Reader question: If a fund must sell a large block by the end of the day, why not trade it all immediately — or, at the other extreme, split it into tiny pieces and trade as slowly as possible?
Because those two extremes solve different problems. Trading immediately reduces the risk that the market moves against the remaining position, but it can consume liquidity aggressively and create large market-impact cost. Trading very slowly can reduce temporary impact, but leaves the position exposed to random price movement for longer. The Almgren–Chriss optimal-execution framework turns that tension into a mathematical optimisation problem.
The basic model chooses a liquidation trajectory that minimises a weighted combination of expected execution cost and execution-cost variance. Temporary impact penalises trading too fast; volatility risk penalises trading too slowly. Under the model’s linear-impact assumptions, the solution is explicit and forms an efficient frontier of schedules.
What this page owns — and what it does not
This page owns the computational question:
large parent order + horizon + volatility + impact parameters + risk aversion → optimal deterministic execution trajectory.
It does not replace exchange matching algorithms, which decide how submitted orders interact with the market; market-risk VaR and Expected Shortfall; or volatility forecasting. Those are adjacent owners.
This is mathematical market-microstructure education. It is not a recommendation to execute any real security order in any particular way.
The parent order becomes an inventory path
Suppose a trader begins with X shares to sell over a horizon from time 0 to T.
Let:
- x(t) = shares remaining at time t;
- x(0)=X;
- x(T)=0.
The trading rate is approximately:
v(t) = −dx(t)/dt.
Every feasible liquidation schedule is therefore a path from X down to zero.
The two forces pulling the schedule in opposite directions
The Almgren–Chriss model separates price impact into permanent and temporary components.
A stylised permanent-impact function is:
g(v) = γv.
A stylised temporary-impact function is:
h(v) = ε·sign(v) + ηv.
Here:
- γ controls lasting impact;
- η controls temporary rate-dependent impact;
- ε can represent a spread-like fixed cost.
Temporary impact makes fast trading expensive. The market-risk term makes slow trading risky.
The expected-cost term
In the linear discrete model, a useful structural form for expected cost is:
E[C] = permanent term + spread term + temporary-impact term.
The temporary component is proportional to something like:
(η/τ) Σ nk2,
where nk is the number of shares traded in interval k and τ is the interval length.
The square matters. Splitting one large slice into two equal smaller slices reduces the sum of squared trade sizes, which is why convex temporary impact rewards spreading the order through time.
The timing-risk term
Let the unaffected asset price evolve with volatility σ. While shares remain unexecuted, their value is exposed to market movement.
In the discrete model, execution-cost variance has the structural form:
Var[C] = σ² Σ τ xk².
Large inventory held for a long time creates more timing risk than small inventory held briefly.
This term creates the opposite incentive from temporary impact: it rewards faster liquidation.
The optimisation objective
The classic mean–variance objective is:
minimise E[C] + λ Var[C].
The parameter λ ≥ 0 is risk aversion.
- λ near zero: expected impact dominates; the schedule becomes relatively slow and even.
- larger λ: timing risk matters more; the schedule becomes increasingly front-loaded.
The family of solutions across different λ values forms an efficient frontier: no other feasible schedule can achieve lower expected cost for the same risk under the model.
Why the risk-neutral solution approaches TWAP
Under constant liquidity and linear temporary impact, setting λ=0 removes the timing-risk penalty. The problem then minimises the quadratic temporary-impact cost subject to the requirement that all shares are sold by time T.
Convexity makes equal-sized slices optimal:
n1 = n2 = … = nN.
That is the mathematical core of a constant-rate TWAP-like liquidation.
This result depends on constant impact parameters and no intraday volume profile. If liquidity varies strongly through the day, equal clock-time slices need not be cost-minimising.
The closed-form continuous trajectory
Under the simplified linear-impact continuous approximation, the optimal remaining inventory has the hyperbolic-sine form:
x(t) = X · sinh[κ(T−t)] / sinh(κT).
A simplified continuous expression for the urgency parameter is:
κ ≈ √(λσ²/η).
This equation exposes the mechanism directly:
- higher volatility σ → larger κ → faster liquidation;
- higher risk aversion λ → larger κ → faster liquidation;
- higher temporary-impact cost η → smaller κ → slower liquidation.
