Machines vs. Machines: High Frequency Trading and Hard Information

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Machines vs. Machines: High Frequency Trading and Hard Information Yesol Huh 2014-33 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Machines vs. Machines: High Frequency Trading and Hard Information Yesol Huh∗ Federal Reserve Board First draft: November 2012 This version: March 2014 Abstract Intoday’smarketswherehighfrequencytraders(HFTs)bothprovideandtakeliquidity,whatinfluencesHFTs’liquidityprovision? Iarguethatinformationasymmetryinducedbyliquidity-takingHFTs’ use of machine-readable information is important. Applying a novel statistical approach to measure HFT activity and using a natural experiment of index inclusion, I show that liquidity-providing HFTs supply less liquidity to stocks that suffer more from this information asymmetry problem. Moreover, whenmarketsarevolatile,thisinformationasymmetryproblembecomesmoresevere,andHFTssupply less liquidity. I discuss implications for market-making activity in times of market stress and for HFT regulations. ∗I would like to thank Stefan Nagel, Arthur Korteweg, and Paul Pfleiderer for their feedback and discussions on this paper. I would also like to thank Shai Bernstein, Darrell Duffie, Sebastian Infante, Heather Tookes, Clara Vega, and the seminar participants at the Stanford GSB, Federal Reserve Board, Vanderbilt University, and University of Rochester for helpfulsuggestions,andNASDAQforprovidingthelimitorderbookdata. Theviewsexpressedinthisarticlearesoleythose oftheauthorandshouldnotbeinterpretedasreflectingtheviewsoftheFederalReserveBoardortheFederalReserveSystem. Allcommentsandsuggestionsarewelcome. email: yesol.huh@frb.gov 1

1 Introduction Financial markets are much faster and more automated than they were a decade ago. There has been a rise inanewtypeofalgorithmictradingcalledhigh frequency trading (IuseHFTtodenotebothhighfrequency tradingandhighfrequencytrader),whichemploysveryfastconnectionandcomputingspeed. Inthelast5–8 years, HFTshaveexplodedingrowthandareestimatedtoaccountforanywherebetween40to85%ofdaily volumeinequitymarkets.1 Inlightofthe2010FlashCrash,HFThasbeenasubjectofintensepublicdebate, with various controversial aspects such as price manipulation, order anticipation, fairness, and liquidity. In this paper, I study what influences HFTs’ liquidity provision. HFTs have taken over the market-making business to a large extent, raising concerns that since HFTs have no market making obligations, they might fleewhenneededthemost.2 Thus,understandingthedeterminantsofHFTs’liquidityprovisionisimportant. While we know from a long-standing empirical microstructure literature much about what affects liquidity, given the differences between HFTs and traditional market makers, an interesting question is whether there are additional factors that are now relevant. In particular, I focus on how liquidity-taking HFTs affect liquidity-providing HFTs, and show that the information asymmetry induced by liquidity-taking HFTs’ use of machine-readable information is an important factor. Toillustratetheintuition,let’sconsiderasimpleexampleoftheApplestock. Assumethereisaliquidityproviding HFT that quotes in the Apple stock and a liquidity-taking HFT that trades in this stock. I define liquidity-providing HFTs as HFTs that supply limit orders and liquidity-taking HFTs as those that submit market orders.3,4 Assume the two HFTs have same reaction speed on average, the liquidity-providing HFT is faster 50% of the time, and the liquidity-taking HFT is faster other 50% of the time. Assume further that the liquidity-taking HFT uses the NASDAQ-100 ETF price as an input, and the liquidity-providing HFT, knowingthis,alsowatchestheETFclosely. WhentheETFpricemoves,halfofthetime,theliquidity-taking HFT sees this first and submits market orders before the liquidity-providing HFT has a chance to adjust his quotes. At those times, the liquidity-providing HFT is adversely selected. If he takes this into account ex ante, he will increase spreads and decrease liquidity. Therefore, the higher the information asymmetry associated with ETF price information, the less liquidity HFTs provide. 1For example, TABB Group estimates that 73% of volume comes from HFTs. Table 1 of Brogaard et al. (2012) indicates thatHFTsparticipateinsomewherebetween43and85%oftradesinNASDAQ. 2“One of the primary complaints of traditional, slower investors, like mutual funds, is that high-speed trading firms flood themarketwithordersexceptintimesofcrisis,whentheordersaremostneeded.”,TheNewYorkTimes,‘ToRegulateRapid Traders,S.E.C.TurnstoOneofThem’,Oct7,2012 3Marketablelimitorderswillbereferredtoasmarketordersthroughoutthepaper. 4Hagstr¨omerandNord´en(2012)documentthatsomeHFTsactmostlyasmarketmakerswhileotheractmostlyasliquidity takers or opportunistic traders. If a same entity sometimes submit limit orders and use market orders at other times, he is thoughtofasliquidityproviderwhenheuseslimitorders,andaliquiditytakerwhenheusesmarketorders. 2

Intheaboveexample,twodefiningcharacteristicsofHFTsareimportant. First,HFTscanreceivemarket data and can submit, cancel, and replace orders extremely quickly, in order of milliseconds or even faster. Secondly, HFTs utilize machine-readable information that are relevant for prices. This type of information is referred to as hard information. A common example of hard information for HFTs are prices of indexed products such as index futures or ETFs (Jovanovic and Menkveld (2011) and Zhang (2012)), as they tend to lead the cash market. In this paper, I use a specific example of hard information, the price of the NASDAQ-100 ETF (QQQ), and show that the associated information asymmetry problem exists and that the market-making HFTs supply less liquidity to stocks that are more prone to this problem. Moreover, I show that when market volatilityincreases,thisinformationasymmetryincreasesandHFTssupplylessliquidity. Liquiditysupplyby HFTs is estimated using Hawkes’ self-exciting model, and the degree of information asymmetry is measured as the fraction of transactions triggered by ETF price movement. My results also indicate that stocks with low spreads, high beta, and low volatility have a greater information asymmetry. Another contribution of this paper is the methodology used to measure HFT activities. By directly estimating reaction time, the Hawkes self-exciting model characterizes limit order book dynamics in a way that easily separates the activities of HFTs from those of slower traders. An advantage of this method is that it does not require any trader classification data, which is not readily available. Moreover, this method also allows the liquidity-providing and the liquidity-taking HFT activities to be identified and measured separately, whereas many of the existing empirical studies may capture more of the liquidity-providing HFTs or more of the liquidity-taking HFTs depending on the event or the measure used. Because HFTs playamajorroleinliquidityprovisionintoday’smarkets,isolatingandstudyingliquidityprovidedbyHFTs is relevant. Although the Hawkes self-exciting model has been used in a few other papers to study liquidity replenishment (Large (2007), Toke (2011)), to the best of my knowledge, this is the first paper to utilize a model of this sort to study HFT activities. The information asymmetry studied here differs from the usual information asymmetry based on private informationinthatmosthardinformationisgenerallyconsideredpublic. However,duringthebriefperiodof timeinwhichsometradersobserveaparticularpieceofhardinformationandothersdonot,thatinformation caneffectivelybethoughtofasprivate. Giventhatvirtuallyallmodelsofinformationasymmetryrequirethe privateinformationbeatleastpartiallyrevealedatalaterorterminalperiod,mechanismsforthisparticular type of information asymmetry are not different. However, this specific type of information asymmetry is distinct and important for a couple of reasons. 3

Thedegreeofthisinformationasymmetryincreaseswhenmarketsarevolatile,asindexedproductsexperience morefrequentpricechangesduringsuchperiods. Therefore,theinformationasymmetryassociatedwithhard information will increase, and market-making HFTs will supply less liquidity. Empirically validating this is important since as mentioned before, a major concern about HFTs dominating the market making business the quality of liquidity they provide at times of market stress. I show that when market volatility increases by one standard deviation, the information asymmetry measure increases by 0.86 standard deviations and liquidityreplenishmentdecreasesby0.67standarddeviations,oralternatively,increasesby19%anddecreases by 10%, respectively. In contrast, with the traditional information asymmetry studied in the literature, whether it increases with market volatility is unclear. Even before the advent of HFTs, there were traders who used hard information to trade in individual stocks,5 but the information asymmetry stemming from this activity has rarely been studied. Today, with the growth of HFTs, the information asymmetry associated with hard information is much more important. Moreover, the usual measures of information asymmetry (the PIN measure of Easley et al. (1996), etc.) are not equipped to measure this specific type of information asymmetry. Fortunately, the fact that the aggressive HFTs are the ones that engage in exploiting hard information allows one to measure the degree of this information asymmetry directly. Finally, many governments around the world are considering regulating HFTs. Analyzing how each proposal might affect this type of information asymmetry is important, as it would influence the liquidity provision by HFTs. For example, a regulation that handicaps liquidity providers without affecting liquidity takers may have the unintended consequence of increasing information asymmetry and decreasing liquidity provision. This paper is organized as follows. Section 2 reviews the relevant literature, and Section 3 introduces the Hawkes model and the data, as well as providing the baseline results of the Hawkes model. In Section 4, I construct the information asymmetry measures and the liquidity replenishment measure, and study the relationship between the two in Section 5. Section 6 studies which stock characteristics are important determinants of information asymmetry. Section 7 concludes. 5Ithasbeenempiricallywelldocumentedthatindexfuturesleadthespotmarket,andalargeamountofcapitalisinvolved inthearbitragebetweenfuturesandspotmarket. Althoughthisfutures-spotarbitragedoesnotnecessarilymeantradersare usingfuturesinformationtotradeinindividualstocks,arbitrageactivitycombinedwiththelead-lagrelationshipwillimplicitly entailsuchbehavior. 4

2 Literature Review A growing body of research directly models a trading game between slow human traders and fast HFTs to studytheeffectofHFTs’speedadvantage. MostdirectlyrelevanttothispaperareJovanovicandMenkveld (2011), Foucault et al. (2012), and Martinez and Rosu (2011), which study how HFTs’ speed advantage in reading hard information affects liquidity. In these papers, as well as most other papers studying HFTs, HFTs are modeled to be liquidity providers or liquidity takers, but not both.6 Interestingly, depending on whether HFTs are modeled as liquidity providers or takers, one reaches an opposite conclusion about whether they increase liquidity. One side argues that speed gives HFTs a natural advantage in market-making: HFTs can adjust their quotes faster when hard information comes out and consequently are less subject to adverse selection (Jovanovic and Menkveld (2011)). For example, an HFT with the best outstanding ask quote in a stock can cancel the quote when he observes an increase in index futures price before others have a chance to transact against his quote. Therefore, market-making HFTs increase liquidity and decrease bid-ask spreads. On the other hand, if HFTs are modeled as liquidity takers, they submit market orders when they receive hard information (as in Foucault et al. (2012)), thereby imposing an adverse selection problem on limit order providers.7 For instance, when an HFT sees an increase in futures price, he can submit a market buy order in a stock before the ask quote gets cancelled. If limit order providers incorporate this ex ante, then spreads willbehigherandliquiditywillbelower. Insum,HFTscaneithersolveorimposetheexactsameproblemof adverse selection associated with hard information, and increase or decrease liquidity depending on whether they are liquidity providers or takers. In reality, however, neither scenario completely captures the effects of HFTs since some HFTs act as liquidity providers and others act as liquidity takers. Instead, liquidity-taking HFTs exacerbate the information asymmetry problem, while liquidity-providing HFTs mitigate it. Then, the question is whether the information asymmetry problem still exists, and if so, whether the liquidity-providing HFTs take this into account when providing liquidity. Intuitively, if the liquidity-providing HFTs were always faster than the liquidity-taking HFTs, this information asymmetry problem would be completely resolved. However, if no 6For example, Jovanovic and Menkveld (2011) and Gerig and Michayluk (2010) models HFT as a liquidity provider, and Foucault et al. (2012) and Martinez and Rosu (2011) models it as a liquidity taker. Empirical papers with trader identities distinguishbetweenliquidity-providingandliquidity-takingHFTs,buttheymostlyfocusontheaggregategamebetweenHFTs andnon-HFTs. 7Insomeofthepapersinthiscategory,HFTsareassumedtogetmorepreciseinformation. Thisbothincreasesanddecreases liquidity, becauseinthestaticsense, itinducesadverseselectionanddecreasesliquidity, butitalsomakespricesdynamically more informative, so market makers quote a tighter spread (Martinez and Rosu (2011)). Foucault et al. (2012) compares a model where HFTs are both faster and receive more precise information to a model where they only receive more precise information,andfindthatliquidityislowerinthefirstmodel. 5

onetypeoftraderisnecessarilyfaster,8 market-makingHFTshavetobeawareofbeingadverselyselectedby aggressive HFTs. This paper can be thought of as an empirical test of whether this information asymmetry problem exists and whether the market-making HFTs are sensitive to it, explicitly taking into account both types of HFTs. Althoughboundbydataavailability,theempiricalliteratureonHFThasbeensteadilygrowinginrecent years, and it has yielded a few themes pertinent for this paper. Jovanovic and Menkveld (2011) and Zhang (2012) are especially relevant, who find that HFTs act on hard information faster than non-HFTs. Also related are Brogaard et al. (2012) and Menkveld (2012), who show that when HFTs provide liquidity, on average, their limit orders are adversely selected and yet they do make a positive profit. Hagstr¨omer and Nord´en (2012) empirically document that there are diverse types of HFTs, some that act mostly as market makers and others that engage in statistical arbitrage or momentum strategies. Usingdifferentnaturalexperiments,RiordanandStorkenmaier(2012)andHendershottetal.(2011)find that an increase in algorithmic trading decreased adverse selection cost. On the other hand, Hendershott andMoulton(2011)findsthatincreaseinspeedandautomationattheNewYorkStockExchangeincreased adverse selection cost. These seemingly opposite conclusions may be the result of each ‘natural experiment’ increasing different types of algorithmic trading. In general, in most of the empirical studies on HFT mentioned above, the same issue appears: depending on the data or the HFT classification used, these studies might be capturing more of liquidity-providing or -taking HFTs. Hagstr¨omer and Nord´en (2012) do look at the two types separately, but do not consider how they affect each other. In contrast, I capture the liquidity-providing and liquidity-taking HFT activities separately and study their joint dynamics. Also, this paper empirically tests a specific channel by which HFTs affect the market. Most empirical papers use proprietary data provided by exchanges that identify HFTs in some way. Several papers (Brogaard et al. (2012), Zhang (2012), and Hirschey (2011)) use a dataset from NASDAQ that identifies each side of the trade as HFT or non-HFT. Menkveld (2012) and Jovanovic and Menkveld (2011) use data from Chi-X with anonymous IDs, and Hagstr¨omer and Nord´en (2012) use similar data from the NASDAQ OMX Stockholm. One advantage of the methodology used in this paper is that it does not require knowing the trader type, but simply infers it using data. Limit order book data is more readily available than trader identity data, which are not only proprietary and rarely available, but also rely on 8One scenario is that there are multiple HFTs of different types in the fastest group. Given that HFT companies invest heavilytogaincomparativeadvantageoverotherHFTs,thisscenarioisverylikely. Forexample,SpreadNetworks,anetwork provider that caters to HFTs, built a new route from Chicago to New York to shave a couple milliseconds off the latency of theoriginalcableline. AnothercompanyisbuildinganunderseatransatlanticcablefromNewYorktoLondontodecreasethe latencyby5milliseconds. 6

