. Given the large differences, we report results for the two months separately in our subsequent analysis. It is notable from Table 1 that the number of cancellations is higher than the number of limit order submissions. This is due to that the volume of limit orders is often partially cancelled. 2.2 HFT and non HFT trader classification The use of algorithms in trading is nowadays widespread. In a survey conducted by the Swedish Financial Supervisory Authority, 20 (7 banks and 13 institutional investors) out of the 24 financial firms that participated claimed that they use algorithms in their trading (Finansinspektionen, 2012). In order to investigate the nature and impact of AT it is thus necessary to break down the term into subcategories. However, categorization of traders is in general complicated by the fact that traders do not stick to any one strategy. On the contrary, traders adapt and change their strategies in accordance to their expected returns and risk taking. An important distinction, however, which may also be observed in the data, is whether traders apply their strategies to their own holdings or as services to clients. Applying this distinction to algorithmic trading, we have the two subgroups of AT, agency algorithms and proprietary algorithms, where the 10

11 latter is what is typically referred to as HFT. Agency algorithm firms provide execution services for clients, typically using their infrastructure and market knowledge to minimize price impacts of trading. HFT strategies, on the other hand, may be further subdivided into market making and opportunistic trading, such as arbitrage and directional (momentum) trading. 9 In order to analyze the behavior of HFTs at NOMX St, we classify market member firms into three categories: (1) members who are primarily HFT, i.e., engage in proprietary trading only and who use algorithms in their trading strategies; (2) members who primarily trade for clients; and (3) members who engage in both proprietary and client trading. The categorization is done with the aid of NASDAQ OMX in house expertise about member activities. Including all MPIDs, we classify 29 members as HFT firms, 49 members as non HFT firms, and 22 members as hybrid firms with both proprietary and agency activities. 10 Our data set does not allow us to isolate the activities of HFTs accessing the market through SA. Thus, such activity ends up among the hybrid firms. To confirm that the HFT firms are indeed using algorithms in their proprietary trading we filter their activity with respect to USERID. Messages originating from algorithms have USERIDs starting with either PRT (program trading) or AUTD (automated trading). In the HFT group, 98.2% of all messages (96.7% of the trading volume) originate from algorithmic USERIDs. The high propensity of such USERIDs is taken as a proof of validity of the qualitative categorization process. In the non HFT group and the hybrid group, 9 See SEC (2010) and Gomber et al. (2011) for definitions of and details about these strategies. 10 Confidentiality requirements do not allow us to disclose the categorization, but the complete member list is available publicly online: 11

12 13.5% and 63.6% of the messages (14.0% and 53.0% of the trading volume), respectively, originate from algorithmic USERIDs, confirming that the use of AT is widespread in the market. Our HFT categorization is similar to the procedure applied to a data set of NASDAQ stocks utilized by Brogaard (2011a,b; 2012) and Hendershott and Riordan (2011b). The HFT categorization used for the US data set is richer in the sense that it is dynamic (continuously updated) whereas ours is done at one point in time (May 2012). One advantage of our categorization relative to the US data set is that along with the HFT group, we also identify a group that is free from HFT activities. This feature allows us to benchmark our HFT results to a group of non HFT members. Furthermore, whereas the US data set contains quote information at BBO prices only (the inside quotes), we access order information at all levels in the order book. Having full information of the limit order book allows a richer analysis of order submission strategies, though the picture is incomplete in the sense that we do not observe activity at other exchanges. The key advantage of our data set, however, is that we are able to observe the activity of individual HFT members, whereas the US data set only indicates whether a given trade or quote is associated with one (not which) out of 26 HFT firms. To our knowledge, the only previous analysis of AT and HFT with access to trader identities is Kirilenko et al. (2011), who study trading behavior before and during the flash crash on May 6, Their data contains trading records from three days of extraordinary volatility in one asset (S&P 500 E mini stock index futures). We take a broader approach, studying both trades and quotes, in 30 different stocks, during two months with both volatile and less volatile days. 12

