High-Frequency Trading Competition
Jonathan Brogaard and Corey Garriott*
May 2015
Abstract
High-frequency trading (HFT) firms have been shown to influence market quality by exploiting a speed advantage. Using a firm-identified limit-order book dataset we show it is also competition among HFT that influences market quality, and we evaluate the trading dynamics through which the influence occurs. HFT entrants (exiters) increase (decrease) liquidity and improve (harm) price efficiency. In markets with fewer incumbents the impacts are larger. We observe specific behavioral changes consistent with competitive pressure: After Passive HFT entry (exit), incumbent HFT spreads tighten (widen). After Aggressive HFT entry, incumbent order flow loses informedness. HFT trading revenue suggests competition reduces the profitability of HFT. The results show that part of the value of HFT comes from its competitiveness.
JEL classification: G20, G14, L10
*Contact: Jonathan Brogaard, Foster School of Business, University of Washington: [email protected]. Corey Garriott, Market Infrastructure Division, Bank of Canada: [email protected]. We are grateful to Michael Brolley, Terry Hendershott, Talis Putnins, Ryan Riordan, and Joshua Slive for advice and suggestions, and to Omer Mohammad, Adrian Eng and Faith Chin for their research assistance. This article is an extension of the Bank of Canada working paper by the same name. All errors are our own.
1. Introduction
Direct access to equity markets had once been the reserve of a small number of well-connected specialists: access to order books was restricted by the trading venue, and a select few market-makers and seat members were located on-site. Today, a computer-based trader need only become a client of a sponsoring broker, as the cost of computing hardware has fallen steadily, and there is no requirement to register as a market-maker. This has led to a mixture of technology and strategy referred to as high-frequency trading (HFT). HFT is the combination of low-latency connectivity, short holding periods and low inventory positions. Two features stand out in HFT: speed and competition. The speed of HFT is an increasingly well-understood contributor to market liquidity and price efficiency (e.g., Brogaard et al. 2013; Baron, Brogaard and Kirilenko 2014; Weller 2013). This paper examines the role of competition among HFT firms and its relationship to market liquidity and price efficiency.
We study a market in which the industrial organization changes over time. Successive HFT entrants and exiters interact with incumbents and other participants. HFT entry and exit disturbs the trading environment and leads HFT incumbent firms to adapt. Following a HFT entry, HFT incumbents behave more competitively. Following an exit, the remaining HFT become less competitive. The competitive changes are largest following the first few entrants. Overall, market liquidity and price efficiency improves after HFT entry and deteriorates after HFT exit.
We observe evidence of competitive pressures in a number of dimensions. First, incumbents lose volume share to new entrants. Second, entry by Passive HFT firms, those
that mainly provide liquidity, leads incumbents to tighten their spreads. Third, following entry by Aggressive HFT firms, those that mainly take liquidity, incumbents experience a loss in their ability to trade in the direction of future price movements. Finally, after new HFT entry, HFT incumbents earn less revenue. Each of these effects diminishes as the HFT entry number increases. As more HFT firms begin trading a stock, they have a decreasing effect on volumes, trades, incumbents' spreads, incumbents' price impact and revenues.
The study uses data from the Alpha trading venue. Alpha commenced business as a Canadian alternative trading system and became an exchange in April 2012.1 It is now owned by the TMX Group, which also owns the Toronto Stock Exchange. Alpha provided data from its first day of operation in November 2008 to 25 September 2012. The data set for orders, messages and trades contains sufficient information to identify the dates on which participants first began to trade on the venue.
We use a differences-in-differences methodology around the staggered entry and exit dates of new HFT firms to isolate the impact of entry and exit (combined referred to as an “event”) from other cross-sectional and time-series variation. We use a 10-day centered window around an event to isolate the role of HFT competition. A nice feature of the setting is that the HFT event dates vary by stock, so there is a control group of stocks that do not experience any change in HFT competition. We use two empirical criteria to identify participants as HFT: low overnight inventory holdings and frequent switching of the direction of their trades. Similar to much of the literature, we find that HFT competition is associated with improvements in liquidity.
1
For the purposes of the application of trading rules in Canada, there is effectively no difference in treatment between an alternative trading system and an exchange.
We find that HFT firms compete for volume. After HFT entry into a stock, the proportion of trading volume conducted by HFT increases. Moreover, the new HFT entrants take away volume from the incumbent HFT firms. This suggests that there are particular trades that HFT firms want to engage in and the new entrants compete vigorously to capture trading volume. The reverse is true for HFT exits.
We study different types of HFT participants. Similar to Hagstromer and Norden (2013), we define two HFT subgroups: Passive HFTs and Aggressive HFTs. Passive HFTs are those that use marketable orders less than 33% of the time; Aggressive HFTs use marketable orders more than 66% of the time. The entry of Passive HFTs, whose behavior is consistent with a market-making strategy, leads incumbents to quote more aggressively. After entry by Aggressive HFT firms, whose behavior is consistent with high-speed price forecasting, incumbents experience a loss in their ability to trade in the direction of future price changes. Entry is associated with HFTs generating fewer revenues. For exits, the reverse occurs.
Furthermore, there are diminishing marginal effects of entry. Later-arriving entrants have less of an impact than earlier arrivals. After a third HFT begins trading a stock, the impacts on HFT share of volume and trades, the competitive pressure on HFT quoted spreads and trade informedness, and HFT revenues lose statistical significance. Later entrants do not generate new business for HFT overall but do displace incumbent market shares with statistical significance.
As more Passive HFTs enter, the best bid-ask spreads quoted by incumbent firms tighten less and less, diminishing not only in magnitude but also in statistical significance. The new entrant typically undercuts the best incumbent's prices; entrants quote, on
average, slightly lower than incumbent HFTs. Entry impact similarly diminishes for Aggressive HFT entrants. After the entry of a new Aggressive HFT, incumbents are less able to forecast short-term price movements. More firms are trading to impound information into prices, deteriorating the predictability of prices.
Microstructure theory is mixed on how competition should affect market participants. Economic intuition suggests competition improves outcomes for other market participants. Bondarenko (2001) studies a trading game in which several market makers compete for the order flow and finds market-maker competition is good for price impact. On the other hand Dennert (1993) models market-makers and demonstrates that competition can result in higher bid-ask spreads because of the higher riskiness of each individual market-maker interacting with an informed trader.
Empirically, competition is well-studied in financial markets. For instance, when exchanges compete, market participants usually benefit (Huang 2002; Mayhew 2002; Battalio and Jennings 1997), but Pagnotta and Philippon (2011) show competition can be harmful. Others study competition among formal market-makers. Market-making competition increases liquidity (Klock and McCormick 1999; Weston 2000; Van Ness, Van Ness and Warr 2005; Battalio 1997). The focus of this paper is to examine competition among HFT firms.
Competition specifically among HFT firms is highlighted in the theoretical literature. Biais, Foucault and Moinas (2013) explain why slow traders may choose to become fast and that increased HFT participation will increase adverse selection costs. Pagnotta and Philippon (2011) show that competition among HFTs decreases profits, but that the
competition is of Cournot type and so profits remain positive, while Jovanovic and Menkveld (2010) model an environment in which HFTs compete away any profits.
At least two other empirical papers examine slightly different questions regarding HFT and competition. Menkveld (2013) finds that when an HFT firm begins to participate on the European Chi-X, bid-ask spreads decrease 50%. However, Breckenfelder (2013) finds that when multiple HFT firms enter the Swedish Nasdaq, market quality deteriorates. We extend both papers by examining how the result varies by the number of HFT incumbents present and how HFTs change their trading behavior as a result of the entry or exit of HFT competitors.