the data provider (usually the exchange) to correctly identify which traders are HFTs. Hasbrouck and Saar (2012)isanexception: theyconstructameasureofHFTthatusesthesamelimitorderbookdataasinthis paper, by detecting a particular HFT order submission strategy. But their measure does not separate out the liquidity-providing and liquidity-taking HFT activities, and my measure most likely captures a broader set of activities than theirs. Theempiricalmethodologyusedherecanbethoughtofasanextensionofstudyingtheconditionalevent frequencies which has often been employed to study limit order book dynamics (Biais et al. (1995), Ellul et al. (2007)). These empirical papers have mostly found that after an order, the next order is likely to be of the same type, and that the limit order book moves in the direction of restoring its balance in the longer term (Biais et al. (1995), Ellul et al. (2007)). As Gould et al. (2010) point out, most of the studies in this area use older data, mostly from the 1990s, so it is unclear the extent to which their results hold true today. Moreover, the Hawkes model has the added benefit of measuring reaction time, allowing one to explicitlyaccountforHFTs. Thus, comparingthebaselineresultsoftheHawkesmodelpresentedinSection 3 with the results from older literature can provide us an insight as to how the limit order book dynamics have changed. For example, while the existing studies show that market buy order is most likely after a market buy, in today’s markets, improvement in bid and improvement in ask are both more likely. Results are discussed further in Section 3.4 and Appendix D. 3 Model and Data 3.1 Intuition Figure 1 illustrates the basic idea of measuring HFT activity. Let’s say we are interested in the degree to which HFTs replenish liquidity. The idea is to look at how likely it is and how long it takes the best bid or ask to revert back towards the original level after a market order moves it. Initially, the stock in this figure has a best bid of 37.92 and a best ask of 38. Let’s say a market buy transacted against the best ask and exhausted the liquidity at that price, so the best ask is now 38.05. We want to answer the question of how fast the best ask comes back down towards 38, and how likely it is. To do so, we need to know how fast and likely will a new sell limit order be submitted below the current best ask of 38.05. As long as the new sell limit order is below 38.05, I count it as liquidity replenishment. Thus, all three scenarios <1>, <2>, and <3> in the figure qualify. If and only the new sell limit order is submitted very soon and changes the best ask, I label it as liquidity being replenished by HFTs. Hawkes’ self-exciting model in the following section 7

38.05 38.05 38.02 38.00 38.00 37.98 37.92 37.92 37.92 37.92 37.92 initial mktbuy <1> <2> <3> Figure 1: Illustration of measuring HFT liquidity replenishment Blue ticks are bids, and brown ticks are asks. After a market buy, the limit order book looks like the second figure. All three figures on the right count as liquidity replenishment. will let us formalize the intuition outlined here. 3.2 Hawkes’ self-exciting process The goal in this section is to statistically model limit order book dynamics in a way that can identify HFT activities. Also, the model should be able to incorporate ETF events to study the particular information asymmetry in question. Limit order book events happen at irregular intervals, sometimes multiple events in the same millisecond, and sometimes as far apart as a few minutes. If we want to divide data by regular intervals, we have to use very fine intervals in order to study HFTs, which is highly impractical. On the other hand, using event time does not let us incorporate the ‘wall-clock’ time (‘real’ time), which is crucial for thinking about HFTs. Thus, I treat the data as a realization of a point process and model the intensity in continuous time.9 Thisallowsmetopreserveboththeorderingofeventsandthewall-clocktime. Theintensityisparametrized as a function of all past events and the duration between now and each past event (‘backward recurrence time’). This model, first proposed by Hawkes (1971), allows us to directly measure the following: when an event of type m happens, how likely and how fast will an event of type r happen as a reaction. ThereareM typesoflimitorderevents, andeachpoint(‘event’)isofaparticulartype. Forexample, “a 9Generally speaking, the dynamic intensity of a point process (PP) can be formulated as an autoregressive conditional intensity(ACI)modelorasaself-excitingmodel. Theformerapproach,proposedbyRussell(1999),specifiestheintensityas anautoregressiveprocesswithcovariatesupdatedwhenaneventoccurs. Theoneusedhereisthelatterapproach. 8

new sell order that decreases the best ask” is an event type. I specify the intensity of the arrival of type r as M (cid:90) (cid:88) λ (t)=µ (t)+ W (t−u)dN (u) r r rm m m=1 (0,t) M (cid:88) (cid:88) =µ (t)+ W (t−tm), (1) r rm i m=1tm<t i where N (t) is the counting process for event type m, W (t − u) captures the effect of event m that m rm happened on time u<t on the intensity of event r at time t, and tm is the ith occurrence of type m event. i W is parametrized as rm W rm (u)=a( r 1 m )b( r 1 m )e−b( r 1 m )u+a( r 2 m )b( r 2 m )e−b( r 2 m )u, (2) whereb(1) >b(2) >0anda(1), a(2) >0.10 µ (t)isthebaselineintensityandismodeledasapiecewiselinear rm rm rm rm r functionwithknotsat9:40, 10:00, 12:00, 1:30, 3:30, and3:50tocaptureintradaypatterns. Aversionofthis model with M = 1 and a constant µ(t) was first proposed by Hawkes (1971). Bowsher (2007) studies the Hawkes model with M >1, and Large (2007) uses a specification similar to the one proposed here to study the resiliency of the limit order book. Conditions for existence and uniqueness are discussed in Appendix A. The parameters a and b have an intuitive interpretation. For simplicity, first consider a case in which the function W rm (u) has only one term, thus W rm (u)=a rm b rm e−brmu. Let’s say type m event occurred at time s. Then the intensity of type r will jump up by a b and this jump will decay exponentially at a rm rm rate b . Figure 2 illustrates this effect. This model is called self-exciting because when there is only one rm typeofevent(M =1),whenaneventhappens,itsintensityincreases,thusitismorelikelytohappenagain. In the above scenario, what is the difference in the expected number of event r occurrences with respect tothecounterfactualworldinwhichtheeventmdidnotoccurattimes? DefineF asthenaturalfiltration s up to, but not including s, and F+ as the filtration including s. Then the value in question is s G (t−s)=λ(t|F+)−λ(t|F ), (3) rm s s where t > s. G can be defined as a M ×M matrix where the element (r,m) is G (t−s) from the above rm equation. Then, as shown in Proposition 4.1 of Large (2007), this is a function of (t−s) (invariant of s), 10One restriction is that a’s cannot be negative, thus we cannot allow event m to hinder event r. Bowsher (2007) extends the model to allow for negative a’s while ensuring that intensity stays non-negative. I do not adopt this approach because it makesestimationsmoredifficultandtime-consuming,anditisunlikelythatatemporarydecreasewouldreverseveryquickly. 9

λ r armbrm a rm s t Figure 2: Illustration of a Hawkes model and can be written as (cid:90) G(u)=W(u)+ W(u−z)G(z)dz. (4) (0,u) The first term in the above equation is the direct effect, and the second term captures the chain reaction. I focus on the direct effect because the value in interest is the immediate reaction to each event. Since (cid:90) W(u)du=a , (5) rm (0,∞) theexpectednumberofeventrhappeningdirectlybecauseofaspecificoccurrenceofeventmisa . Iloosely rm alsocallthis“probability”(“Afteraneventoftypemoccurs,eventr willoccurwitha probability.”11) or rm the“sizeoftheeffect.” Thehalflifeoftheeffectislog(2)/b duetoexponentialdecay. Becauseb(1) >b(2), rm rm rm the first term in W (u) (in (2)) captures the faster reaction. I thus focus on the first term in the analysis rm later to study HFTs. The second term is included to capture the actions of slower agents; without it, the first term might be confounded by the actions of non-HFTs.12 11Preciselyspeaking,thisisnotacorrectstatementbecauseinsomestatesoftheworld,eventrmightoccurtwiceduetothe specificoccurrenceofeventm,sotheprobabilityofr occurringwillactuallybelessthanarm. However,sincethisstatement capturesthemainintuitionandmakestheinterpretationeasierandlesswordy,Iuseitinaloosesense. 12ItisalsopossibletoincludemorethantwotermsinWrm(u)(in(2))tocaptureadditionalspeeddifferences,forinstance, ifwebelievetherearethreedifferentspeedgroups. Preliminaryanalysiswiththreetermsshowsimilara(1) andb(1) estimates. Adding more terms can make the interpretation tricky as the half-life of the fastest group and the second fastest group may become too close, so that the second term may also pick up HFT actions. Wrm(u) can also be other functional forms, but exponential decay is the easiest to estimate and gives the most natural interpretation. One can also think about slightly modifyingthefunctionalformtoincorporateminimumlatency,butthisrequiresagoodestimateofwhattheminimumlatency is. 10

If we observe data points in (0,T], the log likelihood function is M (cid:40) (cid:90) (cid:90) (cid:41) (cid:88) l(θ)= (1−λ (s|θ))ds+ logλ (s|θ)dN (s) , (6) r r r r=1 (0,T] (0,T] whereθisthevectorofmodelparametersdefinedasθ =(θ ,··· ,θ )andθ ={a ,··· ,a ,b ,··· ,b ,µ (t)}. 1 M r r1 rM r1 rM r Since λ is independent of all θ for r (cid:54)=r(cid:48), the log likelihood function can be rewritten as r r(cid:48) M M (cid:40) (cid:90) (cid:90) (cid:41) (cid:88) (cid:88) l(θ)= l (θ )= (1−λ (s|θ ))ds+ logλ (s|θ )dN (s) . (7) r r r r r r r r=1 r=1 (0,T] (0,T] This brings down the number of variables to be estimated jointly as we can estimate θ separately by maxr imizing l (θ ). Ogata (1978) establishes that, under certain regularity conditions, the maximum likelihood r r estimator is consistent and asymptotically normal for a univariate Hawkes process with a constant base intensity of µ(t) = µ. Bowsher (2007) extends it to multivariate and non-stationary processes. Expression of l (θ ) in terms of a and b is in Appendix B. For derivation of likelihood function and the asymptotic r r convergence of the maximum likelihood estimator, also see Appendix B and citations within. This approach allows conditioning on all past events, but because the effects are additive, it does not allow for a more complicated conditioning such as “if event 1 follows event 3.” Such a possibility might be useful as many HFTs use “fleeting orders” (see Hasbrouck and Saar (2009), for example), in which a trader submits a limit order and cancels it quickly. In the current framework, it is impossible to distinguish the effect of a limit order addition that is not cancelled quickly from the one that is cancelled quickly. To consider such effects, one can expand the event space. 3.3 Data NASDAQ TotalView-ITCH is a comprehensive datafeed from NASDAQ that many market participants, especiallyalgorithmicandhighfrequencytraders,subscribetoinordertoreceivelimitorderbookinformation in real time with low latency. I use the historical record of this datafeed. It contains all order additions, cancellations, executions, and modifications, as well as trades of hidden orders and other system messages. Each action is time stamped in milliseconds during my sample period. Because each order can be tracked throughout the day using the order reference number, the visible limit order book can be reconstructed for any point in time. Whether a transaction was initiated by a buyer or a seller is clear, thus eliminating the need to sign each trade. Data structure and order book matching rules are similar to those described 11

in Hasbrouck and Saar (2009).13 Orders are matched in price-visibility-time priority basis. This is an extremely active market; on a random day (September 9th, 2008) in the sample, there were 530 million electronic messages. I use a sample period of 43 trading days between August 15 and October 15, 2008. This period is specificallychosen foritslarge variationinmarketvolatility. Iuse92 stocksinthe NASDAQ100 indexthat continuously traded throughout the sample period and did not experience any major events such as stock splitsormergerannouncements. IndexmemberstocksareusedbecauseETFpriceinformationwillcertainly matter for these stocks. For the ETF, I use QQQ, the most popular ETF that tracks the NASDAQ-100 index. I reconstruct the limit order book and record the best bid and ask throughout the trading day for the sample stocks, and for each change in the best bid or ask price, I record whether it was caused by a market order, a cancellation, or a new limit order. I also include the market orders that did not change the best bid or ask. In sum, there are 8 types of events: 1. MB (Market Buy): Market buy that increases the ask 2. MS (Market Sell): Market sell that decreases the bid 3. AB (Add Bid): Incoming buy limit order that increases the bid 4. AA (Add Ask): Incoming sell limit order that decreases the ask 5. CB (Cancel Bid): Cancellation of an outstanding buy limit order that decreases the bid 6. CA (Cancel Ask): Cancellation of an outstanding sell limit order that increases the ask 7. MB2: Market buy that does not move the ask 8. MS2: Market sell that does not move the bid A single market order that is transacted against multiple limit orders show up as multiple observations in the original data; I combine them into one observation.14 Events that have the same millisecond time stamp are evenly spaced out over the same millisecond.15 13HasbrouckandSaar(2009)usethelimitorderbookdatafromINET.NASDAQmergedwithINETandBrutin2006and integratedtheplatform; thecurrentplatformisbasedontheINETsystem. TheITCHdatahavefinertimestampsandfiner datacategoriesinmysampleperiod,butthecoreremainsthesame. 14I am able to do this quite accurately because a single market order transacted against multiple limit orders appear as consecutive observations in the data with the same time stamp. ITCH data is sequenced in time order for all securities that tradeintheNASDAQplatform;thusitisunlikelythattherearemultipleconsecutiveobservationsforthesamesecurityunless theyaretrulyconsecutive. 15Oneexceptioniswhenamarketablelimitorderwithasizelargerthanthecurrentbestbid/askdepthexecutesagainstthe currentbestbidoraskandtheremainderisaddedasalimitorder. SeeAppendixCfordetails. 12

