Does Algorithmic Trading Improve Liquidity?

Transcription

1 Does Algorithmic Trading Improve Liquidity? Terry Hendershott 1 Charles M. Jones 2 Albert J. Menkveld 3 1 Haas School of Business, UC Berkeley 2 Graduate School of Business, Columbia University 3 Tinbergen Institute, VU University Amsterdam WFA, June 2008

2 Time trend bid-ask spread (ctd) Spreads for the long run Relative bid-ask spread Dow Jones stocks Figure 1. Bid-ask spreads on Dow Jones stocks (all stocks , (all DJ stocks DJIA , stocks DJIA stocks ) 1929-present) 1.60% 1.40% 1.20% Proportional spread 1.00% 0.80% 0.60% 0.40% 0.20% 0.00%

3 Time trend bid-ask spread (ctd)!"#$%"%&'()*+%'((,%"-%+*&*#-%./)-(+0 Figure NYSE 1. NYSE value-weighted average proportional effective spreads 0.25% 0.20% Proportional effective spread 0.15% 0.10% 0.05% 0.00%

4 Institutional trading How do institutions trade prior to algorithms? To buy 100,000 IBM shares, they hire a broker-dealer to take down or shop a block hire NYSE floor broker who uses judgement to slowly work the order Broker-dealers now offer algos that minimize price concession through a dynamic trading strategy that optimizes over price, quantity, time, and venue And, broker-dealers and hedge funds supply liquidity with algos (e.g. D.E. Shaw, Getco,... )

5 Related literature IO of liquidity supply competition: Kyle (1985), Biais, Martimort, and Rochet (2000) Free trading option of limit orders (Copeland and Galai (1983)) monitoring public information flow is costly (Foucault, Roëll, and Sandas (2003)) AT may raise costs of non-at limit orders (Rock (1990)) Optimal execution of large orders (Keim and Madhavan (1995), Bertsimas and Lo (1998), Almgren and Chriss (2000)) market vs. limit, aggressiveness (Harris (1998), Griffiths, Smith, Turnbull, and White (2000), Lo, MacKinlay, and Zhang (2002), Boehmer, Saar, and Yu (2005), Hasbrouck and Saar (2007), Obizhaeva and Wang (2005))

6 What do we do? We measure algo trading through normalized (electronic) message traffic at the NYSE message traffic is electronic order submissions, cancels, and trade reports Panel regressions associate time-series increases in algo trading with more liquid markets we exploit the exogenous, staggered introduction of autoquote at the NYSE as an instrument to establish causality

7 250 messages it (#electronic messages per minute i.e. proxy for algorithmic activity (/minute)) Q1 Q2 Q3 Q4 Q5 95% conf. interval

8 algo_trad it (dollar volume per electronic message times ( 1) to proxy for algorithmic trading ($100)) Q1 Q3 Q5 Q2 Q4 95% conf. interval

9 Autoquote Decimals in 2001 shrink inside quote depth In October 2002 NYSE proposes liquidity quote firm bid and offer for substantial size (> 15, 000 shares) Autoquote is proposed simultaneously to free up the specialist to concentrate on the liquidity quote specialists had been manually disseminating the inside quote software would now autoquote any change to book Liquidity quote delayed, autoquote immediate Autoquote is important for AT immediate feedback about terms of trade algo liquidity suppliers see abnormally wide inside quote algo liquidity demanders access quote more quickly

10 200 #stocks that trade in Autoquote Q1 Q3 Q5 Q2 Q /27/03 last stocks to Autoquote /29/03 first stocks to Autoquote /2/02 start of sample 7/31/03 end of sample

