The Need for Speed: Minimum Quote Life Rules and Algorithmic. Trading

Transcription

1 The Need for Speed: Minimum Quote Life Rules and Algorithmic Trading Alain Chaboud Erik Hjalmarsson Clara Vega [PRELIMINARY AND INCOMPLETE. PLEASE DO NOT CITE.] May 4, 2015 Abstract We study the impact that a minimum-quote-life (MQL) rule of 250 milliseconds in the eurodollar currency pair had on trading volume, bid-ask spreads, market depth, and adverse selection costs for slow traders. We find that the MQL rule causes trading volume to decline in the EBS trading platform, suggesting that some market participants may have moved to other trading venues without MQL rules. In addition, we find that the MQL rule did not have an impact on average bid-ask spreads, which suggests that the counterbalancing effects of lower adverse selection costs and higher market-making costs for HFTs play a role in a way that they cancel each other out. However, the MQL might have increased bid-ask spreads in volatile times. The impact on depth is inconclusive. We also find some (weak) evidence that HFTs are less willing to provide liquidity during volatile times after the MQL rule was implemented. Finally, consistent with the recent theoretical literature on HFT, we find that slow traders pay a premium (adverse selection cost) when they transact with better informed high-frequency traders, and that these costs declined after the MQL rule was implemented. JEL Classification: F3, G12, G14, G15. Keywords: Algorithmic trading; Price discovery. Chaboud and Vega are with the Division of International Finance, Federal Reserve Board, Mail Stop 43, Washington, DC 20551, USA; Hjalmarsson is with University of Gothenburg, Department of Economics and Centre for Finance, Vasagatan 1, SE Gothenburg, Sweden, and Queen Mary University of London, School of Economics and Finance, Mile End Road, London E1 4NS, UK. Please address comments to the authors via at [email protected], [email protected], and [email protected]. We are grateful to EBS/ICAP for providing the data. We have benefited from the comments of Ola Bengtsson, Johannes Breckenfleder, Jungsuk Han, Chester Spatt, Per Strömberg, and of participants in the Lund University Finance Seminar, and the Swedish House of Finance Seminar. The views in this paper are solely the responsibility of the authors and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System or of any other person associated with the Federal Reserve System.

2 1 Introduction Recent theoretical papers suggest that fast or high-frequency traders (HFTs) may have a negative impact on market quality. Specifically, Foucault, Hombert and Roşu (2013), Budish, Cramton and Shim (2013), and Biais, Foucault, and Moinas (2013), among others, argue that the participation of HFTs may reduce market liquidity due to increased adverse selection costs. Consistent with this negative view, financial market regulators and trading platforms have recently implemented policies with the intention of slowing down HFTs. In this paper, we analyze the impact of one of these policies, the implementation of a minimum-quote-life (MQL) rule of 250 milliseconds for the euro-dollar currency pair, on June 15, 2009, by the Electronic Broking Services (EBS) FX trading platform. We rely on a novel data set consisting of several years (Jan 2008 to December 2010) of highfrequency transaction and quote data from EBS. The data represent a large share of spot interdealer transactions across the globe in the euro-dollar currency pair, with EBS widely considered to be the primary site of price discovery in this exchange rate during our sample period. A crucial feature of the data is that, on a second-by-second frequency, the volume and direction of transactions of manual (slow), Bank-AI (fast), and PTC-AI (fastest) traders are explicitly identified, allowing us to measure changes in their respective behavior due to the MQL rule. 1 We first use our data to study the impact of MQL on trading volume. Theory predicts counterbalancing effects. The posting of a limit order essentially provides a free option to the market to trade at the price posted. The value of an option is increasing in its time-to-maturity, and an MQL rule effectively imposes a lower bound on this time-to-maturity. Thus, the MQL rule makes the option provided by the limit order more valuable, and market making thereby becomes more expensive (or risky). The MQL rule we study, a 250 milliseconds minimum resting time, only affects the ability of algorithmic and high-frequency traders to cancel their posted quotes. 2 We thus expect the MQL rule to have the biggest impact on the behavior of fast traders. Some fast traders, 1 In the EBS trading platform, traders can enter trading instructions manually, using an EBS keyboard, or, upon approval, via a computer directly interfacing with the system, what EBS labels automated interface (AI) traders. EBS further breaks down AI traders into two groups, Professional Trading Community (PTC) traders and Banks. PTC traders are hedge funds and CTAs (commodity trading advisors). These traders, according to EBS, are traders with the lowest latency, or what the media refers to as high frequency traders. In our analysis we classify manual traders as slow, Bank-AI as fast, and PTC-AI as the fastest. 2 EBS indicated to us that there is no manual trader in the EBS trading platform who is able to post and cancel an order in less than 250 milliseconds. This makes sense as the blink of a human eye lasts approximately 400 milliseconds. 1

3 for example, may decide to go to other trading venues, where market making is less expensive from their point of view. As a result, some of the slow market participants may also leave because the fast market participants, who provided a market making service, left. Thus, the MQL rule may cause a decline in overall trading volume because of the departure of several types of market participants. In contrast to the option theory, Foucault, Hombert and Roşu (2013), Budish, Cramton and Shim (2013), and Biais, Foucault, and Moinas (2013) highlight that the speed advantage of HFTs the ability to post and cancel quotes quickly gives HFTs an informational advantage. The MQL rule takes away this informational advantage from HFTs, thereby lowering adverse selection costs for slow traders. The impact that lower adverse selection costs for slow traders has on trading volume is ambiguous. Trading volume may increase if the arrival of new slow traders due to lower adverse selection costs is bigger than the departure of HFTs because they do not have an informational advantage any more. To investigate the impact MQL had on trading volume, we use a difference-in-differences methodology using trading volume in the futures market as the comparison group. We find that trading volume in EBS declines relative to trading volume in the futures market. Interestingly, we find evidence of an increase in trading volume in the futures market, suggesting that some market participants may have moved to the futures market following the implementation of the MQL rule in the spot market. We next study the impact of MQL on bid-ask spreads and market depth (i.e., the quantity available to trade at posted quotes). Similar to the predictions related to trading volume, theory predicts counterbalancing effects. Option theory predicts that the MQL rule will cause market making to be more expensive or risky for HFTs, which in turn implies that the HFT market-makers, who decided to stay in the EBS market, may demand a higher compensation for their marketmaking services by posting wider bid-ask spreads and perhaps providing less depth at the top of the limit order book. Also, market-making competition may be lower after the implementation of the MQL rule, because some HFTs may leave the market as conjectured above, which in turn would increase bid-ask spreads (Li (2013)). However, the MQL rule might also lower competition amongst aggressive HFT liquidity demanders, as the MQL rule takes some of their informational advantage away from them, and lower competition amongst this type of HFTs would lower bid-ask spreads, as Breckenfelder (2013) shows. As explained above, this type of HFTs leaving the market would also 2

4 lower adverse selection costs. These latter two effects would cause bid-ask spreads to decline and market depth to increase. Our difference-in-difference estimation shows that, on average, MQL had little, or no, impact on bid-ask spreads, suggesting that the counterbalancing effects play a role in a way that they exactly cancel each other out. 3 The results for market depth are harder to interpret, with a large increase in depth on the EBS market subsequent to the MQL implementation, but an even larger increase is seen in the Futures market, resulting in a negative difference-in-difference effect. Taken at face value, the difference-in-difference would thus suggest that the MQL rule had a (large) negative impact on market depth. However, given that both the EBS and futures market in fact exhibited great increases in depth around the implementation of the MQL rule, it is diffi cult to put much faith in the difference-in-difference result for depth. Option theory predicts that the effect of an MQL rule might be particularly noticeable during volatile times. The value of an option increases with time to maturity and volatility. Thus, the value (or cost) of the imposed minimum time-to-maturity will be largest in times of high volatility. Since volatile times might be associated with times of market stress, these are arguably the periods during which the access to liquidity is most important. An MQL rule, however, is likely to make liquidity provision in volatile times relatively scarcer. We find some weak evidence that high-frequency traders are less willing to provide liquidity during volatile times, but these results are only seen in the shortest time window immediately following the MQL implementation. Over longer windows, this results is no longer present. We do, however, find relatively strong and consistent evidence that bid-ask spreads are slightly higher during volatile times after the MQL rule is implemented in the system. Finally, we investigate further the idea that slow traders pay a premium (adverse selection cost) when they transact with better informed high-frequency traders, as predicted by the theoretical papers of Hombert and Roşu (2013), Budish, Cramton and Shim (2013), and Biais, Foucault, and Moinas (2013). To expore this possibility we estimate the effective spread (difference between the transacted price and the mid-quote) paid by different types of traders, slow, fast and fastest, when they are trading with high-frequency or the fastest traders. We find that both the slow 3 Another possibility is that the 250 ms minimun-resting-time for orders is too short to have any impact on the market. We rule this possibility out by showing that MQL did have an impact on trading volume. We also show that the 250 ms MQL rule is binding, i.e., prior to the implementation of the rule there were several market participants who cancelled their quotes within 250 ms. Below, we also show that MQL did have an impact on adverse selection costs for slow traders. 3

5 traders and the fast traders pay an effective spread that is approximately 0.2 pips higher than the spread paid by the fastest traders (the average spread is of an order of 1 pip). 4 One can view these differences in effective spreads as compensation for HFT s investment in technology, and a social planner may not find the same effective spread across all trader categories to be optimal, as previous studies show that traders benefit from the presence of algorithmic traders (e.g., Chaboud, Chiquoine, Hjalmarsson, and Vega (2014) find that algorithmic trading participation increases the informational effi ciency of prices and Hendershott, Jones, and Menkveld (2011) find that algorithmic traders improve market liquidity). Nevertheless, we test the hypothesis, suggested by the theoretical literature, that an MQL rule may lower adverse selection costs for slow market participants, and compare effective spreads paid by different types of traders before and after the MQL rule. We find evidence that there is a small decline in effective spreads paid by slow traders after the MQL rule implementation. In other words, the MQL rule might have accomplished what EBS intended to do, namely to lower adverse selection costs for slow traders. The paper proceeds as follows. In Section 2, we briefly discuss the related empirical literature on the impact of policies aimed at slowing down high frequency traders. Section 3 introduces the high-frequency data used in this study, and offers a short description of the structure of the market. Section 4 shows that the 250 millisecond rule was indeed binding for some market participants and in Section 5, we test whether there is evidence that the MQL rule had a causal impact on trading volume, bid-ask spreads, and depth. Section 6 discusses whether trading volume by computers were differently affected by the MQL rule than human trading volume. Section 7 investigates conditional volatility effects of the MQL rule and, Section 8 consider the possibility that slow traders pay an adverse selection cost premium when transacting with high-frequency traders, and whether this adverse selection cost goes down after the MQL rule was implemented. Finally, Section 9 concludes. 2 Related literature The posting of a limit order essentially provides a free option to the market to trade at the price posted. Classical option theory therefore provides some basic ideas of the effects of an MQL rule. 4 An alternative explanation is that high-frequency traders provide liquidity to other market participants when liquidity is expensive. We rule out this explanation because we compare effective spreads paid by different types of traders after controlling for the inside quoted spread. 4

6 The value of an option is increasing in its time-to-maturity, and an MQL rule effectively imposes a lower bound on this time-to-maturity. Thus, the MQL rule makes the option provided by the limit order more valuable, and market making thereby becomes more expensive (or risky). Since the primary compensation for market making activities is collected through the bid-ask spread, all else equal, one would expect the bid-ask spread to increase after the imposition of an MQL rule. Somewhat more speculatively, one might also expect that depth decreases, as a result of a decreased willingness to post quotes. To the extent that alternative trading venues exist, one might also posit a decrease in overall transacted volume, as market participants move to cheaper platforms. The implications from classical theory are thus quite negative, and this view is reflected in normative statements by many leading academics. 5 Another immediate implication of classical option theory is that the effect of an MQL rule might be particularly noticeable during volatile times. The value of an option increases with time to maturity and volatility. Thus, the value (or cost) of the imposed minimum time-to-maturity will be largest in times of high volatility. Since volatile times might be associated with times of market stress, these are arguably the periods during which the access to liquidity is most important. An MQL rule, however, is likely to make liquidity provision in volatile times relatively scarcer. The above insights would appear quite general, in the sense that they derive from essentially model-free results in option theory. However, these implications are likely best seen as the outcome of a comparative static analysis, where the liquidity conditions in a market with pre-dominantly high-frequency market makers is compared before and after an MQL rule is imposed. In a recent theory paper, Han, Khapko, and Kyle (2014) analyze the equilibrium bid-ask spread in a simple model where there may or may not be high-frequency market makers present in the market. If the probability of high-frequency order posting is equal to one, this results in bid-ask spreads that are always smaller or equal to those observed in the case without high-frequency market makers (i.e., probability of high-frequency order posting equal to zero). 6 In a situation where high-frequency 5 The following quotes sum up well the scepticism of MQL rules among many academics. The independent academic authors who have submitted studies are unanimously doubtful that minimum resting times would be a step in the right direction... (Linton, O Hara, and Zigrand in Foresight); This will result in a dry up of liquidity in the book and might even cause an increase in volatility. To be honest, we think this proposal is a terrible idea... (Farmer and Skouras, in Foresight); However, the minimum time-in-force appears to be a particularly blunt, poorly considered tool. (Jones, 2013). 6 This essentially confirms the discussion above, since the known absence or presence of high-frequency market makers can be viewed as analogous to a real-world situation with or without an MQL rule in place. In practice, the exact time-constraints imposed on order withdrawals by the MQL rule will, of course, have differing implications. 5

