A Novel Deep Reinforcement Learning Strategy in Financial Portfolio Management

Abstract

Reinforcement learning algorithms are used in various fields widely, such as cryptocurrency market forecasting, image recognition, and natural language processing. In this research, we use the Reinforcement learning algorithm to solve the portfolio management problem. In the Reinforcement learning algorithm, we adopt the squeeze-and-excitation in the neural network neural network to realize the Ensemble of Identical Independent Evaluators proposed by Jiang et al. "The Squeeze-and-Excitation" block works by adaptively recalibrating channel-wise feature responses, which improves the ability of the network in extracting information from the financial environment. To further improve the performance of the network, we adopt the soft thresholding function as nonlinear transformation layers to effectively eliminate the noise-related features. The cryptocurrency market is used to test the efficacy of our strategy along with eight traditional portfolio management strategies as well as Jiang et al.’s strategies. In our experiments, we use the Accumulated Portfolio Value, Sharpe Ratio, and Maximum Downward to assess the efficacies of the strategies. In conclusion, our strategy outperforms Jiang et al.’s strategies and other traditional strategies. Although our strategy has a nearly 30% Maximum Downward as metrics in back-tests, Accumulated Portfolio Value can reach nearly 170% in two different two-month back-tests, which is about 160% greater than the traditional strategy and Jiang et al.’s strategy. Moreover, our result Sharp ratios are 393.88% and 96.18% higher than the traditional strategy and Jiang et al.’s strategy, respectively.