SignalTrace

Optimal Execution

John Montgomery5 min read

Optimal Execution

Optimal execution is the problem of buying or selling a fixed, already-decided quantity of an asset over a trading horizon so as to minimise total trading cost. Its defining tension is the speed dilemma: trade fast and you minimise exposure to price volatility but pay heavy market impact for demanding liquidity quickly; trade slow and you minimise impact but leave the unliquidated inventory exposed to adverse price moves. The cost benchmark is the price at the moment of the trading decision, so the quantity minimised is Implementation Shortfall. Optimal execution appears in this vault as the single most important trading application of Markov methods — and, equally important, as the application that most clearly shows what those methods can and cannot do.

It is the canonical sequential decision problem cast as a Markov Decision Process Trading Model. The state is remaining inventory and time-to-deadline (optionally augmented with volatility, price level, or Limit Order Book imbalance); the action is how many shares to trade in the current period; the reward is the negative incremental cost — impact plus a penalty on cost-variance; the transition decrements inventory by the action while the price follows a stochastic process. Because the objective is a cumulative cost over a horizon and the next state depends only on the current state and action, optimal execution satisfies the Markov property by construction and is solvable by dynamic programming — value iteration on the Bellman equation — or, for the linear-impact mean-variance case, in closed form.

Optimal Execution [part-of] Markov Decision Process Trading Model Almgren Chriss 2000 [defines] Optimal Execution Square-Root Law of Market Impact [relates] Optimal Execution

The Almgren Chriss 2000 model is the classic formulation: liquidate X shares over horizon T by minimising E + λV, the expected execution cost plus a risk-aversion-weighted variance, yielding a smooth convex efficient frontier of cost-versus-risk and a front-loaded optimal trajectory with a characteristic trade half-life. Pedersen 2023 revisits exactly this model and solves it four ways — analytically, by numerical optimisation, by dynamic programming, and by reinforcement learning (DDPG) — confirming that the dynamic-programming solution “align[s] with the model intuition” and reproduces the analytical efficient frontier. That paper also pins down the MDP/solution distinction this vault cares about: dynamic programming solves the execution MDP exactly but “is infeasible for large portfolios” because the state space grows combinatorially (the Curse of Dimensionality); reinforcement learning solves the same MDP from sampled experience when the transition dynamics are unknown, the realistic live-market case. The cost inputs the model needs are not arbitrary — the Square-Root Law of Market Impact supplies the empirically validated impact function, and Almgren Thum Hauptmann Li 2005 estimated the permanent/temporary impact coefficients directly from equity-desk data.

Pedersen 2023 [tests_strategy] Optimal Execution Curse of Dimensionality [opposes] Optimal Execution Reinforcement Learning Trading Policy [optimises_policy] Optimal Execution

Optimal execution is the one Markov/MDP trading application with genuine, large-scale practical adoption. VWAP and TWAP Execution algorithms — the workhorses of institutional trading desks — are limiting cases of the Almgren-Chriss framework; investment banks and electronic trading platforms run calibrated variations of it to schedule large orders and to route between dark pools, limit orders and aggressive fills. This is real, deployed technology, not a backtest. But the reason its profitability_evidence_grade is inconclusive — out of scope for alpha is the central point of the vault: optimal execution minimises the cost of a trade that has already been decided. It takes the buy/sell direction and the size X as given inputs. It does not forecast price direction and it does not generate alpha. Its job is to make a chosen trade cheaper to implement — to shave basis points off Transaction Costs and Slippage — not to decide what or whether to trade.

That distinction matters for grading every Markov trading claim in this vault. A profitable execution algorithm and a profitable trading strategy are different objects: execution is cost optimisation on an exogenous decision, while a trading strategy must produce a positive signal about future returns. The genuine adoption of optimal execution is therefore not evidence that Markov methods generate tradeable edge — it is evidence that Markov methods are excellent at a cost-reduction problem where the objective is unambiguous and the dynamics are well-behaved. Indeed, optimal execution’s main contribution to the rest of the vault is defensive: by quantifying how impact scales with order size and turnover, it sets the realistic cost bar that high-turnover regime-switching strategies must clear — and frequently fail to, once Transaction Costs and Slippage are modelled honestly.

Optimal Execution [relates] Transaction Costs and Slippage Optimal Execution [contradicts] Markov Decision Process Trading Model Almgren Thum Hauptmann Li 2005 [supports] Optimal Execution

Connections

Sources

Graph View

Backlinks