A Portfolio Model Based on BiLSTM Prediction and Improved PPO Algorithm

Abstract

This paper develops a portfolio investment model to maximize traders’ returns of gold and bitcoin. We establish a Bidirectional Long ShortTerm Memory Network based Proximal Policy Optimization (BiLSTM-PPO) algorithm. Then, we improve the PPO algorithm by setting the penalty factor to implement a strategy in which gold does not trade on non-trading days. Finally, with the proposed BiLSTM-PPO algorithm to learn the state vector composed of historical data, covariance and BiLSTM prediction results, we obtain the optimal trading strategy. A portfolio selection case is given to illustrate the application process and effectiveness of the method. We compare the BiLSTM-PPO with traditional PPO to prove the ef-fectiveness of it. Even in the worst case, the final income increases by 6.28% than the traditional PPO. And then, we compare the BiLSTM-PPO with five common investment de-cision algorithms such as Min-Variance, Deep Deterministic Policy Gradient, and etc by six financial metrics to prove the optimality of our model. The experimental results show that the BiLSTM-PPO achieves the highest final revenue and strategy provided by the algorithm is adjusted adaptively to ensure the maximization of returns.