Reinforcement learning for deep portfolio optimization

Portfolio optimization is an important financial task that has received widespread attention in the field of artificial intelligence. In this paper, a novel deep portfolio optimization (DPO) framework was proposed, combining deep learning and reinforcement learning with modern portfolio theory. DPO not only has the advantages of machine learning methods in investment decision-making, but also retains the essence of modern portfolio theory in portfolio optimization. Additionaly, it was crucial to simultaneously consider the time series and complex asset correlations of financial market information. Therefore, in order to improve DPO performance, features of assets information were extracted and fused. In addition, a novel risk-cost reward function was proposed, which realized optimal portfolio decision-making considering transaction cost and risk factors through reinforcement learning. Our results showed the superiority and generalization of the DPO framework for portfolio optimization tasks. Experiments conducted on two real-world datasets validated that DPO achieved the highest accumulative portfolio value compared to other strategies, demonstrating strong profitability. Its Sharpe ratio and maximum drawdown also performed excellently, indicating good economic benefits and achieving a trade-off between portfolio returns and risk. Additionally, the extraction and fusion of financial information features can significantly improve the applicability and effectiveness of DPO.