The approach in this paper helps to isolate the role of competition from the role of speed and aims to understand the channel by which competition affects markets. Our findings complement HFT market-quality papers. Most papers show a relationship between HFT and improved market quality (Brogaard, Hendershott and Riordan 2014; Hasbrouck and Saar 2013; Hagstromer and Norden 2013; Carrion 2013), while others find a mixed or negative relationship (Brogaard, Hendershott and Riordan 2013; Gai, Yao and Ye 2012). We are interested in the marginal effect of competition among HFT firms and how it might help explain some of the HFT and market-quality findings.
The rest of the paper is structured as follows: Section 2 describes the data and empirical design. Section 3 evaluates how liquidity and price efficiency change after HFT entry and exit. Section 4 reports how HFT events influence market dynamics. Section 5 concludes.
2. Data and Methodology
We use order message and trade data from the Canadian Alpha equity exchange. Alpha opened for trade on 7 November 2008. It was independently operated until merging with the TMX Group as part of the Maple transaction on 1 August 2012. By January 2009, Alpha had the second-largest share of trading volume in Canada, and by January 2010, it had over 20% of Canadian trading volume. Alpha became an exchange April 2012 but has not commenced listing operations. Figure 1 displays the time-series distribution of equity trading volume across exchanges for the 279 stocks sampled in this study.
INSERT FIGURE 1 ABOUT HERE
The data set contains all order messages received by Alpha and trade notifications sent by Alpha between 7 November 2008 and 25 September 2012. The data include all limit-order quotations, updates, fills and cancellations, and they are millisecond time stamped. In addition to the order-book information that is publicly available, the data include a field populated by Alpha on an ad hoc basis with identifiers based on Alpha’s knowledge of its customer base, in the form of numerical codes giving certain counterparty information: a code identifying orders from smart-order routers, a code identifying the brand of order-management system used to disseminate an order, and an anonymous identification code that identifies a unique market participant firm.
The anonymous code allows us to identify the date each market participant firm begins trading. We use data between 1 January 2009 and 25 September 2012. We include
in the analysis stocks whose average market capitalization is greater than the lowest average market capitalization of any stock in the S&P TSX 60 index in September 2012 and whose average daily trading volume is higher than the lowest average daily trading volume of any stock in the S&P TSX 60 index in September 2012. 279 stocks meet the criteria.
Identifying HFT firms
A market participant firm is eligible to be identified as an HFT firm if it is not identified as a smart-order router and if none of its orders are managed by an order-management system.2 An eligible market participant firm is identified as an HFT firm if the market participant’s behavior satisfies two criteria for at least two contiguous calendar weeks (10 contiguous trading days) for any of the stocks in the sample. First, the participant must switch the sign of its trade (follow a buy with a sell or a sell with a buy) on average 40% of its trades. Second, it must hold on average no more than 15% of its daily volume overnight. Participants who satisfy these criteria are called HFT firms. If a participant is identified as an HFT firm for any stock, it is identified as an HFT firm for every stock.
The criteria were chosen to be consistent with the literature on how to distinguish HFT from other kinds of algorithmic trading. The first criterion distinguishes HFT from algorithmic execution. HFT firms trade frequently but do not build a position in a stock, so their trading direction should frequently reverse between buying and selling. The second criterion distinguishes HFT from strategies that bear significant longer-term risk. HFT firms hold very little inventory overnight as they do not wish to expose themselves to the
2
risk that the asset value might change. We choose a relatively high inventory threshold of 15%, since there are multiple markets for the stocks traded on Alpha. An HFT firm may open a position on Alpha and close it on another trading venue, or vice versa, resulting in what would appear to be an inventory position from the researcher's perspective, even if the trader is neutral. The firm must satisfy the HFT criteria for a sustained period of time (two weeks) to confirm its identification. A firm that is identified once as HFT is HFT always so that its impact can be studied even for periods of time when its behavior does not meet our definition of HFT identity.
Identifying HFT entry and exit dates
A participant stock-entry date is the calendar day on which a participant first trades at least ten lots (1000 shares) on Alpha. The threshold is more than the minimum lot to exclude days on which the participant is merely testing its connection. Stock-entry events are stock-specific, so the same participant can have different entry dates for different stocks. We study only first entries.
The first day an HFT firm trades zero volume in a stock and continues to do so for at least a calendar month is an HFT exit date. In principle the same HFT firm can exit a stock more than once. There are 45 second exits, five third exits, and one fourth exit.
There is variation in how many stocks HFT-identified entrants begin trading or stop trading at a time. Most HFT-identified entrants start trading several stocks on the same day; the same is true when they stop trading. The day a firm begins trading it starts trading a median of 7 stocks and a mean of 20.6 stocks. The largest number of stocks a firm begins trading on the same day is 81. When a firm exits on average it stops trading a median of 2
stocks and an average of 4.3 stocks. The largest number of stocks a firm stops trading on the same day is 17.
Some entrants start by trading just one or two stocks and, after some time, begin trading many stocks the same day. The same is true for exits. Figure 2 Panel A displays a histogram of the number of HFT entry events over the time series. Panel B does the same for exits.
INSERT FIGURE 2 ABOUT HERE
The entrants and exits are staggered over the four-year window, providing a time-series control. There is some clustering of entry (exit) events at specific dates, meaning that some HFT firms choose to enter (exit) multiple stocks at the same time.
Classification of HFT entry and exit events
We classify the HFT event dates by the aggressiveness of the entrant. Aggressiveness is the percentage of a participant's orders that are marketable. Aggressive HFT firms are those for whom trades are executed using marketable orders at least 66% of the time. Passive HFT firms are those for whom trades are executed using marketable orders less than 33% of the time. There are 241 Aggressive entries and 565 Passive entries. 87 entries are neither passive nor aggressive. Figure 3 depicts the distribution of aggressiveness over the set of entry events.
The distribution is bimodal with large concentrations of events in which participants are only passive (only limit orders) or only aggressive (only marketable orders). HFTs that are mixed do not cluster around any particular percentage of aggressiveness.
Table 1 provides summary statistics for the HFT events for all events and for Passive and Aggressive events. Panel A reports entry events and Panel B exit events. Each row of the table gives the average, standard deviation, 25%, median and 75% statistics of various measures. Daily volume is the average number of shares the entrant traded daily. Daily trades is the average number of trades the entrant executed daily. Per cent daily volume is the average per cent of total daily volume the entrant contributed. Overnight volume, % of daily volume is the average per cent of its own daily volume the entrant held at the end of day. Switching rate is the percentage of the entrant's trades for which the entrant switched the direction of its trade—the per cent of the time it bought after it had sold or sold after it had bought. Percent aggressive is the percentage of the entrant's trades for which the entrant initiated the trade (either through a market order or marketable limit order). N is the number of events.
INSERT TABLE 1 ABOUT HERE
Entrants trade on average 23,345 shares a day and on average 7.1% of daily volume. For comparison, 7% of the average daily volume of IBM on NASDAQ is roughly 300,000
shares. Most participants held less than 21% of their daily trading volume overnight, and most participants switched the direction of trade more than 30% of the time.