7 participation is possible but not certain, these unambiguous implications no longer hold, and bid-ask spreads might at times, and on average, end up higher than in the case where there is no highfrequency participation for certain. This reflects the additional adverse selection risk faced by the low-frequency makers, since they are not able to update their quotes in the case of new information arrivals, unlike the high-frequency makers. In this case, the imposition of an appropriate MQL rule could lower average bid-ask spreads by essentially bringing the market (closer) to the situation where there is no high-frequency participation. Another intuitive implication of an MQL rule is that prices might become less informationally effi cient, as market makers are restricted in updating their quotes given new available information. Given the very short time-horizons that might be affected by any MQL rule in modern markets (e.g., 250 milliseconds in the EBS case), this potential ineffi ciency is perhaps not of great importance, but highlights once again the trade-off between effi ciency and adverse selection that has been a common theme throughout much of the theoretical literature on high-frequency trading. 3 Market structure and data description 3.1 The EBS interdealer market Global interdealer spot trading in major currency pairs (exchange rates) is dominated by two different trading platforms: Electronic Broking Services (EBS) and Reuters. The euro-dollar trade primarily on EBS, and the price-discovery process for this currency pair is therefore concentrated to the EBS trading platform. Prices on EBS constitute the reference for derivative pricing and dealer quotes in these currencies. The EBS system is an interdealer system accessible to foreign exchange dealing banks and, under the auspices of dealing banks (via prime brokerage arrangements), to hedge funds and commodity trading advisors (CTAs). EBS controls the network and each of the terminals on which the trading is conducted. The minimum trade size over our sample period is 1 million of the base currency (i.e., the euro in our case), and trade sizes are only allowed in multiple of millions of the base currency. On June 15, 2009, EBS imposed a so-called minimum-resting-time, or minimum-quote-life However, in a highly stylized model such as the one studied by the Han, Khapko, and Kyle (2014), the analogy seems valid, and the authors themselves partly intepret their results in light of minimum quote life rules. 6

8 (MQL) rule, of 250 milliseconds for the euro-dollar currency pair. Specifically, this new rule specified that any limit order posted on the euro-dollar market could not be withdrawn within the first 250 milliseconds of its posting. The MQL rule was implemented on a Monday (June 15, 2009), and took full effect immediately; i.e., there was no transition period of any sort. The MQL rule and its implementation date was, as far as we are aware, known in advance to market participants. However, as described below, there is no empirical evidence that market participants changed their behavior prior to the actual implementation of the MQL rule on June 15, It should also be pointed out that no other changes to the EBS market structure was implemented at, or around, the date of the MQL implementation. Concurrent media reports suggested that EBS implemented a minimum quote life in order to (i) avoid, or at least limit, flickering quotes, (ii) promote genuine intent-to-trade, (iii) improve the fill-ratio (i.e., the ratio of executed quotes to submitted quotes), and (iv) level the playing field between human and computer traders. All of these arguments can be viewed as soft considerations that were likely motivated by a desire to appease human traders on the EBS trading platform. 3.2 Data description Our primary data are obtained from the EBS trading platform, but in order to conduct a comparative difference-in-difference analysis, we also use data from the Chicago Mercantile Exchange (CME) futures market. These data sets are described in detail below High-frequency EBS quote and volume data From Jan 2008 until June 2010, we have detailed intra-daily data on quotes and transacted amounts (i.e., the volume of trade measured in the base currency). 7 The quote data provide the highest bid quote and the lowest ask quote on the EBS system in a given currency pair. All quotes are executable, and therefore represents the market price at that instant. In addition, the amount available for trade at both the highest bid and the lowest ask price are recorded (i.e., the depth at the inside quotes), as well as the amounts available, on the bid and ask side separately, for quotes at 2, 3, and 5 basis points away from the inside bid and ask quotes. The total depth (sum of the depths at the bid and the ask quotes) at the inside quotes will be referred to as depth0, and the 7 The euro-dollar currency pair is quoted as an exchange rate in dollars per euro, with the euro the base currency. 7

9 depths 2, 3, and 5 basis points away are labeled depth2, depth3, and depth5, respectively. The transactions data provide detailed information on the volume and direction of trade. In particular, for each time-interval, we observe both the total volume of trade, as well as the buy and sell amounts from the perspective of the taker. That is, the direction of trade is based on actual trading records: A trade is recorded as, for instance, buyer-initiated, if it is the result of a hit on a posted ask quote. On the EBS trading platform, traders can enter trading instructions manually, using an EBS keyboard, or, upon approval, via a computer directly interfacing with the system. EBS, in turn, records whether a trade was placed by an ordinary keyboard interface, or by a direct computer interface, allowing for a classification of trades into human or computer trades. This classification into human and algorithmic (computer) trading formed the basis for the recent study by Chaboud, Chiquoine, Hjalmarsson, and Vega (2014). While we are also able to classify trades as human or computer trades, our data actually allows for an even finer classification of trades. In particular, in this more recent data (starting in 2008), EBS decomposes the computer (algorithmic) volume into two different categories, referred to as Bank-AI and PTC-AI, where PTC stands for Professional Trading Community and AI for Automated Interface, which is how EBS labels the direct computer interfaces. The Professional Trading Community essentially refers to hedge funds and CTAs (commodity trading advisors), which can trade directly on EBS under the auspices of dealing banks via prime brokerage arrangements. That is, in our data we are not only able to distinguish whether trading was conducted by a human or a computer, but also, in case of a computer, whether that computer was in the employ of a bank or the professional trading community. Specifically, for each observed time-interval we observe the volume attributable to all nine possible combinations of human, Bank-AI, or PTC-AI maker or taker, with the direction of trade from the perspective of the taker explicitly identified. The transactions data and depth data described above are aggregated by EBS and are available at a one-second frequency throughout our sample period. The actual quote data are also available at the second-by-second frequency throughout the sample period, although we also have access to quote data that are sampled every 100ms during the year These latter 100ms data are used in analyzing transaction level data, as described in the following subsection. 8

10 3.2.2 Millisecond time-stamped EBS transaction records In addition to the longer sample of one-second data on quotes and transacted volumes, we also have access to a somewhat shorter sample of individual time-stamped trades, spanning all of These data record the time-stamp, with millisecond precision, of each trade that occurred, along with the actual transaction price. The amount transacted, and the direction of the trade from the taker perspective, is also recorded. Importantly, the nature of the maker and taker of the trade is explicitly identified, using the three different categories described above (Human, Bank-AI, and PTC-AI). Thus, each trade is classified according to nine different possible combinations of maker and taker, as well as with an indicator of whether the taker was the buyer or the seller. These millisecond time-stamped transactions data are lined up with the 100ms quote data, such that for a given transaction, the best bid and ask prices prevailing at the nearest previous whole 100ms interval are also available. That is, the quote data are lagged up to 99ms for a given recorded transaction. 8 In additions, the depth available at the nearest previous whole second is also lined up with the transaction level data, in order to provide as complete and timely a picture as possible of prevailing market conditions at the time of a given transaction Daily CME futures data Further, we also use data from the CME futures exchange, where futures contracts on the euro-dollar currency are traded. In particular, we have data on daily trading volume (in the base currency), the daily average bid-ask spread at the best bid and ask quotes, and the daily average of the depth available at these quotes (i.e., the sum of the depths at the bid and ask quotes, corresponding to depth0 in the above EBS data). These data are also available from Jan 2008 through June The trading activity on the futures market is affected by the roll-cycle of the futures contracts, which leads to a very clear intra-cycle seasonality in futures trading volume. Since changes in trading volume due to this roll-cycle is not of explicit interest, we also construct roll-adjusted futures volume series. In particular, as specified above, at a given point in time, we use trade data for a specific futures contract with a specific expiry date. We can therefore create a series of dummy variables that indicate the number of weeks to expiry at a given observation date. In particular, 8 Given inherent latencies in the system, the 100ms-frequency quote data lined up with a given transaction is likely as representative a bid and ask price that one can feasibly achieve. 9

11 for any observation, there is between 1 and 13 weeks to expiry for the contract currently used to construct our data. Our series of futures volume is regressed on these 13 dummy variables, and the residuals are computed. The roll-adjusted volume series is defined as these residuals plus the average non-adjusted volume; the adjusted and non-adjusted series thus have identical means. The daily futures data is merged with aggregated daily EBS data. In particular, the one-second EBS quote and volume data are aggregated into a daily data set, where the bid-ask spread and depth variables are averaged across all observations within the day. The traded volume is summed up over the day, to correspond with the daily futures volume. This merged data set spans from Jan 3, 2008 to June 30, 2010, for a total of 618 daily observations. 9 4 Was the MQL rule binding? Before conducting any further analysis, the most fundamental question that needs answering is whether the MQL rule was actually binding. That is, prior to the implementation of the MQL rule, did market makers on EBS to any relevant degree cancel quotes within 250 milliseconds of their posting? In order to address this question, we use a separate data set on quote submissions and quote interruptions on the EBS platform. In particular, we have access to daily data, from January 5, 2009 to December 30, 2009, on the total number of submitted and cancelled quotes in the euro-dollar market on the EBS system. Crucially, in terms of evaluating whether the MQL rule was truly binding, we also have daily data for the number of quotes cancelled within 250ms of their posting. As seen from Figure 1, it is very clear that the MQL rule was indeed binding. Figure 1 show time-series plots of the daily fraction of all submitted quotes on EBS that were cancelled prior to 250 milliseconds, before and after the implementation of the MQL rule. In particular, daily data from January to December 2009 are shown, with the MQL implementation occurring in the middle of the sample, on June 15, Due to the highly proprietary nature of these data, the actual 9 We exclude data collected from Friday 17:00 through Sunday 17:00 New York time from our sample, as activity on the system during these non-standard hours is minimal and not encouraged by the foreign exchange community. Trading is continuous outside of the weekend, but the value date for trades, by convention, changes at 17:00 New York time, which therefore marks the end of each trading day. We also drop certain holidays and days of unusually light volume: December 24-December 26, December 31-January 2, Good Friday, Easter Monday, Memorial Day, Labor Day, Thanksgiving and the following day, and July 4 (or, if this is on a weekend, the day on which the U.S. Independence Day holiday is observed). 10

12 scale cannot be shown in the figure, but rather the data are indexed such that the euro-dollar series starts at an index value of 100 on the first day of the sample. However, to give an approximate idea of the scale in the graph, the fraction of all submitted quotes that were cancelled within 250 milliseconds of their posting was typically between 10 and 15 percent, prior to the implementation of the MQL rule. After the implementation of the MQL rule, this cancellation rate naturally drops to (almost) zero. After the MQL rule was put in place, there is still a tiny fraction of quotes that are recorded as cancelled before the 250 milliseconds mark. According to EBS, this is essentially a recording error that can sometime occur when quotes are cancelled as close to the 250 milliseconds mark as possible. Thus, prior to the implementation of the MQL rule, a sizeable fraction of all quotes (of the order of 10 percent) were cancelled within 250 milliseconds of their posting. The MQL rule was, in this sense, clearly binding to many market participants. Figure 1 also shows that there is no evidence that market makers altered their behavior prior to the actual enforcement of the MQL rule. That is, there is no gradual drop in cancellation rates in the period leading up to the MQL implementation date, but rather a very steep fall from the day prior to the day of the enactment of the rule. Taken together, these findings motivate a further analysis of the impact of the MQL rule, on market characteristics beyond the mere cancellation rates documented here. The sharp drop, and lack of gradual adjustment, in quote cancellations, also motivate a before-after analysis, where market conditions right before the implementation of the MQL rule is compared to conditions immediately afterwards. Figure 2 provides some further evidence on the quoting and trading behavior on EBS, in the period around the implementation of the MQL rule. In particular, the graphs in Figure 2 show daily aggregates of (i) the number of submitted quotes and the (ii) the number of interrupted quotes (i.e., quotes that do not result in a trade). Similar to Figure 1, in each panel the series are indexed such that they begin at Given the evidence in Figure 1, the most striking feature of the graphs in Figure 2 is the lack of any clear effect at the time of the MQL implementation. Indeed, neither of the two variables plotted in Figure 2 show any clear and consistent reaction to the MQL implementation. There is perhaps some suggestive evidence of a break around the MQL 10 The series in the two panels in Figure 2 look virtually identical. This is due to the fact of the high order cancellation rates in electronic limit order markets, where the vast majority of orders are cancelled before leading to a trade. 11

