high-frequency trading: empirics · 2016. 11. 1. · outline 1 introduction to empirics of hft...

High-Frequency Trading: Empirics

Jakub Rojcek

University of Zurich and Swiss Finance Institute

November 3, 2016University of Sankt Gallen

Jakub Rojcek High-Frequency Trading: Empirics 1 / 58

Outline

1 Introduction to Empirics of HFTObjectivesAlgorithmic Trading and High-Frequency TradingLiterature Overview

2 Market Quality and High-Frequency Trading

3 Evaluating the Impact of Regulations

4 Strategies of High-Frequency Traders

5 Conclusion and Outlook


What Can We Face while Studying HFT?

Figure: Bid size of Pfizer stock on October 10, 2011.


Objectives

1 Learn how to measure high-frequency activity.

2 Explore data sources.

3 Answer pressing questions about high-frequency trading.

4 Evaluate impact of regulations.


Algorithmic Trading vs High-Frequency Trading

Figure: Source: Johnson (2010).Jakub Rojcek High-Frequency Trading: Empirics 5 / 58

Some of the Streams of Empirical Inquiry

Impact on Market Quality

Impact of Regulations Strategies of HFTs

Optimal Execution/LOB EmpiricsHF Econometrics


Outline

1 Introduction to Empirics of HFT

2 Market Quality and High-Frequency TradingHasbrouck and Saar (JoFM 2013)Boehmer, Fong and Wu (2015)Brogaard, Hendershott and Riordan (RFS, 2014)Lin Tong (2015)





Low-latency trading

Joel Hasbrouck and Gideon Saar

Journal of Financial Markets 16 (2013) 646-679

Joel Hasbrouck and Gideon Saar Low-latency trading 8 / 58

Measuring Low Latency Trading: Strategic RunsHasbrouck and Saar proxy HFT/low latency activity using Strategic Runs, which arelinked submissions, cancellations, and executions that are likely to be parts of a dynamicalgorithmic strategy.They use only longer runs–runs of 10 or more messages–to construct the measure.They construct the measure of RunsInProcess, as the time-weighted average of thenumber of strategic runs of 10 messages or more the stock experiences in an interval.

The measure is highly correlated with the NASDAQ measure of HFT trading (0.7-0.8).


Measuring Low Latency Trading: Strategic RunsThey estimate the following model using GMM, Driscoll-Kraay standard errors.

Moreover, controlling for index return & volatility and time of the day dummies.

MktQualityi,t = α1RunsInProcessi,t + α2TradingIntensityi,t + εi,t . (1)

Market Quality Measure α1 α2

Spread -0.214 (<0.001) 0.096 (<0.001)EffSprd -0.170 (<0.001) 0.103 (<0.001)NearDepth 0.246 (<0.001) -0.104 (<0.001)HighLow -0.056 (<0.001) 0.322 (<0.001)

There might be a problem of simultaneity or endogeneity in the above model.

They use RunsNotINDi,t as an instrument. Average of RunsInProcessi,t excluding theINDividual stock and the same INDustry.

2 SLS results are as below.

Market Quality Measure α1 α2

Spread -0.682 (<0.001) 0.104 (<0.001)EffSprd -0.493 (<0.001) 0.109 (<0.001)NearDepth 0.460 (<0.001) -0.107 (<0.001)HighLow -0.437 (<0.001) 0.329 (<0.001)


International Evidence on Algorithmic Trading

Boehmer, Fong and Wu

2015

Boehmer, Fong and Wu International Evidence on Algorithmic Trading 11 / 58

Study AT Proxy on 42 Equity Markets Around the World

Main data source is Thomson Reuters Tick History (TRTH) database, which containsintraday trades and quotes data for many markets around the world. They combine thesedata with TAQ, and merge the result with firm-level data in Datastream and Center forResearch and Security Prices (CRSP) database. Data on buy-side transaction costs comefrom the Ancerno database.

In construction of their AT proxy, they follow Hendershott, Jones, and Menkveld (JF

2011): algo tradi,t = − $ Volumei,t# of Messagesi,t

.

Their model is

MktQualityi,t = αi,t + γt + βalgo tradi,t−1 + δXi,t−1 + εi,t . (2)

AT proxy could again suffers from simultaneity/endogeneity, they use a co-locationdummy as intrumental variable in 2SLS estimation.

They document decreased spreads and increased efficiency, in line with the Swedish

co-location study of Brogaard, Hagstromer, Norden and Riordan (RFS, 2015).