There are some differences between the sample of Passive HFT entry events and Aggressive HFT entry events. Passive HFT firms trade more shares than Aggressive HFT firms. After entry a Passive HFT firm trades 28,731 shares daily compared with only 9,699 by an Aggressive HFT firm. Another difference is that Passive HFT firms use aggressive orders seldom, 5.6% of the time, while Aggressive HFTs more frequently use passive orders, 9.3% of the time.
Panel B shows that exits are different. The average trading volume is about half that of entrants, with 12,278. The percent of daily volume is slightly higher at 7.4%, suggesting that exits occur in less actively traded stocks. The percent aggressiveness is much lower for exits. 404 of the exits are by Passive HFT firms, whereas only 26 of the exits are by Aggressive HFT firms.
Table 2 shows the average and standard deviation of the summary statistics by each event-order grouping. Panel A is for entrants, Panel B for exits. The same measures reported in the rows of Table 1 are represented in the columns of Table 2. The rows of Table 2 report the measures by the entrant grouping in a stock, from the first entrants to the sixth and later entrants.
INSERT TABLE 2 ABOUT HERE
A few trends stand out. Later entrants generate more daily volume than earlier entrants. The first entries trade on average 18,391 shares, while the sixth and later entries
trade on average 46,298 shares. The percent of daily volume of new entrant trades decreases, from 14.7% for the first entrants to 4.4% for the sixth and later entrants. Both statistics are consistent with a market growing in the number of participants and in size. Also, later entrant groupings are generally composed of more aggressive HFTs. The average aggressiveness of the first event group is 18.81%, whereas it is 53.5% for the last.
HFT exits show similar trends. Exits when there are 6+ incumbents are done by incumbents with 25.3% aggressiveness, whereas when it is the only HFT left, the exiting firm is 9.9% aggressive. For daily trades the 6+ incumbent exiting firm traded 96 times daily, whereas when there was zero incumbents it was only 13.
Empirical Design
This paper uses a differences-in-differences methodology. For each HFT event (treatment) stock observation, we include contemporaneous observations from a matched control stock not experiencing an HFT event. A stock is ineligible to be a control if it experiences an HFT event in a 20-day window around the participant event date. From the eligible set of control stocks, the selection of the control stocks is based on a nearest-neighbor ranking. Control stocks are ranked by the sum of the squared distance from the treatment stock’s log average market capitalization, log average trading volume, average closing price, average 10-day daily price volatility and average relative bid-ask spread during a 10-day centered window around the participant event date. If any control stock is the nearest neighbor to more than one experimental stock for a particular calendar event date, we match control stocks to experimental stocks using the Gale-Shapley algorithm.
After matching control stocks to treatment stocks, we collect observations of every variable of interest during a 10-day centered window around the HFT event date. Observations are winsorized at three standard deviations from the window average to eliminate outliers. We then construct a dataset of differenced differences. For each window of observations of a variable of interest, we subtract the average before the event date from the average after the event date, minus the same difference for the control stock. There is one differenced difference per event for a total of 815 entry observations and 293 exit observations per variable of interest.
We estimate event impacts by computing the weighted sample average of the differenced differences. We estimate statistical significance by performing a weighted t-test on the sample average. The weight for both statistics is the log of the HFT firm’s percent contribution to daily trading volume for the period in the window that it is active (the period after entry for HFT entrants, the period before exit for exiting HFT firms). The weighting diminishes the influence of HFT events for which the HFT does not contribute much volume and hence cannot be expected to influence trading for the stock.
The methodology of reporting the sample average differenced difference and a t-test on the average is a simple one and avoids dependence on some of the complexities of event-study methodology (Bertrand, Duflo and Mullainathan 2004; Petersen 2009; Thompson 2011). Because there is only one observation per event, there is no need to double cluster the standard errors over a time series and cross section. Because the observations are twice differenced, any fixed effects over stock or over time have been removed through subtraction. The finding of significant impacts is a robust one since we have given up statistical power in reducing the sample size: We lose 19 out of 20
observations when collapsing each panel of ten treatment and ten control observations to a single differenced difference.
Observations of the differenced differences are again winsorized at three standard deviations to eliminate HFT entries that have extreme positive effects—for certain entry dates there is scant trading before a Passive HFT entry, and entry to such stocks can cause spreads to tighten by more than 2000 bps (in one case, a tightening of $2.90).
3. HFT Competition and Market Quality
We identify how stock market quality changes before and after an HFT event by examining the difference in market-quality measures in the differenced differences setting described in the previous section.
We focus on two liquidity measures: Relative Spread and the Dollar Depth to 3 Ticks. Relative Spread is the time-weighted average difference between the best bid and ask price on Alpha. Dollar Depth to 3 Ticks is the time-weighted average dollar volume available at the three price points closest to the midquote on both the bid and the ask side. Markets with lower relative spreads and higher dollar depth indicate greater liquidity.
The results are reported in Table 3 Panel A.
INSERT TABLE 3 ABOUT HERE
The columns to the left report the results for entrants, the column to the right report them for exits. The first row considers all events. The next two rows break down the analysis by
Passive and Aggressive events. Row (4) and (5) consider events without and with incumbents. Rows (6) and (7) consider Passive events without and with incumbents. Rows (8) and (9) do the same but for Aggressive events. The first column reports the number of events in each analysis. The second reports relative spread. Column (3) reports the Dollar depth to 3 ticks. Columns (4) through (6) repeat the analysis for exit events.
The results generally show HFT entry and exit are followed by changes in liquidity. For all events, when an HFT enters, the relative spread decreases by 60.4 basis points, whereas the dollar depth to 3 ticks increases by $4,539. Both are statistically significant at the 1% level. After exits the relative spread increases by 28.4 basis points and depth decreases by $7,881 and both changes are statistically significantly different than zero at the 1% level. The takeaway is that liquidity improves with HFT competition.
Passive events drive the results. When a Passive HFT enters, liquidity improves and the relative spread narrows by 87.4 basis points, whereas when an Aggressive HFT enters the change in relative spread is a 32.0 bps increase. The overall Depth result is also driven by the Passive HFT entrants.
We find that the changes are much larger when there are no HFT incumbents. For entrants, the first HFT entrant is followed by the spread narrowing 145.7 basis points and depth increasing $13,035, whereas when there are incumbents the spread narrows by only 41.3 basis points and the depth increases by only $2,630.
In Table 3 Panel B we examine two measures of market price efficiency: |30-second autocorrelation| and Delay. The |30-second autocorrelation| is the absolute value of the average correlation of the midquote and the midquote in thirty seconds, estimated daily. For Delay we follow Comerton-Forde and Putnins (2014). We estimate a 1-minute
midquote return regression for each stock on each day with the independent variable being the TSX 60 1-minute return, including the contemporaneous return and the 1 – 10 minute lagged returns. The R-squared is captured. The regression is repeated with the coefficients on the lagged TSX 60 returns being restricted to zero. Delay is then 100x(1-(R2constrained / R2unconstrained)). The measure captures how long it takes for market wide information to be reflected in stock prices. The Delay variable varies from 0 to 100 and the larger the value the less efficient the price. The table layout is structured the same as Panel A.
The results in Table 3 Panel B are consistent with HFT entry improving price efficiency and HFT exit preceding a decrease in price efficiency. A lower autocorrelation and delay measure are evidence of more efficient prices.
For all events, when an HFT enters, autocorrelation decreases by .008, and the delay falls by 1.71, both are statistically significant at least the 5% level. After exits the autocorrelation increases by .012 basis points and delay increases by 1.90 and both changes are statistically significantly different than zero at the 1% level. The takeaway is that price efficiency improves with HFT competition.
