Skip to content

Optimal Execution

Definition

Optimal execution is the problem of trading a large order (the "metaorder" or "parent order") over a finite time window while balancing three intrinsically competing objectives:

  • Completion — finish the order inside the window.
  • Market impact — minimise the adverse price move caused by your own trading.
  • Opportunity / timing cost — minimise slippage versus a benchmark (arrival price, VWAP, TWAP, or a custom reference) while avoiding exposure to adverse price moves.

It sits at the intersection of stochastic control, market microstructure, and production engineering.

The canonical tradeoff

Two extreme policies frame the problem:

Policy Completion risk Market impact Opportunity cost
Full market order up front None Maximal None (done in one shot)
Tiny passive limit-order slices over the full window High (may not fill) Minimal High (exposed to drift)

Good policies interpolate; that's where the modelling happens.

Core frameworks

Static schedules (classical era)

  • Bertsimas & Lo (1998) — DP formulation under linear impact → optimum is TWAP; the baseline for flat linear impact.
  • Almgren & Chriss (2001) — mean-variance formulation penalises cost uncertainty; produces an efficient frontier of execution paths analogous to Markowitz. Static schedule set once before trading. Remains the industry-default baseline.
  • VWAP-tracking — industry default; follows the intraday U-shaped volume profile. Introduced by Berkowitz et al.

Dynamic / online methods

  • Cartea & Jaimungal — mix of limit and market orders: post passively when ahead of schedule, cross the spread to catch up when behind. Rule-based online adjustment.
  • Busetti & Boyd (2005) — LQG stochastic control under VWAP benchmark, handles revealed-volume uncertainty via DP.
  • Reinforcement Learning (Nevmyvaka 2006; Hendricks–Wilcox; Moallemi–Wang; Li et al. hierarchical RL) — model-free execution policies trained directly on LOB data; action spaces vary from raw aggressiveness scalar to fractions of an Almgren-Chriss schedule.
  • Model Predictive Control — treats execution as approximate DP with one-step lookahead and a rollout base policy; solves a fast QP per decision step. See mpc-trade-execution.

Signal-aware frameworks

  • Lehalle-Neuman (2019) — incorporate a Markovian (e.g. Ornstein-Uhlenbeck) signal into the Gatheral-Schied-Slynko transient-impact problem. Closed-form optimal schedule for OU signal + exponential-decay impact, linear in both initial inventory and initial signal value. See signal-aware-optimal-execution and lehalle-neuman-signals-optimal-trading.
  • Cartea-Jaimungal (2015-) — continuous absolutely-continuous trading under instantaneous impact + bounded Markov signal. The \(\rho \to \infty\) limit of Lehalle-Neuman.
  • Key empirical input: order-book imbalance is an OU-like signal, actively used by HFT market makers to tilt their trading rate.

Slippage definitions (caveat lector)

There is no single "slippage". Common variants:

Slippage type Benchmark
Arrival slippage Mid-price at order arrival
Interval slippage Mid-price at each child-order placement
VWAP slippage Volume-weighted average price over the window
Implementation shortfall Arrival mid − execution average, including opportunity cost of unfilled shares

Papers and brokers use different conventions — always check which benchmark is being reported before comparing algorithms.

Connection to price impact

The cost of executing a metaorder is determined by the impact function. Two workhorses:

  • Linear impact on OFI (price-impact-order-book-events): \(\Delta p \propto \text{OFI}\) with slope inversely proportional to market depth.
  • Square-root impact on volume: \(\Delta p \propto \sqrt{Q}\) for trade size \(Q\). Cont et al. derive this from the linear OFI model via a scaling argument; Almgren-Chriss-style models take it as a primitive.
  • Power-law decay kernels (reality-gap-lob-simulation; Bacry et al.) — accumulate signed flow with a memory kernel so that impact decays toward mean reversion after execution, a key feature for realistic simulation.

See price-impact for the full picture.

Open questions

  • How much of the Almgren-Chriss efficient frontier survives once realistic, non-linear, non-stationary impact is used?
  • What is the right base policy for MPC rollout, and can it be learned rather than handcrafted?
  • Are RL policies trained in simulation actually robust in live trading, or do they overfit to the simulator's impact model?
  • Cross-venue routing (fragmented markets, dark pools) introduces a much richer action space — how do classical frameworks extend?
  • ~~How should executed volume be sized against short-horizon order-flow-imbalance signals?~~ — largely addressed by signal-aware-optimal-execution for OU-type signals; open for non-OU signals and for cross-asset OFI inputs.
  • Does signal-aware execution admit transaction-triggered price manipulation? Lehalle-Neuman show non-monotone optimal strategies exist; the conditions on impact kernel and signal that rule out manipulation are open.
  • How to handle signal-aware execution's time inconsistency under transient impact? Re-planning vs commit-at-\(t=0\) vs collapse to CJ limit — no principled resolution.

Connections