List of Co-location Events across Countries


Impact of AT on Market Quality Using IV


High-frequency trading and price discovery

Brogaard, Hendershott and Riordan

Review of Financial Studies, 27(8), 2267-2306, Aug 2014

Brogaard, Hendershott and Riordan High-frequency trading and price discovery 15 / 58

NASDAQ HFT Dataset vs TAQ DatasetHFTD is a sum of {HH,HN} types, HFT S of {HH,NH}.


State Space Model of HFT and Prices

The state space model assumes that the price can be decomposed into permanent andtransitory component (Menkveld, Koopman and Lucas 2007):

pi,t = mi,t + si,t . (3)

The permanent (efficient) component follows a martingale:

mi,t = mi,t−1 + wi,t . (4)

Impact of HFT demanding liquidity is captured by surprises from autoregression HFTD

:

wi,t = κDi,HFT HFTD

i,t + κDi,nHFT nHFTD

i,t + µi,t . (5)

Transitory component is assumed to be stationary and modelled as:

si,t = φsi,t−1 + ψDi,HFTHFT

Di,t + ψD

i,nHFTnHFTDi,t + νi,t . (6)

Assumption: innovations are uncorrelated— cov(µt , νt) = 0.


Demanding Liquidity in the State Space Model


Supplying Liquidity in the State Space Model


HFT Trading Around Macro News Announcements


Summary and Related Literature

Overall, HFT facilitate price discovery by trading in the direction of permanent pricechanges,

and in the opposite direction of transitory pricing errors.

However, the interpretation of the results is not straightforward, which lead many to study,whether HFTs lean against the wind or go with the wind.

Boulatov, Bernhardt and Larionov (2016) study Nash equilibria when multiple traders seekto minimize transaction cost when trading continuously in a fixed time period. They findthat traders who have a small position to trade, either lean against the orders oflarge-position traders or prey on them. Both types of behavior can occur in equilibrium.The large-position traders’ optimal response to preying is to delay trading.

Korajczyk and Murphy (2016), and van Kervel and Menkveld (2016) find that HFTs leanagainst the order in the first hour, but turn around and trade with the order in the case ofmulti-hour executions.

Breckenfelder (2013) finds that increased competition leads to increase in liquidityconsumption by HFTs (going with the wind).

Tong (2015) find by combining data on institutional trades and HFT trades, that HFTincreases traditional institutional investors’ trading costs measured by execution shortfall.


A Blessing or a Curse? The Impact of High FrequencyTrading on Institutional Investors

Lin Tong

2015

Lin Tong A Blessing or a Curse? The Impact of High Frequency Trading on Institutional Investors 22 / 58

Are Institutional Investors Suffering Higher ExecutionCosts?

Uses the standard NASDAQ HFT dataset to measures the HFT intensity.

The measure of HFT daily activity on stock i, denoted as HFT Intensityi,t , is defined asthe aggregate HFT volume for stock i on day t divided by the stock’s average dailytrading volume in the past 30 days.

The measure of institutional trading costs is constructed from the Ancerno proprietarydataset. The execution shortfall is measured as

P1 − P0

P0× D, (7)

where P1 is the value-weighted execution price of the ticket, P0 is the price at the timewhen the broker receives the ticket, and D is a dummy variable that equals 1 for a buytrade and -1 for a sell trade.


HFT Activity Coincides with Increased Execution Shortfall


Institutional ES does not Granger-cause HFT Activity


Outline



3 Evaluating the Impact of RegulationsColliard and Hoffman (JF, 2016)More Regulation Event Studies




Financial Transaction Taxes, Market Composition, andLiquidity

Jean-Edouard Colliard and Peter Hoffmann

Forthcoming in The Journal of Finance

Jean-Edouard Colliard and Peter Hoffmann Financial Transaction Taxes, Market Composition, and Liquidity 27 / 58

On August 12, 2012, France Introduced FTT

On August 1st, 2012, France introduced an FTT of 20 bps on stock purchases. This taxapplies to shares of all listed companies incorporated in France with a marketcapitalization above one billion euros.

The tax is payable on daily net position changes (i.e., ownership transfers), which impliesthat pure intraday trading is de facto exempted.

Simultaneously, the government introduced a tax of 1 bp on the notional amount ofmodified or cancelled messages by HFTs exceeding an order-to-trade ratio of 5:1. Unlikethe FTT, this tax applies to trading in all French stocks. However, it is only levered onHFTs residing in France, which excludes all major HFT firms. Moreover, message trafficdue to market-making is exempt.

They utilize difference-in-difference approach summarized by the regression

yi,t = αi + γt + βAugDAugi,t + βSep/OctD

Sep/Octi,t + εi,t . (8)

They use high-frequency data of Euronext French stocks for treated group andcomparable Euronext non-French stocks for control group.


