A Deep Reinforcement Learning Model for Portfolio Management Incorporating Historical Stock Prices and Risk Information

At present, more and more studies have begun to apply deep reinforcement learning (DRL) to portfolio management problems. However, most cutting-edge DRL models have some limitations in processing time series data, while also not considering the extent to which historical risks impact the current market. To solve this problem, this paper proposes a DRL model that takes into account historical stock prices and risk information. Based on the double-delay depth deterministic strategy gradient (TD3) model, this study incorporates a convolutional short term neural network (CNN-LSTM) to enable TD3 to handle time series data and account for the interrelation between multiple features. To consider the degree of influence of historical risk on the current market, a risk assessment unit is introduced, which processes the risk indicators through long and short term neural network (LSTM) and dynamically weights the results. The experiments reveal that the proposed model is capable of attaining higher returns when accounting for risks.