Stock portfolio decision-making: An integrated deep reinforcement learning approach based on differential sharpe ratio

The stock market is an important investment channel and asset allocation tool for investors. How to use portfolio strategy to combine different stocks together to maximize investment return and risk control is the most important concern for investors and a very important issue in the field of finance. In this paper, we propose a deep reinforcement learning (DRL) -based machine learning approach (DSRRL) for stock portfolio selection and prediction of investment returns. An innovative financial model that integrates the differential Sharpe ratio of multiple stocks as the agent reward function is proposed. In addition, we propose an integrated learning framework with an alternate training scheme (ILAT) that ensembles the decisions of multiple reinforcement learning algorithms to approximate a theoretically optimal portfolio. Compared with classical portfolio strategies and existing DRL baselines, DSRRL achieves annualized returns of around 80% on our selected stock portfolios, while ILAT further improves the best-case annualized return to 94. 23% (on the rising-stock portfolio) ; meanwhile, the average Sharpe ratio is improved by about 0. 15. The effectiveness of the proposed method is demonstrated on U. S. equity portfolios under different market regimes and financial cycles.