Causal Impact of the FTT on all Stocks.


Causal Impact of the FTT on all Stocks II.


Causal Impact of the FTT on all Stocks. III


Causal Impact of the FTT on Trading Volume and OrderFlow Composition

HFTs are strongly impacted by the FTT despite the effective exemption of intradaytrading.


Related Literature on the Impact of RegulationsThe empirical evalution of the impact of transaction tax includes the following:

French FTT: Meyer, Wagener, and Weinhardt (2015) document a reduction in thenumber of quote and price updates by liquidity suppliers and a decline in top order bookdepth. Gomber, Haferkorn, and Zimmermann (2015) report a drop in both liquiditydemand and supply, an increase in spreads, and a decline in top order book depth.Becchetti, Ferrari, and Trenta (2014) find a reduction in turnover and intraday volatility,but mixed effects on liquidity.

Umlauf (1993) investigates the impact of the introduction of the Swedish transaction taxand documents a large shift in trading volume from Sweden to the U.K.

Cancellation fees

Malinova, Park, and Riordan (2013), find that following their introduction on the TorontoStock Exchange, quoting activity decreased by 30% and spreads increased by 9%.

Friederich and Payne (2015) study the introduction of excessive order-to-trade ratio fee onApril 2nd, 2012, by the Milan Borsa. They observe decreased depth and worsened pricediscovery.

Make/Take fees

Malinova and Park (2015) use a change in Toronto Stock Exchange fees to investigate theimpact of make-take fees on market quality. They find that rebates improve quotedspreads, decrease adverse selection costs, and increase the aggressiveness of retail traders.They also observe that the effective spread plus total fee retained by the exchangeremained constant, confirming the conjecture of Colliard and Foucault (2012) that onlytotal fees (spread plus exchange fees) affect liquidity and trading volume.


Outline




4 Strategies of High-Frequency TradersBaron, Brogaard, Hagstromer and Kirilenko (2016)Gencay, Mahmoodzadeh, Rojcek and Tseng (2016)



Risk and Return in High-Frequency Trading

Baron, Brogaard, Hagstromer and Kirilenko

2016

Baron, Brogaard, Hagstromer and Kirilenko Risk and Return in High-Frequency Trading 35 / 58

Measuring Latency of a HFT Based on his ActionsFor each HFT in each month, they record all cases where a passive trade is followed by anaggressive trade by the same HFT, in the same stock and at the same trading venue,within one second. The time-stamp difference between the two trades of each case formsan empirical distribution of response times. To capture the fastest possible reaction time,they define Decision Latency as the 0.1% quantile of the aforementioned distribution.


Profitability Statistics across HFTs


Impact of Latency on ProfitabilityThey measure influence of the performance in the following model:

Pi,t = αt+β1 log Latencyi,t+β21top 1i,t +β31top 2−3

i,t +β41top 4−5i,t +γ′Ci,t+month−FEs+εi,t .

(9)


Price Impact and Bursts In Liquidity Provision

Gencay, Mahmoodzadeh, Rojcek and Tseng

2016

Gencay, Mahmoodzadeh, Rojcek and Tseng Price Impact and Bursts In Liquidity Provision 39 / 58

Market Activity Exhibits Bursts

The quotation arrival process directly reflects market makers’ activity.

Number of quotes for Alcoa between 10:31:15 and 10:32:15 on October 10, 2011.


Research Questions

We analyze the periods during which the arrival intensity of quotes is largerelative to what would be normal in a given interval. We call such periodsbursts in quotation activity.

We investigate the functioning of markets and price formation during thesebursts.

What is the impact of bursts in quotation activity on market quality?

Is there a change in the informational relationship between marketmaker and market taker? We will examine this from two perspectives.

1 Price impact of trades – ex ante adverse selection2 Realized spread based empirical measure – ex post adverse selection


Main Findings

On market characteristics during bursts:

I Quoted spread, effective spread, and volatility increase.I Size of incoming orders decreases.

On informational relationship between maker and taker during bursts:

I The informational dichotomy breaks down. Market makers no longerpassively impound information from order flow into quotes.

I Ex ante adverse selection (price impact) rises significantly, on averagefrom 0.002% to 0.011% per trade.

I Ex post adverse selection also increases.

Adverse selection is in fact asymmetric, depending on whether the marketmaker occupies the same side of the LOB as the burst.

I On the burst side, there is negative price impact.I On the opposite side, the severity of adverse selection rises significantly.


