Algorithmic trading Equilibrium, efficiency & stability Presentation prepared for the conference Market Microstructure: Confronting many viewpoints Institut Louis Bachelier Décembre 2010 Bruno Biais Toulouse School of Economics
Outline 1) Definition 2) Benefits of algos 3) Costs of algos 4) Equilibrium investment in algos 5) Policy implication 6) Conclusion/summary
1) Definition Algorithms use computers to => collect & process information => reach investment & trading decisions => route orders without human intervention Computers faster than humans => Algorithms // High Frequency Trading Hedge funds, investment banks, algo trading firms => HFT, prop trading brokers, fund managers => order execution > 50 % of trades in major stock markets
2) Benefits of algos
2.1) Algos mitigate cognition limits Traders must analyze risk exposure, gross positions, net aggregate, compliance with regulation & limits Especially tough when market hit by shock While humans collect & process this info, can t make trading decision: algos can
Trading & liquidity with limited cognition Biais, Hombert & Weill, 2010 price Ask Successive limit sell orders undercut each other Algo traders who bought unwind position as their limit orders to sell get filled Transaction price Liquidity shock time Algos trigger buy orders that accomodate liquidity demand of investors hit by shock
Consistent with stylized facts Order splitting Short--term momentum No withdrawal after shock (Brogaard, 2010: Nasdaq) Benefit from price reversals (Brogaard, 2010) Provide liquidity when scarce & rewarded (Hendershott Riordan, 2010: Xetra)
2.2) Algos reduce search costs Fragmentation need to search for trading opportunities, compare prices, etc Algos reduce search costs & increase search speed More trading opportunities can be identified and gains from trade reaped
3) Costs of algos
3.1) Adverse selection Computers better than humans at collecting & aggregating info from + colocation => asymmetric information between slow & fast traders Hendershott Riordan (2010), Brogaard (2010): Algos have > permanent price impact Algos lead price discovery
Equilibrium algorithmic trading Biais, Foucault & Moinas, 2010 Asset common value θ + ε Institution s private value v + δ v 1/2 1/2 θ ε v δ
Fast & slow traders α Fast Observes v & finds counterparty Slow ρ Finds counterparty Or not 2 countervailing effects of algos on welfare: trading opportunities information asymmetry
Extensive form of trading game Nature draws v & trader s type (+/- δ, fast/ slow) Trader observes private value and, if fast, v If trader finds counterparty, she places order (buy or sell 1unit) Price = E(v order): A if order = +1, B if -1 v realized
Equilibrium strategies & prices If δ > ε, mild adverse selection: first best If δ < ε, severe adverse selection: M1: fast mix M2: fast mix θ θ + ε δ θ + δ θ + ε A P1 : Fast buy if + ε Slow buy if + δ P2 : Fast buy iff + ε + δ Slow buy if + δ P3: Fast buy iff + ε + δ Slow don t trade
Multiple equilibria M2 P3: only fast with good news and +δ buy P1 M1 P2 Slow don t trade Slow with +δ buy «market breakdown» α 0 ρ(ε δ) ρ(ε δ) ρδ ρ(ε δ)+δ ρ(ε δ)+δ/2 ρδ+(ε δ)/2 1
Algos create adverse selection & possibly market breakdown Fast traders observe information before others => their profits = slow traders losses Increase in algo trading => increase adverse selection costs => slow traders evicted from market Algos = negative externality for slow traders
Empirical evidence consistent with algos evicting slow traders Jovanovic & Menkveld (2010): Fast algo trader enters Chi-X (Dutch stocks) => Drop in volume Due to slow traders pulling away from market?
Why multiplicity? For given fraction of algorithmic investors, if large fraction of trades expected to stem from fast traders => large adverse selection expected => slow stay away from market => self-fulfilling expectation Multiple equilibria in trading game: => liquid vertuous circle <= illiquid vicious circle
Flash crash: May 6, 2010 CFTC SEC Report Investors worry about Greece, volat rises Morning: depth of book declines on bid side 2:32: Large investor triggers algo sell program: 2:32-2:45: sells 35,000 E-mini S&P contracts => 5% drop in S&P 2:45: circuit breaker triggered on CME 3:00: most prices revert back to normal
A (partial) theory of the flash crash BFM 2010 Before May: Market participants coordinate on liquid equilibrium: algos programmed to play that May 6: Growing tension (Greece) => animal spirits change: humans switch to bad equilibrium Algos don t have animal spirits: => continue to play as before => sell aggressively => large price drop Corrected when markets understand they overestimated the information content of trades
3.2) Operational & systemic risk Operational risk Computer can go crazy: bug in program, connections breakdown, unexpected event Humans too can err, but algos leveraged by volume => can amplify human mistake (error in parametrisation of program) Problem with one (or a few) algo(s) could propagate to whole market (systemic risk?)
Revisiting the flash crash Systemic risk Initial selling pressure triggered by algo HFTs first bought then sold back => high volume & price drop Initial algo reacted by increasing rate at which it was feeding orders => further price drop HFTs reacted by pulling out of market (to investigate what was happening) Dangerous spiral
4) Investing in algorithmic trading: Equilibrium and welfare Biais, Foucault, Moinas 2010 Fixed investment in hardware, code, colocation. Algos increase profitability of each share traded Large institutions reap these benefits on large scale: profitable for them to invest in algos Small institutions don t invest in algos => bear adverse selection costs when matched with superiorly informed fast traders
Algo trading is contagious BFM 2010 If I expect majority of others to invest in algos, I anticipate severe adverse selection costs if I remain slow => I also invest in algos (self-fulfilling expectation) But if I expect others remain slow, no adverse selection cost if I remain slow (self-fulfilling again) Multiple levels of equilibrium investment in algos.
Overinvestment in algos BFM 2010 Equilibrium investment in algos: i) to reap gains from trade, ii) to make profits against slow traders i) Socially useful. Not ii)! Equilibrium investment > Social optimum (equilibrium differs from optimum because of externality)
5) Tentative policy implications A role for exchanges Monitor algos and disconnect or investigate them if fishy (e.g. NYSE Euronext this summer) Circuit breakers to stop spirals & give market time to think when equilibrium miscoordination
Fragmentation can be dangerous Multiple trading venues : hard to aggregate info, dilute responsibility => Exchanges less effective at monitoring Multiple trading venues: circuit breaker can be inefficient if applied in one market and not the other(s) Need for integrated circuit breaker and for market surveillance across venues
Pigovian tax? If colocation => negative externality on slow traders, excess investment in algo trading: optimal to tax colocation If algo trading => systemic risk for others: again optimal to tax Tax proceeds could be invested in insurance/stability fund: available to bear costs arising in case of operational risk
Competition policy? Fixed costs associated with algos => market concentration => less competition Monitor market: investigate if excessive concentration Policy moves to level playing field: minimum latency?
6) Conclusion / summary Benefits: mitigate cognition limits-search frictions Costs: information asymmetry imperfect competition systemic & operational risk Investment in algos factors in benefits not costs: excessive Policy: Integrated market surveillance Integrated circuit breakers Pigovian tax Competition policy