13 Kirilenko et al. (2011) classify traders with small deviations from their target inventories as intermediaries, and the 7% most actively trading intermediaries are labeled as HFTs. We think that a qualitative categorization technique is preferable to a data driven approach, as it allows us to study HFT behavior without imposing a prior on what they do. In the next section, we employ various metrics of trading behavior to see whether the HFTs in our data set conform to common conceptions about such traders. 3 CHARACTERISTICS OF HIGH FREQUENCY TRADERS Given the plethora of HFT definitions in the current literature (see Gomber et al., 2011) it is interesting to study several dimensions of HFT activity. According to SEC (2010), HFTs tend to (i) end the day with close to zero inventories; (ii) frequently submit and cancel limit orders; (iii) use colocation facilities and highly efficient algorithms allowing them to minimize different types of latencies; and (iv) have short holding periods. The aim of our analysis in this section is to see whether such conceptions of HFT behavior apply to our sample of HFT members. Along with measures of trading and quoting volumes, we report various metrics for the HFT sample as well as the control group of non HFT members. The group of residual (hybrid) members is likely to contain substantial HFT activity, entered through, e.g., SA or trading desks of banks that also have clients. As this group is a mixture of HFT and non HFT trading, and any results related to their activity would be difficult to interpret, we do not explicitly analyze its activities. 3.1 Metrics of trading activity Our first metric of HFT behavior is the absolute day end inventory (the sum of all signed trading volumes) divided by daily trading volume, denoted Inventory. Market makers 13

14 and opportunistic traders alike strive to have low Inventory as positions held overnight are subject to clearing and capital costs. However, as we are unable to observe trades at other exchanges, this metric may falsely indicate high inventory of firms that utilize central clearing for trades at several exchanges (e.g., cross market arbitrage strategies) or several asset classes. Hence, high Inventory of HFTs may be seen as an indication of arbitrage activities. A second feature often associated with HFT is intense submissions and cancellations of limit orders. As the algorithms continuously scan the markets for news about fundamentals, order flows, and related prices, the optimal quotes are subject to continuous changes. Intense order submissions and cancellations force other market participants to store and analyze huge amounts of data. Thus, HFTs are sometimes claimed to carry a negative externality. Accordingly, many exchanges apply fines to market participants who have excessive limit order submissions relative to executions. The most common metric for this, which is also applied by NOMX St, is the order totrade ratio, q/t, defined as the number of quotes (limit order submissions) divided by the number of trades (executions) during the continuous trading session for a given day and stock. 11 Hendershott et al. (2011) use a related metric, the number of messages (limit orders submissions and cancellations) per $100 trading volume as a proxy for market wide AT activity. Boehmer et al. (2011) use two other variations of quoting intensities to approximate the amount of AT. The European Commission (2010) has considered imposing limits on order to trade ratios in its review of the Markets in 11 At NOMX St the limit for fines is q/t = 250, measured at a monthly frequency. 14

15 Financial Instruments Directive (MiFID). German legislators are considering similar limits. 12 A dimension of trading activity that is, at least semantically, closely related to HFT is the speed of trading. Millions of dollars are invested in trading system infrastructure to cut the latency of information and order submissions. It is also well known that financial firms pay rents to be able to co locate their computers at the exchanges. At NOMX St the current trading system, INET, was introduced in February, 2010, to increase the order processing capacity and to allow co location of data servers. Hasbrouck and Saar (2011) argue that low latency trading strategies is a hallmark of HFTs, and show that HFTs at NASDAQ in the United States are able to respond to news events within 2 3 milliseconds (in a sample from ). Several theoretical models use latency as the defining characteristic of AT and HFT, finding that fast traders profit at the expense of slow traders (Jarrow and Protter, 2011; McInish and Upson, 2011). According to Biais et al. (2011) such potential profits lead HFTs to overinvestment in technology. To see whether the members included in our HFT sample are faster than other members, we measure the lifetime of each limit order in our sample, from submission to the first cancellation (limit order volume can be partially cancelled). Our measure of Minimum latency is the minimum limit order lifetime for each member, each day, and each stock. Furthermore, we use the limit order lifetime to calculate averages for each member, stock, and day, referred to as Limit order duration. Deviating from the HFT features reported by SEC (2010), we also report the fraction of trades where a member is on the passive side, i.e., where his limit order is hit by a 12 empt_mifid_ii _with_new_hft_law.aspx 15

