Benchmarking Reinforcement Learning Approaches in Stock Portfolio Management

Abstract

Reinforcement Learning has wide applications across multiple fields, from robotics to gaming [1]. However, the applications of reinforcement learning in stock portfolio management were explored narrowly, with researchers building their own work in reference to state-of-the-art stock portfolio management approaches that are not machine learning-based. This paper explores two reinforcement learning approaches in stock portfolio management through benchmarking their performance in terms of Annual Percentage Return, volatility, and Sharpe ratio. The AlphaStock and the Deep Reinforcement Learning(DRL) for dynamic portfolio optimization models were selected because of their similar architecture and Sharpe ratio maximization objective. The DRL model outperforms the AlphaStock model by averaging about 9% in Annual Percentage Rate(APR), while the AlphaStock model delivers about 7.7% in APR. Overall, the AlphaStock model selects portfolios with a higher volatility than DRL. For the same time horizon, the DRL model outperforms the market by more than half a percentage point, indicating that fewer stocks can deliver market returns when carefully selected.