Adaptive risk-based control in financial trading

Abstract

A critical part of an automated trading strategy is its ability to adapt to changing market conditions. However, many state-of-theart approaches fail to include risk management as part of the core algorithm. In this work, we propose a Distributional Reinforcement Learning approach that considers action confidence as part of the trading process. To achieve this we utilize the structure of the TD3 algorithm, replacing the critic network with a Distributional RL agent. Furthermore, we introduce the idea of a Volatility-Prioritized Replay Buffer which improves training by utilizing more suitable market conditions. We test our approach on a set of 30 Assets over 4 years. Overall, our implementation improves risk-adjusted performance, achieving a statistically larger Sharpe ratio, small drawdown periods, lower volatility, and improved consistency compared to other popular methods.