Risk-Averse Trader: A Deep Reinforcement Learning-Based Portfolio Optimization Method for Risk-Averse Investors

Portfolio management requires continuous adjustments to asset allocation based on market conditions and investor goals, aiming to optimize returns at a given risk level. Risk-averse investors prefer stable returns with lower risk, which poses challenges for standard portfolio management. This paper proposes a deep reinforcement learning-based method, named Risk-Averse Trader, to control investment risk for such investors. We first employ a feature fusion extraction algorithm that combines Multi-scale CNN, U-Net structures, and Multidimensional Attention Mechanisms to enhance the model’s ability to dynamically process financial market data. Additionally, we introduce a downside risk evaluation mechanism specifically designed for risk-averse investors, using the target return as a benchmark to penalize suboptimal daily returns, thereby enhancing risk control capabilities. The proposed model is evaluated on the Dow Jones Industrial Average (2019) and the CSI 300 Index (2023). Results demonstrate improved performance in terms of annualized return, maximum drawdown, and Sharpe ratio, effectively balancing risk and return under varying economic conditions.