16 market order. Passive trades denotes the fraction of all member trades being passive, and Passive volume is its volume weighted version (volume is measured in SEK). In general, market makers are expected to trade more passively than other traders. Thus, if our HFT data set is dominated by market makers we expect the HFTs to be more passive than the non HFTs. Similar metrics are reported by Kirilenko et al. (2011) and Jovanovic and Menkveld (2011). 3.2 HFT and non HFT trading activity Table 2 displays trading and order volumes as well as the metrics introduced above, calculated for HFTs and non HFTs. Results for August, 2011, which was a month of high volatility, are given in Panel A; and results for February, 2012, which was a calmer month, are in Panel B. The bottom three rows of each panel contain p values of tests for differences in mean, median, and standard deviations between HFTs and non HFTs. During August, 2011, the HFTs in our sample represent on average about 30% of all trades in the OMXS 30 stocks. The fraction of trades is somewhat lower in February, 2012; 26%. Both figures are much lower than the almost 70% that Brogaard (2011b) reports for US markets (NASDAQ and BATS). Any comparison of such numbers is however complicated by the fact that hybrid member firms (with both HFT and agency services) are excluded. Hybrid members represent on average about 40% of the trading volume in both August, 2011, and February, Given that 53.0% of the hybrid firm trading volume is entered using algorithmic USERIDs (see Section 2.3), we can conclude that the HFT trading activity in the OMXS 30 stocks ranges between 26% and 52%. Interestingly, the share of limit order submissions traced to HFTs is on par with the trading activity at around 30%. This indicates that the HFTs in our sample do not overflow the market with limit orders at a rate higher than average. Still, the q/t ratio is 16

17 significantly higher for HFTs than for non HFTs (both in terms of mean and terms in medians). For both trader groups the q/t ratio is slightly higher in the more volatile month (August 2011). The two ratios indicating liquidity supply both show that HFTs are more passive than non HFTs. The differences are statistically significant for both months, and both for means and medians. This is a sign of market making activity among HFTs. As all trades have one passive and one active party, the average ratio across all traders should be equal to unity. The ratios reported indicate that HFTs supply more liquidity than they demand, that non HFTs supply roughly as much as they demand, and accordingly that the third group of traders are net demanders of liquidity. The HFT group liquidity supply ratios around 55 59% are much lower than the 74 79% observed by Jovanovic and Menkveld (2011) for one HFT market maker trading Dutch stocks at Chi X. Our results are on average slightly higher, however, than those of Kirilenko et al. (2011), who report trade weighted (volume weighted) liquidity supply ratios of around 50% (54%) for their HFTs. The Inventory metric also shows that HFTs to some extent are market makers, as the ratio is significantly lower than that of non HFTs in both months. The fact that the Inventory ratios on average lie in the interval of indicates, however, that market making is not the sole activity in the HFT group. Jovanovic and Menkveld (2011) report that their Chi X market maker close with zero inventory in 33 60% of the trading days. Finally, our investigation clearly shows that HFTs are indeed faster than their non HFT peers. Both the minimum latency and the average order lifetime are significantly lower for HFTs. According to theory (see, e.g., Biais et al., 2011), the ability to adapt faster to 17

18 news (in fundamentals as well as order flows) puts HFTs at an informational advantage, allowing them to extract an adverse selection cost from slower traders. Such profits are dividends paid towards HFT investments in information technology. What is important to note when interpreting the time dimension is that there are huge differences between means and medians. Limit orders that are left in the order book for a whole day can have a substantial impact on the mean, but typically leave the median unaffected. When focusing on the medians, we note that both HFTs and non HFTs improved their latency from August 2011 to February 2012, reflecting the continuous investments in information technology. The gap between HFTs and non HFTs decreased but remained statistically significant in February The limit order duration increased for both groups, perhaps because lower volatility decreases the need for order updates. The standard deviations presented in Table 2 describe variability across days and stocks. On several accounts, the variability is high. Given that the results reported are aggregate measures across a large number of member firms, this variability is not likely to be due to individual member behavior. Furthermore, there is ample evidence that HFT firms tend to have correlated trading strategies (Brogaard, 2011b; Chaboud et al., 2009; Hendershott and Riordan, 2011b). In the next subsection, we study how variability in aggregate HFT activity is related to market quality. 3.3 How HFT activity correlates to market quality We employ panel regressions to investigate how HFT trading activity is related to market quality measures. In the regression analysis, we seek to explain the variability in four dependent variables: HFT total SEK trading volume, HFT active SEK trading volume (volume of trades where the HFT initiates the trade by posting an executable order), 18