Burst Detection

”one man’s noise is another man’s signal”

We detect bursts based on the following market activityvariables:

Number of trades

Number of quotes

Number of quotes on different sides of the book

These quantities are approximately exponentiallydistributed (see also Zhu and Shasha 2003, 2005. Vlachos et al 2004, 2008.)

1 For each 15 min. interval, we estimate the mean of the distribution µ.

2 We record a burst for each interval (100ms, 1s, 5s), in which the variable exceedsa threshold x = −µ log(P), where P is the tail probability (e.g. P = 99.9%).

Our definition of bursts coincides with that of unified outliers in Knorr (1997).


Example: Alcoa on October 10, 2011, 1 Second Frequency

Quotes Trades


Market Behavior Around Bursts: Quoted Spread

100ms Interval 1s Interval 5s Interval

The black line represents the average quoted spread across all 100ms (1s, 5s)intervals before, during and after bursts, as well as across all stocks in our sample.

The initial value was normalized to 100.

The shaded area represents one standard deviation from the mean.


Market Behavior Around Bursts: Effective Spread


The black line represents the average effective spread across all 100ms (1s, 5s)intervals before, during and after bursts, as well as across all stocks in our sample.




Market Behavior Around Bursts: Return Volatility


The black line represents the average volatility across all 100ms (1s, 5s) intervalsbefore, during and after bursts, as well as across all stocks in our sample.




Market Behavior Around Bursts: Order Size


The black line represents the average order size across all 100ms (1s, 5s) intervalsbefore, during and after bursts, as well as across all stocks in our sample.




Estimating Permanent and Temporary Price Impact

We use the price impact methodology of Hasbrouck (1991).

Let

rt denote the change in the market maker’s estimate of the fundamental value.

qt be the signed order flow (-1 if more trades at bid, +1 if more trades at ask and0 if equally many or no trades.)

The VAR system is

rt = α1rt−1 + α2rt−2 + · · · + β0qt + β1qt−1 + · · · + εt︸︷︷︸Market maker’s public information

, (10)

qt = γ1rt−1 + γ2rt−2 + · · · + δ1qt−1 + δ2qt−2 + · · · + νt︸︷︷︸Market taker’s private information

. (11)


Price Impact is Higher During Bursts

We add dummy variables BQt and BT

t to identify periods with bursts in quotes and trades.

rt =10∑i=1

αi rt−i +10∑i=1

βiqt−i + β0qt + φQBQt qt + φTBT

t qt + φQ&TBQt B

Tt qt + ε′t .

α1 β0 φQ φT φQ&T R2-Adj R2-Adjwith bursts w/o bursts

AA 0.925 0.002 0.046 0.007 -0.012 0.585 0.489BAC 0.988 0.001 0.058 0.007 -0.009 0.593 0.510GE 0.917 0.001 0.028 0.007 -0.007 0.567 0.484IBM 0.247 0.003 0.006 0.005 -0.001 0.416 0.416JNJ 0.445 0.002 0.004 0.004 -0.001 0.411 0.397JPM 0.875 0.006 0.009 0.011 -0.002 0.502 0.482MRK 0.851 0.002 0.012 0.007 -0.004 0.518 0.465PFE 0.827 0.000 0.023 0.005 -0.001 0.541 0.461PG 0.619 0.002 0.005 0.005 -0.001 0.403 0.381PM 0.336 0.002 0.004 0.006 0.000 0.407 0.407T 0.742 0.001 0.015 0.006 -0.006 0.484 0.430VZ 0.846 0.001 0.009 0.007 -0.003 0.505 0.437WFC 0.922 0.005 0.013 0.013 -0.008 0.551 0.509WMT 0.705 0.002 0.007 0.005 -0.002 0.462 0.430XOM 0.463 0.002 0.004 0.006 0.002 0.459 0.441

Median coeff 0.827 0.002 0.009 0.006 -0.002 0.502 0.441Median p-value 0.000 0.000 0.000 0.000 0.125

Median daily # of events 419.3 706.1 318.0

% of significant firm-days at 5 % 99.67% 93.00% 81.67% 91.67% 30.00%


Breakdown of Informational Dichotomy

Recall the price and order flow equations:

rt = α1rt−1 + α2rt−2 + · · ·+ β0qt + β1qt−1 + · · ·+ εt︸︷︷︸Market maker’s public information

qt = γ1rt−1 + γ2rt−2 + · · ·+ δ1qt−1 + δ2qt−2 + · · ·+ νt︸︷︷︸Market taker’s private information

If the market maker can predict the order flow, εt and νt will be positivelycorrelated.