13 implementation, although this is diffi cult to establish given the high day-to-day variance in the data. In any case, any such potential effect appears to have been short-lived. Table 1 explores this notion further. The table shows the change in averages of the logtransformed variables from 10, 30, or 60 days prior to the implementation of the MQL rule to 10, 30, or 60 days after the implementation. That is, the table shows the average log-values in the 10 (30, 60) days immediately following the MQL implementation, minus the average log-values in the 10 (30, 60) days immediately prior to the MQL implementation. This after-minus-before value thus provide an estimate of the change from before to after the MQL implementation. Although some of the estimated changes are statistically significant, there is little consistency across time horizons (10, 30, or 60 days before and after). In particular, both the number of submitted and interrupted quotes seem to have dropped by about 15 percent in the period immediately following the implementation of the MQL rule, consistent with the idea that the orders that were previously posted and cancelled in less than 250 milliseconds did not get posted at all after the implementation of the MQL rule. However, if such an effect did occur, it appears to have been very short-lived, as the same pattern cannot be seen at the 30- and 60-day horizons. Thus, apart from some evidence of a short-run effect, the results in Table 1 mostly confirm the visual evidence in Figure 2, namely that the MQL rule did not have any clear and obvious impact on the quoting behavior on EBS. 5 Daily before-after analysis In the first part of the empirical analysis, we compare the daily average bid-ask spread, daily depth, and total daily volume before and after the implementation of the MQL rule. This is done both in a simple before-after manner, as well as in a difference-in-difference manner, where the beforeafter outcome on the EBS trading platform is compared to the before-after outcome on the CME Futures market (which was not subject to an MQL rule). By differencing out any common effects in these two markets, a clearer estimate of the impact of the MQL rule is obtained. After a graphical presentation of the data, we present comparisons of unconditional averages in a simple differencein-difference analysis, and then further verify those findings in a regression-based formulation of the difference-in-difference model, where secular trends and observable factors are controlled for. 12

14 5.1 Graphical evidence The top panel in Figure 3 plots the average daily (inside) bid-ask spreads on EBS from Jan 2008 to June The series are logged and standardized to each start at a value of 100, such that a comparison of relative changes over time can easily be done. The EBS spread is on a downward trend at the time of the MQL implementation. There is no strong indication that this trend is altered around the MQL implementation, although there is perhaps some suggestion that the decrease in spreads become more prominent after the implementation of the MQL rule. As a point of comparison, the spreads in the Futures market are also shown in Figure 3. The observed patterns in the futures market are qualitatively similar to those on EBS, with a downward trend around the MQL implementation. There is again no clear sign of any break around the MQL implementation, which might be expected since the futures market was not directly affected by the MQL rule. Although an impact on the bid-ask spread is probably the clearest prediction coming out of the classical option theory, one might also conjecture that the posted depth in the market is affected by an MQL rule. That is, the increased cost of posting quotes might lead to a smaller depth. The bottom panel of Figure 3 show graphs of the standardized average daily depth available at the inside spread (i.e., depth0 in the above notation). Again, there is no obvious impact around the MQL implementation, although there might be some indication that depth is rising a bit more sharply after the MQL implementation. As is seen, depth is also trending strongly upwards on the futures market around the time of the MQL implementation. Overall, there is thus no completely clear graphical evidence that the MQL rule affected average bid-ask spreads or depths. However, the graphs presented in Figure 3 still leaves open the possibility that the MQL rule did have some impact on spreads and depths, an we present more formal results in the difference-in-difference analyses below. The top panel of Figure 4 show time-series plots of the daily total volumes on EBS and the Futures market, from Jan 2008 to June The series are logged and standardized to each start at a value of 100. The bottom panel of Figure 4 show the same plots, but using the rolladjusted Futures volume, described in the data section above. As is evident from both figures, the EBS and the Futures volumes exhibit very similar patterns over time, and co-vary strongly with each other. However, after the implementation of the MQL rule on EBS, there appears to have 13

15 been a noticeable shift in these patterns, with an increase in Futures volume that is not matched by a similar increase in EBS volume. The graphs in the top panel of Figure 4, which show the unadjusted futures volume, suggest that this shift started before the implementation of the MQL rule, but the roll-adjusted volumes in the bottom panel indicate that the shift appears to actually have happened right around the implementation of the MQL rule. 11,12 These graphical evidence are certainly suggestive of an impact of the MQL rule on overall market activity. In particular, they suggest that, at least in relative terms, the trading volume on EBS might have decreased as a result of the MQL implementation. 5.2 Difference-in-difference mean analysis Let y EBS,t represent either the average bid-ask spread, depth, or the total trading volume in the EBS market at time t = 1,..., T, and let y F ut,t be the corresponding daily measure in the CME Futures market. In order to gauge the impact of the MQL rule, a difference-in-difference analysis is used, dd = (ȳ EBS,after ȳ EBS,before ) (ȳ F ut,after ȳ F ut,before ). (1) Here, ȳ i,after, i {EBS, F ut}, is the mean of y i,t over the period after the MQL implementation, and ȳ i,before is the mean of y i,t over the period before the MQL implementation. In order to isolate the effect of the MQL rule, only data relatively close to the event date are used to calculate ȳ i,after ȳ i,before. In particular, windows spanning from 10, 30, or 60 days before the MQL implementation to 10, 30, or 60 days after, respectively, are considered. Table 2 provides results on the log-changes in the data around the implementation of the MQL rule. In particular, the table shows the change in average log-spread (log-depth, log-volume) from 10, 30, or 60 days prior to the implementation of the MQL rule to 10, 30, or 60 days after the implementation. That is, the table shows the average log-spread (log-depth, log-volume) in the 10 (30, 60) days immediately following the MQL implementation, minus the average log-spread (logdepth, log-volume) in the 10 (30, 60) days immediately prior to the MQL implementation. Changes 11 The implementation of the MQL rule on EBS coincided with an expiry date in the futures market. The difference between the un-adjusted and roll-adjusted futures volume is therefore quite noticeable on the date of the MQL implementation. As discussed in the data section, the change over time in the roll-adjusted volumes should provide a better measure of changes in trading activity in the futures market that is not explicitly linked to the roll-cycle. 12 The effect of the roll-cycle on the bid-ask spread and the depth is much more limited, and the graphs for the bid-ask spread and depth therefore only show the unadjusted futures data. 14

16 are shown both for EBS and Futures data. In the results for trading volume, the Futures results are either based on unadjusted (left-hand-side column) or roll-adjusted (the right-hand-side column) volumes. In addition, the difference between the log-change in the EBS data and the Futures data are shown in the bottom rows of each panel in the table. These difference-in-difference results thus correspond to estimates of equation (1). Standard errors are shown in parentheses below each estimate; the rows labeled Historical 90% conf. int. are described in Section below Bid-ask spreads and depth The results for the bid-ask spread and depth0 are shown in the first two columns of Table 2, and mostly reflect the graphical evidence presented in Figure 3. Over the short 10-day before to 10-after period around the MQL implementation, there is no statistically significant change in either of these variables, but over the longer 30- and 60-day before-after windows, there is a clear decrease in the EBS spreads and increase in the EBS depths. The same pattern holds for the futures data, however, where the log-changes in spreads and depths are greater in absolute magnitude than on EBS. The difference-in-difference estimates are therefore positive for spreads and negative for depth, suggesting that when viewed relative to the futures market, there might have been an increase in spreads and a decrease in depth on EBS in the 30- and 60-day windows following the MQL implementation. Since all estimates represent log-changes, they can be interpreted as approximate percentage changes. In the 30- and 60-day windows following the MQL implementation, the bid-ask spread decreased by approximately 4 and 8 percent, respectively, compared to the 30- and 60-day windows prior to the MQL implementation. During the same periods, the corresponding decrease in the Futures spread was around 9 and 12 percent, such that the futures market spread decreased by approximately 4 percentage points more than the EBS spread. The changes in depth on the EBS market were around 4 and 12 percent in the 30- and 60-day before-after windows. The futures market, on the other hand, exhibited 24 and 47 percent increases in depth over the same periods. Depth on EBS, relative to the futures market, therefore decreased by 20 and 35 percent in the 30 and 60 before-after window. These massive increases in depth in the futures market over the event period makes it rather diffi cult to interpret the difference-in-difference estimates. Moreover, in absolute terms, the EBS spread decreased and the depth increased following the MQL implementation. The purpose, of course, of a difference-in-difference analysis is to eliminate 15

17 any common factors that might be present, and thus isolate the actual treatment effect. Still, the interpretation becomes a bit more diffi cult when the absolute change on the treated market goes in the opposite direction of the difference-in-difference estimate Total volume The estimates of the change in average volume on the EBS platform around the MQL implementation, shown in the top rows in each panel in Table 2, tell a fairly clear story. The implementation of the MQL rule appears to have been associated with a fairly large drop in overall volume on EBS. Average euro-dollar volumes appears to have dropped around 16 to 18 percent in the period around the MQL implementation. The unadjusted volumes in the futures market also drop substantially in the 10 days following the MQL implementation, compared to the 10 days before. However, if one considers the 60-day before to 60-day after period, there is instead a substantial and statistically significant increase in the futures volume after the implementation of MQL. These conflicting results for the futures market are reconciled if one instead considers changes in the roll-adjusted futures volume. As seen in the right-most column of Table 2, there is now strong evidence of a substantial increase in futures volume at all horizons. The unadjusted and roll-adjusted results are, in fact, very similar for the 60-day before-after window. This is to be expected since the roll-cycle of futures contracts lasts for three months, or approximately 60 trading days. When considering averages over 60-day horizons, there is therefore hardly any impact of the roll-cycle, since it gets averaged out across the observations. The final rows in each panel in Table 2 show the difference-in-difference estimates. That is, these are the differences between the volume change on EBS and the volume change in the futures market. Using the roll-adjusted volume (shown in the right-most column), this difference-in-difference is large and statistically significant. This, of course, follows from the decrease in EBS volume and increase in futures volume over the same period. With the unadjusted futures volume, the difference-indifference is only clearly statistically significant for the 60-day before-after window, which again is no surprise given the above discussion. 16

18 5.2.3 Statistical significance The results in Table 2, as well as the graphical evidence in Figure 4, suggest that the MQL rule had a negative impact on the trading volume on EBS. Perhaps a bit less convincingly, the evidence in Table 2 also suggest that relative to the futures market, the bid-ask spread on EBS increased and the depth decreased following the MQL implementation. Overall, the estimates of the mean changes shown in Table 2 tell a fairly clear story, but it is of course diffi cult to properly ascertain their statistical significance. The standard errors shown in ordinary parentheses below the point estimates in Table 2 are based on data for the 10- (30-, 60-) before-after period, and concerns about changes in the variance of the data over time might raise doubts about the validity of the standard errors. In order to provide more robust evidence of significant changes in the EBS market in association with the MQL rule, we construct the empirical distribution of the change in the EBS variables and the futures variables, using data only prior to the implementation of the MQL rule. In particular, starting at the beginning of the sample in Jan 2008, we calculate the change in average log-volume (log-spread, log-depth) from the first 10 days of the sample to the following 10 days; i.e., the change between days 1 through 10 and days 11 through 20. This is then repeated for every set of 10 plus 10 days in the sample, prior to the implementation of the MQL rule; i.e., for days 2-11 and days 12-21, days 3-12 and days 13-22, and so forth. The resulting collection of estimates provides an empirical distribution of the 10-day mean log-changes in trading volume at EBS. The last subsample end 10 days prior to the MQL event, such that there is no overlap with the data used to estimate the 10-day before-after results in Table 2. The same procedure is repeated for the futures volume, and for the 30-day and 60-day before-after windows, with the sampling ending 30 and 60 days prior to implementation of the MQL rule, respectively. Finally, the empirical distribution of the difference-in-difference estimates is also constructed analogously. That is, in this case, the difference between the volume change on EBS and the volume change in the futures market is calculated for all 10- (30-, 60-) plus 10- (30-, 60-) day samples prior to the MQL rule, and the empirical distribution of these difference-in-difference estimates is obtained. The same procedure is applied to the bid-ask spread and the depth. These empirical distributions, which are constructed using data unaffected by the MQL rule, 17