Surprisingly passive events drive the results. When a Passive HFT enters, autocorrelation decreases by .009, whereas when an Aggressive HFT enters the decrease is larger with a coefficient of -0.01, but it is not statistically significantly different than zero. The decrease in the overall Delay coefficient is also driven by the Passive events.
We find that the changes are much larger when there are no HFT incumbents. For entrants, the first HFT entrant is followed by an autocorrelation decline of .032 and a decrease in delay of 8.01, whereas when there are incumbents the changes for entrants are
not statistically significant. For exits, events with incumbents where an HFT exits experience an autocorrelation increase of .01 and a delay increase of 1.41.
4. HFT Competition and Trading Dynamics
In this section we examine the role of HFT entry on market share, competitive pressures and revenues in order to understand how the industrial organization of HFT firms influences the trading dynamics among them.
Market Shares
We first examine the influence HFT entry has on the fraction of volume generated by the 687 entry events and 327 exit events. The number of events is lower now as we remove from the sample those entries into stocks for which none of the HFT firms were present. It is necessary to remove these entries because there is no incumbent in them to measure. In the following regressions and in all the regressions in this paper, we use the differences-in-differences event-study methodology described in Section 2. Here the dependent variable is HFT volume share, the percent of trading volume (HFT shares traded / 2x total shares traded) executed by HFT firms for stock i on day t. The results are reported in Table 4, Column 1. Each row is a separate analysis with the t-statistic in parentheses. The coefficient can be interpreted as the percent change in HFT activity in stock i following a new HFT entry. The first row shows the results for all events combined. The remaining rows include subgroups of the observations based on their entry number. Events with 1 and 2 incumbents are reported separately. Events with 3, 4, or 5 incumbents are combined
together in the analysis. A similar combined evaluation is performed for when there are 6 or more incumbents.
INSERT TABLE 4 ABOUT HERE
An HFT entry increases the HFT share of volumes. In Column 2, HFT volume share increases by an average of 5.5%, and for the group of first entrants it increases by its maximum of 6.9%. Later HFT entrants contribute less and less to total HFT volume share, and the sixth and later entrants have essentially no impact.
In Table 4, Column 3, we repeat the regression analysis but focus only on incumbent HFTs. The dependent variable is Incumbent HFT volume share, the percent of share trading volume executed by incumbent HFTs (e.g., all HFTs except the new entrant).
While entry increases an HFT's total volume share, it has the opposite effect on the share of incumbents. Incumbent volume share decreases by an average of 2.3%, and over the various entry groups it falls by varying amounts between 0.6% and 4.9%. The early entrants are generating new volume and displacing existing volume from incumbents.
Column 4 reports the average share of volume generated by HFT during the 10 days before entry. The influence of HFT entrants on volume share diminishes according to the amount of volume that is already generated by HFT.
We repeat the exercise for exits in Panel B. The results are generally symmetric. Focusing on the all group (the subgroups have relatively small sample sizes), the results show that after an HFT exit, HFT share of volume decreases by 2.9%, but incumbent HFT volume share increases by 4.5%.
Table 4 measures HFT activity before and after the entry events. Only the first few entrants increase the level of overall HFT volume. There are two takeaways: First, the results suggest competitive behavior among HFT participants. When new entrants join, they displace other market participants by taking their volume, forcing them to become more willing to trade and possibly crowding them out by sometimes trading with them. Second, there appears to only be a certain portion of order flow with which HFT firms want to trade. On Alpha, the threshold portion seems to be around 40% of order flows. Even in competitive markets with more than 6 HFTs, market share by HFT remains around this level.
Liquidity Provision
We have shown that HFT entrants compete against incumbent HFT firms for order flow. Now we examine whether HFT entry is associated with the exercise of competitive pressure. HFT can be used to take or provide liquidity. HFT strategies that use primarily limit orders are associated with market making or dealing and they derive profits off the spread between bid limit orders and ask limit orders (Ait-Sahalia and Saglam 2013; Hoffman 2013). We study whether incumbent HFTs react to Passive HFT events by changing their spreads. In contrast, Aggressive HFT strategies are associated with informed trading, since they trade in the direction of future price movements (Biais, Foucault and Moinas 2013; Foucault, Hombert and Rosu 2012; Martinez and Rosu 2013). With more market participants monitoring for arbitrage opportunities, price impact should decline. We study whether incumbent HFTs are more or less able to trade in the direction of future price changes.
To study the competitive pressure of HFT, we compute two behavioral statistics for all HFT firms whenever they appear in the sample. We compute the best bid-ask spread for HFT incumbents (not the market best bid-ask spread, but the best spread removing all limit orders but those of HFT) and the price impact for the incumbent HFT order flow. We then separate the events that are passive from those that are aggressive. As before we remove from the sample those entries into stocks for which none of the HFT firms were present, and furthermore we remove passive events for which no incumbents are passive and aggressive events for which no incumbents are aggressive.
The sample selection criteria reject many entries, which makes the set of HFT entry events a conservative estimate. For example, there remain only 62 first Passive HFT entries and only 4 first Aggressive HFT entries.
For Passive HFT events, we focus on bid-ask spreads of the HFT firms. Table 3 shows that after HFT entry (exit) market-wide best bid-ask spreads fall (rise). Table 5 shows analysis of how incumbent HFTs adjust the HFT-only best bid-ask spread after HFT events. Since each incumbent may not always quote competitively, we focus on the best HFT quotes among incumbents; that is, conditional on a Passive HFT incumbent being on the limit-order book, we remove all nonHFT orders and measure the bid-ask spread remaining at five-second snapshots (Tightest incumbent spread). The differences-in-differences analysis with the Tightest incumbent spread as the dependent variable are shown in Table 5, Column 1. Each row is a separate regression, with the first considering all observations, and each subsequent row considering different subgroups of Passive HFT events based on the number of previous events.
INSERT TABLE 5 ABOUT HERE
Passive HFT entry leads to a tightening of the best incumbent bid-ask spread. As shown in Column 2, Row 1, incumbents tighten spreads by approximately 3.3 basis points on average. The first entrants trigger the maximum change of -12.2bps, which is statistically significant at the 5% level. The effect on incumbents derives almost entirely from the first entrants. The remaining entrants have smaller coefficients and are statistically insignificant.
The analysis in Column 2 evaluates the change, but does not provide insights into the level of spreads, which is also important if we hypothesize that more competition should lead to lower spreads. We also would like to compare incumbent spreads with new entrant spreads. In Column 3, we report the average tightest incumbent spread before entry, conditional on the presence of a Passive HFT incumbent. For comparison, in Column 4, we report the average new entrant's spread over the 10 days after its entry.
The incumbents post quotes with a 37.4 bps spread, on average, before the new entrant, while the new entrant posts bid and offer prices with a spread of 35.8 bps, on average. Interestingly, the average new entrant spread is only slightly tighter than the incumbents’ spread.
From the regression analysis in Column 2, we see that this competition is good for other market participants, since it makes the incumbent HFT firms decrease their bid-ask spreads. The results are consistent with diminishing returns, with the first few entrants exercising the most competitive pressure. Moreover, columns 3 and 4 show that the best incumbent spreads tighten over time (with exception of the 2 incumbents row). The
competitive quoting by the new HFT entrants is consistent with the earlier result that new entrants take market share from HFT incumbents.