Our hypothesis is that {εt , νt} are correlated during, and only during, burstsin quotes. We test the correlation between εt and νt .

In addition to correlations during burst, {εt , νt}, and non-burst periods,{ε′t , νt}, we estimate the linear specification

νt = ηQBQt εt + η¬Q(1− BQ

t )ε′t + ν′t .


Information Correlation During Bursts in Quotes

ρQ ρ¬Q ηQ η¬Q ρAll

AA 0.487 -0.002 7.779 -0.130 -0.002BAC 0.431 0.007 4.230 0.508 0.005GE 0.514 -0.002 12.735 -0.237 -0.002IBM 0.193 0.000 7.789 0.000 0.000JNJ 0.323 0.000 20.476 -0.021 0.000JPM 0.413 0.000 13.601 -0.037 0.000MRK 0.516 -0.002 22.172 -0.247 -0.002PFE 0.478 0.000 13.568 -0.087 -0.001PG 0.339 0.000 22.921 -0.049 0.000PM 0.022 0.000 0.526 -0.001 0.000T 0.478 -0.001 21.501 -0.202 -0.001VZ 0.417 -0.003 21.597 -0.332 -0.002WFC 0.493 -0.001 15.465 -0.060 0.000WMT 0.410 0.000 26.169 -0.017 0.000XOM 0.206 0.000 15.633 -0.046 0.000

Median coeff 0.417 0.000 15.465 -0.049 0.000Median p-value 0.000 0.862 0.000 0.871 0.884

Median daily # of events 101.3 20,789 101.3 20,789 20,890

% of significant firm-days at 5 % 100% 0% 100% 0% 0%

Market maker in fact predicts price movement/order flow during bursts in quotes.

Basic market measures show no sign that a burst is anticipated by the market.

Therefore bursts are market maker-initiated informational events.


Adverse Selection and the Side Where the Burst Happens

We investigate the behavior of the adverse selection proxy (Bessembinder (2003))

ADt = ESt − RSt ,

where, letting Pt denote the transaction price and Mt the mid-quote,

RSt = 2qt(Pt −Mt+s)

is the realized spread (profit) of the market maker, and

ESt = 2qt(Pt −Mt).

is the effective spread measuring the price of immediacy and the adverse selectioncomponent.

ADt is the magnitude of the ex post loss due to adverse price movement pluseffective spread.


Asymmetric Adverse Selection cont’d

Adverse selection

Realized spread

Effective spread

Burst at Ask Burst at Bid

Market makers on the opposite side of the burst charge higher cost ofimmediacy thus partially compensating for the negative realized spread.Despite this partial compensation, they suffer more severe adverse selection.Gencay, Mahmoodzadeh, Rojcek and Tseng Price Impact and Bursts In Liquidity Provision 54 / 58

MMs Burst in the Direction of the Price ImpactTo the basic specification we add B

Q,samet and B

Q,oppt representing periods when a burst occurs on the side of the book where

the order flow arrives or on the opposite side, respectively;

rt =10∑i=1

αi rt−i +10∑i=1

βiqt−i + β0qt + φQ,sameBQ,samet qt + φQ,oppBQ,opp

t qt + φTBTt qt + ε′t .

α1 β0 φQ,same φQ,opp φT R2-Adj R2-Adjwith bursts w/o bursts

AA 0.929 0.002 -0.030 0.041 0.031 0.545 0.489BAC 0.988 0.001 -0.024 0.040 0.043 0.558 0.510GE 0.913 0.001 -0.025 0.030 0.018 0.534 0.484IBM 0.247 0.003 0.003 0.007 0.007 0.416 0.416JNJ 0.445 0.002 -0.006 0.008 0.005 0.413 0.397JPM 0.875 0.006 -0.008 0.010 0.013 0.504 0.482MRK 0.848 0.002 -0.014 0.011 0.009 0.507 0.465PFE 0.826 0.001 -0.019 0.024 0.017 0.519 0.461PG 0.619 0.002 -0.003 0.005 0.006 0.397 0.381PM 0.336 0.002 0.000 0.001 0.006 0.408 0.407T 0.742 0.001 -0.013 0.016 0.011 0.472 0.430VZ 0.845 0.001 -0.010 0.010 0.009 0.486 0.437WFC 0.920 0.005 -0.013 0.016 0.014 0.547 0.509WMT 0.704 0.002 -0.008 0.009 0.006 0.455 0.430XOM 0.446 0.002 0.002 0.002 0.006 0.455 0.441

Median coeff 0.826 0.002 -0.010 0.010 0.009 0.486 0.441Median p-value 0.000 0.000 0.002 0.000 0.000