19 can then be used to evaluate how abnormal the change in volume (spread, depths) was around the MQL implementation. In Table 2, 90% confidence intervals based on this historical empirical distribution are reported in square brackets. When viewed against the usual standard errors, shown in ordinary parentheses below the estimates in Table 2, the diff-in-diff estimates for the 30- and 60-day before-after windows are significant at the 5 or 1 percent level for both the spreads and the depths. If one instead compares the diff-in-diff outcome to the arguably more robust confidence intervals based on the historical empirical distribution, the spread results are no longer significant whereas the depth results remain statistically significant also under this more stringent evaluation. As seen in the final column of Table 2, the diff-in-diff estimates for volume, using the detrended futures volume as control, are all statistically significant according to the historical distribution. Figure 5 provides further evidence along these lines. In particular, Figure 5 shows plots of the empirical densities for the changes in euro-dollar EBS and roll-adjusted futures volumes, as well as the empirical densities of the difference-in-difference estimates. These plots also indicate the actual change around the MQL implementation, along with the percentile of the empirical distribution into which this change falls. It is evident that the drop in EBS volume around the MQL implementation was unusual but not extreme. The change in volume around the MQL implementation is roughly at the 15th percentile of the empirical distribution. However, this drop in EBS volume was associated with a large increase in futures volume, and the resulting difference-in-difference estimate falls in the extreme left tail of the difference-in-difference distribution. The results in Figure 5 thus suggest that the implementation of the MQL rule on EBS was associated with a moderate to large decrease in EBS volume and a moderate to large increase in Futures volume. Looked at separately, the changes in volumes on neither EBS nor the futures market appears extreme compared to the observed historical variation. However, when taken together, the simultaneous volume drop on EBS and volume increase in the futures market, does appear very extreme and highly statistically significant. That is, when comparing the difference-in-difference estimates with the historical empirical distributions, they are all in the extreme left tail. Figures 6 and 7 show the corresponding density plots for the historical changes in the log-spread and the log-depth, respectively. As seen in both of these figures, and reflecting the results in Table 2, there is seemingly no impact of the MQL implementation of the 10-day before-after window. For the spread, this is true also for the 30-day before-after window, when judged against the historical 18

20 distribution. In the 60-day before-after window, however, both the EBS spread and depth exhibit large to extreme changes, but so do the corresponding Futures variables. In the spread case, this results in a difference-in-difference estimate that is in the middle of the historical distribution. In the depth case, the difference-in-difference estimate ends up in the extreme left tail. Based on the historical distribution, there is thus no statistically significant evidence to suggest that the spread changed as a result of the MQL implementation. Although the effect on depth appears statistically significant, the results presented in Figure 7 provide a much less convincing case of an actual MQL impact than does the results for volume presented in Figure 5. In the volume case, the absolute change on EBS is in the same direction as the diff-in-diff estimate, and significance of the diff-in-diff estimate comes about through an unusual combination of a fairly large drop in EBS volume and a fairly large increase in Futures volume. In the depth case, both EBS and futures depth exhibit large to extreme positive changes, but the futures change is the most extreme, resulting in a negative and statistically significant difference-in-difference estimate for depth. 5.3 The difference-in-difference analysis in regression format The above difference-in-difference analysis strongly suggests that the implementation of the MQL rule was associated with a significant simultaneous drop in EBS volume and increase in futures volume. There is little evidence to suggest a significant and substantial impact on the average bid-ask spread, whereas the impact on depth is formally statistically significant but open to interpretation. To provide further evidence along these lines, we re-write the difference-in-difference analysis in a regression format, which allows for the control of other possible factors that may have affected trading volume around the date of the MQL implementation. Using the same notation as previously, y i,t, i {EBS, F ut}, represents the log-volume (logspread, log-depth) in the EBS and futures market. The basic difference-in-difference analysis can be equivalently formulated in a regression format as follows, y i,t = α + β MQL MQL i + β After After i,t + β DD (MQL i After i,t ) + γx i,t + ɛ i,t. (2) Here, MQL i is a dummy variable taking on values of 1 for the group affected by the MQL rule 19

21 (i = EBS) and zero otherwise. After i,t is a dummy variable taking on values of 1 after the MQL implementation (i.e., from June 15, 2009, onwards), and zero before. The difference-in-difference coeffi cient is given by β DD, the coeffi cient on the interaction between the MQL dummy and the before-after dummy. Additional control variables can be included as regressors X i,t, which control for other factors that may have affected changes in volume around the MQL date. If X i,t is excluded from the regression, the coeffi cient β DD will be numerically identical to the dd estimate in equation (1). The regression in equation (2) would typically be estimated using data only for a limited timeperiod around the MQL event; e.g., 10, 30, and 60 days before and after the event. As in the estimation of the direct difference-in-difference formulation captured by equation (1), this is done in order to identify the effect of the MQL rule. That is, the difference-in-difference coeffi cient (β DD ) captures the difference in mean from before to after the event, now after controlling for the factors in X i,t. In order for this comparison to be relevant, the time-span before and after the event must be limited. However, given access to a long data set that stretches far beyond the immediate period around the MQL event, it makes sense to use this longer sample to better pin down the impact of the control variables in X i,t ; i.e., to more precisely estimate the coeffi cient γ. Use of such an extended sample is particularly useful to estimate any long-run secular trends that may be present in the data. Using the same notation as before, After i,t is a dummy variable taking on values of 1 in the period immediately after the MQL implementation (e.g., in the 10, 30, or 60 days following the MQL event). Now define Long_After i,t as a dummy variable which takes on values of 1 in the period following the period defined by the After i,t dummy until the end of the sample. Similarly, let Before i,t be a dummy variable taking on values of 1 in the period immediately before the MQL implementation (e.g., in the 10, 30, or 60 days before the MQL event), and define Long_Before i,t as a dummy variable which takes on values of 1 in the period from the beginning of the sample to the period defined by the Before i,t dummy. Using this notation, the following regression allows for the full use of the sample, while still estimating a local difference-in-difference effect around the 20

22 MQL event, y i,t = α + β MQL MQL i,t + β After After i,t + β DD (MQL i,t After i,t ) + γx i,t +Long_After i,t β LA + Long_Before i,t β LB + (MQL i,t Long_After i,t ) β DD_LA + (MQL i,t Long_Before i,t ) β DD_LB + ɛ i,t. (3) The coeffi cient β DD is now the difference-in-difference effect, comparing right after the event to right before the event. β DD_LA provides the difference-in-difference coeffi cient between the longafter period and the right-before period; β DD_LB provides the difference-in-difference coeffi cient between the long-before period and the right-before period. 13 The Before period constitutes the base-line period, and the Before i,t dummy is therefore excluded from the regression. Equation (3) is estimated using the full sample period from Jan 3, 2008 to June 30, 2010, for differing specifications of X i,t and differing before-after windows defining the After i,t dummy variable (and implicitly the Long_After i,t and Long_Before i,t dummies). The results are shown in table 3 for 10-, 30-, and 60-day before and after periods. Each column in each panel of the table represents a different set of controls included in X i,t. Since the roll-adjustment of the futures volume can now be conducted in the actual difference-in-difference regression, the unadjusted futures volume is always used in the regressions. The estimated main difference-in-difference effect (β DD ) for the regressions with the bid-ask spread are statistically insignificant in virtually all specifications. The discussion below therefore focuses on the volume and depths results. In specification (1), all controls (including the roll-adjustment) are excluded. The before-after difference-in-difference coeffi cient, shown in the top row of each panel, is in this case identical to the corresponding difference-in-difference estimate in table 2; the standard errors are different, however, since Newey and West (1987) standard errors are used in table 3. In specification (2), linear and quadratic time trends are included in X i,t, separately for the EBS and futures variables. As is seen, this has no, or a very small, effect on the estimated before-after difference-in-difference coeffi cient in the volume regressions. It does impact the magnitude of the coeffi cients in the depth 13 Although it might perhaps not be intuitively obvious that this is indeed the effect that the three DD coeffi cients capture, it is straighforward to show this analytically by explicitly defining the entire multiple regressor matrix, and show that the corresponding regression coeffi cients are equal to the correct mean differences. 21

23 regressions, although the trends do not affect the statistical significance. Specification (3) shows the results when X i,t is made up exclusively of the futures expiry dummies which constitute the roll-adjustment of the futures volume, as described previously in the data section. As might be expected given previous results, the roll-adjustment substantially alters the difference-in-difference estimates for volume in the 10 and 30-day before-after windows. As before, there is hardly any impact on the 60-day before-after results. Column 4 shows the results from including both the linear and quadratic trends, as well as the roll-adjustment in X i,t. The estimates are virtually the same as without the secular trends in the volume case. In the final two columns (5 and 6), daily realized volatility is included in X i,t, in addition to the secular trends and the roll-adjustment. In column 5, the contemporaneous realized volatility is included, whereas in column 6 the realized volatility is lagged by one day; this latter specification is used to alleviate concerns regarding the simultaneous determination of volatility and volume. The realized volatility is calculated as the sum of all 1-min squared returns over the day, using mid-quote prices from EBS. Including volatility as a control variable only marginally affects the difference-in-difference estimates for volume, but has a fairly substantial impact on depth. Overall, Table 3 shows that the difference-in-difference results are mostly robust to the inclusion of separate trends in the EBS and futures markets, as well as controlling for volatility. For volume, the main difference-in-difference effect (β DD ) is only clearly statistically significant in the 60-day before-after window, but is of a similar magnitude and marginally significant in the 30-day window. For depth, the β DD coeffi cient is fairly consistently significant in both the 30-day and 60-day beforeafter windows. However, as mentioned above, there are some doubts to the interpretation of the depth results, given the large upward moves in both the EBS and futures depth around the MQL implementations. In addition, the results for the difference-in-difference coeffi cient between the long-after period ) and the right-before period (β DD_LA and the long-before period and the right-before period ) (β DD_LB, highlight some additional interesting findings. The long-after minus before differencein-difference coeffi cient measures the difference in volume (or depth) in the period following the immediate event period, relative to the volume (depth) right before the event. This estimate thus provides an indication of the long-run effects of the MQL, and thus whether the initial negative impact on EBS volume is persistent or not. Interestingly, in the volume case, the magnitude of the 22

24 long-after difference-in-difference coeffi cient is quite similar to the right-after coeffi cient, suggesting that the impact of the MQL rule on volume was permanent or, at least, long lasting. This is, of course, only a tentative conclusion, since other factors that we do not control for might also have had an important effect over such long time spans. Finally, the long-before minus before differencein-difference coeffi cient is generally insignificant, once secular trends are controlled for, which is what one would expect. In the case of depth, similar results are seen, although the long-after ) coeffi cient (β DD_LA is often substantially larger than the immediate diff-in-diff coeffi cient. Again, the long-before coeffi cient is typically not significant. 6 Computer and human volume The analysis in the previous section suggests that overall trading volume on EBS decreased following the implementation of the EBS rule, at least relative to the volume on the futures market. One immediate reaction to this result might be that one thinks that the volume due to computer trading decreased, since the MQL rule was specifically aimed at computer trading, and human market makers are arguably not directly affected by a 250 milliseconds minimum quote life. Recall the three possible type of trader categories on EBS: Human, Bank-AI, and PTC-AI. Bank-AI represents algorithmic trading by dealing banks, whereas PTC-AI represents algorithmic trading by hedge funds and CTAs. According to EBS, the latter category is made up, to a great extent, of what would typically be referred to as high-frequency traders. Thus, both the Bank-AI and PTC-AI categories of traders might have been affected by the MQL rule, but the effect is likely to be most noticeable for PTC-AI. The top panel in Figure 8 shows the trading volume on EBS broken down according to the type of maker: H-Make (Human maker), B-Make (Bank-AI maker), and P-Make (PTC-AI maker). The bottom panel of Figure 8 shows the corresponding graphs when volume is instead classified according to the type of taker. The graphs span the full sample available for these data, from January 2008 to June As before, volumes are log-transformed and indexed such that H-Make and H-Take volumes start at 100 in each panel, respectively; the graphs show actual (indexed) values for each trade category, not proportional ones. Both figures clearly show that the volumes attributable to each category, both on the maker and taker sides, tend to move in close lock-step 23

