SignalExec
Given the same trading signal, can RL execute an order better than TWAP?
Research question
Given the same trading signal and target position, can a learned execution policy reduce implementation shortfall relative to TWAP?
Experiment
- Compared TWAP, a hand-built rule, SAC, PPO, and Recurrent PPO across five decision intervals.
- Trained 45 models across three random seeds, then evaluated frozen policies with matched signals, costs, and deterministic replay.
- Isolated execution quality from signal quality with common-noise signals and mirrored buy and sell targets.
- Tested order-flow features through matched state ablations rather than changing the signal or cost model.
Finding
On the locked May–August 2026 test, 15-minute PPO beat TWAP in all three seeds by 8.15 basis points on average. The broader result was conditional rather than universal: performance changed materially with decision interval, policy class, and state representation.
Evidence boundary
The controlled signal intentionally contains future-return information. This project compares execution policies after a target has been supplied; it does not claim alpha discovery, live profitability, or reproduction of a proprietary trading strategy.