% of significant firm-days at 5 % 99.0% 93.33% 67.33% 82.67% 96.0%Median daily number of events 24.8 60.7 318


Conclusion

We have shown that during bursts in quotes:

1 The price impact rises markedly from 0.002% to 0.011% per trade.It is higher than during bursts in trades.

2 Market makers do not passively update quotes based oninformation inferred from order flow.

3 Limit orders on the opposite side of the burst suffer adverseselection.

4 Price movement during bursts is likely not part of the normal priceformation process.

5 Whether bursts are sufficiently disruptive to warrant concern is aquestion of interest to regulators, given recent attempts by exchangesto curtail high frequency spamming.


Outlook

Model extensions:

1 Change to business clock instead of 1s intervals.

2 Vary the tail probability for computing threshold to see if there isnon-linearity in price impact.

Obtain data to:

1 Investigate heterogeneity in the cross-section of firms.

2 Distinguish bursts around news announcements from others.


Conclusion and Outlook

We have answered many questions about HFT...1 We learned how to measure the HFT activity...2 ... and how it impacts traditional measures of market quality.3 We learned that regulations can introduce frictions, evidence of which

supports theoretical predictions.4 And we learned that two papers studying the same thing can lead to

different conclusions, dwindling our confidence...5 ... and that there might exist strategies systematically manipulating

the prices.

... but we still have not much evidence on1 The true transaction costs impact of HFT for institutional investors.2 The ability of HFTs to predict order flow or extract private information.3 The relation of HFTs and systemic risk.4 The impact and influence of HFTs outside of the public equity or FX

markets.5 ...


Selected References

Aıt-Sahalia, Y. and J. Jacod. 2014. High-Frequency Financial Econometrics. Princeton, New Jersey, USA: Princeton UniversityPress.Baron, M. D., J. Brogaard, B. Hagstromer and A. A. Kirilenko. 2016. Risk and return in high frequency trading. Available atSSRN 2433118.Baruch, S. and L. R. Glosten, 2013. Fleeting orders. Working paper, Columbia University and the University of Utah.Becchetti, L., Ferrari, M. and U. Trenta. 2014. The impact of the French Tobin tax, Journal of Financial Stability 15, 127–148.Bernales, A. 2014. How fast can you trade? High frequency trading in dynamic limit order markets. Working paper, Banque deFrance.Biais, B., T. Foucault and S. Moinas. 2015. Equilibrium fast trading. Journal of Financial Economics 116(2), 292–313.Boehmer, E., K. Y. L. Fong and J. J. Wu. 2015. International evidence on algorithmic trading. In AFA 2013 San DiegoMeetings Paper.Brogaard, J. 2010. High Frequency Trading and Its Impact on Market Quality. Working paper, Northwestern University.Brogaard, J., T. J. Hendershott and R. Riordan. 2014. High-frequency trading and price discovery. The Review of FinancialStudies 27, 2267–2306.Brogaard, J., A. Carrion, T. Moyaert, R. Riordan, A. Shkilko, and K. Sokolov (2015). High-frequency trading and extreme pricemovements. University of Washington Working Paper.Budish, E. B., P. Cramton and J. J. Shim. 2013. The high-frequency trading arms race: Frequent batch auctions as a marketdesign response. Fama-Miller Working Paper, 14–03.Chacko, G. C., J. W. Jurek and E. Stafford. 2008. The price of immediacy. The Journal of Finance 63 3, 1253–1290.Chakravarty, S., P. Jain, J. Upson, and R. Wood. 2012. Clean sweep: Informed trading through intermarket sweep orders.Journal of Financial and Quantitative Analysis 47, 415–435.


