Logic-guided Deep Reinforcement Learning for Stock Trading

Zhiming Li, Junzhe Jiang, Yushi Cao, Aixin CUI, Bozhi Wu, Bo Li, Yang Liu

Keywords: stock trading, quantitative finance, program synthesis

Abstract: Previous state-of-the-art trading strategy proposes using ensemble reinforcement learning to combine the advantages of different subpolicies. Despite its improved performance, we observe that this policy is still quite sensitive to market volatility. In this work, we propose a novel framework called SYENS (Program Synthesis-based Ensemble Strategy) which aims to improve the trading strategy's robustness via the program synthesis by sketching paradigm. SYENS is a hierarchical strategy that uses a program sketch as the high-level strategy. The program sketch embeds human expert knowledge of market trends. And based on the program sketch, we adopt the program synthesis by sketching paradigm to synthesize the detailed ensemble strategy. Experimental results demonstrate that SYENS achieves the highest return while retaining low drawdown.

Submission Number: 72