In Panel B the same analysis is carried out for HFT exits. The sample sizes are small so we emphasize the overall result, which is statistically significant. Spreads rise by 11.6 basis points after an HFT exit. Exiters with fewer incumbents have larger impacts than exiters with many incumbents. On average before the exit occurs, HFTs tightest spread is 80 basis points. After the exit it widens to 86.7.
Informativeness
For Aggressive HFT entries, we focus on the predictability of the price path given the record of Aggressive HFT trades. Hirschey (2013) shows that HFT trades predict future returns. Following Hirschey, we examine the ability of HFT firms to trade in the direction of future price changes. We measure the τ-second price impact, which is the number of basis points by which the midquote price changes τ seconds after a trade. A positive price impact indicates informedness. The price-impact measure is created for each HFT stock day by estimating the τ-second midquote-price-impact coefficient of Aggressive HFT incumbent trades using the linear model,
%Δ𝑀𝑖𝑑𝑝𝑜𝑖𝑛𝑡𝑖,𝑗,𝑡,𝜏 = 𝛽𝑖,𝑡,𝜏𝑆𝑖𝑧𝑒𝑖,𝑗,𝑡+ 𝜖𝑖,𝑗,𝑡,𝜏 (3)
where i is the event, j is the trade, t is the day of trade, and τ is the time interval of the price impact. %Δ𝑀𝑖𝑑𝑝𝑜𝑖𝑛𝑡𝑖,𝑗,𝑡,𝜏 is the signed percent price change for trade j over a time interval τ in basis points, and the independent variable 𝑆𝑖𝑧𝑒𝑖,𝑗,𝑡 is the signed trade volume. We
repeat for the 1-second, 5-second, 30-second, and 5-minute price impacts. There is no constant since we assume a zero-volume trade has a price impact of zero. We use the midquote-price impact rather than the execution-price impact to isolate price predictability by avoiding measuring liquidity variation. We report the variable Incumbent price impact as the price impact of the Aggressive HFT incumbents.
A positive price-impact coefficient for event i, day t, and time interval τ, 𝛽𝑖,𝑡,𝜏, indicates informedness at the time horizon τ. If an entrant were to increase price efficiency, incumbents would be left with fewer and smaller price discrepancies to arbitrage.
Since we are interested in informedness, we focus our analysis on Aggressive HFT firms. Because there are so few Aggressive HFT exits with Aggressive HFT incumbents (20) we only evaluate entries. There are 118 entries. In Table 6, we use the differences-in-differences methodology to look at the maximum incumbent price-impact change after a new Aggressive HFT entrant. The coefficient from the differences-in-differences regressions with Incumbent price impact as the dependent variable is reported in Column 2. Each row is a separate regression, with the first considering all observations, and each subsequent row considering different subgroups of Aggressive HFT entries based on the number of previous entries.
INSERT TABLE 6 ABOUT HERE
Panel A reports the 1-second results, Panel B the 5-second, Panel C the 30-second, and Panel D the 5-minute results. In the text we report the one-second results but the findings generally hold for all periods. Aggressive HFT entry decreases the price impact of
the most informed incumbents' trades. The average incumbent price impact decreases on average by 0.8 basis points. Due to the small sample size, interpreting the broken down results is difficult. However, by the time the results are aggregated into the 3-5 incumbent and 6+ incumbent groups the price impacts fall by very small amounts.
Column 3 shows the average incumbent price impact over the 10 days before the HFT entry. The average price impacts generally decrease, albeit not uniformly. Column 4 shows the average new HFT entrant price impact over the 10 days after entry. The new entrants’ price impact generally decreases with as the market crowds. Relative to the HFT incumbents, the new HFT entrants are no more informed. Even with 6+ incumbents the price impact remains above 1 basis point, suggesting that aggressive HFTs as a whole remain informed. The information is not short lived either, as the price impact remains above 1 basis point after 5 minutes (Panel D).
Revenues
The results so far have shown that when an HFT firm enters, the new entrant competes for order flow and, depending on its strategy, either tightens HFT incumbent spreads or erodes the ability of HFT incumbents to predict prices. Both effects should be associated with a decline in revenue for HFT firms. In Table 7, we examine the revenue of HFT firms. We find that the competitive pressures are associated with HFT-sector revenue losses.
We return to using the entire set of entry events. HFT revenues are calculated as the value of sales minus the value of purchases, less taker fees plus maker rebates, with any end-of-day inventory valued at the average daily midquote price. Because some HFT firms
do much more business than others, we take a ratio of an HFT firm's daily revenue to its daily value traded to calculate its revenue per value traded. Column 2 shows the change in total HFT revenue per $10,000 traded after entry using the differences-in-differences methodology. A negative regression coefficient suggests that as more HFT firms enter, revenues decrease.
Column 3 shows the average incumbent HFT revenue per $10,000 for the 10 days before the entry event. Column 4 reports the average HFT revenue per $10,000 for the 10 days after the entry event.
Each row in Table 7 shows a separate regression; the first considers all observations, and each subsequent row considers different subgroups of HFT entries based on the number of previous entries.
INSERT TABLE 7 ABOUT HERE
HFT entry decreases both overall HFT revenues and HFT incumbent revenues. Column 2 shows that entry events reduce incumbents’ revenues, on average, by $130 per $10,000 traded. For exits, overall, HFTs make $124 more per $10,000 traded after an exit. This is driven mainly by the change in profitability when the HFT industrial organization shifts from a duopoly to a monopoly. In these instances, revenues rise $380 per $10,000 traded.
The revenue statistics are noisy and not all of the results by incumbent group are statistically different than zero. Nonetheless, the coefficient sizes are informative. There is a trend in the order of entry. Early HFT entrants increase the revenues that HFT participants
accrue, whereas later HFT entrants decrease the revenues. The impact of entry on incumbents is again competitive: entrants reduce the revenue of incumbents. The impact is largest in the second entry event. While the revenue results are noisy, the size of the coefficients are consistent with the hypothesis that increasing competition reduces revenues.
5. Conclusion
This paper shows the same technological changes that have enhanced trading speed have also changed the industrial organization of the trading landscape. In the last ten years financial markets have adopted computer technology. Today an individual or firm can purchase a server colocated at the exchange and begin trading directly with the market with full access to all available information. It is no longer the case that a select group of traders has exclusive access to information or proximity. HFTs have been among the most prevalent adopters of technology that processes information quickly and reacts quickly. A number of studies already examine the role of speed in markets. In this paper, we analyze a segment of technology adopters, HFT firms, but instead of focusing on their speed, we focus on their competitiveness.
We exploit a four-year microstructure data set that provides anonymous firm-level resolution from the Canadian Alpha trading venue. We isolate hundreds of HFT entry and exit events and study their entry into trading Canadian stocks on the Alpha exchange.
HFT entry (exit) is associated with an improvement (deterioration) in liquidity and price efficiency. When an HFT firm begins trading a stock, it disturbs the trading environment and leads incumbent HFT firms to change their behavior. Part of the incumbents' volume share is lost to the entrant. Competition in providing liquidity leads
incumbents to tighten their spreads. Entry results indicate that incumbent HFT price predictability decreases, consistent with markets becoming more efficient. The culmination is that revenues fall with competition. The influence of both Passive and Aggressive entrants diminishes with each subsequent entry.
The approach in this paper helps to isolate the role of competition from the role of speed and aims to understand the channel by which competition affects markets. Our findings complement papers on HFT market quality. We show that competition among HFT firms, not just speed, plays a role in how they behave in the market and consequently may be partially responsible for the documented relationships between HFT and market quality.