11 analysis. It is based on daily observations in the period when autoquote was phased in, i.e. December 2, 2003, through July 31, For variable definitions, we refer to Table 1. We exploit the exogenous Autoquote dummy as instrument for algo trad autoquote dummy (0 before the autoquote introduction, 1 after) to instrument for algo trad it it in order to identify causality from algo trad it to our liquidity measures. In the IV estimation, we exclude identification off of a time trend (by adding time dummies) and thus solely rely on the nonsynchronous introduction of autoquote (see Figure 5). Before we report the IV estimation results in subsequent tables, this table reports correlations between the instrument (auto quote it) and the endogenous variable (algo trad it) after removing the time trend. messa algo share volatility it ket cap it 1/price it ln mar ges it trad it turnover it Panel A: Overall, between, and within correlation after removing the time trend auto quote it ρ(overall) 0.15* -0.05* 0.02* 0.03* 0.02* 0.10* ρ(between) 0.23* -0.16* * * ρ(within) 0.08* 0.03* -0.01* * -0.01* Panel B: Within correlation by quintile after removing the time trend auto quote it Q1 ρ(within) 0.15* 0.03* 0.01* * -0.03* auto quote it Q2 ρ(within) 0.03* 0.04* -0.01* * 0.01* auto quote it Q3 ρ(within) 0.05* 0.03* * auto quote it Q4 ρ(within) 0.01* auto quote it Q5 ρ(within) * -0.02* * -0.04* : Based on the time means i.e. x i = 1 T T t=1 xi,t. b : Based on the deviations from time means i.e. x i,t = xi,t xi. *: Significant at a 95% level. F -tests reject null that instruments do not enter first-stage regression for all our IV regressions

12 IV regression including T/O, volatility, price, and size This table regresses various measures of the (half) spread on our algorithmic trading proxy. It autoquote was phased in, i.e. December 2, 2003, through July 31, We use the exogenous no for the endogenous algo trad it to identify causality from algorithmic trading to liquidity. We est L it = α i + γ t + βa it + δx it + ε it where L it is a spread measure for stock i on day t, A it is the algorithmic trading measure algo tra share turnover, L it volatility, = α i + 1/price, γ t + and βalog it market + δxcap. it + Weε always it include fixed effects and time all explanatory variables, except that we replace algo trad it with auto quote it. We regress by q that are robust to general cross-section and time-series heteroskedasticity and within-group auto : We report the Dickey-Fuller test statistic based on the residuals in order to diagnose nonstati Coefficient on algo trad it Coefficients on Q1 Q2 Q3 Q4 Q5 share vola turnover it tility Panel A: quoted spread, quoted depth, and effective spread qspread it -0.52** -0.42** ** 0.90* (-3.23) (-2.21) (-1.44) (-0.05) (1.22) (-3.01) (9.70 qdepth it -3.47** (-2.50) (-1.16) (-1.07) (0.39) (0.19) (-0.64) (-1.87 espread it -0.18** -0.32** ** 0.69* (-2.65) (-2.23) (-1.56) (-0.42) (1.16) (-2.32) (9.51 Panel B: spread decompositions rspread it 0.35** 0.76** 1.03** * -1.06* (3.53) (3.97) (2.06) (0.46) (1.36) (1.92) (-2.15 adv selection it -0.53** -1.07** -1.39** ** 1.75* (-3.57) (-4.08) (-2.06) (-0.47) (-1.33) (-2.24) (3.29 #observations: 1082*167 (stock*day) */**: Significant at a 95%/99% level. a : We use quintile-specific coefficients for the control variables and time dummies. For brevity, cap-weighted coefficient for the control variables and its t-statistic.

13 2.5 stdev_tradecorr_comp it (stdev of trade correlated component of eff. price innovations cf. Hasbrouck (1991a,199 Q1 Q2 Q3 Q4 Q5 95% conf. interval

14 3.5 stdev_nontradecorr_comp it (stdev of non trade correlated component of eff. price innovations cf. Hasbrouck ( Q1 Q2 Q3 Q4 Q5 95% conf. interval

