Risk-Adjusted Deep Reinforcement Learning for Portfolio Optimization: A Multi-reward Approach

Portfolio optimization is a widely studied topic in quantitative finance. Recent advances in portfolio optimization have shown promising capabilities of deep reinforcement learning algorithms to dynamically allocate funds across various potential assets to meet the objectives of prospective investors. The reward function plays a crucial role in providing feedback to the agent and shaping its behavior to attain the desired goals. However, choosing an optimal reward function poses a significant challenge for risk-averse investors aiming to maximize returns while minimizing risk or pursuing multiple investment objectives. In this study, we attempt to develop a risk-adjusted deep reinforcement learning (RA-DRL) approach leveraging three DRL agents trained using distinct reward functions, namely, log returns, differential Sharpe ratio, and maximum drawdown to develop a unified policy that incorporates the essence of these individual agents. The actions generated by these agents are then fused by employing a convolutional neural network to provide a single risk-adjusted action. Instead of relying solely on a singular reward function, our approach integrates three different functions aiming at diverse objectives. The proposed approach is tested on daily data of four real-world stock market instances: Sensex, Dow, TWSE, and IBEX. The experimental results demonstrate the superiority of our proposed approach based on several risk and return performance metrics when compared with base DRL agents and benchmark methods.