25 with each other over time. Since the volumes are log-transformed, this suggests that when overall volume in the market changes, each trade category changes roughly in proportion to its overall size. Interestingly, this pattern appears stable around the MQL implementation, when viewed both from the maker-side and the taker-side. Table 4 provides further results on the trading volume attributable to Human, Bank, and PTC- AI traders, respectively. Using the transaction level data, the first four columns of Table 4 reports the trading activity of these three type of traders in 10, 30, and 60-day windows before and after the implementation of the MQL rule. We focus on data sampled between 3am and 11am New York time, which represent the most active trading hours of the day (see Berger et al., 2008, for further discussion on trading activity on the EBS system), leaving a total of 625, 480, 1, 704, 596, and 3, 219, 341 unique trades in the 10,30, and 60-day before-after windows, respectively, The table shows the total number, as well as fractions, of trades attributable to the three different makers and takers in the market. The total number of trades decreased for all trade categories following the implementation of the MQL rule, in line with the overall volume analysis in the previous section. 14 As is seen, around the time of the MQL implementation, in June 2009, approximately 50 percent of all trades in the euro-dollar currency pair were the result of human market making (H-Make). Bank-AI (B-Make), and PTC-AI (P-Make) made up approximately 15 and 35 percent, respectively, of market making. Confirming the results in Figure 8, there is no evidence to support the notion that the MQL rule led to a disproportionately large decrease in computer trading on EBS. In fact, for all three event windows (10, 30, and 60 days before and after), there is a small increase in the fraction of trading attributable to P-makers, and a corresponding small decrease in H-making; B-making remains virtually flat. The fraction of taking attributed to the three trader categories is also documented in Table 4. These mostly mirror the fractions seen on the maker side, although the numbers suggest that Bank-AI traders are more prone to taking than making. Similar to the make-side, there is a small increase in P-taking after the MQL event, and a corresponding small decrease in H-taking. There is thus no obvious evidence to support the notion that high-frequency computer making 14 Traders on EBS can transact in amounts ranging from one to 999 million of the base currency. In practice, however, as large deals are routinely broken down, most transactions are for amounts of one to five million, and the average trade size varies little over time during our sample period. As a result, there is a very high correlation between the trading volume in a given time period and the number of transactions in that same period. 24

26 (P-Make) decreased disproportionately much after the implementation of the MQL rule. If anything, P-making (and P-Taking) appears to have increased in proportion after the implementation of the MQL rule. To the extent that the MQL rule is a binding constraint on the behaviour of computer trades, as is clearly suggested by Figure 1, this result might seem puzzling. However, to the extent that overall volume on EBS decreased after the MQL implementation, it is quite possible that even if the MQL rule triggered computer traders to leave the market, the subsequent general equilibrium effects were such that human traders ended up trading less in a similar proportion as well. 7 Volatility effects The empirical analysis thus far has mostly been of an unconditional nature. That is, daily averages before and after the implementation of the MQL rule are compared to each other. Of course, the primary underlying idea of event studies and difference-in-difference analyses is that with a short enough window around the event, and a good enough control group, there is no need to control for other factors; by differencing out the common impact that also affected the control group (i.e., the futures market), one implicitly controls for the other factors that may have impacted the outcome over the observed period. However, if the impact of an event is conditional on certain other market conditions, it might still be diffi cult to detect such conditional effects in a comparison of daily averages, regardless of whether the control group is well specified. In fact, the conditional effect might very well completely wash out in the daily averaging, depending on the distribution of the conditioning variable. In the current study, the key theoretically motivated conditioning variable is volatility. The posting of a quote is a free option to the market, and the value of an option increases with the level of volatility. Thus, the effects of the MQL rule should be particularly noticeable during high-volatility periods. In the regression-based difference-in-difference design (equation (3)), there is an attempt to control for this effect by including daily realized volatility as a control variable. The inclusion of volatility as a control variable does seem to have some impact on the estimated MQL effect, but the daily aggregation makes it diffi cult to draw any strong conclusions. In this section, we therefore explicitly try to examine whether the sensitivity to volatility of market makers changed from before to after the implementation of the MQL rule. To this end, we analyze market-making on EBS, 25

27 using the trade-by-trade high-frequency data set described in the data section. The corresponding data for the futures market is not available, and a difference-in-difference type analysis is therefore not possible. However, given the conditional nature of the analysis, and the high-frequency of the data, which allows for the use of data covering only a short temporal time span around the MQL implementation, while retaining a large sample, the lack of an explicit comparison group does not seem crucial. In the below analysis, we continue to make use of the trade-by-trade observations on EBS. As described previously, in the transaction level data, we observe for each individual trade the type of maker and taker (i.e., human (H), Bank-AI (B), and PTC-AI (P)), the direction of the trade (i.e., buy or sell) from the point of the taker, the traded amount in millions of the base currency, and the exact time of the trade recorded with millisecond precision. These trade data are merged with the 100ms frequency quote data, and the second-by-second depth data. In particular, for a given trade, we line up the data recorded for that trade with the most recent quote and depth data available. That is, each trade record is lined up with the quote data available at the previous whole 100 millisecond, and the depth data available at the previous whole second. The quote data is therefore lagged up to 99ms in comparison with the trade data, and the depth data is lagged with up to 999ms. We do this because the trade records do not provide any quote or depth information, which is only available at these somewhat lower frequencies. Again, we focus on data sampled between 3am and 11am New York time, the most active trading hours of the day. That is, each day in our sample is made up of the intra-daily observations between 3am and 11am New York time. As in the above analysis, we focus on 10, 30, and 60-day before and after windows. 7.1 Market-making before and after the MQL implementation To address the question whether market makers become more sensitive to high levels of volatility after the implementation of the MQL rule, we specify a multinomial logit model of the probability of observing a given type of market maker M {H-Make, B-Make, P-Make} for trade τ. In particular, 26

28 we consider the following specification of the probability of observing a maker of type M, Pr (M τ = M) = β (M) 15min V ol15min τ + β (M) ( ) 15min,After V ol 15min τ After τ +γ (M) X τ + γ (M) After (X τ After τ ) +δ (M) G τ + δ (M) After (G τ After τ ) + ɛ (M) τ. (4) As in the previous specifications, After τ is a dummy variable taking on values of 1 if trade τ occurs after the MQL implementation. τ period leading up to trade τ. In particular, τ is measure of the observed volatility during the 15-minute is defined as the sum of absolute 1-minute returns during the previous 15 minutes. The returns are based on mid-quote prices, calculated from the 100ms quote data. In terms of timing, the returns used to calculate τ whole 100ms preceding trade τ. τ ends at the last thus represent the realized absolute variation in returns over the past 15 minutes, using 1 minute returns data. We rely on realized absolute variation rather than realized variance (i.e., sum of squared returns) in order to reduce sensitivity to individual returns in the relatively short temporal time spans used to calculate V olτ 15min. Measuring volatility over the 15 minutes immediately preceding a given trade (rather than over a longer time period) is intended to capture the prevailing market conditions at the time of the trade; Hendershott and Riordan (2013) use a similar specification when characterizing the sensitivity of computer trading to volatility. Additional control variables are included in X τ. In particular, X τ includes dummy variables indicating whether the previous trade had a B-Maker or a P-Maker (H-maker is the omitted category), analogous dummies for the previous trade s taker, the traded volume, in millions of the base currency, over the previous 15 minutes (also ending on the most recent whole 100ms), the inside spread at the most recent whole 100ms, the depth available at the inside spread (recorded for the most recent whole second), as well as the depths available at quotes 2, 3, and 5 basis points away from the inside quotes. In addition, we also include the amount traded in trade τ, and the time elapsed since the last trade. G τ represents half-hour dummy variables, included to control for any intra-daily patterns, as well as a linear time trend in the number of trading days from the event (i.e., taking on values of, say, 10, 9,..., 9, 10 in the specification using date from 10 days before 27

29 to 10 days after the MQL implementation). 15 Finally, we also test whether the type of taker for the current trade impact the estimate of volatility effects. All included regressors, including the time trend and half-hour intra-daily dummies, are interacted with the After τ dummy, such that a separate coeffi cient is estimated from the data before and after the implementation of the MQL rule. In particular, all Af ter coeffi cients measure the differing impact in the period after the implementation of the MQL rule, compared to the period before. The main coeffi cients of interest are the volatility coeffi cients β (M) 15min and β(m) 15min,After. In particular, β (M) 15min,After measure the differing impact of volatility on market making decisions, after the implementation of the MQL rule. Statistical significance of this coeffi cient thus indicate that there was a significant change in the response to volatility following the implementation of the MQL rule. 16 The multinomial logit model captured by equation (4), M {H-Make, B-Make, P-Make}, is estimated using maximum likelihood, with the probabilities of H-Make used as a baseline. That is, all reported coeffi cients should be interpreted relative to the probability of human making (i.e., M =H-Make). In particular, the coeffi cients are expressed as relative risk ratios, such that a coeffi cient greater (smaller) than unity indicates a positive (negative) effect of the corresponding regressor on the likelihood of observing a trade with maker M, relative to the likelihood of observing M =H-Make. For instance, a coeffi cient equal to, say, 0.9, for some regressor in the P-Make equation, would suggest that a one unit increase in the corresponding regressor would reduce the likelihood of observing a P-Make trade by 10 percent relative to an H-Make trade. The estimation results are shown in Table 5. In the interest of space, only the coeffi cients for the two volatility measures along with the constant and main Af ter effect are shown. The coeffi cients for all other control variables are omitted. The bottom rows in the tables indicate which of the above mentioned controls that were included in a given specification. As discussed previously, trading by PTC-AI (P-Make) is likely the cleanest measure of high-frequency trading in our data, 15 Controlling directly for changes in daily volatility is not straightforward. One can easily calculate the daily realized volatilty (or realized absolute variation), using intra-daily data for a given day in the sample. However, since such a measure uses data from the entire day, it becomes forward looking, which might bias the results. One could include an estimate of yesterday s volatility, but it is not clear what such a lagged measure would capture. Result not shown indicate that including a measure of the daily realized volatity does not alter the estimates for the coeffi cients of interest. 16 All control variables are demeaned prior to estimation (and prior to being interacted with the After indicator). This ensures that the level After effect on market-making is directly comparable across specifications. 28

30 and thus the class of traders that is most likely to have been affected by the MQL rule. The Bank-AI category is likely more mixed, and the effects of the MQL rule on this category more ambiguous. We therefore focus our discussion on the coeffi cients in the P-Make equation, although the corresponding coeffi cients in the B-Make equation are also shown in the second part of Table 5. The table report results for a number of different specifications, where each specification adds additional control variables. Specification (1) is simply a regression with the Af ter dummy as the single explanatory variable, thus giving an estimate of before-after level effect in market making, without controlling for any additional factors. Specification (2) includes the main variables of interest, τ τ After τ, along with the half-hour dummy variables and the linear time trend. In the following specifications, subsequent control variables are added. In particular, specifications (3) and (4) add controls for the maker and taker kind in the previous trade, and the trading volume over the previous 15 minutes, respectively. Specifications (5)-(8) add controls for the depth available at the previous whole second, the amount traded in the current transaction and the time since the last transactions, the most recent inside spread, and finally the current taker, respectively. Clearly, all of these control variables cannot be viewed as exogenous with regards to the current maker of the trade. The previous maker and taker, as well as the volume over the previous 15 minutes would appear to be mostly exogenous, but the remainder of the control variables added in specifications (5)-(8) are arguably not fully exogenous with regards to the making outcome. Some care is therefore due when interpreting the results, and whether the coeffi cients change beyond specification (4). The first rows in Table 5 show the plain After-effect, which provides an estimate of whether the level of P-making increased or decreased after the implementation of the MQL rule. As seen in Specification (1) in the first part of Table 5, the unconditional probability of P-making increased after the MQL rule (the After coeffi cient is greater than one), which is entirely in line with the summary statistics shown previously in Table 4. However, once one controls for intra-daily patterns (half-hour dummies) and a linear time trend, this result no longer holds generally. In particular, in the 10- and 30-day before-after windows, the Af ter coeffi cient is below one, and statistically significant, suggesting that P-Making became less likely subsequent to the MQL rule. Additional controls do not alter this result in a qualitative way. This pattern does not hold up in the 60-day before-after window, however, where the Af ter coeffi cient remains greater than one for any set of controls. and 29

31 Turning to the volatility effects, the coeffi cients on V olτ 15min in the P-Make equation are above one, or statistically indistinguishable from one, in specifications (2)-(4), suggesting that P-Making is more likely, relative to H-Making, in times of high volatility. Once additional, and possibly endogenous controls, are added, this pattern is reversed, however. The coeffi cient on the After- MQL interaction for volatility, τ After τ, is less than unity, and typically statistically significant, in the 10-day before-after sample. In the 30 and 60 day windows, this pattern becomes less clear, however, with insignificant coeffi cients, or coeffi cients greater than unity. There is thus no clear additional effect of volatility on P-making after the MQL implementations. The After-MQL interaction for B-making (seen in the second part of Table 5) is statistically significantly greater than unity in the 10-day window, but mostly insignificant in the longer samples. Immediately following the MQL implementation, B-makers thus became relatively more active and P-makers relatively less active, compared to human makers, in times of high volatility. 17 This might suggest that P-makers, likely relying on relying on speed advantages to extract a surplus from market making, scaled down their market making activities in volatile times after the MQL implementation, and that this relative lowering of P-making was absorbed by B-makers. Over the longer 30- and 60-day before-after windows, this effect disappears, however. One could speculate that this lack of a long-run effect might be due to learning by the P-makers. That is, the P-makers required some time to adept to the new environment with an MQL rule. However, this is not a story that can be empirically tested in our data. Table 6 reports results from an identical model to the one described by equation (4), but with taker-type, instead of maker-type, used as the dependent variable. These results can be seen as a form of robustness check of the previous results. In particular, if the results in Table 5 is truly driven by the MQL rule, one would not expect to see the same effect on P-taking as one did on P-making. This is also not the case. It would appear that P-taking increases in volatile times immediately after the implementation of the MQL rule (i.e., in the 10-day window following the MQL event). This is consistent with the idea that the decreased P-making is, at least partly, compensated for by increased P-taking during volatile periods. Again, these effects disappear after the initial 10-day 17 A coeffi cient greater (smaller) than unity in either the P-Make or B-Make equations signals a positive (negative) impact of the corresponding regressor on the likelihood of observering a P-Make or B-Make, relative to an H-Make trade. However, for a given regressor, a greater coeffi cient in, say, the P-Make equation than in the B-Make equation, does not necsarrily imply that an increase in that regressor would lead to a larger probability of observing a P-Make trade relative to a B-Make trade. 30

