Market microstructure as a control problem
Reframing execution as feedback control yields cleaner guarantees than statistical priors alone.
By Finance Programme
Execution is usually framed as a forecasting problem: predict short-horizon price, then trade against the prediction. We argue that this framing is backwards. The quantity an execution system actually controls is its own participation rate, and the environment responds to that control with measurable, repeatable dynamics.
Control, not prediction
Treating the order book as a plant and the schedule as a controller lets us borrow the full apparatus of control theory — stability margins, observability, and bounded-error guarantees — instead of leaning entirely on statistical priors that decay the moment they are deployed.
- The state is inventory and realised impact, not a price forecast.
- The cost function penalises variance of slippage, not just its mean.
- Guarantees hold under adversarial fills, where priors quietly fail.
A controller that is honest about its uncertainty beats a forecaster that is confident and wrong.
What this buys us
In backtests across NYSE and three on-chain venues, the control formulation
reduced realised slippage variance by roughly a third at equal mean cost. More importantly,
it produced diagnosable failures: when it underperformed, we could point to
a specific violated assumption rather than shrug at noise.