← Research logMachine Learning
PARKED · sim-to-real
#

A PPO agent learns to slice a parent order to minimize implementation shortfall versus TWAP/VWAP, adapting to live book conditions.

PPO maximizes a clipped policy ratio against the advantage estimate:

$$ L^{CLIP}(\theta)=\mathbb{E}_t\big[\min(\rho_t A_t,\ \mathrm{clip}(\rho_t,1-\epsilon,1+\epsilon)A_t)\big] $$

Train in a simulated LOB with a market-impact model; reward = negative implementation shortfall. Evaluated in-sim only.

PARKED
Execution (not alpha) is the right place for RL — the reward is well-defined and dense. Parked on the sim-to-real gap: the agent overfits the simulator’s impact model, and we lack a faithful live LOB simulator to trust it with real orders. Promising, not deployable yet.
RL belongs to execution, not direction. But an RL agent is only as honest as its simulator — a wrong impact model trains a confidently wrong agent.

We publish the failures too.

This is one of 100+ documented hypotheses. Browse the full lab notebook, or see the strategies that survived into production.