19 HFT passive SEK trading volume, and HFT limit order supply. Each variable is defined as the fraction of the sum of HFT and non HFT activity, rather than as a fraction of all activity. The motivation for excluding the hybrid trader group is that it contains HFT activity. Hence, benchmarking to the non HFT group gives a cleaner measure of HFT activity. In our general panel regression framework, each dependent variable is denoted,, referring to observations for stock i (i=1,...,30) on day t (there are 23 and 21 trading days in August 2011 and February 2012, respectively, yielding t=1,,44). Each panel regression model takes the following form: (1),,, where, is a k vector of market quality variables, and, are error terms. The parameter represents the overall constant term in the model, whereas represents stock specific fixed effects. The vector contains k regression coefficients. The error terms are allowed to follow a general auto regressive (AR) process according to: (2),,, where is the autocorrelation coefficient of order r, and the innovations, are independently and identically distributed. The coefficients in Equations (1) and (2) are estimated using the two step cross section seemingly unrelated regression (SUR) technique, allowing the error terms to be both cross sectionally heteroscedastic and contemporaneously correlated. In addition, the standard errors are computed with a White type technique, where the coefficient covariance estimator is robust to cross section (contemporaneous) correlation as well 19

20 as arbitrary, unknown forms of different error variances in each cross section (see Arellano, 1987, and White, 1980). The AR lag length is chosen in a step wise fashion, adding coefficients until the Ljung Box (LB) test results in a non rejection of the hypothesis that the residuals are not auto correlated up to 10 lags (at the 5% level). The daily market quality variables included in, are SIXVX, a volatility index for the OMXS 30 index; stock specific one minute realized volatility; average relative bid ask spread; trading volume in million SEK; and depth, defined as the average volume required to move the price by 1%. We also include log market capitalization (MC, expressed in million SEK) as measured at the end of each month. In Table 3, we present panel regression results for HFT trading and quoting activities. For aggregate trading volume, HFT activity is positively correlated to volatility. This effect is seen both for market wide (SIXVX) and stock specific volatility. HFT trading volume is negatively related to liquidity (increasing in spreads and decreasing in depth), trading volume, and size (MC). In interpreting these results it is important to emphasize that we analyze contemporaneous correlations. Thus, the results presented here is neither evidence that HFTs increase their trading activity in volatile and illiquid markets, nor that market quality falls when HFTs increase their trading activity. The conclusion that we can draw is that among the OMXS 30 stocks, HFTs trade relatively more (higher volumes) in smaller stocks with less liquid order books, less volume, and more volatile prices (all these results are statistically significant). Separating the trading volumes into active and passive trading, we see that the volatility effect is due to active volume. That is, HFTs tend to demand more liquidity (active volume) in volatile markets, whereas the liquidity supply (passive volume) is unaffected by volatility. The effects related to bid ask spreads, size, and trading volume are, in 20