Characteristic liquidation time
The quantity 1/κ behaves like a characteristic execution timescale.
If κT is small, the hyperbolic-sine path is close to a straight line. If κT is large, inventory decays rapidly near the start of the horizon.
This gives a compact diagnostic: the dimensionless product κT says more about schedule shape than κ alone.
A simple schedule comparison
Suppose a 1,000,000-share order must be liquidated in 10 equal time intervals.
A risk-neutral constant-rate schedule trades 100,000 shares per interval.
A more risk-averse schedule might trade larger early slices such as 170,000, 150,000, 130,000 and then taper. Its expected temporary-impact cost rises because squared trade sizes are larger, but its timing-risk variance falls because fewer shares remain exposed later in the horizon.
The model does not say the aggressive schedule is “better” in isolation. It says the choice lies on an explicit cost–risk trade-off.
Implementation shortfall as the observable cost target
Execution is commonly benchmarked against an arrival or decision price.
A stylised sell-order implementation shortfall is:
IS = initial mark-to-market value − realised sale proceeds.
Normalised implementation shortfall can be expressed in basis points relative to the initial notional.
The model tries to understand the expected value and variance of that cost under specified price and impact dynamics.
Permanent impact is schedule-insensitive only under the linear model
In the classic linear specification, the aggregate permanent-impact cost is largely determined by total quantity rather than how the order is sliced.
That makes temporary impact the main schedule-sensitive expected-cost term.
This is a model result, not a universal market law. If permanent impact is nonlinear, decays, interacts with information or depends on participation rate, the schedule can affect it more strongly.
Why a static schedule can be optimal inside the model
The unaffected price is modelled as a random walk with independent increments and no predictive signal. Under those assumptions, a price move during execution contains no information about the expected direction of the next price move.
Therefore reacting to realised price noise does not improve the expected objective merely because the price happened to move.
That is why a precomputed schedule can be optimal in the basic framework.
If the trader has predictive alpha, changing liquidity, volume shocks or state-dependent impact, the argument breaks and adaptive control can be useful.
Inputs and outputs
An Almgren–Chriss execution engine can require:
- parent-order quantity X;
- execution horizon T;
- number and length of execution intervals;
- price volatility σ;
- temporary-impact parameter η;
- permanent-impact parameter γ;
- spread/fixed-cost parameter ε where used;
- risk-aversion parameter λ;
- price and quantity units;
- optional participation, minimum-slice or liquidity constraints.
Outputs can include:
- shares remaining at each time step;
- child-order quantity per interval;
- expected temporary and permanent impact;
- expected total implementation cost;
- execution-cost variance or standard deviation;
- efficient-frontier coordinates;
- sensitivity to volatility, impact and risk aversion.
Evidence polarity: what supports confidence?
Evidence for a correct implementation includes recovery of TWAP as risk aversion approaches zero under constant-liquidity assumptions, more front-loaded paths as λ or σ increases, slower paths as η rises, expected cost and variance matching independent closed-form calculations, and Monte Carlo simulations agreeing with the analytical mean and variance.
Evidence against confidence includes schedules that become slower when volatility rises, inventory that does not finish at zero, negative child-order sizes in a pure liquidation problem without permission to buy back, permanent impact changing with slice pattern when the implemented model says it should be schedule-invariant, or simulated execution costs that systematically disagree with the analytical objective.
Counterexample: lowest expected impact is not necessarily lowest execution risk
A very slow schedule can minimise temporary impact but leave most of the order exposed to price moves for much of the day.
Optimising expected cost alone can therefore choose a schedule with intolerably high uncertainty.
Counterexample: fastest execution is not automatically safest economically
Dumping the entire order at once nearly eliminates timing risk but can produce extreme temporary impact and poor realised prices.
Reducing one risk can increase another cost sharply.
Counterexample: linear market impact can fail empirically
The original closed form uses linear impact for tractability. Empirical market-impact research often finds concave or square-root-like scaling over important size ranges.
If true impact is concave while the model assumes linear impact, the schedule and predicted cost can be materially miscalibrated.
This is a core falsifier: the execution optimiser cannot be more accurate than its impact model.
Counterexample: constant liquidity ignores the trading day
Real market volume and spreads often vary strongly by time of day. A constant-η model treats 10:00 and 12:30 as economically identical if the other state variables match.