Selected References cont’d

J.-E. and T. Foucault. 2012. Trading Fees and Efficiency in Limit Order Markets, Review of Financial Studies 25, 3389–3421.Colliard, J. E., and P. Hoffmann. 2015. Financial Transaction Taxes, Market Composition, and Liquidity. Working paper,European Central Bank.Committee of European Securities Regulators. 2010. Micro-structural issues of the European equity markets; Ref.:CESR/10-142, Call for Evidence.Commodity Futures Trading Commission (CFTC). 2014. Adoption of Rule 575 (“Disruptive Practices Prohibited”). CMESubmission No. 14 – 367.Conrad, J., S. Wahal and J. Xiang. 2015. High-frequency quoting, trading, and the efficiency of prices. Journal of FinancialEconomics 116(2), 271–291.Du, S. and H. Zhu. 2014. Welfare and Optimal Trading Frequency in Dynamic Double Auctions. Working paper, NationalBureau of Economic Research.Dufour, A. and R. F. Engle. 2000. Time and the price impact of a trade. Journal of Finance 55, 2467–2498.Egginton, J. F., B. F. Van Ness and R. A. Van Ness. 2014. Quote stuffing. Louisiana Tech University Working Paper.European Parliament. 2011. Directive of the European Parliament and of the Council on markets in financial instrumentsrepealing Directive 2004/39/EC of the European Parliament and of the Council, COM(2011) 656 final.Foucault, T., O. Kadan and F. Kandel. 2013. Liquidity Cycles and Make/Take Fees in Electronic Markets, The Journal ofFinance 68, 299–341.Foucault, T., J. Hombert and I. Rosu. 2016. News Trading and Speed. The Journal of Finance 71, 335–381.Friederich, S. and R. Payne. 2015. Order-to-trade ratios and market liquidity. Journal of Banking & Finance 50, 214–223.



Goettler, L. R., C. A. Parlour and U. Rajan. 2005. Equilibrium in a Dynamic Limit Order Market, The Journal of Finance 60,2149–2192.Goettler, L. R., C. A. Parlour and U. Rajan. 2009. Informed traders and limit order markets, Journal of Financial Economics 93,67–87.Hasbrouck, J. 1991. Measuring the information content of stock trades. Journal of Finance 46, 179–207.Hasbrouck, J. and G. Saar. 2013. Low-Latency Trading. Journal of Financial Markets 16, 646–679.Hasbrouck, J. 2015. High Frequency Quoting: Short-Term Volatility in Bids and Offers. Working paper, SSRN.Hendershott, T. J., Jones, C. M. and Menkveld, A. J. 2010. Does Algorithmic Trading Improve Liquidity? The Journal ofFinance 66, 1–33.Hoffmann, P. 2014. A dynamic limit order market with fast and slow traders. Journal of Financial Economics 113, 156–169.Johnson, B. 2010. Algorithmic Trading & DMA: An introduction to direct access trading strategies. Vol. 200. London:4Myeloma Press.Knorr, E. M. and R. T. Ng. 1997. A unified notion of outliers: Properties and computation. KDD Conference, 219–222.Korajczyk, R. A., and D. Murphy. 2015. High frequency market making to large institutional trades. Available at SSRN2567016.Malinova, K. and A. Park. 2015. Subsidizing Liquidity: The Impact of Make/Take Fees on Market Quality, The Journal ofFinance 70, 509–536.Malinova, K., A. Park and R. Riordan. 2013. Do retail traders suffer from high frequency traders? Working paper.Maskin, E. and J. Tirole. 2001. Markov perfect equilibrium: I. Observable actions. Journal of Economic Theory 100, 191–219.Menkveld, A. J. 2013. High Frequency Trading and The New-Market Makers. Journal of Financial Markets 16, 712–740.



Menkveld, A. J. and M. A. Zoican. 2016. Need for speed? Exchange latency and liquidity. Review of Financial Studies.O’Donoghue, S. M. 2015. The Effect of Maker-Taker Fees on Investor Order Choice and Execution Quality in US StockMarkets. Working paper.O’Hara, M. 2015. High frequency market microstructure. Journal of Financial Economics 116, 257–270.Pakes, A. and P. McGuire. 2001. Stochastic Algorithms, Symmetric Markov Perfect Equilibrium, and the ’curse’ ofDimensionality, Econometrica 69, 1261–1281.Riordan, R. and A. Storkenmaier. 2012. Latency, liquidity and price discovery. Journal of Financial Markets 15, 416–437.Rostek, M. and M. Weretka. 2015. Dynamic thin markets. Review of Financial Studies..Rosu, I. 2014 Fast and slow informed trading. AFA 2013 San Diego Meetings Paper.Securities and Exchange Commission (SEC). 2010. Concept release on equity market structure, Release No. 34-61358; File No.S7-02-10, Concept release.Tong, L. 2015. A Blessing or a Curse? The Impact of High Frequency Trading on Institutional Investors. Working paper.Vlachos, M., K. L. Wu, S. K. Chen and S. Y. Philip. 2008. Correlating burst events on streaming stock market data. DataMining and Knowledge Discovery 16, 109–133.


Role of Market Maker: View from the Market

“The role of the Market Maker on TSX is to augment liquidity, while maintainingthe primacy of an order-driven continuous auction market based on price-timepriority. TSX’s Market Maker system maximizes market efficiency and removes theinterfering influence of a traditional specialist. In the TSX environment, a MarketMaker manages market liquidity through a passive role. Market Makers are visibleonly when necessary to provide a positive influence when natural market forcescannot provide sufficient liquidity.”

