Risk-Sensitive Portfolio Management by using Distributional Reinforcement Learning

Abstract

In recent years, many studies applied deep reinforcement learning in portfolio management. However, few studies have explored the use of value-based reinforcement learning as it is unclear how the risk of a portfolio can be incorporated. In this research, we proposed an agent called C21-SR by adapting the 21-bin categorical reinforcement learning and action-selection strategy based on Sharpe ratio to control the risk of investment and maximize profit. Our results revealed that a C21-SR agent could outperform buy&hold and constant rebalance strategies, and the action-selection strategy based on the Sharpe ratio could enhance the performance of categorical reinforcement learning in the financial market.