Learned Participation-Rate Policies

A percentage-of-volume (POV) algorithm trades a fixed fraction of whatever volume trades in the market — say, participate at 10% of volume — and it's popular precisely because it's simple and adapts automatically to how busy the market is: it naturally speeds up when volume is heavy and slows down when it's thin. But a single fixed participation rate chosen at the start of the order still leaves real decisions on the table: should the rate be higher when the market is moving favorably, lower when the spread is unusually wide, or adjusted based on how the order is tracking relative to its schedule? A learned participation-rate policy makes the rate itself a function of real-time conditions instead of a constant.

What a fixed rate misses

Choosing 10% at the start bakes in a single assumption about the tradeoff between speed and impact that gets applied uniformly regardless of what actually happens during the day. If volatility spikes and the price starts drifting away from the order's side, a fixed rate doesn't respond — it keeps trading at the same pace even though delaying now risks a much worse average price, or conversely, even though slowing down now would meaningfully reduce impact cost with only modest additional risk. A learned policy conditions the rate on the same state variables a human trader would watch: how far ahead of or behind schedule the order currently is, recent price drift relative to the arrival price, current spread and depth, and time remaining.

Setting it up as a policy problem

The rate at each interval, $\rho_t$, becomes an output of a function of the current state rather than a fixed number: $\rho_t = f(\text{state}_t)$. This is naturally framed as a reinforcement learning problem, since the "reward" — total implementation shortfall over the whole parent order — depends on a whole sequence of rate decisions, not any single one in isolation, and the effect of speeding up now (more impact, less residual risk) trades off against speeding up later. The policy is trained on historical data (or a realistic fill simulator) to minimize expected shortfall across many simulated order paths, learning, in effect, when the desk should lean into the market and when it should hang back.

Worked example: reacting to adverse drift

An order targeting a fixed 10% participation rate is 40% complete at the midpoint of its trading window, exactly on schedule. The price has drifted 30 basis points against the order's side over the last 20 minutes, and volatility has picked up. A learned policy, conditioned on that adverse drift and elevated volatility, might raise the effective rate to 15% for the next interval — accepting somewhat higher market impact now to reduce the risk of an even worse average price if the adverse drift continues — whereas the fixed 10% policy keeps trading at the original pace regardless of what the price has just done. If the drift instead reverses back in the order's favor, the learned policy would lower the rate rather than continuing to chase a price that's already moving the desk's way.

What this means in practice

Learned participation-rate policies typically improve on fixed rates most in orders that run long enough for real intraday conditions to change meaningfully — short orders in stable markets see little benefit, since there's not much for the policy to react to. The main practical risk is training on historical data generated under the desk's existing fixed-rate policy, which never observed the market's response to the more aggressive or more passive actions the new policy wants to take, making honest evaluation — via simulation or careful counterfactual methods — essential before deploying a learned rate live.