A volume-profile-aware execution model can outperform the simple constant-liquidity schedule when intraday liquidity is predictable.
Counterexample: market prices may contain information
If the trader has a short-horizon predictive signal, a static schedule that ignores price moves can be suboptimal.
The basic model’s “do not adapt to random price moves” conclusion is valid only when those moves are genuinely uninformative about future expected returns.
Weak links in implementation
Impact calibration drift. η and γ are estimated from an old liquidity regime.
Volatility mismatch. Daily volatility is inserted into a model expecting interval volatility without scaling.
Unit inconsistency. Shares, dollars, percentages and basis points are mixed.
Discrete/continuous mismatch. A continuous κ formula is used inside a coarse discrete schedule without checking approximation error.
Numerical overflow. Direct evaluation of very large hyperbolic sines loses stability.
Wrong boundary conditions. Inventory does not begin at X and finish exactly at zero.
Linear-impact overconfidence. Model cost is presented as truth despite nonlinear observed impact.
Liquidity blindness. Participation limits and intraday capacity are omitted.
Diagnostics: how to test the engine
- λ→0 test: recover an approximately equal-slice schedule under constant liquidity.
- high-λ test: verify the trajectory becomes more front-loaded.
- σ shock test: higher volatility should accelerate liquidation.
- η shock test: higher temporary-impact cost should slow the schedule.
- boundary test: require x(0)=X and x(T)=0 exactly.
- quantity conservation: sum of child-order quantities equals the parent order.
- frontier test: as risk aversion rises, expected cost should generally rise while timing-risk variance falls.
- Monte Carlo test: simulate unaffected-price paths and compare sample mean/variance with analytical values.
- impact-misspecification test: score schedules under nonlinear “true” impact after optimising under linear impact.
- TWAP benchmark: compare the optimised schedule with a simple equal-slice schedule under both assumed and stressed parameters.
What would falsify confidence?
Confidence should be withdrawn if simulated and analytical execution-cost moments fail to reconcile; if the optimiser violates inventory constraints; if small parameter changes create unexplained schedule discontinuities; if out-of-sample market impact systematically follows a different functional form; or if the optimised schedule persistently underperforms a simple benchmark such as TWAP after realistic transaction-cost and liquidity controls.
Alternatives and extensions
VWAP-style execution follows expected market volume rather than equal clock time.
Participation algorithms constrain the trade rate to a fraction of observed volume.
Adaptive stochastic-control models update the path when liquidity, volatility or alpha changes.
Transient-impact models allow price impact to decay rather than divide cleanly into permanent and temporary components.
Limit-order-book models represent queue depth and replenishment explicitly.
The Almgren–Chriss framework remains valuable because it makes the cost–risk trade-off transparent enough to test.
How this connects to the surrounding knowledge estate
Volatility forecasting supplies one major urgency input. Market matching determines how the child orders actually interact with the venue. Exact money arithmetic keeps cost attribution numerically consistent. Market-risk measurement asks a broader portfolio-loss question; Almgren–Chriss asks how much of that exposure should remain unexecuted through time.
Verification and update triggers
Preserve the model version, impact specification, calibration window, volatility convention, time discretisation, units, risk-aversion policy and benchmark schedule. Revalidate after liquidity-regime shifts, exchange-microstructure changes, large changes in tick size or spread, volatility jumps, evidence of impact-model nonlinearity, or persistent out-of-sample execution-cost misses.
Primary and high-quality references
- Robert Almgren and Neil Chriss, Optimal Execution of Portfolio Transactions, Journal of Risk, the foundational efficient-frontier execution model.
- Robert Almgren, Optimal Execution with Nonlinear Impact Functions and Trading-Enhanced Risk, extending the framework beyond the simplest linear-impact specification.
- Robert Almgren et al., Direct Estimation of Equity Market Impact, on empirical impact calibration.
- U.S. Securities and Exchange Commission, market-structure and execution-quality materials, for the broader execution-quality context.
- FINRA, best-execution and order-handling materials, for the distinction between mathematical scheduling and actual execution obligations.
Educational boundary: This article explains optimal-execution mathematics. It does not recommend an execution schedule, venue, security transaction or trading strategy for any reader.