Oftenlargemarketordersarethoughttobeinformed,whereassmallerordersareconsidereduninformed. Biaisetal.(1995)findevidenceconsistentwiththis. However, inlinewithFarmeretal.(2004), preliminary analysis using a subset of the data shows that a single market order that transacts at multiple price levels (‘walk up the book’) is quite rare. Therefore, I do not distinguish between them and market orders that only take out the best bid or ask. To observe the effects that the ETF has on individual stocks, I add four ETF events to the earlier list: 9. ETF↑: Increase in ETF ask 10. ETF↓: Decrease in ETF bid 11. ETF2↑: Increase in ETF bid 12. ETF2↓: Decrease in ETF ask Event types will be referred to by their acronyms. 3.4 Estimation of Hawkes self-exciting model The effect of event m on event r has two dimensions: the size of the effect, a , and the speed of the effect, rm which is characterized by its half-life log2/b . However, comparisons are difficult if both dimensions vary rm across stocks, so I fix the half-lives to be the same across stocks and days in the following way. First, for each stock, I estimate the full model for each r.16 I take the cross-sectional median of the half-life log2/b(j) rm for each (r,m,j) combination, and set all b(j)’s such that the half-life equals the median value. The model rm is then reestimated for each stock-day pair to obtain a(j) and the base intensities. rm Here I fix the half-lives to be constant across stocks and days and look at how the size of the effect changes. But what if the size of the effect stays constant but the effect happens more slowly? Intuitively, in this scenario, if we estimate the misspecified model of constant half-lives, it will appear as if the size of the effectdecreasesbecauseIhaveassumedthatthehalf-lifeisshorterthanthetruevalue. AppendixEconfirms this using simulations. Since both the decrease in the size of the effect and the increase in the half-life can be thought of as a decrease in the effect overall, inferences using the current specification should be robust. 16Since this step involves calculating a likelihood value for a point process spanning over multiple days, I assume that each dayisanindependentrealizationofamultivariateHawkesprocesswiththesameparameters. 13

3.5 Other variables Market capitalization and price are taken from CRSP as of the end of July 2008, and beta and volatility are calculated using daily stock return data from CRSP between August 2007 and July 2008. Quoted spreads are calculated as time-weighted bid-ask spreads and constructed for each stock daily using the limit order bookdata. PanelAofTable1providesthesummarystatisticsofvariousstockcharacteristicsforthesample. 3.6 Results of the baseline model Estimates from the Hawkes model are presented in Table 2a and 2b. Table 2a presents the estimates of the faster effect (first term), Table 2b the slower effect. Each row represents the triggering event type, and each column represents the affected event type. For each pair, I present the mean estimated half-life, mean estimated a (effect size), and the percentage of the observations that are significant at the 5% level. Since I am interested in HFT activities, I mostly focus on the faster effect. Rowsrepresenttriggeringeventtypes,andcolumnsareaffectedeventtypes. Forexample,meana(1) AA,MB is 19.9% with half life of 1.7ms. In other words, after an event of type MB (market buy) happens, event of type AA happens with half life of 1.7ms roughly 19.9% of the time. From Table 2a, we can see that reactions to market buys and sells that change the best price tend to be fastest and strongest. After a market buy order increases the ask, there is 19.9% probability that the ask comes back down very fast (event type AA), 21.9% likelihood that bid increases very fast (event type AB), 7% likelihood that there is another market buy although slightly slower (event type MB), and 5% likelihood that the ask increases further due to a cancellation (event type CA).17 Appendix D discusses the results in more detail and compares to the results from the existing literature on limit order book dynamics. 4 Measures of information asymmetry and liquidity replenishment In this section, I first verify that HFTs use ETF price information to trade in individual stocks, and that this use induces an information asymmetry problem. I then measure the degree of information asymmetry and HFTs’ liquidity provision. 17Asnotedbefore,theseareroughlyclosetoprobabilities,butnotquite,andthusdoesn’tnecessarilysumupto100%. 14

4.1 ETF price as hard information Price of index futures is a natural candidate for an example of hard information that HFTs might use in individual stocks, since index futures tend to lead price discovery. Kawaller et al. (1987) and Stoll and Whaley(1990)confirmthatindexfuturesgenerallyleadthecashindex,andJovanovicandMenkveld(2011) and Zhang (2012) show that HFTs use index futures price information to quote and trade in individual stocks. Unfortunately,becausetheexactsequencingandrelativetimedifferencearecrucialinmyanalysis,using futurescanbeproblematicsincefuturesandstockstradeondifferentplatformsandthushaveunsynchronized timestamps. For this reason, I use ETF data instead of futures. QQQ is listed and mostly traded on NASDAQ,18 whereas the NASDAQ-100 E-mini futures trades on Chicago Merchandise Exchange’s Globex platform. One concern in using the ETF price is that unlike index futures, ETFs might not lead individual stocks inpricediscovery,especiallysinceithasbeendocumentedthatETFsgenerallylagindexfutures(Hasbrouck (2003)). However, Tse et al. (2006) find that while E-mini futures do lead price discovery most frequently, but that ETFs also do lead price discovery a significant fraction of the time and more frequently than the cashindex. EvenifthefuturesoftenleadETFs,usingETFdatashouldworkaslongasETFsgenerallyreact faster to price changes in futures than individual stocks do. Unfortunately, there are no empirical studies I am aware of that directly looks at whether ETFs lead the cash index, but it seems plausible that ETFs respond to futures price changes before stocks do, because ETFs are more closely related to index futures. To verify that the ETF price is indeed a good example of hard information, I show first that traders can benefit from using ETF information when trading in individual stocks (thus, liquidity providers that do not condition on ETF information sufficiently fast face higher adverse selection costs), and secondly that liquidity-takingHFTsdoindeedtradeonETFinformation. Ifirstdividealltradesthatchangethebestbid oraskintoETF-drivenandnon-ETF-driventrades,asillustratedinFigure3. ForeachpricechangeinETF, I characterize the market orders of individual stocks that are in the same direction within 50 milliseconds after the ETF price change as ‘ETF-driven’, all others as ‘non-ETF-driven’.19 In order to measure the adverse selection cost for each trade, I use the standard spread decomposition 18QQQ was originally listed on the American Stock Exchange under the ticker QQQ (hence often called “cubes”), but in December2004itmovedtoNASDAQandchangedthetickertoQQQQ.InMarch2011itchangedthetickerbacktoQQQ. 19Table 2a indicates that after ETF price changes, a market order of same direction follows in a stock with 2% probability and33.8mshalf-life. 50msischosentolettheintervalbecomparabletobutslightlylongerthanthehalf-life. 15

50ms buy sell sell buy ETF↑ ETF non-ETF non-ETF non-ETF -driven -driven -driven -driven Figure 3: Characterizing ETF-driven orders Each “X” marks a market order, with the “buy” and “sell” above it indicating the direction. and calculate the permanent price impact for the j’th transaction of stock i that occurs at time t as PImpact(k) =q (m −m )/m , (8) i,j,t i,j i,t+k i,t i,t where q is 1 for buyer-initiated and -1 for seller-initiated transactions. m is the midpoint at time t− i,j i,t (rightbeforethetransactionj),andm istheprevailingmidpointattimet+k. Forexample,ifthetrade i,t+k was a buy and the midpoint increased by 1% permanently after the trade, that is how much the trader who provided the limit order lost if he held the position permanently. Typically, 5 or 30 minutes are used for k, the measurement duration, but since HFTs tend to have a shorter holding period, I calculate the measure using various shorter durations (50ms, 100ms, 500ms, 1s, 5s, 10s, 1m, 5m). The effective half spread is measured as ESpread =q (p −m )/m , (9) i,j,t i,j i,j,t i,t i,t where p is the transaction price. i,j,t For each stock and day, I separately calculate the mean permanent price impact and the mean effective half spread for ETF-driven transactions and for non-ETF-driven transactions. To test whether ETF-driven trades pose an adverse selection problem for the limit order providers, I run PImpact(k) =γ 1(l is ETF-driven)+A +B +(cid:15) , (10) i,l,τ 1 i τ i,τ ESpread =γ 1(l is ETF-driven)+A +B +(cid:15) , (11) i,l,τ 2 i τ i,τ where l is either ETF-driven or non-ETF-driven. For each stock i and day τ, there are two observations: one is the mean of ETF-driven market orders and another is that of non-ETF-driven ones. A and B are i τ stock and day fixed effects. Results are presented in Table 3. ETF-driven orders have larger permanent price impact, both statistically and economically, over all 16

measurement durations. The permanent price impact is 3–24% higher for ETF-driven orders, and the difference is largest when measured over a 5-second window. ETF-driven orders are a bigger problem for liquidity-providing HFTs than for slower market makers—when the holding period increases, the difference decreases. Thus, ETF-driven orders make a profit and pose an adverse selection problem for liquidityproviding HFTs. The second aspect to be tested is whether HFTs actually act on ETF price information. In other words, ifETFpricechanges,doHFTstradeinthesamedirectioninindividualstocksrightaway,aboveandbeyond theirnormalfrequency? Toanalyzethis,IreturntotheestimatesfromtheHawkesmodelthatarepresented in Table 2a. a , a , a , and a correspond to this effect. For example, MB,ETF↑ MB,ETF2↑ MS,ETF↓ MS,ETF2↓ the first term, a measures the probability that event ETF↑ (increase in ETF ask) triggers event MB,ETF↑ MB (market buy). a and a are fast and statistically significant with an average value of MB,ETF↑ MS,ETF↓ 2.2% and a half-life of 34ms, and 83% of the values are significant at a 5% level. An average value of 2.2% is fairly low, but given that the ETF price changes much more frequently than most stocks, this is not too surprising. 4.2 Measures of information asymmetry 50ms buy sell sell buy ETF↑ ETF opposite driven Figure 4: Characterizing market orders for PET calculation With the validity ofETF price as hard information established, I now construct measures of information asymmetryassociatedwithaggressiveHFTsusingETFpriceinformation. Foreachstock-daypair,Icalculate the probability of ETF-driven trading (PET), that is, the probability that a market order that changes the best bid or ask is an ETF-driven trade. As before, I characterize the market orders of underlying stocks that are in the same direction within 50 milliseconds after an ETF price change as ‘ETF-driven’, but I additionally characterize the ones that move in the opposite direction as ‘opposite’. Figure 4 illustrates this classification. PET is calculated as the difference in the fraction of market orders that are ETF-driven and the fraction of opposite ones. The difference is taken to control for the mechanical effect that can arise from the number of ETF price changes. A single market order can be counted as both ETF-driven and opposite 17

if the ETF moved in both directions during the last 50 milliseconds. In sum, PET is measured as # of ETF-driven mkt orders−# of opposite mkt orders PET = . (12) i,τ # of total price-changing mkt orders The PET measure, unfortunately, may present some measurement issues. Imagine a situation in which ETF buys and individual stock buys are correlated for some other reason. In this case, the PET measure is positive, although ETF information is not driving the trades in the stock. A likely scenario is that futures price changes represent an unobserved common factor that leads HFTs to trade in the ETF and the underlying stocks with about the same response time. This is not much of a concern, since I can rename the futures price to be the hard information, and the economic mechanism will remain the same. More worrisome is a case in which the common driver of the ETF and the stock trade is not hard information. For instance, a possible source of such “soft” information could be news about company fundamentals. To address this issue, I introduce another information asymmetry measure that is not affected by such sources and check the PET measure against it. Reaction to ETF change (REC) is measured as a(1) +a(1) REC = MB,ETF↑ MS,ETF↓, (13) 2 where a(1) and a(1) are estimates from the Hawkes model. As mentioned before, this measures MB,ETF↑ MS,ETF↓ the probability of a same-direction price-changing market order happening in a stock after an ETF price change. a(1) and a(1) are excluded since they are slower and weaker, but the results are MB2,ETF↑ MS2,ETF↓ similar when they are included. Both the PET and REC measures are constructed for each stock for each day. Panel B of Table 1 provides summary statistics for these measures. Theoretically, PET is a more correct measure for my purposes than REC. Because I want to study how theinformationasymmetryaffectstheamountofliquidityprovidedbymarket-makingHFTs,measuringthe information asymmetry from the point of view of the liquidity provider would be most appropriate. Since PET measures the probability that a given price-changing market order is driven by an ETF price change, PETbetterfitsthisdescription. Incontrast, fortwostockswiththesameRECvalue, ifonestockhasmuch less uninformed trading, then adverse selection cost will be higher in that stock for the liquidity provider. However,RECservestwousefulpurposes: checkingPETagainsttheconcernslistedaboveanddecomposing PET into two components, which will be described later. The REC measure is derived from the Hawkes model, and the estimates for the fast term in the Hawkes 18

model are robust to the concerns on correlated latent intensity. The intuition is that the only scenario that a common latent driver of ETF and market orders is captured in the a(1), the fast term in the Hawkes model,iswhenthecommonlatentintensityprocessismean-revertingrapidly,inasimilarspeedasthemean reversion of the fast process measured in the Hawkes model. If the common latent process does mean-revert in milliseconds, what drives that common intensity process must also be hard information. The unobserved hard information can be relabeled as the new hard information, and REC will still measure the information asymmetry associated with such hard information. Appendix F further formalizes and shows this intuition through simulations. I confirm that PET is a sensible measure by looking at whether PET is higher for stocks with a higher REC measure. PET =γ +γ REC +(cid:15) (14) i 0 1 i i ItaketheaveragePETandRECvaluesforeachstocki,winsorizethemata5%leveltocontrolforoutliers, and then standardize them by subtracting the cross-sectional mean and dividing by the cross-sectional standard deviation to get PET and REC . As a result, both of them have a sample mean of zero and a i i standard deviation of one. I get γ = 0.44 with t value of 3.97. (Similar result with non-winsorized, non-standardized PET is 1 presented in the fourth column of Table 9.) In other words, for a stock with a REC one standard deviation i higher, PET is 0.44 standard deviations higher. Therefore, the concerns about correlated latent intensity i are somewhat mitigated. Appendix G provides another robustness check. Another advantage of constructing the REC measure is that it allows a natural decomposition of the PET measure that will come in handy later for understanding where the variation in PET comes from. If we define ϕ as the intensity of the ETF price change on day τ (q index stands for QQQ), and ϕ as q,τ i,τ non-ETF-driven trading intensity on stock i on day τ, the ETF-driven trading intensity in stock i will be ϕ REC . The PET measure should then theoretically be q,τ i,τ ϕ REC PET = q,τ i,τ . (15) i,τ ϕ REC +ϕ q,τ i,τ i,τ Thus, holding everything else equal, stocks with a high REC and low non-ETF-driven trading will have a higher PET. Non-ETF-driven trading includes uninformed trading, informed trading on soft information, and informed trading on other hard information. 19

