ekkehart boehmer, edhec business school kingsley fong, unsw julie wu, university of georgia
TRANSCRIPT
INTERNATIONAL EVIDENCE ON ALGORITHMIC TRADING
Ekkehart Boehmer, EDHEC Business SchoolKingsley Fong, UNSWJulie Wu, University of Georgia
What is algorithmic trading?
Using computer algorithms to generate, monitor, and cancel orders automatically process prices, LOB, news predict flows, liquidity, prices
Has been around for at least 25 years (ATD) but saw enormous growth over last decade milestones: Limit order display rules (NYSE
1993, Nasdaq 1997), Reg ATS (1998), Reg NMS (2005)
U.S. Trading centers (SEC 2010)
Trading centers and estimated % of share volume in NMS stocks (September 2009)
Agency vs. proprietary algos
Agency algos – used by buy-side to minimize costs of portfolio turnover
Proprietary algos goal is to profit from trading
environment rather than investing a subset of these proprietary algo users
we consider high frequency trading (HFT) – traders with response times measured in milliseconds
Who are HFT?
Proprietary trading firms, potentially organized as B/D
Proprietary trading desk of large B/D firm
Hedge funds
Four broad categories of HFT strategies (SEC 2010)
Passive market making strategies – provide liquidity, earn the spread
Arbitrage strategies – seek short-horizon patterns
Structural strategies – exploit features of trading protocol or regulation
Directional strategies – order anticipation, momentum ignition, …
None of these strategies is new, but the technology to implement them is.
Issues
HFT is for real – between 60% and 80% of volume, depending on the market
Proportion of good vs bad strategies is unknown
What are consequences of HFT for• market quality?• welfare (of traders, society, …)?• systemic risk?
Our priors regarding HFT’s effect on market quality
Depends on strategies passive market making should improve liquidity stat arb should improve efficiency structural and directional strategies could be
wealth transfers
WE DON’T KNOW ACTUAL STRATEGIES, SO THE BEST WE CAN DO IS LOOK AT THE AGGREGATE EFFECT of AT/HFT
(Ideally, we should care about aggregate welfare: competition among HFT and with other traders may be desirable, IT arms races are probably not)
Theoretical literature
Modeling HFT
Standard approach: design a model with (typically slow) liquidity
traders and/or market makers, add HFT typically assume that HFT is informed or sees
information first not surprisingly, result is a wealth transfer from
slow to fast traders (e.g., Cartea and Penalva 2011, Jarrow and Protter 2011)
Instead, we may need models where the choice to become HFT is endogenous or liquidity traders can penalize HFT
Social welfare
Probably most important: under what conditions can HFT increase welfare Hoffman 2012 lets traders endogenously invest in
fast trading technology. Welfare gains when markets are sufficiently efficient. Otherwise IT overinvestment.
Biais, Foucault, and Moinas 2010 show that HFT can generate gains from trade and gains from adverse selection, but a social planner would only consider gains from trade, not from adverse selection. Again, HFT overinvest in technology.
Ambiguous effects in Jovanovic and Menkveld 2011
Empirical literature
Available data sources
Transactions data where we infer the presence of AT/HFT indirectly from periods of very high message traffic
Transactions data with trader category information that have a set of transactions attributed to AT/HFT
Transactions data with trader account information
Results based on inferred AT/HFT activity: Message counts Hendershott, Jones, and Menkveld
2011 use order-level message counts as a proxy for AT sample covers NYSE activity 2001-2005 algo trading is positively related to
market quality (improves spreads and price discovery, and reduces information asymmetry)
AT improve price discovery (Hendershott and Riordan 2009)
Results based on inferred AT/HFT activity: Millisecond episodes Hasbrouck and Saar 2011 use
2007/2008 INET Itch data infer HFT from millisecond responses to
market events such as quote updates identify fast trades that are linked in
time to identify episodes of HFT find that HFT improves short-term
volatility, spreads, and depth
Results based on inferred AT/HFT activity: Spikes in quoting activity Egginton et al. 2011 use 2010 TAQ data
select one-minute intervals where quotes-per-minute exceed 20 historical 20-day s.d. (on days where daily quoting is less than two s.d. away from the mean)
discard days with information events
Episodic spikes of (TAQ) quoting activity are quite frequent in many stocks are associated with degraded liquidity
and elevated short-term volatility
Results based on the Nasdaq sample with trader categories Brogaard (2010) uses a 2008-2009 Nasdaq
sample covers 120 Nasdaq stocks (selected by
academics) 26 HFT firms/traders that account for 77% of
trades/74% of dollar volume in this sample (selected by Nasdaq, excludes large prop desks)
HFT activity is associated with better liquidity, mixed effect on volatility, better price discovery (Brogaard 2010, Hendershott and Riordan 2011)
Problems with the Nasdaq sample with trader categories
Problem with the Nasdaq sample: potential selection bias less than sparkling clean traders would rationally veto
reporting included transactions do not include same-trader
activity on other markets and may not include all their activity on Nasdaq (?)
exchange may purposely select liquidity suppliers
This sample probably contains HFT activity that is more benign than that in a random sample of HFT
Results based on data with individual trader IDs
Kirilenko, Kyle, Samadi, and Tuzun 2011 see individual strategies in S&P500 e-minis find that HFT worsened (but did not cause)
the Flash Crash of 2010.