References
Ait-Sahalia, Yacine, and Saglam, Mehmet. 2013. High frequency traders: Taking advantage of speed. Working Paper.
Baron, Matthew, Brogaard, Jonathan, and Kirilenko, Andrei. 2014. The risk and return in high frequency trading. Working Paper.
Battalio, Robert. 1997. Third market broker-dealers: Cost competitors or cream skimmers. Journal of Finance 52 (1), 341-352.
Battalio, Robert, Greene, Jason, and Jennings, Robert. 1997. Do competing specialists and preferencing dealers affect market quality? Review of Financial Studies 10 (4), 969-993.
Bertrand, Marianne, Duflo, Esther, and Mullainathan, Sendhil. 2004. How much should we trust differences-in-differences estimates? Quarterly Journal of Economics 119 (1), 249-275.
Biais, Bruno, Foucault Thierry, and Moinas, Sophie. 2013. Equilibrium fast trading. Working Paper.
Bondarenko, Oleg. 2001. Competing market makers, liquidity provision, and bid–ask spreads. Journal of Financial Markets 4.3 (2001): 269-308.
Breckenfelder, Johannes. 2013. Competition between high-frequency traders, and market quality. Working Paper.
Brogaard, Jonathan, Hagstromer, Bjorn, Norden, Lars, and Riordan, Ryan. 2013. Trading fast and slow: Colocation and market quality. Working Paper.
Brogaard, Jonathan, Hendershott Terrence, and Riordan, Ryan. 2014. High frequency trading and price discovery. Review of Financial Studies, forthcoming.
Brogaard, Jonathan, Hendershott, Terrance, and Riordan, Ryan. 2013. High Frequency Trading and the 2008 Short-Sale Ban. Working Paper.
Carrion, Allen. 2013. Very fast money: high-frequency trading on the nasdaq. Journal of Financial Markets 4, 680-711.
Comerton-Forde, Carole, and Tālis Putniņš. 2014. Dark trading and price discovery. Working Paper.
Dennert, Jurgen. 1993. Price competition between market makers. Review of Economic Studies 60, 735-751.
Foucault, Thierry, Hombert, Johan, and Rosu, Ioanid. 2012. News trading and speed. Working Paper.
Gai, Jiading, Yao Chen, and Ye, Mao. 2012. The externalities of high-frequency trading. Working Paper.
Hagstromer, Bjorn, and Norden, Lars. 2013. The diversity of high-frequency traders. Journal of Financial Markets 16, 741-770.
Hasbrouck, Joel, and Saar, Gideon. 2013. Low-latency trading. Journal of Financial Markets 16, 646-679.
Hirschey, Nicholas. 2013. Do High Frequency Traders Anticipate Buying and Selling Pressure? Working Paper.
Hoffmann, Peter. 2013. A dynamic limit order market with fast and slow traders. Working Paper.
Huang, Roger. 2002. The quality of ECN and nasdaq market maker quotes. Journal of Finance, 57 (3), 1285-1319.
Jovanovic, Boyan, and Menkveld, Albert. 2010. Middlemen in limit-order markets. Working Paper.
Klock, Mark, and McCormick, Timothy. 1999. The impact of market maker competition on nasdaq spreads. Financial Review 34, 55-74.
Martinez, Victor, and Rosu, Ioanid. 2013. High frequency traders, news and volatility. Working Paper.
Mayhew, Stewart. 2002. Competition, market structure, and bid-ask spreads in stock option markets. Journal of Finance 57 (20), 931-958.
Menkveld, Albert. 2013. High frequency trading and the new-market makers. Working Paper.
Pagnotta, Emiliano, and Philippon, Thomas. 2011. Competing on Speed. Working Paper. Petersen, Mitchell. 2009 Estimating standard errors in finance panel data sets: Comparing
approaches. Review of Financial Studies 22 (1), 435-480.
Thompson, Samuel. 2011. Simple formulas for standard errors that cluster by both firm and time. Journal of Financial Economics 99 (1), 1-10.
Van Ness, Bonnie F., Van Ness Robert A., and Warr, Richard S. 2005. The impact of market maker concentration on adverse-selection costs for nasdaq stocks. Journal of Financial Research 28 (3), 461-485.
Weller, Brian. 2013. Intermediation chains and specialization by speed: Evidence from commodity futures markets. Working Paper.
Weston, James P. 2000. Competition on the nasdaq and the impact of recent market reforms. Journal of Finance 55 (6), 2565-2598.
Figure 1: Trading in Canada. This figure plots the dollar-volume market share of the Alpha exchange in the TSX 60 over the sample period, as well as the total trading dollar volume of the TSX 60 in Canada. The market shares are calculated using Bloomberg data on the monthly trading volume of the 279 sample stocks.
Dollar Trading Volume and Alpha’s Market Share
$0 $50,000 $100,000 $150,000 $200,000 $250,000 $300,000 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 11/ 2008 1/2 009 3/2 009 5/2 009 7/2 009 9/2 009 11/ 2009 1/2 010 3/2 010 5/2 010 7/2 010 9/2 010 11/ 2010 1/2 011 3/2 011 5/2 011 7/2 011 9/2 011 11/ 2011 1/2 012 3/2 012 5/2 012 7/2 012 9/2 012 $ -Vo lu m e in M ill io n s A lp h a % o f $ -Vol u m e
Figure 2: Time series of entrants. This figure plots the time series of HFT entries and exits from 01/01/2009 to 09/25/2012. There are 895 entry events and 453 exit events. Panel A displays the entry events, Panel B the exit events.
Panel A: HFT Entry Events
Panel B: HFT Exit Events
0
20
40
60
80
100
1/1/2009
1/1/2010
1/1/2011
1/1/2012
0
5
10
15
20
1/1/2009
1/1/2010
1/1/2011
1/1/2012
Figure 3: Aggressiveness over the entry and exit events. This figure plots the number of events into buckets based on the HFT firms’ use of marketable trades as a proportion of all trades it submits (aggressiveness). There are 895 entry events and 453 exit events. Panel A displays the entry events, Panel B the exit events.
Panel A: HFT Entry Events
Panel B: HFT Exit Events
0
100
200
300
400
500
10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Number
of
HF
T
En
tr
ie
s
% of Volume from Aggressive Trades
0
50
100
150
200
250
300
350
10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Num
ber
of
HFT
Exit
s
Table 1: Sample summary trading statistics for entrants. This table gives sample summary trading data for the entrants during the 10 days after their entry. Daily volume is the average number of shares the entrant traded daily. Daily trades is the average number of trades the entrant executed daily. Per cent daily volume is the average per cent of total daily volume the entrant contributed. Overnight volume, % of daily volume is the average per cent of its own daily volume the entrant held at the end of day. Switching rate is the percentage of the entrant's trades for which the entrant switched the direction of its trade— buying after it had sold or selling after it had bought. Per cent aggressive is the percentage of the entrant's trades for which the entrant initiated the trade (either through a market order or a marketable limit order). Panel A shows statistics for entries, and Panel B shows statistics for exits.