4.3 Liquidity replenishment by HFTs In this section, I measure HFTs’ liquidity provision. While liquidity can be defined and measured in various ways, I focus on liquidity replenishement. Here liquidity replenishment is defined somewhat narrowly as the probability of a bid (ask) coming back up (down) after a market sell (buy) that changes the best bid (ask). Under this definition, a and a measure liquidity replenishment in the Hawkes model. Figure 1 AA,MB AB,MS presented earlier is an example of liquidity replenishment, and a measures the size of this reaction. AA,MB Since the terms a(1) and a(1) capture the fast response, liquidity replenishment of HFT market AA,MB AB,MS makers (LR) is defined as 1 LR= (a(1) +a(1) ). (16) 2 AA,MB AB,MS LR is calculated daily for each stock. Summary statistics for the cross-section are provided in Panel B of Table 1. Liquidity replenishment is closely related to the concept of “resiliency,” which is “the speed with which prices recover from a random, uninformative shock” (Kyle (1985)). While resiliency is less studied than the other dimensions of liquidity such as spread and depth, Foucault et al. (2005) and others focus on this particulardimension. However,mydefinitionofliquidityreplenishmentisdifferentinthatitdoesnotinclude anincreaseinbidafteramarketbuyasliquidityreplenishmentwhereasinFoucaultetal.(2005)andothers, they count it towards resiliency. In a Glosten and Milgrom (1985)-type model, a market buy order has positive information, and thus after a market buy, the bid will increase to reflect the revealed information. Since the increasing of bid after a market buy may be mostly due to revelation of information than liquidity provision, I do not include it. 5 Relationship between liquidity replenishment and information asymmetry 5.1 Cross-sectional relationship The main objective of this paper is to examine the relationship between the adverse selection problem that liquidity-takingHFTsposeandtheliquidityprovidedbymarket-makingHFTs. HereItestthecross-sectional relationship on whether HFTs provide less liquidity to stocks that face greater information asymmetry. As noted earlier, PET is the more relevant measure because it directly measures the probability that a limit 20

order is adversely selected. I thus use PET as a measure of information asymmetry. I run the cross-sectional regression LR =γ +γ PET +γ lmktcap +γ beta +γ volatility +γ price +γ spread +(cid:15) . (17) i 0 1 i 2 i 3 i 4 i 5 i 6 i i LR and PET are the average LR and PET values for stock i, respectively. PET, price, and spread are i i winsorized at a 5% level (summary statistics in Table 1 indicate that these variables have large outliers in the cross section), and all right-hand-side variables are standardized to have a mean of zero and a standard deviation of one to make interpretation easier. If HFTs provide less liquidity to stocks with higher information asymmetry of this particular type, γ , 1 the coefficient for PET, should be negative. Table 4 presents the results. LR is 0.199 on average, and one standard deviation increase in PET decreases LR by 0.027, which is 12.7% of the mean value. This effect persistsaftercontrollingforotherstockcharacteristics, andtheincrementalR2 is5.2%. Therefore, liquidity replenishmentisindeedlowerforstockswithahigherPETmeasure,andtheeffectisbotheconomicallyand statistically significant. 5.2 Natural experiment using index inclusions Although the above tests have controlled for various stock characteristics, it still is possible that the results are driven by correlations due to missing variables. I thus use the rebalancing of the NASDAQ-100 index as a natural experiment to test for causality. The NASDAQ-100 index is rebalanced every December in the following way. All non-financial stocks listed on NASDAQ are ranked by market value. For the stocks that are currently in the index, they have to satisfy the following criteria to remain in the index: they should either be ranked in the top 100, or currently ranked in 101–125 and have ranked in the top 100 in the previous year. Stocks that do not satisfy either are replaced with those that are not yet in the index and highest in the rank. After market close on December 12, 2008, NASDAQ announced that it would replace 11 securities, and the change went into effect before the market opened on December 22. When a stock is newly added in the index, its price will have a direct mechanical relationship with the index value and consequently the ETF price, thus increasing the REC measure. Whether it will increase PETisunclear. Indexinclusionwilltendtoalsoincreaseuninformedtrading,soasisevidentfromthePET decompositionin(15),PETcouldmoveineitherdirectiondependingonwhicheffectdominates. Ultimately, it therefore is an empirical question. 21

I construct data for the period 12/15/2008–12/31/2008, the 10 days surrounding the index change, for stocks that ranked 10th to 109th in market capitalization as of the end of November. This size range is chosen to exclude the stocks that leave the index since the majority of them were already very small and thinly traded before the rebalancing. The nine largest stocks are excluded in an attempt to keep the newly joined stocks and existing index stocks comparable in size. Since the sample period starts after the index rebalancing was announced, there are no information effects related to the news about the change. REC, PET, and LR are constructed daily for each stock as before. The first stage is to confirm that REC increases when a stock has been added, and to test whether PET increases as well. I run the regression y =A +B +γ 1(i newly included)1(τ after inclusion)+(cid:15) , (18) i,τ i τ 1 i,τ where y is REC or PET, 1(i newly included) is an indicator variable that is 1 if i is a newly included stock, and1(τ after inclusion)isanindicatorvariablethatis1ifτ isaftertherebalancedate. Iusethreedifferent normalizationsforthey’s: oneusesrawvalues(‘level’),onesubtractsthestock-specificmeananddividesby the stock-specific mean (‘proportional’), and onesubtracts thestock-specific meanand dividesbythe stockspecific standard deviation (‘standardized’). Results are presented in Panel A of Table 5. REC increases 80% or 58% depending on the measure used, and PET increases 25–44% and is statistically significant. Since PET does increase with index inclusion, I now progress to the 2SLS estimation LR =A +B +γ PET +ξ , (19) i,τ i τ 2 i,τ i,τ wherethefirststageis(18). PanelBofTable5presentstheresults. Anincreaseof0.1inPETdecreasesLR by0.126,orequivalently,aonestandarddeviationincreaseinPETdecreaseLRby0.23standarddeviations. Thus,informationasymmetryisnotmerelynegativelycorrelatedwithHFTliquiditysupply,butdoesindeed cause HFTs to supply less liquidity. 5.3 Time series relationship with market volatility Since the ETF or other securities that are sources of hard information are based on broad market indices, their prices will change more frequently when the market is volatile. Hard information, in turn, will arrive more often, increasing information asymmetry. This more frequent arrival of hard information corresponds to an increase in ϕ , the intensity of ETF q 22

price change, from the PET decomposition (15), which increases PET if all other variables stay the same. However, ϕ , the non-ETF-driven trading, is likely to increase when markets are volatile, as trading volume i in general increases during those times. REC may decrease as some of the increase in hard information may actually be uninformative for the stocks, or it simply may not be feasible to respond to all ETF price changeswiththesameprobability. BothofthesemoveinthedirectionofdecreasingPET,althoughitseems unlikely that these effects will completely counteract the increase in PET driven by the increase in ϕ . If q information asymmetry increases, market-making HFTs should respond and decrease liquidity supply. I first study whether information asymmetry increases and liquidity replenishment decreases on average, when markets become volatile. Market volatility on day τ is measured as the standard deviation in oneminute returns of the ETF. This has direct connections to both market volatility and the number of price changes in the ETF. Figure 5 shows the relationship between the daily average PET and market volatility, and Figure 6 shows the relationship between the daily average LR and market volatilty. These graphs strongly suggest that on average, PET increases and LR decreases with volatility. To statistically confirm the above results, I run the following regression y =γ +γ ETFvol +(cid:15) , (20) τ 0 1 τ τ where y is the average PET or LR on day τ, and ETFvol is the standardized ETF volatility (subtract the τ mean and divide by the standard deviation). I present two sets of results in Table 6, one using raw values for y and another using standardized values. Since the latter one is just a scaled version of the former one, the two sets are statistically equivalent, but I present both for interpretation. Regression results presented in Table 6 is consistent with the hypothesis. A one standard deviation increaseinETFvolatilityincreasesPETby0.86standarddeviation,orfrom7.9%to9.4%. Italsodecreases LR by 0.67 standard deviations, or from 19.9% to 18.0%. Although this does not necessarily prove that the decrease in LR during volatile markets is caused by the increase in PET, it seems to be the most likely explanation,and,extrapolatingfromthenaturalexperimentinSection5.2,areasonableone. Inthefollowing section, I aim to establish causality in a somewhat indirect manner by utilizing cross-sectional differences. 5.4 Sensitivity to market volatility Ifmarketvolatilityincreasesinformationasymmetry,andinformationasymmetryinturndecreasesliquidity replenishment, it must be the case that stocks subject to a greater increase in information asymmetry 23

experience a greater decrease in liquidity replenishment. I first estimate each stock’s sensitivity of PET to ETF volatility. For each stock i, I run the following regression to estimate the sensitivity γ . 1i PET =γ +γ ETFvol +(cid:15) . (21) i,τ 0i 1i τ i,τ I run three different specifications: one using raw (level) values for PET, one using proportional values (subtractthestock-specificmeananddividebythestock-specificmean),andthelastoneusingstandardized values(subtractthestock-specificmeananddividebythestock-specificstandarddeviation). Afterobtaining γ values for all 92 stocks, I standardize them by subtracting the mean and dividing by the standard 1i deviation. Standardized values are named γ˜ . Then I run the main regression 1i LR =A +(δ +δ γ˜ )ETFvol +ξ . (22) i,τ i 0 1 1i τ i,τ Since γ˜ has a mean of zero, δ measures how much LR moves for an average γ stock for a one unit change 1 0 1 in ETF vol. We should get δ < 0 in line with results from the previous time-series regression. If stocks 0 with higher γ experience a larger decrease in liquidity replenishment when markets become more volatile, 1 δ isnegative. Tobeconsistent, ifγ˜ isestimatedusinglevelvaluesin(21), IuselevelvaluesforLRin(22), 1 1 proportional values if the first stage used proportional values, and standardized values if the first stage used standardized values. Standardized values are used for ETF vol. Results are presented in Table 7. δ and δ are both negative and statistically significant. The second 0 1 columntellsus,forexample,thataonestandarddeviationincreaseinmarketvolatilitydecreasesLRby8.5% on average, and that stocks with γ one standard deviation higher experience an additional 4.8% decrease. 1 For which stocks is information asymmetry more sensitive to market volatility? Also, consistent with the above results, do stocks with greater sensitivity also encounter a larger decrease in LR? To check this, I estimate y =γ +(γ +γ x )ETFvol +(cid:15) , (23) i,τ 0 1 2 i τ i,τ whereyisPET,REC,orLRmeasuredforeachstockforeachday. x ’sarestockcharacteristics(standardized i cross-sectionally) such as size, beta, volatility, and ex ante REC and PET. For the latter two, I use the first five days of the sample to calculate the average ex ante values. (23) is estimated using data from day 6 onwards. Two variables arise as significant: size and REC. I present the results with these two variables in 24

