Phil Pearson, CFA, Director
Matthew Wigle, CFA, Director
Fangyi Li, Assistant Vice President
Yichu Li, Assistant Vice President
We present a new approach to trade scheduling using reinforcement learning, an area of machine learning. By providing a long-term reward function that balances risk and cost, we train a model to select from a series of actions. This allows us to incorporate complex (non-linear) relationships, for which traditional models typically do not account. We tested the reinforcement learning-based algorithm using a randomized controlled trial that compared the reinforcement learning-based approach to a traditional implementation shortfall-based algorithm. The trial involved approximately 30,000 orders equating to $4 billion in notional value. The results showed a 2.3bps outperformance for the reinforcement learning-based strategy versus an impact-adjusted arrival-price benchmark.