Panel A: Entries
All Events Passive Events Aggressive Events
Mean Std. 25% Med. 75% Mean Std. 25% Med. 75% Mean Std. 25% Med. 75% Daily volume 23345 98139 1100 4800 18600 28731 121034 800 5000 23600 9699 19708 1400 3700 8800 Daily trades 92 188 6 27 85 105 215 4 25 97 55 102 11 26 56 Percent of daily volume 7.1% 8.9% 2.0% 3.8% 7.8% 8.7% 10.6% 2.1% 4.3% 10.0% 3.9% 2.9% 1.7% 3.2% 5.2% Overnight vol., % of daily vol. 20.6% 17.8% 4.4% 16.4% 32.1% 20.6% 16.5% 8.4% 16.4% 29.6% 19.1% 19.7% 1.8% 6.2% 34.9% Switching rate 34.4% 19.5% 19.1% 35.1% 46.0% 28.2% 15.8% 13.0% 30.9% 41.9% 49.2% 20.4% 30.5% 55.7% 65.5% Percent aggressive 32.9% 38.6% 0.6% 9.9% 72.2% 5.6% 8.5% 0.1% 1.3% 6.9% 90.7% 10.8% 81.8% 97.0% 100.0%
N 895 567 241
Panel B: Exits
All Events Passive Events Aggressive Events
Mean Std. 25% Med. 75% Mean Std. 25% Med. 75% Mean Std. 25% Med. 75% Daily volume 12278 63476 0 500 5100 13326 67079 0 500 5500 5158 10382 200 700 4700 Daily trades 60 204 0 2 21 64 215 0 2 21 36 67 1 6 36 Percent of daily volume 7.4% 11.0% 0.8% 2.9% 9.2% 7.7% 11.2% 0.9% 3.0% 9.8% 5.2% 8.0% 0.2% 2.7% 7.1% Overnight vol., % of daily vol. 8.3% 8.7% 2.5% 5.4% 10.8% 7.0% 7.2% 2.3% 5.0% 9.2% 21.1% 14.1% 5.3% 24.2% 33.6% Switching rate 7.2% 6.6% 2.3% 5.1% 11.0% 6.8% 6.1% 2.2% 4.9% 10.1% 13.3% 9.8% 5.0% 10.5% 25.1% Percent aggressive 11.3% 22.2% 0.0% 1.1% 10.6% 4.4% 7.3% 0.0% 0.5% 5.6% 86.2% 11.8% 78.9% 86.7% 100.0%
Table 2: Sample summary trading statistics for entrants by event order. This table gives summary trading statistics for the entrants during the 10 days after their entry (exit), grouped by the order in which the entrant arrived (exit left) relative to the other entry events for a stock. Daily volume is the average number of shares the participant traded daily. Daily trades is the average number of trades the entrant executed daily. Per cent daily volume is the average per cent of total daily volume the entrant contributed. Per cent volume held at end of day is the average percent of its own daily volume that the entrant held at the end of day. Switching rate is the percentage of the entrant's trades for which the entrant switched the direction of its trade—the per cent of the time it bought after it had sold or sold after it had bought. Percent aggressive is the percentage of the entrant's trades for which the entrant initiated the trade (either through a market order or marketable limit order).
Panel A: Entries
Before Event # Daily vol. (1000s) Daily trades % Daily vol. % Vol. EOD Switching rate % aggressive of HFT N Mean Std. Mean Std. Mean Std. Mean Std. Mean Std. Mean Std.
0 204 18391 48498 47 82 14.7% 0.14 23.1% 0.16 32.4% 0.15 18.8% 0.31 1 188 27575 190743 52 113 5.5% 0.06 27.9% 0.18 24.7% 0.15 26.5% 0.34 2 131 13493 29682 49 81 4.3% 0.04 22.4% 0.18 30.8% 0.20 30.4% 0.37 3 117 20097 34935 110 188 4.5% 0.04 16.0% 0.18 39.4% 0.22 39.8% 0.39 4 90 14732 23654 87 124 4.5% 0.04 19.0% 0.19 40.8% 0.21 47.1% 0.42 5 89 33649 78137 194 311 5.2% 0.04 13.1% 0.14 42.8% 0.18 41.0% 0.40 6+ 76 46298 70557 244 347 4.4% 0.04 10.7% 0.12 45.5% 0.21 53.5% 0.44 Panel B: Exits
Before Event # Daily vol. (1000s) Daily trades % Daily vol. % Vol. EOD Switching rate % aggressive of HFT N Mean Std. Mean Std. Mean Std. Mean Std. Mean Std. Mean Std.
2 114 3040 10100 13 34 9.2% 0.15 7.1% 0.07 5.5% 0.05 9.9% 0.20 3 54 4647 10324 12 31 7.2% 0.08 7.3% 0.07 5.9% 0.05 6.9% 0.15 4 64 8699 83957 9 24 3.9% 0.06 7.6% 0.08 4.7% 0.05 8.8% 0.20 5 59 3407 7667 19 38 6.4% 0.09 8.4% 0.10 7.2% 0.07 7.7% 0.21 6 49 14512 36434 94 279 8.5% 0.12 8.2% 0.07 8.8% 0.07 8.4% 0.14 7 61 34630 70491 224 392 11.5% 0.10 10.2% 0.09 13.0% 0.07 14.6% 0.24 8 52 26598 133798 96 243 3.7% 0.06 11.0% 0.13 7.3% 0.06 25.3% 0.33
Table 3: HFT competition and market-quality measures. The columns of this table show the sample average differenced difference HFT volume shares and associated t statistics. For each treatment stock, a nearest-neighbor control stock is matched by the sum of squared log market capitalization, log daily trading volume, closing price, 10-day price volatility and relative bid-ask spread during the event window. The first row shows the results for all events combined. The remaining rows show subgroups of the observations based on the number of incumbents. The Columns to the left are for entrant events; the Columns to the right are for exit events. Panel A considers liquidity measures. Relative Spread isthe time-weighted average difference between the best bid and ask price on Alpha. Dollar Depth to 3 Ticks is the time-weighted average dollar volume at the three bid and ask price points closest to the midquote. Panel B considers efficiency measures. The |30-second autocorrelation| is the absolute value of the average correlation of the midquote and the midquote in thirty seconds estimated daily. Delay is 100x(1-(R2constrained / R2unconstrained)). The R-squared are taken from a 1-minute midquote return regression for each stock-day with the independent variables being the TSX 60 1-minute return, including the contemporaneous return and the 1 – 10 minute lagged returns. The regression is repeated with the coefficients on the lagged TSX 60 returns restricted to zero. The window of observation is 10 days before and after the event. *, **, *** represent statistical significance at the 10%, 5%, and 1% level.
Panel A: Liquidity Entrants Exits
Group N Relative spread $ depth to 3 ticks N Relative spread $ depth to 3 ticks All events 815 -60.4bps*** $4539*** 293 28.4bps*** -$7881*** (-6.08) (3.37) (5.17) (-3.33) Passive events 487 -87.4bps*** $5832*** 273 30.2bps*** -$8260*** (-6.98) (3.38) (5.2) (-3.29) Aggressive events 242 32.0bps* -$671 0 . . (1.76) (-0.27) No-incumbent events 70 -145.7bps** $13035*** 49 96.7bps*** -$3002*** (-2.26) (3.54) (3.78) (-2.68)
Events with incumbents 745 -41.3bps*** $2630* 244 13.4bps*** -$8951***
(-6.33) (1.81) (4.12) (-3.15)
Passive, no-incumbent events 42 -208.5bps*** $13464*** 43 101.4bps*** -$3024**
(-2.85) (2.84) (3.65) (-2.53)
Passive events with incumbents 445 -53.5bps*** $3689* 230 14.5bps*** -$9420***
(-6.43) (1.96) (4.24) (-3.13)
Aggressive, no-incumbent events 23 274.8bps $13062* 0 . .