Only with trader IDs can we track each trader’s orders across
stocks and markets and infer actual strategies overcome endogeneity problem accurately determine triggers and
consequences of HFT activity
To sum up: we have somewhat mixed results and suboptimal data Broad samples (with inferred AT/HFT
activity) imply mostly positive effects on average (MQ) mixed results on volatility negative effects during periods of extremely
high message traffic and around quote updates
Trader type samples positive effect on MQ and efficiency mixed results on volatility selection concerns
Our motivation
What are the consequences of AT/HFT for markets, traders, firms, countries?
To address gaps in the literature, to answer market structure questions, to evaluate whether LTI or AT/HFT are more important: we need more and broader evidence. from new samples. this paper is a first step in that direction.
Our objectives
Use an international sample to measure the impact of AT/HFT on execution cost for different types of
traders price discovery volatility
Document differences across firms, trading protocols, and countries in these measures (incomplete)
Data
• Intraday quote and trade data from Thomson-Reuters Tick History (TRTH) 39 exchanges, 36 countries, 2001-2009 on average about 13,000 stocks will add NYSE and Nasdaq data from TAQ
Daily data on returns, volume, high-low prices from Datastream
Information about trading protocols from Speedguide, Exchange Handbook, WFE
Accounting data from Datastream and WorldScope
Our variables
Proxy for algorithmic trading
AT = -trading volume / # messages measures negative of trading
volume in USD 100 per message follows Hendershott, Jones, and
Menkveld 2011 messages include trades and quote
updates
Liquidity measures
Time weighted quoted spreadsRQS=(ask-bid)/MQ
Trade weighted effective spreadsRES=2*|P-MQ|/MQ
Amihud = |Return|/volumeAncerno (formerly Abel-Noser) buy-
side trading costs
Information vs. transient price impact
RES captures total price impact of a trade.
Decompose RES into RPI, the permanent component, the
change in MQ from trade to 5 minutes later, measures information content (“toxicity”)
RRS, the transient component, measures reward to liquidity providers, RES = RRS + 2* RPI
Informational efficiency
|autocorrelation| for 10-60 minutes intervals for each stock, compute midquote returns for 5,
10, 20, 30 , 60 minute intervals then compute autocorrelation of those returns ignore overnight and zero returns note that there is no bid-ask bounce in this
measure
The more efficient the stock price (the closer it is to a random walk), the smaller is |ARnn|
Volatility
Use several standard volatility measures |R|, for raw and market-adjusted return R^2, for raw and market-adjusted return Log (Ret10_Var), Log (Ret30_Var) daily relative price range =
(High-Low)/Close
Econometric methodology
Methodology
Have three-dimensional unbalanced panel 39 markets, about 2250 days, about 330 stocks
per market standard unobservable firm-level and time effects unobservable market or country effects
Use daily standardization of all variables so coefficients are comparable across markets
Regress variable of interest on AT and controls
Methodology
All regressions control for Volume (share turnover) Volatility (relative price range, excluded from
volatility regressions) price level (1/price) firm size (ln market cap)
Volatility regressions additionally control for RES and |AR30|
Estimation strategy estimate firm/day panel regression for each market then aggregate across markets
Methodology
Use two different approaches within markets Two-way dynamic panels within each market ▪ use firm and day fixed effects▪ use Arellano and Bond (1991) standard errors for
market-level inference Fama-MacBeth within each market (same
results, not reported)
▪ For global inference, compute means across markets, use cross-sectional t-test for inference
Regression results
The relationship between algorithmic trading and liquidity
More AT activity is associated with lower spreads and smaller temporary and permanent price impacts.
RQS RES RPI RRS AmihudMean coefficient on AT -0.013 -0.013 -0.002 -0.018 -0.009t-stat -9.1 -7.4 -0.7 -9.2 -6.5%positive 5% 8% 54% 5% 15%
Cross-sectional tests
How does the effect of AT differ for stocks with different characteristics?
