Explaining reinforcement learning agent for high-frequency trading in quantitative finance

Abstract

High-frequency trading (HFT) has emerged as a prominent domain within quantitative trading, leveraging advanced algorithms to exploit microsecond-level market inefficiencies, particularly evident in the volatile Cryptocurrency (Crypto) market. Despite its potential, HFT faces challenges such as low data efficiency, dynamic market changes, complexity in decision-making, and a lack of explainability. In response, this study presents a novel approach rooted in Reinforcement Learning (RL) to enhance HFT’s efficiency and explainability. Leveraging a hierarchical Markov Decision Process (MDP) framework and integrating feature importance analysis methods, the proposed methodology demonstrates significant improvements in efficiency across various financial criteria, outperforming existing baselines. Furthermore, the incorporation of explainable methods enhances transparency, especially in complex decision-making scenarios. Observations from feature importance analysis shed light on critical market dynamics, informing trading strategies. Future work includes extending the proposed approach to low-frequency data, feature importance for out-of-distribution (OOD) detection and exploring predictive analysis of feature importance on price patterns, promising avenues for advancing quantitative finance strategies.