32 window. 7.2 Effective spreads before and after the MQL implementation The daily averages of bid-ask spreads, reported in Figure 3 and analyzed in Tables 2 and 3, showed little evidence of any strong reaction to the implementation of the MQL rule. Although this lack of reaction in the daily data certainly suggests that the there was no major change in the unconditional EBS bid-ask spreads following the MQL rule, it does not rule out that there were other more subtle conditional effects. In particular, it is possible that there is a volatility dependent effect on the bid-ask spread, as traditional option theory might suggest. Table 7 provides a more detailed view of the bid-ask spread before and after the implementation of the MQL rule, using the high-frequency transaction level dataset. In particular, summary statistics for the most recent inside spread, prior to any transaction, are reported, along with the same summary statistics for the effective spread for each transaction. The effective is calculated as two times the difference between the transaction price and the mid-quote, in case the taker in the trade is a buyer. If the taker is a seller, it is instead calculated as two times the difference between the mid-quote and the transaction prices. Thus, for instance, if a taker-buyer manages to transact at a price below the mid-quote prices, this is recorded as a negative spread. The multiplication by two is done in order to put the effective spread on the same scale as the quoted spread. The spreads are reported in pips, and thus effectively lives on a grid {0, ±1, ±2,...}. Consistent with the daily averages shown in Figure 3, Table 7 shows that both the average quoted spreads and the effective spreads for transactions in the 10, 30, and 60 day before-after windows went down somewhat around the implementation of the MQL rule. The average quoted spread in the 10 (30, 60) window before the implementation of the MQL rule was (1.409, 1.430) pips and (1.361, 1.365) in the 10 (30, 60) window afterwards. As a mnemonic, at an exchange rate of unity between the euro and the dollar, one pip is equal to one basis points. The average change in spreads from before to after the MQL event is thus of the order of one twentieth of a basis points in absolute terms, and of the order of 3 to 5 percent in relative terms. The effective spread, which has a considerably smaller mean than the quoted spread, exhibit changes of a similar magnitude. The table also reports percentiles for the empirical distribution of the spreads before and after the MQL event. As is seen, these percentiles remain virtually unchanged from before to after, again suggesting that the impact 31

33 of the MQL rule on the unconditional bid-ask spread was small. In order to ascertain whether there was any conditional effect on the bid-ask spread, we use the following regression, Spread τ = α + After τ + β 15min τ +γx τ + γ After (X τ After τ ) + β 15min,After ( V ol 15min τ After τ ) +δg τ + δ After (G τ After τ ) + ɛ τ. (5) The dependent variable is the effective spread for trade τ, and the control variables are defined identically to above. Equation (5) is estimated using OLS, and again only the intercept, main aftereffect (After τ ) and volatility coeffi cients are reported. The results are shown in Table 8. As is seen, the After τ parameter is negative and statistically significant in all specifications. The magnitude is around 0.05 to 0.07, in line with the unconditional mean changes reported in Table 7. Although the absolute magnitude of the coeffi cient is small, the relative effect implies that the effective spread is upwards of 10 percent lower after the MQL rule, which is similar to the daily changes for the EBS spread reported in Table 2. As was seen in the daily difference-in-difference analysis in Table 2, there was a similar change on the Futures market, which makes it diffi cult to attribute this level change to the MQL rule. The main purpose of this analysis, however, is the interaction term with volatility. As is seen, in most specifications, high volatility is associated with a larger spread. Interestingly, this effect appears amplified after the MQL implementation. The coeffi cient in front of the interaction τ After τ is around 0.1 in most specification. This is true across the different estimation windows, as long as one controls for the volume over the past 15 minutes. A unit change in τ thus leads to a 0.1 pips larger increase in the effective spread, after the implementation of the MQL rule. V olτ 15min, which is re-scaled to daily units in the analysis, has a standard deviation of around 0.3. A change from a low to average volatility environment to a high volatility environment, would entail a change in τ of between 0.3 and 0.5, with a subsequent impact of 0.03 to 0.05 pips on the effective spread. Since the average effective spread is around 0.7, and the median is equal to 1, this still suggests a not inconsequential effect on the effective spread. The analysis thus suggests that effective spreads might have become more sensitive to high volatility after the implementation 32

34 of the MQL rule. 8 Adverse selection costs before and after the MQL implementation In the final part of our analysis, we study the cost of trading paid by different types of traders, before and after the MQL implementation. In particular, in our data, we can identify nine unique maker-taker combinations, stemming from the possible combinations of human (H), bank-ai (B), and PTC-AI (P) traders acting either as makers or takers. Using this information, we can estimate the effective spread paid by the taker in each of these combinations before and after the MQL event. In particular, we estimate a regression similar to equation (5) above, with the difference that we now include dummy variables for the exact identity of the maker-taker of the trade and their interactions with the Af ter indicator. The full set of control variables mentioned above are included, apart from changing the inclusion of current maker and taker to the exact identity of the maker-taker pair. 18 The results are reported in Table 9, which presents the coeffi cients for the maker-taker combinations, and their interactions with the Af ter indicator, as well as the volatility effects for comparison with Table 8; the volatility coeffi cients are virtually unchanged. The human-maker human-taker (HH) combination is the excluded category, and the coeffi cients of other maker-taker combinations thus represent deviations from the effective spread paid in the HH transactions. Starting with the main effects (i.e., the non-interacted terms), a fairly clear pattern emerges. P-takers, facing either a H-maker or a B-maker end up paying about 0.35 pips and 0.15 pips less in trading costs, respectively, than would a H-taker or B-taker. P-makers earn about 0.2 pips more when trading with human or bank takers, than with other P-takers. When two P traders meet, the effective spread is virtually identical to the spread in HH transactions, suggesting an offsetting effect. Thus, there is clear and strong evidence that slow traders pay a premium (adverse selection cost) when they transact with better informed high-frequency traders, as predicted by the theoretical papers of Hombert and Roşu (2013), Budish, Cramton and Shim (2013), and Biais, Foucault, and Moinas (2013). In particular, when the fastest traders (P), interact with slower traders (H and B), the appear to earn higher 18 The lagged maker and taker variables are also changed to explictly reflect the exact maker-taker combination. 33

35 spreads as makers and pay lower spreads as takers. 19 One can view these differences in effective spreads as compensation for HFT s investment in technology, and a social planner may not find the same effective spread across all trader categories to be optimal, as previous studies show that traders benefit from the presence of algorithmic traders (e.g., Chaboud, Chiquoine, Hjalmarsson, and Vega (2014) find that algorithmic trading participation increases the informational effi ciency of prices and Hendershott, Jones, and Menkveld (2011) find that algorithmic traders improve market liquidity). The MQL rule puts limitations on the speed advantage of makers, and one might therefore conjecture that the spreads earned by P-makers following the MQL implementation are lower. This is also mostly reflected in the results presented in Table 9. In particular, P-makers earn about 0.05 to 0.03 pips less in their transactions with humans and banks in the 10 and 30 day windows subsequent to the implementation of the MQL rule, compared to same sample windows before the MQL implementation. These effect are mostly gone in the 60-day before-after window, however. Somewhat curiously, the effective spread paid by P-takers to both Human and other P-makers increase in the 10- and 30-day periods following the MQL implementation. This could possibly be suggestive of computer trading strategies relying on delicate interactions between take and make orders, where the curtailing of computer making as a result of the MQL rule also affects the effi cacy of the taking side of their trading. 9 Conclusion Many researchers have stated that imposing an MQL rule will have a negative impact on market quality. For example, Linton, O Hara, and Zigrand in the Foresight study state that the independent academic authors who have submitted studies are unanimously doubtful that minimum resting times would be a step in the right direction, This view is in line with option theory predictions. However, recent theoretical papers (e.g., Foucault, Hombert and Roşu (2013), Budish, 19 An alternative explanation is that high-frequency traders provide liquidity to other market participants when liquidity is expensive. We rule out this explanation because we compare effective spreads paid by different types of traders after controlling for the inside quoted spread. 20 Other similar quotes are, this [MQL] will result in a dry up of liquidity in the book and might even cause an increase in volatility. To be honest, we think this proposal is a terrible idea... (Farmer and Skouras, in Foresight), and however, the minimum time-in-force appears to be a particularly blunt, poorly considered tool. (Charles Jones, 2013). 34

36 Cramton and Shim (2013), and Biais, Foucault, and Moinas (2013)) suggest that the participation of HFTs may reduce market liquidity due to increased adverse selection costs, thus an MQL rule that mainly affects HFTs, as the 250 millisecond rule is intended to, may have a positive impact on market liquidity. We find that the MQL rule causes trading volume to decline in the EBS trading platform, suggesting that some market participants may have moved to other trading venues without MQL rules. However, we find that the MQL rule did not appear to have an impact, on average bid-ask spreads and depth, which suggests that the counterbalancing effects of lower adverse selection costs and higher market-making costs for HFTs play a role in a way that they cancel each other out. However, we do find some evidence that spreads might be more sensitive to volatility after the MQL rule was implemented. Finally, there is also evidence that computer makers collect smaller spreads after the MQL rule, likely by limiting their speed advantage. This study does not provide a definite answer on whether MQL is good or bad for the market, or whether any policy aimed at slowing down HFTs is good or bad. Instead, our study highlights trade-offs. In addition, we note that the actual impact must also depend on the actual length of the minimum resting time that is imposed. That is, the 250 milliseconds imposed on EBS only affects algorithmic traders, and this may explain why there are counterbalancing effects. 35

37 References [1] Berger, D.W., A.P. Chaboud, S.V. Chernenko, E. Howorka, and J.H. Wright, Order Flow and Exchange Rate Dynamics in Electronic Brokerage System Data, Journal of International Economics 75, [2] Biais, B., Foucault, T., and Moinas S., 2011, Equilibrium Algorithmic Trading, Working Paper, Tolouse School of Economics. [3] Biais, B., and P. Woolley, 2011, High Frequency Trading, Manuscript, Toulouse University, IDEI. [4] Boehmer, E., Fong, K. Y. L., Wu, J., 2012, Algorithmic Trading and Changes in Firms Equity Capital, Working Paper. [5] Brogaard, J. A., Hendershott, T., and Riordan, R., 2012, Algorithmic Trading and Price Discovery, working paper, University of California at Berkeley. [6] Foucault, T., 2012, Algorithmic Trading: Issues and Preliminary Evidence, Market Microstructure Confronting Many Viewpoints, John Wiley & Sons. [7] Foucault, Thierry, Johan Hombert, and Ioanid Roşu, 2013, News trading and speed, Working paper, HEC. [8] Foucault, T., and Menkveld, A., 2008, Competition for Order Flow and Smart Order Routing Systems, Journal of Finance, 63, [9] Han, J., M. Khapko, and A.S. Kyle, Liquidity with High-Frequency Market Making, Swedish House of Finance Research Paper No [10] Hasbrouck, J., and Saar, G., Low-Latency Trading, Working Paper. [11] Hendershott, T., C.M. Jones, and A.J. Menkveld, Does Algorithmic Trading Improve Liquidity?, Journal of Finance, 66, [12] Hendershott, T., and Riordan, R., 2012, Algorithmic Trading and the Market for Liquidity, Journal of Financial and Quantitative Analysis, forthcoming. 36