Table 8. Large stocks and low REC stocks experience a greater increase in PET when markets become volatile. REC decreases on average when volatility increases, as hypothesized before, but the decrease is smaller for large stocks and low REC stocks. Consistent with the notion that stocks with higher PET sensitivity to market volatility should see a greater decrease in LR, large stocks and low REC stocks do see a greater decrease in LR. Investigating why PET increases more for these stocks is left for future work. 6 Cross-sectional determinants of information asymmetry So far I have established that the information asymmetry problem associated with hard information is an important determinant of HFTs’ liquidity provision. While I have treated the degree of information asymmetry mostly as given, it is actually determined endogenously. Thus, in this section I study which stocks have a greater information asymmetry problem of this particular type. Ingeneral,informationasymmetryisgreaterforstocksofcompanieswithlessdisclosure(Welker(1995)), higher insider holdings, larger number of shareholders (Glosten and Harris (1988)), and lower institutional ownership (Jennings et al. (2002)). However, for the specific type of information asymmetry studied here, different set of variables are likely to matter. Let’s say the ETF price increased by one tick. Which stock would aggressive HFTs buy? In other words, which stocks have a higher REC measure? I hypothesize that they would buy large, active, high-priced, liquid stocks and stocks that closely follow the index. Larger stocks have a bigger weight in the index, and thus have a stronger mechanical relationship with the ETF. Also, HFTs will have an easier time unwinding the trades later in more active stocks, and transaction costs will be lower in stocks with smaller bid-ask spreads. Because the tick size is discrete, they should be less likely to buy low-priced stocks; a one-cent increase in a $10 stock is fractionally much higher than the same one-cent increase in a $50 stock, so a one-cent increase in the ETF may be enough to buy the $50 stock but not the $10 stock. Stocks that track the index closely are another natural candidate, as HFTs bear less idiosyncratic risk due to higher predictability using ETF prices. “Tracking the index closely” is equivalent to a smaller idiosyncratic volatility and a higher beta, or a higher “R2 in one-factor CAPM regression” as in Jovanovic andMenkveld(2011). Adebatablepointiswhatfrequencydatashouldbeusedtomeasurethis. Giventhat HFTs rarely hold positions for longer than a day, the more relevant data might be intraday data (Jovanovic and Menkveld (2011) use intraday data in this spirit), but there can also be reverse causality in which the 25

stocks that aggressive HFTs act on more often track the index more closely. The finer the data used, the more severe this problem will become. For example, if one uses 1 millisecond data, it will precisely pick up thestocksthatHFTsbuyafterETFpriceincreases. SinceIammoreinterestedinwhichstockfundamentals determine HFTs’ decisions, I use beta and stock volatility calculated from daily data. The above predictions pertain directly to the REC measure, but also have implications for the PET measure since PET and REC are positively correlated. On the other hand, the decomposition in (15) indicates that after controlling for REC, PET is higher for stocks with low non-ETF-driven trading. Thus, holdingRECequal,PETshouldbehigherforstockswithlowuninformedtrading—small,low-spreadstocks. Itisnotclearwhatthepredictionsareforbeta,volatility,andprice. Forexample,uninformedtradingmight belowinstockswithvolatilefundamentals,butitmightalsobethatsomestocksarevolatilebecausetrading volumes are high. The effect of price is most likely non-linear; stocks with very high price are illiquid, but stocks with very low price also trade infrequently. These predictions are tested in the following cross-sectional regressions REC =γ +γ lmktcap +γ beta +γ volatility +γ price +γ spread +(cid:15) , (24) i 0 1 i 2 i 3 i 4 i 5 i i PET =δ +δ lmktcap +δ beta +δ volatility +δ price +δ spread +δ REC +ξ . (25) i 0 1 i 2 i 3 i 4 i 5 i 6 i i REC and PET are the average value of REC and PET for stock i over the sample period. As in Section i i 5.1, price and spread are winsorized at a 5% level, and REC is winsorized at a 5% level for (25) to control for outliers,20 and all explanatory variables are standardized (subtract the mean and divide by standard deviation) to make interpretation easier. Results are presented in Table 9. One standard deviation increase in size increases REC by 29% (=0.0058/0.0199), but once I control for beta and volatility, this becomes insignificant. As predicted, stock with higher beta, lower volatility, higher price, and lower spreads have a higher REC. ThelastthreecolumnsofTable9presenttheresultsfrom(25). Asexpected,PETishigherforhighREC stocks. Controlling for REC, higher beta, lower volatility, lower price, and lower spread stocks have a lower PET.Thecoefficientforsizeismarginallysignificant,withlargerstockshavinglowerPET.Thelastcolumn shows the results for the PET regression with REC omitted. Stocks with higher beta, lower volatility, lower price,andlowerspreadhavehigherPET.ComparedtotheRECregressioninthethirdcolumn,onepossible explanationforwhythecoefficientforpricehastheoppositesignisthattheeffectthroughnon-ETF-trading 20ResultsarequalitativelysamewhenPETandRECarewinsorizedata5%levelwhentheyappearasadependentvariable aswell. 26

dominates. Information asymmetry, measured either as REC or PET, is higher for stocks that track the index closely. The fact that the stocks with lower spreads have greater information asymmetry seems contrary to the usual notion of information asymmetry (think back to Glosten and Milgrom (1985), where information asymmetry is a determinant of bid-ask spread) and almost counter-intuitive. One way to think about this is to first assume that the spreads vary cross-sectionally for reasons unrelated to information asymmetry associated with hard information. Then, since liquidity-taking HFTs are more likely to act on stocks with lower spreads, those will have greater information asymmetry of this specific type. The feedback loop, in which liquidity providers quote higher spreads for stocks with greater information asymmetry, will mitigate the original magnitude. If anything, then, the coefficients on spread in the cross-sectional regressions (24) and (25) are underestimated. Thishasinterestingimplicationsforliquidityreplenishment;stockswithlowerspreadshavelowerliquidity replenishment, as can be seen from the third and fifth columns of Table 4. This also may seem counterintuitive; since liquidity replenishment decreases spreads, holding all else equal, spreads must be higher for stocks with lower liquidity replenishment. However, this is reconciled by the fact that stocks with low liquidity replenishment tend to have high opposite replenishment. In other words, for those stocks, after a market buy that increases the ask, it is less likely for the ask to come back down but more likely for the bid to go up. 7 Conclusion and Discussion Inthispaper,Ishowthatthespecifictypeofinformationasymmetrythatarisesduetospeedandtheuseof hardinformationisimportantintoday’smarkets. Thisproblembecomesmoreseverewithmarketvolatility. These results have interesting implications for the ongoing HFT debate and regulations. AmajorconcernaboutHFTsreplacingtraditionalmarketmakersisthatsinceHFTsdonothavemarket making obligations, they might leave the market when market makers are needed the most. Although my sample period does not cover certain extreme events such as the 2010 Flash Crash (the market turmoil in 2008 is arguably quite extreme as well, albeit in a different way), I do document that the market-making HFTsprovidelessliquidityreplenishmentwhenmarketsarevolatile. However,itistheaggressiveHFTsthat impose higher adverse selection cost on the market-making HFTs; even if market making were solely done by humans, they would still face the same information asymmetry problem and provide less liquidity during 27

volatile times. Human market makers would be even more adversely selected, because they are unlikely to observe hard information before the aggressive HFTs do. Another implication is that the information asymmetry associated with hard information should be consideredwhenwethinkaboutpotentialHFTregulations. Forinstance,variousparties,includingtheSEC and the European Commission, have put forward a proposal for imposing a minimum quote duration.21,22 The main logic behind this suggestion is that HFTs intentionally or unintentionally overflow the exchange system and datafeed by submitting and canceling a large number of orders that do not benefit anyone. Adopting this proposal might have an unintended consequence of increasing information asymmetry, as it woulddisadvantagetheliquidity-providingHFTs’abilitytoquicklyadjusttheirquotesforhardinformation without handicapping the aggressive HFTs. If the sole purpose of the proposal is to limit the amount of data flooding caused by HFTs, other measures such as imposing a fine for a low executions-to-orders ratio might be a better idea. More broadly, in both modeling and empirically studying HFTs, we should think about what changes HFTs’ incentives for liquidity provision and for liquidity taking. While much of the literature has focused on the game between computers and humans, the speed of computers have far surpassed that of human cognition. Thus, when new technology enables computers to become faster, the increase in their relative speed to humans have no material impact anymore; it is the relative speed difference amongst computers that changes. Thus, an interesting question is, when a faster technology becomes available, who would be the first to adopt? That in turn will determine what the impact is on non-HFTs. Finally, the Hawkes model can also be used to study other questions related to HFTs. For example, although only briefly touched upon in this paper, limit order book dynamics in the presence of HFTs are notyetwellunderstood—howcertainHFTstrategiesaffectthelimitorderbook,whyHFTssometimesuse a long chain of fleeting orders that make prices oscillate, how cancellations are used, etc. The methodology presented and used in this paper will hopefully be a useful tool for future research in addressing such questions. 21SecuritiesandExchangeCommission,ConceptReleaseonEquityMarketStructure,January2010. 22EuropeanCommission–DirectorateGeneralInternalMarketandServices. PublicConsultation: Reviewofthemarketsin financialinstrumentsdirective(MiFID).December2010. 28

References Biais, B., Hillion, P. and Spatt, C. (1995), ‘An empirical analysis of the limit order book and the order flow in the Paris Bourse’, Journal of Finance 50(5), 1655–1689. Bowsher, C. (2007), ‘Modelling security market events in continuous time: Intensity based, multivariate point process models’, Journal of Econometrics 141(2), 876–912. Brogaard,J.,Hendershott,T.andRiordan,R.(2012),‘Highfrequencytradingandpricediscovery’,Working Paper . Easley,D.,Kiefer,N.,O’Hara,M.andPaperman,J.(1996),‘Liquidity,information,andinfrequentlytraded stocks’, Journal of Finance 51(4), 1405–1436. Ellul, A., Holden, C., Jain, P. and Jennings, R. (2007), ‘Order dynamics: Recent evidence from the NYSE’, Journal of Empirical Finance 14(5), 636–661. Embrechts, P., Liniger, T. and Lin, L. (2011), ‘Multivariate hawkes processes: An application to financial data’, Journal of Applied Probability 48, 367–378. Farmer, J., Gillemot, L., Lillo, F., Mike, S. and Sen, A. (2004), ‘What really causes large price changes?’, Quantitative Finance 4(4), 383–397. Foucault, T., Hombert, J. and Rosu, I. (2012), ‘News trading and speed’, Working Paper . Foucault, T., Kadan, O. and Kandel, E. (2005), ‘Limit order book as a market for liquidity’, Review of Financial Studies 18(4), 1171–1217. Gerig, A. and Michayluk, D. (2010), ‘Automated liquidity provision and the demise of traditional market making’, Working Paper . Glosten, L. and Harris, L. (1988), ‘Estimating the components of the bid/ask spread’, Journal of Financial Economics 21(1), 123–142. Glosten, L. and Milgrom, P. (1985), ‘Bid, ask and transaction prices in a specialist market with heterogeneously informed traders’, Journal of Financial Economics 14(1), 71–100. Gould,M.,Porter,M.,Williams,S.,McDonald,M.,Fenn,D.andHowison,S.(2010),‘Thelimitorderbook: A survey’, Arxiv preprint arXiv10120349 . 29

Griffiths, M., Smith, B., Turnbull, D. and White, R. (2000), ‘The costs and determinants of order aggressiveness’, Journal of Financial Economics 56(1), 65–88. Hagstr¨omer, B. and Nord´en, L. (2012), ‘The diversity of high frequency traders’, Working Paper . Hasbrouck, J. (2003), ‘Intraday price formation in U.S. equity index markets’, Journal of Finance 58(6), 2375–2400. Hasbrouck, J. and Saar, G. (2009), ‘Technology and liquidity provision: The blurring of traditional definitions’, Journal of Financial Markets 12(2), 143–172. Hasbrouck, J. and Saar, G. (2012), ‘Low-latency trading’, Working Paper . Hawkes, A. (1971), ‘Spectra of some self-exciting and mutually exciting point processes’, Biometrika 58(1), 83–90. Hendershott, T., Jones, C. and Menkveld, A. (2011), ‘Does algorithmic trading improve liquidity?’, The Journal of Finance 66(1), 1–33. Hendershott, T. and Moulton, P. (2011), ‘Automation, speed, and stock market quality: The nyse’s hybrid’, Journal of Financial Markets 14(4), 568–604. Hendershott, T. and Riordan, R. (2012), ‘Algorithmic trading and the market for liquidity’, Working Paper . Hirschey, N. (2011), ‘Do high-frequency traders anticipate buying and selling pressure?’, Working Paper . Jennings, W., Schnatterly, K. and Seguin, P. (2002), ‘Institutional ownership, information and liquidity’, Advances in Financial Economics 7, 41–71. Jovanovic, B. and Menkveld, A. (2011), ‘Middlemen in limit order markets’, Manuscript, VU University Amsterdam . Kawaller, I., Koch, P. and Koch, T. (1987), ‘The temporal price relationship between S&P 500 futures and the S&P 500 index’, Journal of Finance 42(5), 1309–1329. Kyle, A. (1985), ‘Continuous auctions and insider trading’, Econometrica 53(6), 1315–1335. Large, J. (2007), ‘Measuring the resiliency of an electronic limit order book’, Journal of Financial Markets 10(1), 1–25. 30

Martinez, V. and Rosu, I. (2011), ‘High frequency traders, news and volatility’, Working Paper . Menkveld, A. (2012), ‘High frequency trading and the new-market makers’, Working Paper . Ogata, Y. (1978), ‘The asymptotic behaviour of maximum likelihood estimators for stationary point processes’, Annals of the Institute of Statistical Mathematics 30, 243–261. Ogata, Y. (1981), ‘On Lewis’ simulation method for point processes’, IEEE Transactions on Information Theory 27(1), 23–31. Parlour, C. (1998), ‘Price dynamics in limit order markets’, Review of Financial Studies 11(4), 789–816. Ranaldo, A. (2004), ‘Order aggressiveness in limit order book markets’, Journal of Financial Markets 7(1), 53–74. Riordan, R. and Storkenmaier, A. (2012), ‘Latency, liquidity and price discovery’, Journal of Financial Markets 15(4), 416–437. Ro¸su, I. (2009), ‘A dynamic model of the limit order book’, Review of Financial Studies 22(11), 4601–4641. Russell, J. (1999), ‘Econometric modeling of multivariate irregularly-spaced high-frequency data’, Manuscript, GSB, University of Chicago . Stoll, H. and Whaley, R. (1990), ‘The dynamics of stock index and stock index futures returns’, Journal of Financial and Quantitative Analysis 25(04), 441–468. Toke, I. (2011), “‘Market making” in an order book model and its impact on the spread’, Econophysics of Order-driven Markets pp. 49–64. Tse, Y., Bandyopadhyay, P. and Shen, Y. (2006), ‘Intraday price discovery in the DJIA index markets’, Journal of Business Finance & Accounting 33(9-10), 1572–1585. Welker,M.(1995),‘Disclosurepolicy,informationasymmetry,andliquidityinequitymarkets’,Contemporary Accounting Research 11(2), 801–827. Yeo, W. (2006), ‘Persistence in limit order flow’, Working Paper . Zhang, S. (2012), ‘Need for speed: An empirical analysis of hard and soft information in a high frequency world’, Working Paper . 31

