Reinforcement Learning for High-Frequency Trade Execution and Signal Monetization

While many researchers in quantitative finance focus their attention on the generation of alpha and trading signals, trade execution and timing can have a material impact on realized P&L, even in the presence of high-quality signals. In this paper, we study and evaluate the performance of deep reinforcement learning models in optimizing high-frequency trading objectives. We simulate trading using a combination of historical L1 or L2 constructed limit order book snapshots and the trades tape from 2026 BTC-USD data from the Coinbase exchange. The first environment we test is when a favorable trading signal is provided to the agent, and in the second environment the agent must reach a specified net long or short position within a given time frame.

Ultimately, we show that a deep-Q-network (DQN) trading agent outperforms the strongest heuristic baseline when a trading signal is provided on L1 data, with a 31% P&L improvement, performs on par with baselines on L2 data, and slightly underperforms in the position-target environment on both datasets. Conditioning P&L on signal accuracy also demonstrates that on L1 data, the DQN profits 19% more than the strongest baseline when trading with an accurate signal and suffers 15% fewer losses when the signal is incorrect.