38 [13] Hirschey, N., 2011, Do High Frequency Traders Anticipate Buying and Selling Pressure? University of Texas at Austin, Working Paper. [14] Hoffman, P., 2012, A Dynamic Limit Order Market with Fast and Slow Traders, Working Paper. [15] Ito, T., Yamada, K., Takayasu, M., and Takayasu, H., 2012, Free Lunch! Arbitrage Opportunities in the Foreign Exchange Markets, NBER Paper No [16] Jovanovic, B., and A.J. Menkveld, 2011, Middlemen in Limit-Order Markets, Working Paper. [17] Kirilenko, A, A. S. Kyle, M. Samadi, and T. Tuzun, 2011, The Flash Crash: The Impact of High Frequency Trading on an Electronic Market, Working Paper. [18] Kondor, P., 2009, Risk in Dynamic Arbitrage: Price Effects of Convergence Trading, Journal of Finance, 64, [19] Kozhan, R., and W. Wah Tham, 2012, Execution Risk in High-Frequency Arbitrage, Management Science, forthcoming. [20] Li, S., 2013, Imperfect Competition, Long Lived Private Information, and the Implications for the Competition of High Frequency Trading, University of Maryland Working Paper. [21] Martinez, V., and I. Roşu, 2011, High Frequency Traders, News and Volatility, Working Paper. [22] Menkveld, A., 2011, High Frequency Trading and the New-Market Makers, Working Paper. [23] Oehmke, M., 2009, Gradual Arbitrage, Working Paper, Columbia University. 37

39 Table 1: Log-changes in quote behavior around MQL implementation. The table shows the change in the mean of the log of the variables from before to after. Number of Submitted Quotes Number of Interrupted Quotes 10-day before and after window After minus Before (std. err.) (0.054) (0.054) 30-day before and after window After minus Before (std. err.) (0.047) (0.047) 60-day before and after window After minus Before (std. err.) (0.031) (0.031) 38

40 Table 2: Mean log-changes around the MQL implementation. Resullts are based on daily data measured over the entire 24h trading day. Volume Volume Bid-Ask Spread Depth0 (Control: Raw Futures Volume) (Control: Detrended Futures Volume) 10-day before and after window EBS Change (std. err.) (0.024) (0.023) (0.068) (0.068) [Historical 90% conf. int.] [ 0.088, 0.158] [ 0.206, 0.095] [ 0.226, 0.282] [ 0.226, 0.282] Futures change (std. err.) (0.018) (0.115) (0.071) (0.085) [Historical 90% conf. int.] [ 0.100, 0.108] [ 0.401, 0.213] [ 0.407, 0.389] [ 0.240, 0.297] EBS-minus-Futures Change (Diff-in-Diff) (std. err.) (0.029) (0.117) (0.098) (0.109) [Historical 90% conf. int.] [ 0.051, 0.080] [ 0.170, 0.223] [ 0.182, 0.213] [ 0.226, 0.171] 30-day before and after window EBS Change (std. err.) (0.015) (0.020) (0.056) (0.056) [Historical 90% conf. int.] [ 0.151, 0.307] [ 0.409, 0.139] [ 0.239, 0.239] [ 0.239, 0.239] Futures change (std. err.) (0.016) (0.054) (0.065) (0.063) [Historical 90% conf. int.] [ 0.160, 0.269] [ 0.734, 0.324] [ 0.301, 0.289] [ 0.263, 0.302] EBS-minus-Futures Change (Diff-in-Diff) (std. err.) (0.022) (0.057) (0.086) (0.084) [Historical 90% conf. int.] [ 0.014, 0.083] [ 0.190, 0.342] [ 0.219, 0.298] [ 0.282, 0.170] 60-day before and after window EBS Change (std. err.) (0.010) (0.014) (0.038) (0.038) [Historical 90% conf. int.] [ 0.036, 0.391] [ 0.528, 0.124] [ 0.285, 0.227] [ 0.285, 0.227] Futures change (std. err.) (0.009) (0.038) (0.044) (0.042) [Historical 90% conf. int.] [ 0.062, 0.338] [ 0.980, 0.112] [ 0.272, 0.207] [ 0.287, 0.211] EBS-minus-Futures Change (Diff-in-Diff) (std. err.) (0.014) (0.041) (0.058) (0.057) [Historical 90% conf. int.] [0.008, 0.084] [ 0.123, 0.456] [ 0.266, 0.246] [ 0.271, 0.257] 39

41 Table 3: Difference-in-Difference regression analysis for log-changes. Resullts are based on daily variables measured over the entire 24h trading day. Newey-West standard errors with [T 1/3 ] = 8 lags are used. Bid-Ask Spread Depth0 (1) (2) (3) (4) (5) (6) (1) (2) (3) (4) (5) (6) 10-day before and after window Diff-in-Diff: After minus Before (std. err.) (0.023) (0.024) (0.059) (0.049) (0.037) (0.039) (0.099) (0.101) (0.157) (0.124) (0.091) (0.099) Diff-in-Diff: Long-After Minus Before (std. err.) (0.020) (0.025) (0.038) (0.038) (0.034) (0.032) (0.076) (0.113) (0.108) (0.127) (0.102) (0.099) Diff-in-Diff: Long-Before Minus Before (std. err.) (0.042) (0.050) (0.051) (0.049) (0.032) (0.030) (0.111) (0.104) (0.131) (0.108) (0.081) (0.079) 30-day before and after window Diff-in-Diff: After minus Before (std. err.) (0.034) (0.036) (0.055) (0.045) (0.035) (0.032) (0.086) (0.098) (0.135) (0.105) (0.077) (0.073) Diff-in-Diff: Long-After Minus Before (std. err.) (0.030) (0.035) (0.031) (0.036) (0.037) (0.035) (0.073) (0.125) (0.091) (0.131) (0.109) (0.106) Diff-in-Diff: Long-Before Minus Before (std. err.) (0.049) (0.062) (0.052) (0.060) (0.037) (0.035) (0.113) (0.118) (0.127) (0.120) (0.080) (0.076) 60-day before and after window Diff-in-Diff: After minus Before (std. err.) (0.024) (0.032) (0.019) (0.025) (0.028) (0.027) (0.088) (0.120) (0.071) (0.083) (0.061) (0.059) Diff-in-Diff: Long-After Minus Before (std. err.) (0.022) (0.042) (0.019) (0.041) (0.041) (0.039) (0.079) (0.157) (0.074) (0.141) (0.114) (0.110) Diff-in-Diff: Long-Before Minus Before (std. err.) (0.049) (0.063) (0.047) (0.059) (0.035) (0.035) (0.121) (0.141) (0.119) (0.128) (0.077) (0.077) Linear and Quadratic Time trends NO YES NO YES YES YES NO YES NO YES YES YES Futures expiry NO NO YES YES YES YES NO NO YES YES YES YES Realized Volatility NO NO NO NO YES NO NO NO NO NO YES NO Realized Volatility (Lagged) NO NO NO NO NO YES NO NO NO NO NO YES 40

42 Table 3: Difference-in-Difference regression analysis for log-changes (cont.). Volume (1) (2) (3) (4) (5) (6) 10-day before and after window Diff-in-Diff: After minus Before (std. err.) (0.066) (0.066) (0.115) (0.113) (0.108) (0.108) Diff-in-Diff: Long-After Minus Before (std. err.) (0.079) (0.079) (0.095) (0.116) (0.119) (0.120) Diff-in-Diff: Long-Before Minus Before (std. err.) (0.059) (0.090) (0.094) (0.113) (0.103) (0.112) 30-day before and after window Diff-in-Diff: After minus Before (std. err.) (0.138) (0.136) (0.140) (0.140) (0.130) (0.138) Diff-in-Diff: Long-After Minus Before (std. err.) (0.144) (0.154) (0.127) (0.145) (0.146) (0.152) Diff-in-Diff: Long-Before Minus Before (std. err.) (0.132) (0.143) (0.110) (0.123) (0.117) (0.124) 60-day before and after window Diff-in-Diff: After minus Before (std. err.) (0.106) (0.109) (0.098) (0.104) (0.099) (0.104) Diff-in-Diff: Long-After Minus Before (std. err.) (0.118) (0.151) (0.101) (0.139) (0.141) (0.135) Diff-in-Diff: Long-Before Minus Before (std. err.) (0.104) (0.126) (0.086) (0.116) (0.115) (0.113) Linear and Quadratic Time trends NO YES NO YES YES YES Futures expiry NO NO YES YES YES YES Realized Volatility NO NO NO NO YES NO Realized Volatility (Lagged) NO NO NO NO NO YES 41

43 Table 4: Summary statistics for make- and take-volumes, based on trade-by-trade data. 10-day before and after window Total number of trades Fraction of trades Before After Before After H-Make 169, , % 48.36% B-Make 56, , % 16.23% P-Make 120, , % 35.41% H-Take 163, , % 46.16% B-Take 79, , % 24.24% P-Take 102, , % 29.59% 30-day before and after window Total number of trades Fraction of trades Before After Before After H-Make 467, , % 49.01% B-Make 148, , % 16.74% P-Make 297, , % 34.25% H-Take 419, , % 45.40% B-Take 224, , % 24.06% P-Take 269, , % 30.53% 60-day before and after window Total number of trades Fraction of trades Before After Before After H-Make 903, , % 48.69% B-Make 278, , % 17.09% P-Make 544, , % 34.22% H-Take 819, , % 45.44% B-Take 437, , % 23.24% P-Take 469, , % 31.32% 42

44 Table 5: P-Make results from multinomial logit regression, with type of maker as dependent variable. Standard errors are in parantheses below the estimates. P-Make equation (1) (2) (3) (4) (5) (6) (7) (8) 10-day before and after window afterτ (std. err.) (0.006) (0.012) (0.012) (0.012) (0.012) (0.015) (0.016) (0.016) τ (std. err.) (0.014) (0.014) (0.024) (0.022) (0.023) (0.023) (0.022) τ Afterτ (std. err.) (0.024) (0.026) (0.030) (0.031) (0.031) (0.031) (0.033) Constant (std. err.) (0.003) (0.005) (0.005) (0.006) (0.006) (0.006) (0.006) (0.005) N 625, , , , , , , , day before and after window afterτ (std. err.) (0.004) (0.007) (0.007) (0.007) (0.007) (0.009) (0.009) (0.009) τ (std. err.) (0.009) (0.009) (0.015) (0.013) (0.013) (0.013) (0.013) τ Afterτ (std. err.) (0.015) (0.016) (0.021) (0.024) (0.025) (0.026) (0.026) Constant (std. err.) (0.001) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) N 1, 704, 596 1, 704, 596 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, day before and after window afterτ (std. err.) (0.003) (0.005) (0.005) (0.006) (0.006) (0.007) (0.007) (0.007) τ (std. err.) (0.006) (0.006) (0.010) (0.009) (0.009) (0.009) (0.009) τ Afterτ (std. err.) (0.010) (0.010) (0.015) (0.016) (0.017) (0.017) (0.018) Constant (std. err.) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) N 3, 219, 341 3, 219, 341 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 Past Maker and Taker NO NO YES YES YES YES YES YES Past Volume NO NO NO YES YES YES YES YES Depth NO NO NO NO YES YES YES YES Amount and Time Since Last Trade NO NO NO NO NO YES YES YES Inside Spread NO NO NO NO NO NO YES YES Current Maker and Taker NO NO NO NO NO NO NO YES Half-hour Dummies NO YES YES YES YES YES YES YES Linear trend NO YES YES YES YES YES YES YES 43

45 Table 5: B-Make results from multinomial logit regression, with type of maker as dependent variable. B-Make equation (1) (2) (3) (4) (5) (6) (7) (8) 10-day before and after window afterτ (std. err.) (0.007) (0.017) (0.017) (0.017) (0.017) (0.017) (0.017) (0.017) τ (std. err.) (0.015) (0.015) (0.027) (0.027) (0.027) (0.027) (0.027) τ Afterτ (std. err.) (0.046) (0.043) (0.074) (0.075) (0.074) (0.074) (0.075) Constant (std. err.) (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) N 625, , , , , , , , day before and after window afterτ (std. err.) (0.005) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) τ (std. err.) (0.010) (0.010) (0.017) (0.017) (0.017) (0.017) (0.017) τ Afterτ (std. err.) (0.023) (0.023) (0.029) (0.030) (0.030) (0.030) (0.030) Constant (std. err.) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) N 1, 704, 596 1, 704, 596 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, day before and after window afterτ (std. err.) (0.004) (0.006) (0.006) (0.007) (0.007) (0.007) (0.007) (0.007) τ (std. err.) (0.006) (0.006) (0.012) (0.012) (0.012) (0.012) (0.012) τ Afterτ (std. err.) (0.014) (0.014) (0.020) (0.020) (0.020) (0.020) (0.020) Constant (std. err.) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) N 3, 219, 341 3, 219, 341 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 Past Maker and Taker NO NO YES YES YES YES YES YES Past Volume NO NO NO YES YES YES YES YES Depth NO NO NO NO YES YES YES YES Amount and Time Since Last Trade NO NO NO NO NO YES YES YES Inside Spread NO NO NO NO NO NO YES YES Current Maker and Taker NO NO NO NO NO NO NO YES Half-hour Dummies NO YES YES YES YES YES YES YES Linear trend NO YES YES YES YES YES YES YES 44