15 order persistence in decomposing the bid-ask spread. It identifies a fixed (transitory) component (LS (LSB95 adv sel it), and a component due to order persistence (LSB95 order persist it) (see Section details). As the LSB decomposition limits persistence to that of an AR(1) process, we also estimate a H of the trade-related (stdev tradecorr comp it) and trade-unrelated (stdev nontradecorr comp it) comp transactions (see Section and Hasbrouck (1991a, 1991b) for details). For the regressions, we use autoquote to instrument for the endogenous algo trad it to identify causality from algorithmic trading estimate M it = α i + γ t + βa it + δx it + ε it IV regression for LSB and Hasbrouck decompositions M it = α i + γ t + βa it + δx it + ε it where M it is a LSB or Hasbrouck component for stock i on day t, A it is the algorithmic trading me including share turnover, volatility, 1/price, and log market cap. We always include fixed effects and ti in the Hasbrouck component regressions. We regress by quintile and report t-values based on standar : We use quintile-specific coefficients for the control variables and time dummies. For brevity, we o and time-series heteroskedasticity and within-group autocorrelation (see Arellano and Bond (1991)). Coefficient on algo trad it Coefficien Q1 Q2 Q3 Q4 Q5 share v turnover it t Panel A: Lin, Sanger, and Booth (1995) LSB95 fixed it 0.26** 0.59** 0.69** ** (3.63) (4.16) (2.26) (0.46) (1.36) (2.07) ( LSB95 adv sel it -0.26** -0.61** -0.84** * (-3.46) (-3.80) (-2.14) (-0.46) (-1.32) (-1.85) LSB95 order persist it -0.18** -0.30** ** 0 (-3.06) (-3.10) (-1.60) (0.28) (1.21) (-2.33) Panel B: Hasbrouck decomposition stdev tradecorr comp it -0.22** -0.26** -0.30* ** (-2.62) (-3.08) (-1.69) (-0.30) (-2.73) stdev nontradecorr comp it 0.13** 0.13** (2.48) (2.36) (1.47) (0.28) (1.12) #observations: 1082*167 (stock*day) */**: Significant at a 95%/99% level. cap-weighted coefficient for the control variables and its t-statistic.

16 Interpretation: generalized Roll model i.i.d. innovation in efficient price in each of two periods, m t = m t 1 + w t, with w t { ε, +ε} equally likely 1. At t = 0, risk-neutral humans submit a bid and ask quote and, given full competition, the first one arriving bids her reservation price. 2. At t = 1, humans can buy the information w 1 at cost c. If bought, they can submit a new limit order. 3. At t = 2, two informed liquidity demanders arrive, one with a positive private value associated with a trade, +θ, the other with a negative private value, θ. Assume 1. 2c > θ i.e. cost of observing for humans is sufficiently high ( quotes become stale ) 2. ε > θ i.e. large innovations prevent simultaneous transaction by both liquidity demanders (unimportant)

17 A m Interpretation: generalized Roll model uu 0 A 1 2 (ctd) We assume that 2c > θ, i.e., the cost of observing information ε for humans is sufficiently high that they do not updatem their u 1 quotes. The technical assumption 2ɛ > θ is also required so that only one of the two arriving liquidity demanders transacts at t = 2. Humans only m 0 m ud 2 m d 1 A 0 A 1 B 0 B 1 m u 1 m uu 2 m dd 2 ε The important result is that the humans do not buy m m ud the w 1 information and update the quote 0 at t = 1, since they have to quote 2 so far away from the efficient price to make up for c that m d neither liquidity demander will transact at that quote (as 2c > θ). For example, 1 if w 1 is positive, they could put in a new bid of m u 1 ε 2c, but, even if w 2 is negative, the liquidity seller s private value is too small to consume B m dd 0 the bid quote. B 1 2 In this case, we have the following outcomes: probability state efficient transaction There are four equally likely paths through price the binomial price tree: uu, ud, du, and dd, where u represents a positive.25 increment uu of ɛ mto uu 2 the fundamental m uu 2 value and d is a negative.50 ud and du m ud 2 = m du 2 no transaction increment. In equilibrium,.25 humansdd do not buy m dd the w 1 information and update the quote at 2 m dd 2 t = 1, since they have to quote so far away from the efficient price to make up for c that And, estimating the Hasbrouck model gives: neither liquidity demander will transact at that quote (as 2c > θ). Given that they do not P acquire the w 1 information, a P humans b = 2εu, u {1, 0, 1} is the trade sign, (2) at t = 1 public information protect themselves does notbyenter settingquotes the bid price equal to where P a (P b) is the transaction price after (before) the period. Clearly, there welfare is no public loss information due component to possible and allunrealized information comes private from thevalue order flow. 1 21