21 contrast, due to the passive volume (active volume effects are not statistically different from zero). The negative relation between trading volume and depth is consistent across active and passive volumes. In his day level OLS regressions, Brogaard (2011b) find similar effects with respect to market wide volatility, but opposite effect of stock specific volatility. Furthermore, in contrast to our results, he finds that HFT activity is increasing in liquidity and market cap. The results can, however, not be directly compared. The market capitalization of firms in Brogaard s (2011b) data set ranges from 81 to million USD, whereas firms in our data set are less diverse, ranging from roughly 355 to million USD. Furthermore, the panel regression specifications differ substantially, as we consider fixed effects and control for autocorrelation in the error terms. Brogaard s (2011b) results are in line with the stylized fact that HFTs stay away from order books where liquidity and trading volumes are so small that their advantages in speed can not materialize. Our results, on the other hand, show that provided that there is enough liquidity, HFTs have a comparably high share of the activity in slower, more volatile, and illiquid order books. We now turn from trading volumes to the volumes of limit order submissions. By and large, the effects seen for limit orders are in line with the passive trading volume results: decreasing with liquidity, trading volume, and size. Order submissions are also positively correlated to market wide volatility, which is not the case for passive trading volume. The similar effects for passive trading volumes and limit order submissions are in line with market making strategies, where quotes are continuously updated and trading is predominantly on the passive side. In the next section, we seek to distinguish HFTs that are primarily market makers from other HFTs. 21

22 4 Distinguishing market makers from other HFTs So far, all our results are reported for HFTs and non HFTs as groups. This approach is in line with the current HFT literature, but it implicitly assumes that these groups of traders are more or less homogenous with respect to trader types. As our unique data set allows us to observe the behavior of each member firm separately, we now turn to disaggregation of the HFT group. As discussed in the introduction, such disaggregation is encouraged in the regulatory debate. In this section we introduce a measure of market making activity that allows us to separate HFT market makers from opportunistic HFTs. This enables us to study the differences between these distinct HFT trader categories. 13 Furthermore, we investigate how market making and opportunistic trading activities vary across segments of stocks. 4.1 Measuring market maker presence To investigate the degree of market making among HFTs we study the prevalence of each HFT member firm at the best bid and ask prices (the inside quotes). For each stock, we take snapshots of the order book each 10 second period in each trading day in our two month sample. As algorithms may post limit orders cyclically, for example at the turn of minutes or seconds, we randomize the order book snapshot times so that we get observations each 10 seconds uniformly distributed across 10 seconds. The Market making presence for each HFT member is then calculated daily for each stock as the fraction of snapshots where the member has a limit order posted at either side of the inside quotes. 13 Of course, member firms categorized as market makers or arbitrageurs are not homogenous either. Our data set allows observation of each member firm in isolation, but for confidentiality reasons we are required to aggregate our results to member groups. If our disaggregation of HFTs is successful, the heterogeneity is significantly higher in the HFT group than in its subgroups. 22

23 Our measure of market making presence is analogous to a measure applied by NOMX St in a new pricing scheme for market makers introduced in April Similar measures are used by Brogaard (2011b) and Hendershott and Riordan (2011a), but with the important restriction that they measure the aggregate presence of HFTs and algorithmic traders, respectively, rather than the presence of individual members. The results on the degree of market making are presented graphically in Figure 1, with findings for August 2011 in Panel A and for February 2012 in Panel B. The graph can be interpreted as a three dimensional bar chart, with OMXS 30 stocks on the x axis and HFT members on the z axis. The HFT members are sorted by their average degree of market making across stocks and trading days, and the stocks are presented in random order (names of stocks and members are dropped for confidentiality reasons). The market making presence averaged across trading days in each month is displayed on the y axis. The striking result of Figure 1 is that continuous market making in OMXS 30 stocks is concentrated to a handful of HFT members who have orders at the inside quotes more than 20% of the time. This group of market makers is highly active across the board of stocks in our sample. Other HFT members are present at the inside quotes sporadically, but rarely on the continuous basis associated with market making. The general tendency of concentration of market makers to a few members is consistent across August 2011 and February 2012, though there are slight differences in the member identities in the market maker group. We conclude that members with more than 20% market making presence on average across stocks are likely to have market making as their main business model. We classify members with lower degree of market making as 23