scitsitats yrammuS :1 elbaT ,scitsiretcarahckcotsehtstneserpAlenaP .sisylanaehtfoytirojamehtnidesuelpmasehtrofscitsitatsyrammusehtsedivorpwolebelbatehT dradnats ,naidem ,naem tneserp I .erusaem tnemhsinelper ytidiuqil TFH eht dna serusaem yrtemmysa noitamrofni eht stneserp B lenaP dna .elbairav hcae rof elitnecrep ht59 dna ht5 sa llew sa ,noitaived scitsiretcarahc kcots .A lenaP dts xam 59ltcp naem naidem 5ltcp nim desu doirep stinu noitpircsed elbairav 54.0 73.11 50.11 30.01 09.9 74.9 13.9 80’ luJ fo dne 01gol pactkm gol pactkml 54.0 21.21 25.11 95.01 45.01 59.9 08.9 80’ luJ–70’ guA 01gol emulov rallod gol lovdl 54.0 02.01 98.9 20.9 49.8 14.8 51.8 80’ luJ–70’ guA 01gol emulov erahs gol lovl 30.06 57.374 00.18 64.54 93.13 44.11 04.3 80’ luJ fo dne $ ecirp erahs ecirp 82.0 17.1 93.1 59.0 29.0 05.0 82.0 80’ luJ–70’ guA ateb kcots ateb 77.0 40.5 02.4 66.2 74.2 66.1 15.1 80’ luJ–70’ guA % ytilitalov yliad ytilitalov 14.6 03.94 12.7 60.4 60.2 88.1 68.1 51/01–51/8 stnec daerps detouq rallod daerpsqd 71.8 26.47 11.32 11.11 93.9 96.4 97.3 51/01–51/8 .p.b daerps lanoitroporp daerpsq 41.2 15.91 95.6 14.3 38.2 37.1 73.1 51/01–51/8 .p.b daerps flah evitceffe daerpse seitivitca TFH .B lenaP dts xam 59ltcp naem naidem 5ltcp nim desu doirep noitpircsed elbairav 1260.0 2673.0 1623.0 6891.0 0091.0 6511.0 6801.0 51/01–51/8 tnemhsinelper ytidiuqil RL 5610.0 2501.0 0740.0 9910.0 7510.0 2200.0 1000.0 51/01–51/8 egnahc FTE ot noitcaer CER 5120.0 7131.0 1011.0 1970.0 2970.0 5340.0 6110.0 51/01–51/8 gnidart nevird-FTE fo ytilibaborp TEP 29= N 32

ledom sekwaH eht morf setamitsE :2 elbaT ledom sekwaH eht morf setamitse eht tneserp selbat gniwollof owt ehT M . (cid:111) )m i t−t( m )2 r (b−e m )2 r (b m )2 r (a+)m i t−t( m )1 r (b−e m )1 r (b m )1 r (a (cid:110) (cid:88) (cid:88) +)t( r µ=)t( r λ t<mt1=m i setamitse eht stneserp b2 elbaT dna ,)1(b dna )1(a rof setamitse eht stneserp a2 elbaT .tneve gnireggirt eht si m dna ,tneve detceffa eht si r mr mr eht sa llew sa , a evitisop yllacitsitats a htiw snoitavresbo fo noitcarf eht dna a fo eulav naem eht ,riap )m ,r( hcae roF .)2(b dna )2(a rof mr mr mr mr .4.3 noitceS ni denialpxe era erudecorp noitamitse eht no sliateD .detneserp era ,2gol ,efil-flah naidem mrb 33

tceffe retsaF :a2 elbaT tnevedetceffA )2SM(8 )2BM(7 )AC(6 )BC(5 )AA(4 )BA(3 )DM(2 )BM(1 tnevegnireggirT 90700.0 69010.0 92200.0 49200.0 56100.0 23100.0 00548.493 96550.0 efilflah )yuBtekraM(BM 1 95230.0 82820.0 71050.0 71400.0 62991.0 46812.0 44840.0 00070.0 anaem %5.55 %0.39 %4.28 %9.62 %8.99 %1.99 %8.72 %0.18 69.1>tfo% 44900.0 36600.0 52300.0 42200.0 92100.0 66100.0 42650.0 00575.653 efilflah )lleStekraM(SM 2 83520.0 80330.0 33400.0 03050.0 08322.0 98791.0 20070.0 11540.0 anaem %0.98 %4.45 %0.82 %7.18 %9.89 %0.001 %5.08 %4.52 69.1>tfo% 05100.0 96400.0 38300.0 42600.0 55100.0 08400.0 52100.0 81500.0 efilflah )diBddA(BA 3 25530.0 40210.0 24820.0 51050.0 32300.0 73210.0 84950.0 78310.0 anaem %6.39 %5.48 %6.08 %9.88 %7.43 %2.65 %9.99 %9.97 69.1>tfo% 38500.0 54100.0 54600.0 68300.0 09400.0 39200.0 02500.0 13100.0 efilflah )ksAddA(AA 4 00510.0 49430.0 02050.0 68820.0 94210.0 23200.0 15310.0 89850.0 anaem %0.29 %1.59 %1.98 %3.18 %5.65 %6.03 %0.08 %9.99 69.1>tfo% 12200.0 97500.0 75300.0 61300.0 efilflah )diBlecnaC(BC 5 63500.0 89580.0 83280.0 98262.0 anaem %3.93 %4.87 %8.29 %7.79 69.1>tfo% 68500.0 43200.0 23300.0 76300.0 efilflah )ksAlecnaC(AC 6 85780.0 50500.0 66362.0 36480.0 anaem %0.97 %7.63 %8.79 %4.39 69.1>tfo% 93929.0 74600.0 68100.0 06200.0 17200.0 efilflah 2BM 7 81700.0 13461.0 29040.0 86530.0 79821.0 anaem %0.71 %9.99 %4.79 %1.79 %0.001 69.1>tfo% 39500.0 79168.0 38100.0 86200.0 56200.0 efilflah 2SM 8 91461.0 47800.0 59930.0 10530.0 13821.0 anaem %9.99 %0.32 %0.79 %7.69 %0.001 69.1>tfo% 00041.531 70830.0 14070.0 00003.6531 00034.894 79360.0 00000.32801 88330.0 efilflah ↑FTE 9 79500.0 16920.0 92110.0 10600.0 51310.0 03220.0 81100.0 45020.0 anaem %6.4 %6.09 %5.05 %4.4 %2.4 %0.16 %3.0 %4.38 69.1>tfo% 44630.0 00477.78 00000.459 27160.0 52960.0 00051.983 48330.0 00035.638 efilflah ↓FTE 01 49720.0 53600.0 71500.0 66010.0 03120.0 64110.0 23910.0 90500.0 anaem %1.09 %0.6 %6.3 %6.05 %1.75 %2.3 %5.38 %5.4 69.1>tfo% 00099.331 91850.0 51190.0 00015.119 00000.533 28691.0 00003.3905 42260.0 efilflah ↑2FTE 11 37600.0 82110.0 02500.0 53900.0 08210.0 47110.0 90200.0 09700.0 anaem %7.4 %5.15 %7.13 %7.9 %2.6 %9.73 %7.1 %6.54 69.1>tfo% 25370.0 00008.0201 00019.675 73970.0 36711.0 00034.724 24760.0 00004.0347 efilflah ↓2FTE 21 61310.0 47010.0 95900.0 26400.0 94900.0 76800.0 26700.0 22100.0 anaem %8.35 %0.7 %8.9 %2.72 %4.63 %1.3 %3.44 %7.0 69.1>tfo% 34

tceffe rewolS :b2 elbaT tnevedetceffA )2SM(8 )2BM(7 )AC(6 )BC(5 )AA(4 )BA(3 )DM(2 )BM(1 tnevegnireggirT 15742.0 00242.31 75840.0 75380.0 63160.0 05206.01 efilflah )yuBtekraM(BM 1 36040.0 08280.0 61030.0 74460.0 69182.0 26441.0 anaem %3.46 %6.27 %8.07 %4.69 %0.001 %2.89 69.1>tfo% 02646.7 33902.0 15740.0 32160.0 82380.0 05935.01 efilflah )lleStekraM(SM 2 21850.0 23650.0 10030.0 34782.0 95360.0 63341.0 anaem %0.16 %7.57 %3.07 %0.001 %6.59 %2.89 69.1>tfo% 77701.0 49921.0 14353.0 33886.0 05212.12 00380.0 98451.0 efilflah )diBddA(BA 3 53330.0 99510.0 46620.0 65101.0 40560.0 28720.0 94110.0 anaem %0.67 %0.17 %6.27 %4.88 %6.96 %2.19 %1.95 69.1>tfo% 57876.3 00525.5401 57286.0 74053.0 05848.32 00731.0 22390.0 efilflah )ksAddA(AA 4 02600.0 60230.0 81990.0 07620.0 67260.0 84110.0 97720.0 anaem %6.8 %6.9 %5.88 %5.17 %7.86 %9.95 %4.09 69.1>tfo% 41880.0 36651.0 04053.0 efilflah )diBlecnaC(BC 5 53530.0 90051.0 67351.0 anaem %9.15 %9.69 %9.09 69.1>tfo% 01680.0 92673.0 87751.0 efilflah )ksAlecnaC(AC 6 48530.0 36751.0 57741.0 anaem %7.15 %4.09 %7.69 69.1>tfo% 00049.605 89295.0 89670.0 56787.0 80111.0 efilflah 2BM 7 47750.0 52361.0 33620.0 34060.0 89380.0 anaem %5.32 %9.99 %5.98 %4.49 %8.99 69.1>tfo% 28795.0 00005.564 00000.0 96180.0 38728.0 01801.0 efilflah 2SM 8 73761.0 23740.0 00000.0 68520.0 44160.0 02180.0 anaem %9.99 %1.91 %9.98 %1.49 %9.99 69.1>tfo% 00000.0 25223.0 85475.0 00004.8624 00000.7542 45688.0 60744.0 efilflah ↑FTE 9 00000.0 98430.0 60010.0 34200.0 83910.0 43230.0 81220.0 anaem %3.47 %0.33 %5.0 %0.5 %4.85 %4.56 69.1>tfo% 31043.0 00000.0 00000.7014 41775.0 96568.0 00025.305 84254.0 efilflah ↓FTE 01 42830.0 00000.0 14100.0 31110.0 53530.0 41800.0 55220.0 anaem %2.97 %4.0 %6.53 %7.46 %1.1 %2.76 69.1>tfo% 00000.0 20394.0 07924.1 00000.8902 00074.667 00925.1 00000.30154 87505.0 efilflah ↑2FTE 11 00000.0 78410.0 60900.0 28400.0 44800.0 37320.0 45000.0 76800.0 anaem %4.82 %9.22 %2.1 %4.2 %7.04 %0.0 %0.71 69.1>tfo% 54535.0 00000.0 00003.4262 66107.0 00015.1 00006.776 34056.0 00000.000941 efilflah ↓2FTE 21 37210.0 00000.0 42500.0 15700.0 80220.0 82010.0 77700.0 14000.0 anaem %3.22 %3.2 %6.71 %9.93 %2.3 %5.41 %0.0 69.1>tfo% 35

Table 3: ETF-induced orders and adverse selection Foreachstockianddayτ,theaverageeffectivespreadandpermanentpriceimpact(usingdurationsof50ms, 5s, and 5min) are calculated separately for ETF-driven and non-ETF-driven market orders that change the best bid or ask. The table below presents the results from the regression y =γ1(ETF-driven)+A +B +(cid:15) , i,l,τ i τ i,τ wherey iseithertheeffectivehalfspread(ESpread)orthepermanentpriceimpact(PI),and1(ETF-driven) isadummyvariablethatequals1ifthevalueisforETF-drivenorders. Standarderrorsaredoubleclustered by stock and day. (1) (2) (3) (4) ESpread PI (50ms) PI (5s) PI (5m) mean(A +B ) 3.409 3.732 4.980 4.818 i τ 1(ETF-driven) 0.040+ 0.108∗∗ 1.178∗∗ 0.539∗∗ (1.91) (4.91) (15.90) (4.13) Adj. R sq 0.872 0.798 0.781 0.429 N 92 stocks × 43 days × 2=7912 t statisticsinparentheses,usesstderrorsdoubleclusteredbystockandday + p<0.10,∗ p<0.05,∗∗ p<0.01 36

Table 4: Cross-sectional relationship between liquidity replenishment and information asymmetry This table presents the results from the cross-sectional regression LR =γ +γ PET +γ lmktcap +γ beta +γ volatility +γ price +γ spread +(cid:15) . i 0 1 i 2 i 3 i 4 i 5 i 6 i i LR is the average LR value for stock i. All explanatory variables are standardized to have a mean of 0 and i a standard deviation of 1. t values are calculated using heteroskedasticity-consistent standard errors. (1) (2) (3) (4) (5) LR LR LR LR LR constant 0.199∗∗ 0.199∗∗ 0.199∗∗ 0.199∗∗ 0.199∗∗ (33.79) (48.35) (49.39) (41.91) (46.34) PET -0.027∗∗ -0.028∗∗ -0.021∗∗ (-4.38) (-5.94) (-3.49) lmktcap 0.034∗∗ 0.036∗∗ 0.036∗∗ 0.040∗∗ (6.69) (7.21) (6.96) (7.98) beta -0.003 -0.004 -0.013+ -0.011 (-0.37) (-0.54) (-1.67) (-1.50) volatility 0.019∗ 0.017∗ 0.041∗∗ 0.027∗∗ (2.62) (2.58) (5.22) (4.32) price -0.034∗∗ -0.030∗∗ -0.028∗∗ -0.024∗∗ (-5.98) (-5.37) (-4.78) (-4.60) spread 0.013∗ 0.024∗∗ (2.17) (5.74) Adj. R sq 0.177 0.598 0.615 0.465 0.563 N 92 92 92 92 92 t statisticsinparentheses + p<0.10,∗ p<0.05,∗∗ p<0.01 37

Table 5: Natural experiment using index inclusion PanelApresentstheresultsfromthefirststageregressiontestingwhetherinformationasymmetryincreases when a stock is newly included in the index. y is REC or PET. y =A +B +γ1(i newly included)1(τ after inclusion)+(cid:15) . i,τ i τ i,τ Using 1(i newly included)1(τ after inclusion) as an excluded instrument, I run the 2SLS regression LR =A +B +γPET +(cid:15) . i,τ i τ i,τ i,τ Panel B presents the results. t values are calculated from standard errors clustered by stock. Panel A: First stage regressions level proportional standardized (1) (2) (3) (4) (5) (6) REC PET REC PET REC PET mean(A +B ) 0.015 0.048 -0.032 -0.014 -0.044 -0.034 i τ 1(i new)1(τ after) 0.012∗∗ 0.021∗∗ 0.580∗∗ 0.247∗∗ 0.797∗∗ 0.614∗∗ (7.57) (6.12) (6.47) (3.97) (9.20) (4.41) stock f.e. Yes Yes No No No No day f.e. Yes Yes Yes Yes Yes Yes Adj. R sq 0.839 0.764 0.053 0.200 0.082 0.400 N 990 990 990 990 990 990 Panel B: Second stage regressions (1) (2) (3) level proportional standardized mean(A +B ) 0.266 -0.000 -0.000 i τ PET -1.261∗ -0.225∗ -0.227+ (-1.97) (-1.98) (-1.71) stock f.e. Yes No No day f.e. Yes Yes Yes N 990 990 990 t statisticsinparentheses + p<0.10,∗ p<0.05,∗∗ p<0.01 38