18 and there is no public information component. We introduce a machine that can buy the w 1 information at zero cost (c = 0). All above is unchanged except that the machine issues a new quote at t = 1. To illustrate the idea, suppose w 1 > 0. The machine, if alone, enjoys full market power and rationally puts in a new bid of m The model with algorithmic trading u 1 ε θ as depicted in the following Interpretation: generalized Roll model (ctd) Introduce graph: an algo that buys information at zero cost A 0 A 1 A 0 A 1 m uu m uu 2 2 m u 1 m 0 m 0 m u 1 B 1 B 1 θ θ m ud m ud 2 2 B 0 B 0 In this case, we have the following outcomes: probability state efficient transaction Now we introduce an algorithm that can pricebuy the wprice 1 information at zero cost (c = 0)..25 uu m The results at t = 0 remain unchanged. At t = uu 2 m 1, the algorithm uu 2 optimally issues a new quote..25 ud m ud 2 = m du 2 m ud 2 To illustrate the idea, suppose w 1 > 0. The algorithm knows θ.50 du m du 2 = m ud 2 m du 2 + θ that it is the only liquidity provider in possession of w.25 1, and so it dd puts in m dd 2a new bid m equal dd 2 to m 0 θ. If w 2 > 0 as well, And, estimating the Hasbrouck model gives: then a transaction takes place at the original ask of m 0 + 2ɛ. If w 2 < 0, then a liquidity demanderatwill t P= a hit 1P the b public = algorithm s w 1 + εuinformation I [m ud 2 bid.,mdu 2 ] θsign(w This bid enters 1), isu below {1, quotes, 1} theis efficient thebut trademidquote price, sign, (3) so there will eventuallybecomes where a reversal, the noisy first and two terms there measure contain is a temporary information of truecomponent and value the last term in prices. is a transient Contrariwise, if price change (reflecting the machine s market rent). We see that the variance of w 1 < 0, the noalgorithm the unrealized efficientplaces price innovations private a new askis value at unchanged, m 0 + θ, but which the decomposition is traded with nowif assigns it turns out that w 2 > 0. half the value to public information (due to the quote update) and the other half

19 Interpretation: generalized Roll model (ctd) Efficient price is revealed without trades i.e. public information enters quotes without trades Revenue to liquidity suppliers is positive Also matches other findings: more frequent trades, narrower quotes Note: model assumes that algo competition is less intense that human competition

20 Conclusion 1. Panel regressions time-series increases in algo trading correlate with liquidity improvement 2. Staggered introduction of structural change (autoquote) as an instrument confirms algo trading lowers trading cost and increases price informativeness 3. Surprisingly, revenues to liquidity suppliers increase with algo trading. Market power for some period after introduction?

21 Does Algorithmic Trading Improve Liquidity? Terry Hendershott 1 Charles M. Jones 2 Albert J. Menkveld 3 1 Haas School of Business, UC Berkeley 2 Graduate School of Business, Columbia University 3 Tinbergen Institute, VU University Amsterdam WFA, June 2008

22 Almgren, R. and N. Chriss (2000). Optimal execution of portfolio transactions. Journal of Risk 3, Bertsimas, D. and A. Lo (1998). Optimal control of execution costs. Journal of Financial Markets 1, Biais, B., D. Martimort, and J. Rochet (2000). Competing mechanisms in a common value environment. Econometrica 68, Boehmer, E., G. Saar, and L. Yu (2005). Lifting the veil: An analysis of pre-trade transparency at the nyse. Journal of Finance 60, Copeland, T. and D. Galai (1983). Information effects on the bid-ask spread. Journal of Finance 38, Foucault, T., A. Roëll, and P. Sandas (2003). Market making with costly monitoring: An analysis of the soes controversy. Review of Financial Studies 16, Griffiths, M., B. Smith, D. Turnbull, and R. White (2000). The costs and determinants of order aggressiveness. Journal of Financial Economics 56, Harris, L. (1998). Optimal dynamic order submission strategies in some stylized trading problems. Financial Markets, Institutions, and Instruments 7, Hasbrouck, J. (1991a). Measuring the information content of stock trades. Journal of Finance 46,

23 Hasbrouck, J. (1991b). The summary informativeness of stock trades: An econometric analysis. Review of Financial Studies 4, Hasbrouck, J. and G. Saar (2007). Technology and liquidity provision: The blurring of traditional definitions. Technical report, New York University. Keim, D. and A. Madhavan (1995). Anatomy of the trading process: Empirical evidence on the behavior of institutional traders. Journal of Financial Economics 37, Kyle, A. (1985). Continuous auctions and insider trading. Econometrica 53, Lin, J., G. Sanger, and G. Booth (1995). Trade size and components of the bid-ask spread. Review of Financial Studies 8, Lo, W., A. MacKinlay, and J. Zhang (2002). Econometric models of limit-order executions. Journal of Financial Economics 65, Obizhaeva, A. and J. Wang (2005). Optimal trading strategy and supply/demand dynamics. Technical report, MIT. Rock, K. (1990). The specialist s order book and price anomalies. Technical report, Harvard University.