“IMC’s primary role is to help make markets more efficient. An efficient market is

one that is liquid. A liquid market is one that allows investors to buy or sell

securities at any time, in a normal size, and get an instantaneous execution. As a

market maker, we provide that liquidity. An efficient market is one in which

transaction costs for investors are as low as possible. Low costs are the direct

result of the speed at which we trade. The faster we are able to react to adjust our

prices to information and events that move markets, the tighter we are able to

quote, ensuring competitive prices.”


What Does A Market Maker Do?

Traditional definition: a market maker provides liquidity andimpounds information into quotes. For the market maker, order flow(& news) is information flow.

In investigating the impact of algorithmic trading on market quality,by and large the market microstructure literature has retained thedefinition of market maker as a passive liquidity provider.

However, data reveals market maker behavior that cannot beattributed to passive liquidity provision or adverse selection.


First Observations

The spikes—bursts in quotes— in the rate of incoming quotes in factoccurs with regularity throughout the trading day, with averageinter-arrival time of approximately 40 seconds.

The average duration of bursts is approximately 1.1 seconds.

Bursts is therefore a fundamentally high frequency phenomenon.

At first glance, it cannot be attributed to the role of market maker asliquidity provider nor information aggregator.


Related Literature

HFT surveys: Jones (2013), O’Hara (2015), Menkveld (2016) andothers. The rise in quote cancellations in the millisecond environmentis documented, e.g. by Hasbrouck and Saar (2013) and many others

Quotes volatility studied by Hasbrouck (2015) and reflect Edgeworthcycles as in Baruch and Glosten (2013).

Egginton et al. (2014) examine market quality impact of bursts.

Intermarket Sweep Orders (ISO) studied by Conrad et al. (2015) andChakravarty et al. (2012).

Brogaard et al. (2015a) HFTs as net liquidity suppliers during EPMs.Brogaard et al. (2014) find HFT trade in the direction of permanentprice impact and opposite of transitory pricing errors. Brogaard et al.(2015b) HFTs are responsible for price discovery mainly through LOs.

Cont et al. (2014) consider price impact of LOs following the OF priceimpact literature e.g. Hasbrouck (1991a, 1991b), Dufour and Engle(2000), Zhang et al. (2001), Engle and Lunde (2003), Hautsch andHuang (2012), see also Bouchaud (2010), and Pham et al. (2015).


Comments on the VAR model

The lags serve as controls for microstructure frictions: e.g. inventorymanagement, price discreteness, lagged adjustment to information.

Inventory-control effects may cause past quote updates to influencethe market maker’s current quote update. Delayed adjustment toinformation and anticipation of order-splitting means order flow mayhave lagged effects on quote updates.

Order flow {qt} may exhibit autocorrelation due to order splitting. Itis also possible that a market taker who anticipates market maker’sinventory-neutral tendency may adjust his order flow according topast quote updates.


Comments on the VAR model cont’d

Embedded in the exogeneity assumption of the VAR model of {rt , qt}is the traditional definition of a market maker: εt and νt areorthogonal.

Our analysis will suggest a modification of this model that departsfrom this definition and captures the informational relationshipbetween market maker and taker during bursts.

We first carry out a model selection exercise for the appropriatenumber of lags.


VAR Lag Selection

We select the number of lags based on Akaike and Bayesianinformation criteria.

We use 10 lags throughout. Our results are robust for 5 to 30 lags.

This figure represents values of Bayesian information criterion for the returns (left) and order flow equations (right) for onetrading day of Alcoa. We obtained similar results for other stocks and days and by using Akaike information criterion as well.



We will show that, concurrent with bursts, advese selection foropposite sides of LOB undergoes abrupt shifts in severity in oppositedirections.

This explains the apparent contradiction in price impact andinformational correlation.

Price impact is computed based on quotes that were actuallyexecuted against—those on the opposite side of bursts, therefore hasan upward bias.

Information correlation reflects the information of market makers whoinitiate bursts.



Market makers on the opposite side of the burst charge higher cost ofimmediacy for order flow trading in the direction of the price impact,thus partially compensating for the negative realized spread.

Despite this partial compensation, they suffer more severe adverseselection.

While bursts lasts 1.1 seconds on average, the price impact it exertssignificant price impact at 30 second lag.

The asymmetry shown above suggests we adapt the VAR model toconsider price impact for the two sides of LOB separately.


high-frequency trading: empirics · 2016. 11. 1. · outline 1 introduction to empirics of hft...

Documents