Table 6: Market volatility and liquidity replenishment Thetablebelowpresentshowdailyaverageinformationasymmetryandliquidityreplenishmentchangewith market volatility. I run y =γ +γ ETFvol +(cid:15) , τ 0 1 τ τ wherey isthecross-sectionalaverageofPETorLR.ETF vol isthestandarddeviationof1-minutereturns τ of QQQ on day τ. t values are calculated using Newey-West standard errors with 5 lags. level standardized (1) (2) (3) (4) PET LR PET LR constant 0.079∗∗ 0.199∗∗ 0.000 -0.000 (66.02) (45.72) (0.00) (-0.00) ETF vol 0.015∗∗ -0.019∗∗ 0.860∗∗ -0.667∗∗ (13.33) (-7.82) (13.33) (-7.82) Adj. R sq 0.733 0.432 0.733 0.432 N 43 43 43 43 t statisticsinparentheses + p<0.10,∗ p<0.05,∗∗ p<0.01 39

Table 7: Information asymmetry sensitivity and market volatility Below presents the results from the regression LR =A +(δ +δ γ˜ )ETF vol +ξ . i,τ i 0 1 1i τ i,τ γ˜ are standardized value of γ , the sensitivity of stock i’s PET to market volatility. γ is estimated by 1i 1i 1i the following stock-by-stock regression PET =γ +γ ETFvol +(cid:15) i,τ 0i 1i τ i,τ t values are calculated from standard errors that are clustered by stock and day. (1) (2) (3) level proportional standardized mean(A ) 0.159 0.000 0.000 i ETF vol -0.019∗∗ -0.085∗∗ -0.238∗∗ (-8.27) (-7.23) (-7.15) γ˜ ×ETF vol -0.014∗∗ -0.048∗∗ -0.127∗∗ 1 (-6.51) (-4.28) (-5.16) stock f.e. Yes No No Adj. R sq 0.464 0.065 0.072 N 3956 3956 3956 t statisticsinparentheses + p<0.10,∗ p<0.05,∗∗ p<0.01 40

Table 8: Stock characteristics and sensitivity to market volatility I run the following regression y =γ +(γ +γ x )ETFvol +(cid:15) , i,τ 0 1 2 i τ i,τ where y is PET or LR, x are size and REC . REC is calculated using data from the first five days of i 0 0 the sample, and the regression is estimated over the rest of the sample period. t values are calculated using standard errors clustered by stock and day. standardized proportional (1) (2) (3) (4) (5) (6) LR PET REC LR PET REC ETF vol -0.244∗∗ 0.627∗∗ -0.124∗ -0.089∗∗ 0.206∗∗ -0.014 (-6.79) (6.78) (-2.26) (-6.62) (7.20) (-0.59) log mktcap × ETF vol -0.082∗∗ 0.037∗∗ 0.245∗∗ -0.017∗ 0.022∗∗ 0.152∗∗ (-3.82) (20.38) (6.11) (-2.25) (2.84) (6.39) REC × ETF vol 0.100∗∗ -0.053∗ -0.269∗∗ 0.031∗∗ -0.060∗∗ -0.134∗∗ 0 (3.26) (-2.52) (-6.19) (2.81) (-6.00) (-6.38) Adj. R sq 0.072 0.396 0.121 0.060 0.388 0.142 N 3496 3496 3496 3496 3496 3496 t statisticsinparentheses + p<0.10,∗ p<0.05,∗∗ p<0.01 41

Table 9: Cross-sectional determinants of information asymmetry I study the cross-sectional determinants of REC and PET using the following cross-sectional regressions REC =γ +γ lmktcap +γ beta +γ volatility +γ price +γ spread +(cid:15) i 0 1 i 2 i 3 i 4 i 5 i i PET =δ +δ lmktcap +δ beta +δ volatility +δ price +δ spread +δ REC +ξ . i 0 1 i 2 i 3 i 4 i 5 i 6 i i REC andPET aretheaverageREC andPET valuesforstocki. Allexplanatoryvariablesarestandardized i i tohaveameanof0andastandarddeviationof1. t valuesarecalculatedusingheteroskedasticity-consistent standard errors. (1) (2) (3) (4) (5) (6) (7) REC REC REC PET PET PET PET constant 0.0199∗∗ 0.0199∗∗ 0.0199∗∗ 0.0791∗∗ 0.0791∗∗ 0.0791∗∗ 0.0791∗∗ (15.01) (17.00) (22.75) (39.47) (50.08) (52.43) (51.13) lmktcap 0.0058+ 0.0026 0.0010 -0.0022 -0.0032+ -0.0036∗ (1.98) (1.13) (0.51) (-1.24) (-1.89) (-2.28) beta 0.0073∗∗ 0.0065∗∗ 0.0041 0.0053+ 0.0071∗∗ (2.63) (2.76) (1.43) (1.96) (2.84) volatility -0.0077∗∗ -0.0021 -0.0122∗∗ -0.0103∗∗ -0.0109∗∗ (-4.33) (-1.41) (-3.56) (-3.66) (-4.07) price 0.0068∗∗ 0.0054∗∗ -0.0079∗∗ -0.0070∗∗ -0.0052∗∗ (3.24) (3.37) (-4.71) (-3.90) (-3.11) spread -0.0099∗∗ -0.0078∗ -0.0118∗∗ (-6.01) (-2.59) (-5.58) REC 0.0098∗∗ 0.0106∗∗ 0.0057∗ (3.96) (5.90) (2.32) Adj. R sq 0.113 0.309 0.542 0.199 0.502 0.546 0.522 N 92 92 92 92 92 92 92 t statisticsinparentheses + p<0.10,∗ p<0.05,∗∗ p<0.01 42

Figure 5: Information asymmetry and market volatility Daily average PET and ETFvol are plotted over the sample period. ETF volatility are calculated using 1-minute ETF returns, and are rescaled to daily values. 43

Figure 6: Liquidity replenishment and market volatility Daily average LR and ETFvol are plotted over the sample period. ETF volatility are calculated using 1-minute ETF returns, and are rescaled to daily values. 44

Appendix A Existence and Uniqueness The Hawkes self-exciting model is determined by an intensity function of the form M λ r (t)=µ r (t)+ (cid:88) (cid:88) (a( r 1 m )b( r 1 m )e−b( r 1 m )(t−ti m )+a( r 2 m )b( r 2 m )e−b( r 2 m )(t−ti m )). (26) m=1ti <t m As Embrechts et al. (2011) note, this leaves the existence and the uniqueness questions open. Theorem 3.1 of Bowsher (2007) establishes that if, for a given parameter vector θ =(θ ,··· ,θ ) where 1 M θ ={a ,··· ,a ,b ,··· ,b ,µ (t)} r r1 rM r1 rM r M (cid:90) T (cid:88) λ (t;θ)dt<∞, (27) r r=1 0 thereexistsauniquepointprocessforthelimitorderbookarrivalprocesswiththeintensityλ (t;θ)foreach r order type r. Intuitively, the sufficient condition (27) ensures that events do not happen infinitely many times over a bounded interval. For proof, see Bowsher (2007) and papers cited within. Embrechts et al. (2011) also show the existence anduniquenessofDGPifthematrix(a ) hasaspectralradiussmallerthan1. Theyonlylook rm r,m=1,2,···,M at cases with constant µ (t), but we can easily see that this is a sufficient condition for (27) to hold. r B Maximum Likelihood Estimation Ogata (1978) proves that the maximum likelihood estimator is consistent and asymptotically normal for a univariatestationaryHawkesprocesswithaconstantbaseintensityofµ(t)=µ. However,theHawkesprocess Iuseisnonstationarybecauseµ (t)isnotaconstantasintheoriginalHawkesmodel(Hawkes(1971))buta r time-deterministicfunction. Bowsher(2007)extendstheresultstononstationaryandmultivariateprocesses using simulations. 45

Plugging appropriate expressions into (7), the log likelihood function of event r is (cid:90) (cid:90) l (θ )= (1−λ (s|θ )ds+ logλ (s|θ )dN (s) (28) r r r r r r r (0,T] (0,T] (cid:90) M 2 (cid:88) (cid:88)(cid:88) =T − µ (s)ds− a(j)(1−exp(−b(j)(T −tm))) r rm rm i (0,T] m=1 tm j=1 i   M 2 (cid:88) (cid:88) (cid:88) + logµ r (tr l )+ a( r j m )b( r j m )R r (j m )(tr l ), (29) tr m=1j=1 l where tm is the ith event of type m and i (cid:88) R(j)(tr)= exp(−b(j)(tr−tm)). (30) rm l rm l k tm<tr k l If there are multiple days, as noted in Section 3.4, each day is assumed to be an independent realization, thus the log likelihood will be the average of all days. (30) is calculated recursively. I use a modified Newton-Raphson method with an analytical gradient and Hessian to maximize the likelihood function. Ifa(j) iszeroorveryclosetozero,orifb(j) issmall(inotherwords,ifthehalflifeistoolong—typically rm rm longer than the trading day), then it complicates the estimation as the likelihood function becomes flat in that dimension. It generally is not problematic for the point estimates, but it confounds the asymptotics. Therefore, after the preliminary estimation, I take out the terms with a very small median a(j) or b(j). rm rm Empty cells in Table 2a and 2b are terms that are taken out. C Simultaneous events Existence, uniqueness, and maximum likelihood estimation all require the point process (PP) to be a simple PP, i.e., there is zero probability that two events, whether they are of the same type or not, happen at the exact same time. This is a concern when PPs are used to study limit order book data as time stamps are often not fine enough and many events have the same time stamp. For example, Large (2007) has 1-second time stamps, and 40% of the observations have another event falling in the same second. Some papers thin the data so that no two events have the same time stamp; some add random variation in time to the events that happen in the same time stamp; and others keep it as is but with no concurrent excitations. Fortunately, in the NASDAQ TotalView-ITCH data, observations are not only time stamped in milliseconds but are also sequenced. They are in the same sequence that each order was processed and also what 46

investors (or machines) that subscribe and use the feed saw. Thus, even for different tickers I can perfectly order which observation came first. For observations that fall in the same millisecond, I equally space them out over the millisecond. This might lead to slight overestimation of half-lives if there is clustering within the millisecond, but the effect will be minimal. One issue, however, is true concurrent events. For example, if a marketable limit buy order has an order size larger than the current best ask depth but a price lower than the second best ask, it will deplete the current best ask and the remaining order will be added to the book as a buy limit order. This type of event is quite rare (less than 3% of market orders that change the best bid/ask, less than 0.5% of total events), so results are not sensitive to how this issue is dealt with. Similar to Large (2007), I keep the data as is. D Did the limit order book dynamics change? The Hawkes model estimation also allows us to reexamine the limit order book dynamics using recent data andtocompareittotheresultsfromtheexistingliterature. Marketmicrostructurehaschangedsignificantly, yetitisnotclearhowthelimitorderbookdynamicshavechanged. Aretoday’smarketsjustthe“oldmarkets onsteroids,”wherethedynamicsaremostlythesamebutjustfaster? Oraretherefundamentaldifferences? Tostudyconditionaleventfrequencies,therearetwodifferentsetsofstatevariablesthatwecancondition on. First are events such as changes in the limit order book. Second are those that are generally thought of as state variables, such as bid-ask spread or volatility. The methodology in this paper naturally relates to the first approach, but given that most state variables in the second approach change as a result of limit order book events, the two approaches are related. For example, bid-ask spreads change if and only if there are changes in the best bid or ask, which are considered “events” in the first framework. Papers along the first dimension (Biais et al. (1995) and Ellul et al. (2007)) condition on the last order type and examine the conditional probability of the next order being a particular type. They find that there is a diagonal effect in general — market buy tends to be followed by a market buy, limit buy that is within the best quotes tends to be followed by another such order, etc. In addition, Ellul et al. (2007) find a liquidity exhaustion-replenishment cycle in the longer term (from couple seconds to couple minutes). Both papersattributethediagonaleffectinmarketorderstofollow-onstrategiesratherthanorder-splitting,since another market order tends to arrive before liquidity is replenished. In contrast, I find that, on average, after a market order exhausting liquidity, liquidity replenishment happens faster than another market order coming in. For example, a market buy that increases the best ask is followed by a limit sell order that 47

decreases the ask with a 20% probability and a 1.7ms half-life; another market buy that increases the ask with a 7% probability and a 55.7ms half-life; and a market buy that doesn’t move the ask with a 2.8% probability and a 11ms half-life. This difference is driven by faster liquidity provision — perhaps liquidity provisionbecamefastenoughthatitisnowoptimalforfollow-onstrategiestowaitslightlybeforesubmitting orders, ororder-splittingmighthavebecomemoreprevalentrelativetofollow-onstrategiesasdepthbecame lower. The diagonal effect in limit order submission that changes the best quotes found in Biais et al. (1995) and Ellul et al. (2007) is not as strong in the recent data. This is somewhat surprising if the diagonal effect in limit order addition is due to multiple limit order providers undercutting each other for price priority (as inYeo(2006)),aswewouldthinkthatcomputersmadeundercuttingmucheasier. Onepossibleexplanation is that bid-ask spreads are smaller today, making undercutting in succession more difficult, as a limit order that changes the best price by definition decreases the spread. On the other hand, the diagonal effect in limit order cancellation that changes the best quotes is quite strong. This has not been directly studied in earlier research, as all cancellations are generally grouped into one category regardless of whether they change the best quote. Another distinct result that emerges is a pattern of order submissions and cancellations, especially on thesameside, excitingeachother. Thisisnotcompletelynew, asmanypapers(Biaisetal.(1995), Ranaldo (2004), Griffithsetal.(2000))showthatlargebid-askspreadsattractlimitorders. Whatisnew, however,is that this effect is very fast and that cancellation (submission) is more likely to excite same-side submission (cancellation) than the opposite-side one. For example, if there a bid improvement, it is quite likely (5%) to be cancelled quickly (a 6.2ms half life), less so for an ask to be cancelled (2.7% with a 3.8ms half life); and if there is a cancellation on the bid side, it is likely (26%) to be re-filled quickly (a 3.2ms half life) on the bid side compared to the ask side (8.2% with a 3.6ms half life). This phenomenon is likely due to the use of fleeting orders by HFTs. Afteramarketbuythatincreasestheask,thebestbidtendstoincrease. Biaisetal.(1995)alsodocument this“comovement”effect(astermedbyRo¸su(2009)),andthiseffectisquitestrongintheirresults,although slightly less so than the diagonal effect (a market buy is more likely to be followed by a market buy than by an increase in the bid). In my results, if the slower reaction term is included (the half-life of the slower term of this comovement effect is still quite fast at 61ms), a market buy is followed by an increase in bid in 50% of the cases; this is the strongest reaction to a market buy. Biais et al. (1995) argue that this comovement effect is due to the information content of the trade, similar to the market maker updating in 48