24 opportunistic HFTs who are likely to employ arbitrage or directional strategies. 14 Accordingly, we now subdivide our HFT sample into two groups, market makers and opportunistic traders, and compare their activities in terms of trade and quote volumes, inventory ratio, quoting intensity, trading speed, and liquidity supply. 4.2 Market maker and opportunistic trader HFT activity Table 4 provides a comparison of trading activity among market making HFTs and opportunistic HFTs. The trading volume and number of limit orders are presented as fractions of all HFT activity (i.e., for now, we do not consider non HFT activity). The Market maker presence, Inventory, q/t, Minimum latency, Limit order duration, and Liquidity supply metrics are defined as above. To conserve space, we henceforth omit means and standard deviations, reporting medians only. The upper three rows in Table 4 present results for August 2011, whereas the lower three rows hold February 2012 results. The p value of the Wilcoxon rank sum test reported for each metric and each month shows the probability that market makers have the same median as opportunistic traders. To understand our measure of market making presence properly, consider one stock at one trading day. For each such stock day, we calculate the average presence of each member classified as a market maker and opportunistic trader respectively. For market makers, the median of such stock days is 58% in August 2011 and 70% in February That is, in any given instance in any of the OMXS 30 stocks, it is likely that each of the market making members have at least one order posted at the inside quotes. The low market making presence recorded for opportunistic HFTs, less than 1% in both months, indicates that our subdivision of HFT trader types is successful. 14 In future versions of this paper we plan to disaggregate this group further. 24

25 Brogaard (2011b) measure BBO presence on an aggregate level, i.e., the fraction of calendar time when any of the HFTs in his sample has quotes posted at the inside spread. Our measure, in contrast, is the market maker presence averaged across HFTs. For his sample of US stocks in 2010, Brogaard (2011b) reports medians of 56%, 60%, and 94% for small, medium, and large cap stocks respectively. Our evidence show that disaggregation of such numbers into individual members uncovers strong diversity among HFTs. Looking at HFT trading activity, Table 4 shows that the lion share of the HFT activity can be traced to market making. In August 2011, 71.5% of the HFT trading volume (SEK) and 80.5% of the limit order traffic was due to market maker activity. In February 2012, the share of HFT trading volume for market makers was lower (62.8%), but the share of limit orders was higher (86.4%). To our knowledge, this is the first evidence of how HFT activity is distributed between market makers and opportunistic traders. As market making activity is regarded as positive for market quality, whereas opportunistic trading can potentially amplify price fluctuations, the distribution of HFT activity is important for policy making. If the distribution showed here is representative for other markets, any policy making HFT activity in general more expensive, such as the proposed EU transaction tax, would primarily hit market making activity. In accordance to the trading activity figures, q/t ratios are significantly higher for market makers than for opportunistic traders. The difference is increasing from August 2011 to February This confirms that market makers are much more passive than arbitrageurs and momentum traders. Given that Hendershott et al. (2011) and Boehmer et al. (2011) base their AT proxies on quoting activity, our evidence may indicate that 25

26 the market quality effects that they associate with AT is more likely to be effects of market making than with other AT strategies. The Liquidity supply ratios show that market makers are on the passive side in 68% (70%) of their trades in August 2011 (February 2012), compared to the 74 79% reported by Jovanovic and Menkveld (2011). The lower average presence for our market makers may be due to the increasing competition in HFT market making (Jovanovic and Menkveld, 2011, cover 77 trading days in 2007 and 2008). The average Inventory level at 5% and 7% in the two months is in line with Jovanovic and Menkveld s (2011) characterization of HFT middlemen with strong inventory mean reversion. According to their theoretical model, which they back up with empirical evidence, such middlemen have positive impact on social welfare (under all reasonable parameter values). The much higher Inventory recorded for opportunistic traders may be seen as evidence of inter market arbitrage activities. The minimum latency of market makers is the same in August 2011 and February 2012, at 0.1 milliseconds. Opportunistic traders improve their minimum latency over time, but are significantly slower than market makers. The low latency observed for market makers serves as an illustration of the competition in liquidity supply in modern equity markets. Unless market makers respond immediately to news, they risk that their outstanding quotes are picked off by faster traders. The pick off risk translates to adverse selection costs, forcing the market maker to charge wider (uncompetitive) spreads. For HFTs with strategies relying on active rather than passive trading latency may be less critical, which is reflected in our results. The average limit order duration is longer in the less volatile month and longer for market makers than for opportunistic traders. This effect is likely related to the more 26