Sort within each market according to each characteristic
Create dummies LOW and HIGH for low and high tercile
Include interactions with AT in regression models
The effect of AT on liquidity for different firm sizes
Solid colors indicate significance at the 5% level
More AT reduces liquidity in small stocks
-0.020
-0.010
0.000
0.010
0.020
0.030
0.040
0.050
0.060
0.070
RQS RES Amihud
Small cap
Mid cap
Large cap
The effect of AT on liquidity for different share prices
Pale colors indicate no significance at the 10% level
AT does not benefit low-priced stocks
According to Amihud, liquidity even declines significantly for low priced stocks
-0.015
-0.010
-0.005
0.000
0.005
0.010
0.015
0.020
RQS RES Amihud
Low price
Mid price
High price
The effect of AT on liquidity for different volatility levels
Volatility is SD of the 20 most rcent daily returns
AT only benefits liquidity of low and mid volatility firms
There is no AT effect on high volatility firms
-0.020
-0.015
-0.010
-0.005
0.000
0.005
Low volatility
Mid volatility
High volatility
The relationship between algorithmic trading and informational efficiency
More AT activity is associated with better informational efficiency.
|AR10| |AR30|Mean coefficient on AT -0.017 -0.006 t-stat -6.1 -4.5Percent positive 5% 23%
The effect of AT on efficiency in the cross-section
Pale colors indicate no significance at 10%
AT does not improve efficiency for small, low priced firms
-0.010
-0.008
-0.006
-0.004
-0.002
0.000
0.002
0.004
0.006
0.008
0.010
Market cap
Price Volatility
Small
Mid
High
The relationship between algorithmic trading and volatility
• More AT activity is associated with higher volatility.• We control for efficiency and liquidity of each stock.• Thus, the volatility increase does not represent “good” volatility that may arise with very efficient markets.
|ret| Ret^2PriceRang
eLn(Ret10_Var)
Ln(Ret30_Var)
Mean effect of AT
0.033 0.025 0.045 0.017 0.029
t-stat 7.4 6.6 7.9 3.4 4.9%positive 87% 85% 87% 67% 79%
The effect of AT on volatility in the cross-section
Effect on relative intraday price range (Results are robust for several other volatility measures)
0.0000.0100.0200.0300.0400.0500.0600.0700.0800.090
Market cap Price Volatility
Small
Mid
High
Volatility increases most for stocks that are small, low priced, or have volatile returns to begin with
AT = market making?
We do not observe what strategies algo traders use
Liquidity provision for mid and large cap stocks implies that at some AT supply liquidity
How resilient is this liquidity supply over time?
When is market making difficult?
MMs dislike one-sided order flow that moves price. E.g., consider sell imbalance: price moves down, MM is long, faces inventory losses
Will cut back on liquidity provision on such one-way days
Will cut back more when imbalance continues through the next day
Create a proxy for difficult MM days
Select all days where daily return has the same sign as on the previous day
Set HARD dummy to one on these days if the cumulative return exceeds one 20-day standard deviation
Interact with AT as before
How does the effect of AT change on difficult market-making days?
AT increases volatility and efficiency more
AT still increases liquidity, but significantly less than on regular days
AT significantly increases the information content of trades
If AT use MM strategies on average, they tend to resort to other strategies when market making is unusually difficult
Contrasts to traditional MM with affirmative obligations
-0.020 0.000 0.020 0.040
RES
Amihud
RPI
RRS
|AR30|
Price Range
Change on HARD days
Mean coefficient on AT
Which way does causality go?
Use co-location within each market as an instrument for AT
Estimate IV regression at the market level1. compute value-weighted daily averages for
each market2. estimate first stage regressions of AT on CL
dummy and market-day fixed effects3. estimate second-stage IV model with market
and day fixed effects using predicted values from (2)
IV estimation using co-location as an exogenous shock to ATLiquidityDependent
Estimate t
RQS -1.0087 -6.6RES -1.1546 -6.2RPI -0.9285 -7.4RRS -0.2718 -1.8Amihud -0.0001 -6.0
Efficiency
DependentEstimat
e t|AR10| -0.0006 -4.5|AR30| -0.0001 -0.5
Volatility PriceRange 0.0002 10.0|Ret| 0.0001 3.2Ret^2 0.0000 0.1
Results are largely unchanged with IV – evidence that causality originates with AT.
Summary
We contribute evidence from 39 countries to shed some light on how algo trading affects market quality
Algo trading improves liquidity and informational
efficiency worsens volatility, even when controlling
for efficiency and liquidity
Summary
But AT worsens liquidity of the smallest third
of firms in each market AT increases volatility the most for firms
that are small, low priced, and more volatile
on days when market making is more difficult, AT provides less liquidity, increases information content of trades, and increases volatility more.
Conclusions
Results are amazingly consistent across markets
Volatility increases with more algo trading – what exactly are the implications?
In assessing the current market structure, market observers should take into account that the effects of AT are not uniform.
Next steps
Will higher volatility discourage investors in the long run and thus increase firm’s cost of capital? Or affect its ability to raise new capital?
We will use our 11-year panel to assess longer-term effects of AT on
firm/country characteristics distribution of benefits and costs over
good/bad times and across stocks
Thank you for your attention.