Glosten and Milgrom (1985). Parlour (1998) and Ro¸su (2009), however, argue that this can be explained without information asymmetry. Although cancellations are not modeled, since the effect arises from the number of outstanding bid and ask quotes in both papers, a cancellation of a limit sell order that increases the ask (CA) should have the same effect as a market buy that increases the ask (MB). However, in my results, a bid increases much more in reaction to action MB than to action CA; the comovement effect is thus more naturally interpreted as a consequence of the information content of market orders, and market orders must contain more information than cancellations. Finally,liquiditytakersalsoseemtomonitorthemarketcloselyandsubmitmarketbuyordersrightafter ask decreases and market sells after bid increases. This is consistent with Hendershott and Riordan (2012), who finds that algorithmic traders consume liquidity when it is cheap. What emerges is a fast, complicated, and dynamic process, in which both positive andnegative feedback forces are present, and reactions to market orders are strongest. Both the liquidity providers and takers seem to monitor the market actively and react quickly. Most dynamic models of limit order book assume a constant arrival rate of buyers and sellers. Consequently, many of their predictions on dynamics are implicitly derived from the arithmetic identity in which an increase in market buys must mean that limit buy submissions decrease. Although this paper does not offer a formal statistical test, my results suggest that this assumption does not hold for HFTs. E Incorrect Half-life InSection3.4,Ifixthehalf-livestobethesameacrossallstocksanddaysandletthesizeoftheeffect(a ) rm vary. Thesizeandthehalf-lifeoftheeffectaretwoseparatedimensions,andIfixthehalf-lifetobeconstant, assumingthatthesizeoftheeffectistheonlydimensionthatchanges. Iftheprobabilityofeffecthappening stays the same but the effect becomes slower, will this make the interpretations of the estimations from the incorrectly specified model invalid? Intuitively, if the probability of the effect happening remains constant at 10% and the half-life increases from 1s to 3s, but if I assume a 1s constant half-life, the probability of the effect happening would seem smaller because I am looking at a shorter period of time than I should. Thus, a would be estimated to be smaller than 10%. Results from the paper could still be interpreted similarly rm asifIusedthetruemodeltoestimate,sincealongerhalf-lifeandalowerprobabilityoftheeffecthappening both point to less involvement by the HFTs. I test the above idea using simulations. I simulate a simple case with two actions (M = 2) and J = 1. 49

For the first set of simulations, I use the parameters       0.1 0 log(2) 0 0.1 a= , b= 2 , λ = , (31)     0   0.1 0.1 log(2) log(2) 0.1 4 2 where element a(r,m) refers to a and b(r,m) refers to b . The half-lives of both self-excitations are 2, rm rm the half-life of event 1 precipitating event 2 is 4, and event 2 does not precipitate event 1. b is written 1,2 as 0 for convenience, but the value does not matter as a = 0. Now if we have incorrectly assumed that 1,2 b = log(2) (i.e., that the half-life of event 1 precipitating event 2 is 2), thus 2,1 2   log(2) 0 ˆb= 2 , (32)   log(2) log(2) 2 2 and estimated a using ˆb, this means that we have assumed a shorter half-life than the true value. I assume a length of 5,000 for each simulation (T =5,000) and simulate 10,000 times using the above parameters. I also do the same using a different set of parameters that are closer to the estimates from the data.       0.1 0.05 log(2) log(2)×10 0.1 a= , b= 4 , λ =  (33)     0   0.2 0.01 log(2)×2 log(2) 0.1 and   log(2) log(2)×10 ˆb= 4 . (34)   log(2)×10 log(2) In this case, the half-life of the cross-excitation of event 1 on 2 is mistaken as 0.1 instead of 0.5, the true value. Simulations are done using a modified version of Ogata (1981)’s thinning algorithm. A modified version of the algorithm for a multivariate Hawkes process is outlined and proved in Bowsher (2007). Figure 7 presents the distribution of estimated a for both parameters. Comparing (a) with (b) and 2,1 (c) with (d), we can observe that under the incorrect assumption of shorter half-life, a’s are estimated to be smaller than the true values. 50

Figure 7: Histogram of a estimates from simulated data 2,1 Panel (a) and (b) are estimates from simulated data using parameters in (31), (c) and (d) are using (33). Panel (a) and (c) are estimated using incorrect b values, (32) and (34), respectively. Panel (b) and (d) are estimates using correct b values for comparison. True a is 0.1 for (a) and (b), and 0.2 for (c) and (d). 2,1 51

F Potential issues with the Hawkes model and the REC measure OneissuewiththeHawkesmodelisthatitmightseemasifeventmexciteseventriftheirintensityprocesses are correlated even without one really causing the other. If the intensity of event type 1 and 2 are λ (t)=d ψ (t)+ψ (t) (35) 1 1 c 1 λ (t)=d ψ (t)+ψ (t), (36) 2 2 c 2 where ψ (t) is a common latent process, a and a appear positive in the Hawkes model. We thus would c 1,2 2,1 incorrectly interpret the estimates from the Hawkes model as event 1 and 2 mutually exciting each other. Here I divide all possible scenarios by whether the two events are of the same direction and by how fast the ψ (t) mean reverts, and address them one by one. c If event 1 and 2 are both events in the stock, and if the two events are in the opposite direction (one increases price, another decreases), then it seems unlikely that they would be governed by a common latent process unless they directly affect each other. For example, event MB (market buy) tends to be followed by AA(addask),andviceversa,butitwouldnotmakesensetothinkthatwhentraderstendtosubmitmarket buys, traders also tend to improve the ask. Rather, it is because someone submitted a market buy that increasedtheaskthatsomeoneelsesubmitsalimitsellorderthatbringstheaskbackdown, oritisbecause someone submitted an ask that is below the current best ask that someone else seized the opportunity and transacted against the new, better ask quote. If the events are in the same direction, it is plausible that there would be a common latent component. If, in the estimation results, event 1 excites event 2, but not the other way around, it is unlikely that the results are due to a common component. For example, event MB (market buy) triggers event CA (cancel ask), but not the other way around. However, if it is merely the case that one effect is larger than the other, it may be because d and d from (35) and (36) are different. 1 2 In the case where event 1 and 2 appears to mutually excite each other, we can divide the cases by the speed of mean reversion in the common component ψ (t). Recall the Hawkes model specification c M λ r (t)=µ r (t)+ (cid:88) (cid:88) (a( r 1 m )b( r 1 m )e−b( r 1 m )(t−tm i )+a( r 2 m )b( r 2 m )e−b( r 2 m )(t−tm i )), (37) m=1tm<t i where the intensity has three components: slowly changing base intensity µ (t) (in hours), very fast comr ponent (in milliseconds), and slower component (usually in seconds or minutes). Therefore, if the mean 52

reversion rate of ψ (t) is very slow (in hours) or is not mean reverting at all, it will be captured in the base c intensity. If it is faster, it will appear in one of the two interaction terms. Intuitively, if the model specified by (35) and (36) is correct but we estimate the Hawkes model instead, the estimated half-life would be of a similar magnitude to that implied by the speed of mean reversion of ψ (t). (This point will be shown later c via simulations.) If the mean reversion rate is thus moderate (minutes or seconds), it will appear in the secondinteractionterma( r 2 m )b( r 2 m )e−b( r 2 m )(t−tm i ). Onlyifitisinmillisecondswillitappearinthefirstinteraction term a( r 1 m )b( r 1 m )e−b( r 1 m )(t−tm i ), which is the only one that would be relevant since I mostly focus on a( r 1 m ) and b( r 1 m ). Whenwoulditbethecasethatψ (t)meanrevertsveryfast,inmilliseconds? Sincevirtuallyallstoriesof c real news or sentiment would predict a much slower mean reversion, it must also be hard information, such as futures index prices, that is driving ψ (t). For example, it may be the case that both event MB (market c buy) and ETF↑ (increase in ETF ask) are driven by futures price increase, but neither of the two drives the other nor is one faster than the other. In this case, it will still seem as if ETF↑ excites MB, and vice versa, according to the Hawkes model. However, I can redefine the futures price to be the hard information, and the size of the effect estimated would be an underestimate of the true value. In this sense, I can relax the assumption made in Section 4.1 that the information from futures must flow first into ETF and then into stocks. Therefore, the intuition for the REC measure stays the same since the estimates from the Hawkes model measure the reaction to some other hard information, most likely futures price. To show that the common component ψ (t) has to be extremely quickly mean-reverting in order for c the estimated half-life of the Hawkes model to be in milliseconds, I simulate a simple case of the scenario described above by (35) and (36) with two processes having the same intensity λ (t)=λ (t)=ψ (t), (38) 1 2 c where ψ (t) has Poisson-arriving jumps that decay exponentially at a rate b c (cid:88) ψ (t)=µ+ abe−b(t−τi). (39) c τi<t τ istheithjumptime,andτ isaPoissonprocesswithintensityπ. Wecaninterpretthissettingasanevent i happening at each jump time τ, in which the events are observable to the traders but not to the researcher. Traders react to these events with action 1 or 2, and we only observe these actions in the data. 53

I simulate using parameters a=0.5, µ=0.1, π =0.2 (40) b=log2 or log2/10. (41) b of log2 implies that the half-life of each jump is 1, and log2/10 implies a half-life of 10. I estimate the Hawkes model on each simulation, and present the half-life of cross-excitation (event 1 exciting event 2) in Table 10 based on 1,000 simulations. Results show that the slower the mean-reversion of the latent common component,thelongertheestimatedhalf-lifeofthecross-excitation,andthattheyareofthesamemagnitude. Thisresult,ofcourse,mayalsodependontheparametervaluesorthefunctionalformofthelatentcommon process ψ (t). However, it is generally difficult to get a very short half-life with other functional forms of c ψ (t) with a reasonable mean intensity. c Table 10: Estimated half-life of cross-excitation in the Hawkes model 10% 25% 50% 75% 90% b=log2 0.41 0.62 0.91 1.34 2.59 b=log2/10 1.96 4.56 10.24 29.12 157.15 G Potential issues with the PET measure The PET measure faces a similar problem as above: if, for example, the intensities of event MB (market buy) and ETF↑ (increase in ETF ask) are λ (t)=d ψ (t)+ψ (t) (42) MB 1 c MB λ (t)=d ψ (t)+ψ (t), (43) ETF↑ 2 c ETF↑ PET will be positive although ETF↑ does not excite MB. Unfortunately, it is not possible to make the same argument as in the REC measure, because the PET measure is not constructed from the Hawkes model. Even if the mean reversion rate for the common latent process ψ (t) is slow, it will still make PET positive. c One robustness test is to check against the REC measure since REC is not subject to this issue, as shown in Appendix F. As shown in Section 4.1, PET and REC are positively correlated in the cross section, thus mitigating this concern. 54

Checking against the REC measure in the time series, however, is troublesome. When hard information comes out more frequently, aggressive HFTs may act on fractionally less of the information (i.e., REC decreases), but from the point of view of the liquidity provider, a higher fraction of market orders may be informed (i.e., PET increases). I thus present another robustness test using an alternative PET measure that is significant if and only if the mean reversion of the common latent component ψ (t) is very fast. c IfweplotthedistributionofeventMBaroundeventETF↑,thedistributionwouldlookdifferent,depending on the mean reversion rate of ψ (t). As in Figure 8, the distribution would be flat if ψ (t) mean-reverts c c slowlyordoesnotmeanrevertbutwouldexhibitamuchstrongerhump-shapedpatternifψ (t)mean-reverts c rapidly. Therefore, I measure (cid:192)−(cid:193) in Figure 8 to distinguish the two cases. The PET measure used in the paper corresponds to (cid:192) in Figure 8. (a) Slow or moderate (b) Fast (cid:192) (cid:193) (cid:192) (cid:193) ETF↑ 0.05 0.1 ETF↑ 0.05 0.1 Figure 8: Illustration of distribution of event MB around event ETF↑ (a) illustrates the case when the common latency component ψ (t) mean-reverts slowly, and (b) when it c mean-reverts fast. Units are in seconds, thus 0.05 and 0.1 are 50ms and 100ms, respectively. Recall from (12) how PET was measured, and define longPET in the same manner but using a 100ms window to define “ETF-driven” and “opposite” from Figure 4. To measure (cid:192)−(cid:193), I define PET2 as PET2 =PET −longPET . (44) i,τ i,τ If ψ (t) is not driven by hard information, PET2 should be zero on average and independent of PET. If c it is largely driven by hard information, PET and PET2 will be strongly correlated. Figure 9 confirms that the latter is the case. 55

Figure 9: PET and PET2 Panel (a) is the pooled level data using raw values of PET and PET2. Panel (b) focuses on within day variation, where I subtract the daily mean and divide by the daily standard deviation. Panel (c) presents within-stock variation, where I subtract the stock-specific mean and divide by the stock-specific standard deviation. 56