46 Table 6: P-Take results from multinomial logit regression, with type of taker as dependent variable. Standard errors are in parantheses below the estimates. P-Take equation (1) (2) (3) (4) (5) (6) (7) (8) 10-day before and after window afterτ (std. err.) (0.006) (0.012) (0.013) (0.013) (0.013) (0.013) (0.013) (0.013) τ (std. err.) (0.010) (0.011) (0.022) (0.021) (0.021) (0.021) (0.020) τ Afterτ (std. err.) (0.033) (0.033) (0.063) (0.062) (0.060) (0.061) (0.057) Constant (std. err.) (0.002) (0.005) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) N 625, , , , , , , , day before and after window afterτ (std. err.) (0.004) (0.007) (0.007) (0.008) (0.007) (0.008) (0.007) (0.008) τ (std. err.) (0.007) (0.007) (0.014) (0.014) (0.014) (0.014) (0.014) τ Afterτ (std. err.) (0.017) (0.018) (0.022) (0.023) (0.022) (0.022) (0.023) Constant (std. err.) (0.002) (0.003) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) N 1, 704, 596 1, 704, 596 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, day before and after window afterτ (std. err.) (0.003) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.006) τ (std. err.) (0.005) (0.005) (0.010) (0.010) (0.010) (0.010) (0.010) τ Afterτ (std. err.) (0.010) (0.011) (0.016) (0.016) (0.016) (0.016) (0.017) Constant (std. err.) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) N 3, 219, 341 3, 219, 341 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 Past Maker and Taker NO NO YES YES YES YES YES YES Past Volume NO NO NO YES YES YES YES YES Depth NO NO NO NO YES YES YES YES Amount and Time Since Last Trade NO NO NO NO NO YES YES YES Inside Spread NO NO NO NO NO NO YES YES Current Maker and Taker NO NO NO NO NO NO NO YES Half-hour Dummies NO YES YES YES YES YES YES YES Linear trend NO YES YES YES YES YES YES YES 45

47 Table 6: B-Take results from multinomial logit regression, with type of taker as dependent variable. Standard errors are in parantheses below the estimates. B-Take equation (1) (2) (3) (4) (5) (6) (7) (8) 10-day before and after window afterτ (std. err.) (0.007) (0.014) (0.016) (0.016) (0.016) (0.016) (0.016) (0.015) τ (std. err.) (0.012) (0.014) (0.028) (0.027) (0.027) (0.027) (0.026) τ Afterτ (std. err.) (0.034) (0.034) (0.064) (0.062) (0.061) (0.061) (0.059) Constant (std. err.) (0.002) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) N 625, , , , , , , , day before and after window afterτ (std. err.) (0.004) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) τ (std. err.) (0.007) (0.008) (0.016) (0.016) (0.015) (0.016) (0.015) τ Afterτ (std. err.) (0.019) (0.021) (0.025) (0.025) (0.025) (0.024) (0.025) Constant (std. err.) (0.001) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) N 1, 704, 596 1, 704, 596 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, day before and after window afterτ (std. err.) (0.003) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) τ (std. err.) (0.005) (0.005) (0.011) (0.010) (0.010) (0.010) (0.011) τ Afterτ (std. err.) (0.012) (0.013) (0.018) (0.018) (0.018) (0.018) (0.018) Constant (std. err.) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) N 3, 219, 341 3, 219, 341 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 Past Maker and Taker NO NO YES YES YES YES YES YES Past Volume NO NO NO YES YES YES YES YES Depth NO NO NO NO YES YES YES YES Amount and Time Since Last Trade NO NO NO NO NO YES YES YES Inside Spread NO NO NO NO NO NO YES YES Current Maker and Taker NO NO NO NO NO NO NO YES Half-hour Dummies NO YES YES YES YES YES YES YES Linear trend NO YES YES YES YES YES YES YES 46

48 Table 7: Summary statistics for mid-spreads and effective spreads, based on trade-by-trade data. 10-day before and after window Inside Spread Effective Spread Before After Before After Mean Std. Dev th prctile th prctile th prctile th prctile th prctile th prctile th prctile N 346, , , , day before and after window Inside Spread Effective Spread Before After Before After Mean Std. Dev th prctile th prctile th prctile th prctile th prctile th prctile th prctile N 912, , , , day before and after window Inside Spread Effective Spread Before After Before After Mean Std. Dev th prctile th prctile th prctile th prctile th prctile th prctile th prctile N 1, 726, 531 1, 492, 810 1, 726, 531 1, 492,

49 Table 8: Results from regressions for the effective spread. Standard errors are in parantheses below the estimates. (1) (2) (3) (4) (5) (6) (7) (8) 10-day before and after window afterτ (std. err.) (0.003) (0.007) (0.007) (0.007) (0.007) (0.007) (0.007) (0.007) τ (std. err.) (0.009) (0.009) (0.017) (0.017) (0.017) (0.017) (0.017) τ Afterτ (std. err.) (0.017) (0.017) (0.026) (0.027) (0.027) (0.026) (0.026) Constant (std. err.) (0.002) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) N 625, , , , , , , , day before and after window afterτ (std. err.) (0.002) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) τ (std. err.) (0.006) (0.006) (0.010) (0.011) (0.011) (0.010) (0.010) τ Afterτ (std. err.) (0.011) (0.010) (0.014) (0.015) (0.015) (0.015) (0.014) Constant (std. err.) (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) N 1, 704, 596 1, 704, 596 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, 536 1, 704, day before and after window afterτ (std. err.) (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) τ (std. err.) (0.005) (0.005) (0.008) (0.008) (0.008) (0.008) (0.008) τ Afterτ (std. err.) (0.008) (0.008) (0.011) (0.011) (0.011) (0.011) (0.011) Constant (std. err.) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) N 3, 219, 341 3, 219, 341 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 3, 219, 221 Past Maker and Taker NO NO YES YES YES YES YES YES Past Volume NO NO NO YES YES YES YES YES Depth NO NO NO NO YES YES YES YES Amount and Time Since Last Trade NO NO NO NO NO YES YES YES Inside Spread NO NO NO NO NO NO YES YES Current Maker and Taker NO NO NO NO NO NO NO YES Half-hour Dummies NO YES YES YES YES YES YES YES Linear trend NO YES YES YES YES YES YES YES 48

50 Table 9: Bid-ask spreads according to maker-taker type. Standard errors are in parantheses below the estimates. 10-day before and after window 30-day before and after window 60-day before and after window afterτ (std. err.) (0.007) (0.004) (0.003) τ (std. err.) (0.017) (0.010) (0.008) τ Afterτ (std. err.) (0.026) (0.014) (0.011) HBτ (std. err.) (0.008) (0.005) (0.004) HPτ (std. err.) (0.008) (0.005) (0.004) BHτ (std. err.) (0.011) (0.007) (0.005) BBτ (std. err.) (0.011) (0.007) (0.005) BPτ (std. err.) (0.010) (0.007) (0.005) P Hτ (std. err.) (0.009) (0.006) (0.004) P Bτ (std. err.) (0.010) (0.007) (0.005) P Pτ (std. err.) (0.011) (0.008) (0.006) HBτ Afterτ (std. err.) (0.011) (0.007) (0.005) HPτ Afterτ (std. err.) (0.011) (0.007) (0.005) BHτ Afterτ (std. err.) (0.015) (0.009) (0.006) BBτ Afterτ (std. err.) (0.016) (0.009) (0.007) BPτ Afterτ (std. err.) (0.015) (0.009) (0.006) P Hτ Afterτ (std. err.) (0.013) (0.008) (0.006) P Bτ Afterτ (std. err.) (0.015) (0.010) (0.007) P Pτ Afterτ (std. err.) (0.015) (0.010) (0.007) Constant (std. err.) (0.005) (0.003) (0.002) N 625, 460 1, 704, 536 3, 219, 221 Past Maker and Taker YES YES YES Past Volume YES YES YES Depth YES YES YES Amount and Time Since Last Trade YES YES YES Inside Spread YES YES YES Current Maker and Taker YES YES YES Half-hour Dummies YES YES YES Linear trend YES YES YES 49

51 Figure 1: Quotes interrupted within 250m as a fraction of all submitted quotes. The series are standardized and indexed such that the data starts at 100. Figure 2: Number of submitted and interrupted quotes. The series are standardized and indexed such that the data starts at 100 in each panel. Figure 3: Bid-ask spreads and depths on the EBS and Futures market. The series are logtransformed and indexed to each begin at 100. Figure 4: EBS and Futures volumes. The series are log-transformed and indexed to each begin at 100. Figure 5: Empirical distribution with detrended futures volume. Figure 6: Empirical distribution with the log-change in the bid-ask spread Figure 7: Empirical distribution for the log-change in depth0. Figure 8: Make- and Take-Volumes on EBS. These are log-volumes, standardized such that H-Make and H-Take, respectively, starts at

52 120 Quotes Interrupted within 250ms as a Fraction of All Submitted Quotes MQL Implemented Feb2009 Apr2009 Jun2009 Aug2009 Oct2009 Dec2009 Date Figure 1: Quotes interrupted within 250m as a fraction of all submitted quotes. The data are standardized and indexed such that the series starts at 100.

53 Number of Submitted Quotes MQL Implemented Feb2009 Apr2009 Jun2009 Aug2009 Oct2009 Dec2009 Date Number of Interrupted Quotes MQL Implemented Feb2009 Apr2009 Jun2009 Aug2009 Oct2009 Dec2009 Date Figure 2: Combined quote plots. The data are standardized and indexed such that the series starts at 100.

54 EBS Bid Ask Spread Futures Bid Ask Spread Bid Ask Spread MQL Implemented Jan2008 Jun2008 Jan2009 Jun2009 Jan2010 Jun2010 Date Depth at the Best Bid and Ask (depth0) MQL Implemented EBS depth0 Futures depth0 0 Jan2008 Jun2008 Jan2009 Jun2009 Jan2010 Jun2010 Date Figure 3: EBS and Futures bid-ask spreads and depths. The series are log-transformed and indexed to each begin at 100.

55 EBS Volume Futures Volume EBS and Futures Volume MQL Implemented 90 Jan2008 Jun2008 Jan2009 Jun2009 Jan2010 Jun2010 Date EBS Volume Detrended Futures Volume EBS and Detrended Futures Volume MQL Implemented 90 Jan2008 Jun2008 Jan2009 Jun2009 Jan2010 Jun2010 Date Figure 4: EBS and Futures volumes. The series are log-transformed and indexed to each begin at 100.

56 10 Day Change in EBS Volume 10 Day Change in Detr ended Futur es volume 10 Day Diff in Diff 15 MQL Change P ercentile: MQL Change P ercentile: MQL Change P ercentile: Day Change in EBS Volume 30 Day Change in Detr ended Futur es volume 30 Day Diff in Diff 15 MQL Change P ercentile: MQL Change P ercentile: MQL Change P ercentile: Day Change in EBS Volume 60 Day Change in Detr ended Futur es volume 60 Day Diff in Diff 15 MQL Change P ercentile: MQL Change P ercentile: MQL Change P ercentile: Figure 5: Empirical distribution for the log-change in trading volume.

57 10 D ay C hange in EB S Spread 10 Day Change in Futures Spread 10 Day Diff in Diff 25 MQL Change P ercentile: MQL Change P ercentile: MQL Change P ercentile: D ay C hange in EB S Spread Day Change in Futures Spread Day Diff in Diff 25 MQL Change P ercentile: MQL Change P ercentile: MQL Change P ercentile: D ay C hange in EB S Spread Day Change in Futures Spread Day Diff in Diff 25 MQL Change P ercentile: MQL Change P ercentile: MQL Change P ercentile: Figure 6: Empirical distribution for the log-change in the bid-ask spread.

58 10 Day Change in EBS Depth0 10 Day Change in Futures Depth0 10 Day Diff in Diff MQL Change 20 MQL Change 20 MQL Change 15 P ercentile: P ercentile: P ercentile: Day Change in EBS Depth0 30 Day Change in Futures Depth0 30 Day Diff in Diff MQL Change 20 MQL Change 20 MQL Change 15 P ercentile: P ercentile: P ercentile: Day Change in EBS Depth0 60 Day Change in Futures Depth0 60 Day Diff in Diff MQL Change 20 MQL Change 20 MQL Change 15 P ercentile: P ercentile: P ercentile: Figure 7: Empirical distribution for the log-change in the depth0.

59 H Make Volume P Make Volume B Make Volume EBS Make Volumes MQL Implemented 70 Jan2008 Jun2008 Jan2009 Jun2009 Jan2010 Jun2010 Date H Take Volume P Take Volume B Take Volume EBS Take Volumes MQL Implemented 70 Jan2008 Jun2008 Jan2009 Jun2009 Jan2010 Jun2010 Date Figure 8: Make- and Take-Volumes on EBS. These are log-volumes, standardized such that H-Make and H-Take, respectively, starts at 100.