(1.62) (1.75)
Aggressive events with incumbents 219 2.1bps -$2,363 0 . .
Table 3 Continued. Panel B: Price Efficiency
Entrants Exits
Group N |30 sec. autocorrelation| Delay N |30 sec. autocorrelation| Delay All events 815 -0.008** -1.71*** 293 0.012*** 1.90*** (-2.56) (-3.12) (2.64) (2.66) Passive events 487 -0.009** -2.25*** 273 0.012*** 2.41*** (-2.36) (-3.55) (2.59) (3.38) Aggressive events 242 -0.01 0.20 0 . . (-1.49) (0.16) No-incumbent events 70 -0.032** -8.01*** 49 0.0224** 4.04** (-2.24) (-3.39) (2.18) (2.38)
Events with incumbents 745 -0.003 -0.307 244 0.01* 1.41*
(-0.92) (-0.59) (1.93) (1.79)
Passive, no-incumbent events 42 -0.034** -9.08*** 43 0.023** 3.96**
(-2.27) (-3.47) (2.19) (2.18)
Passive events with incumbents 445 -0.002 -0.33 230 0.01* 2.06***
(-0.61) (-0.56) (1.87) (2.65)
Aggressive, no-incumbent events 23 -0.05 3.04 0 . .
(-1.28) (0.44)
Aggressive events with incumbents 219 -0.004 -0.12 0 . .
Table 4: HFT competition and HFT's share of volume. Columns 2 and 3 of this table show the sample average differenced difference HFT volume shares and associated t scores. For each treatment stock, a nearest-neighbor control stock is matched by the sum of squared log market capitalization, log daily trading volume, closing price, 10-day price volatility and relative bid-ask spread during the event window. The first row shows the results for all events combined. The remaining rows show subgroups of the observations based on the number of incumbents. In Column 2, the variable of interest is HFT volume share, the average per cent of trading volume (HFT shares traded / 2x total shares traded) executed by any of the 11 HFT firms. In Column 3, the variable of interest is Incumbent HFT volume share, the average per cent of a stock's total daily trading volume generated by any of the 10 HFT participants present other than the entrant. Columns 4 and 5 show the average HFT volume share before and after the event date. The window of observation is 10 days before and after the event. *, **, *** represent statistical significance at the 10%, 5%, and 1% level, respectively.
Panel A: Entrants Group N % Change, HFT volume share after entry % Change, incumbent HFT volume share after entry HFT volume share before entry HFT volume share after entry Before Event # 687 5.5%*** -2.3%*** 22.8% 29.3% of HFT (18.80) (-8.99) 1 185 6.9%*** -2.3%*** 17.3% 26.7% (11.57) (-4.88) 2 130 6.8%*** -0.6% 15.6% 22.3% (6.71) (-1.16) 3–5 296 5.0%*** -2.3%*** 25.6% 31.1% (11.63) (-5.84) 6+ 76 1.6%*** -4.9%*** 39.6% 41.9% (2.25) (-5.99) Panel B: Exits Group N % Change, HFT volume share after exit % Change, incumbent HFT volume share after exit HFT volume share before exit HFT volume share after exit Before Event # 327 -2.9%*** 4.5%*** 32.0% 28.6% of HFT (-6.29) (10.20) 2 52 -7.3%*** 1.1% 18.8% 12.1% (-5.84) (1.15) 3 63 -7.2%*** 0.9% 21.2% 13.5% (-6.51) (1.25) 4–6 166 -1.1%* 6.3%*** 36.0% 34.2% (-1.82) (9.99) 7+ 46 -3.1%** 2.0%** 40.5% 36.9% (-2.52) (2.04)
Table 5: Passive HFT competition and incumbent spreads. Column 2 of this table shows the sample average differenced difference best HFT relative bid-ask spread and associated t scores for the Passive HFT events. Passive HFTs are those that use marketable orders less than 33% of the time. For each treatment stock, a nearest-neighbor control stock is matched by the sum of squared log market capitalization, log daily trading volume, closing price, 10-day price volatility and relative bid-ask spread during the event window. The first row shows the results for all events combined. The remaining rows include subgroups of the observations based on the number of incumbents. In Column 2, the variable of interest is the Tightest incumbent spread, the best (tightest) relative bid-ask spread provided among the Passive HFT incumbents conditional on a Passive HFT incumbent being present trading. The order book is observed at 5-second snapshots. Columns 3 and 4 show the average tightest incumbent spread for the treatment stock before and after the event date. The window of observation is 10 days before and after the event. *, **, *** represent statistical significance at the 10%, 5%, and 1% level, respectively.
Panel A: Entries
Group N
Change, tightest incumbent spread after entry
Average tightest incumbent spread before entry Average entrant spread after entry Before Event # 283 -3.3bps* 37.4bps 35.8bps of Passive HFT (-1.70) 1 62 -12.2bps** 66.3bps 60.4bps (-2.48) 2 56 -2.5bps 70.7bps 66.8bps (-0.35) 3–5 134 -1.1bps 21.1bps 21.7bps (-0.56) 6+ 31 1.6bps 10.3bps 10.8bps (1.38) Panel B: Exits Group N Change, tightest incumbent spread after exit
Average tightest incumbent spread
before exit
Average tightest incumbent spread after exit Before Event # 218 11.6bps* 80.0bps 86.7bps of Passive HFT (1.92) 2 25 37bps 226.6bps 249.5bps (0.99) 3 41 79.7bps** 367.7bps 427.2bps (3.11) 4–6 121 21.1bps 20.4bps 23.5bps (1.60) 7+ 30 0.80bps 12.3bps 12.2bps (1.02)
Table 6: Aggressive HFT competition and incumbent informedness. Column 2 of this table shows the sample average differenced difference HFT incumbent price impact and associated t scores for the Aggressive HFT events. Aggressive HFTs are those that use marketable orders more than 75% of the time. The price impact of HFT incumbents is estimated by grouping all incumbent trades and fitting 𝛥𝑚𝑖𝑑𝑞𝑢𝑜𝑡𝑒𝑡,𝑡+5 = 𝛽 𝑠𝑖𝑧𝑒𝑡+ 𝜖𝑖,𝑗,𝑡, where Δmidquotet,t+5 is the change in the midquote in basis points five seconds after the
trade and size is the size of the trade in shares (typically 100 or 200). For each treatment stock, a nearest-neighbor control stock is matched by the sum of squared log market capitalization, log daily trading volume, closing price, 10-day price volatility and relative bid-ask spread during the event window. The first row shows the results for all events combined. The remaining rows include subgroups of the observations based on their entry number. In Column 2, the dependent variable is Incumbent price impact, the estimated price impact in basis points five seconds after a trade of size 100 (i.e. we report β*100). Columns 3 and 4 show the incumbent price impact before and after the event date. The window of observation is 10 days before and after the event. *, **, *** represent statistical significance at the 10%, 5%, and 1% level, respectively.
Panel A: Entries, 1 second price impacts
Group N
Change, incumbent price impact after
entry
Average incumbent price impact before
entry
Average incumbent price impact after
entry Before Event # 118 -0.8bps** 1.8bps 1.3bps of HFT (-2.42) 1 4 -1.0bps 2.0bps 2.0bps (-1.51) 2 15 -3.5bps** 3.6bps 0.8bps (-2.59) 3–5 63 -0.1bps 1.4bps 1.5bps (-0.19) 6+ 36 -0.3bps 1.3bps 1.2bps (-0.92)
