A Novel Deep Reinforcement Learning Strategy for Portfolio Management

Portfolio management is an investment strategy to redistribute some given fund into different assets, which aims to maximize the return and minimize the risk over the given period. There has been much research on the application of machine learning techniques to portfolio management. In this paper, we propose a novel investment strategy that uses a framework based on the Depthwise convolution, Squeeze and Excitation module, Residual Block, and Gate Recurrent Unit, called DSRG Network. To test the performance of our strategy, we use eight common trading strategies to obtain the corresponding accumulative portfolio values and Sharpe ratios. In addition, the Sharpe ratio is used to measure the adjusted rate of return for each strategy. Results show that our proposed strategy outperforms at least 40% better than other strategies used in otherwise comparison experiments. It still maintained a rate of return of about 70% in an experiment where all other comparison strategies showed losses.