Deep Reinforcement Learning for Automated Stock Trading: Inclusion of Short Selling

Abstract

Multiple facets of the financial industry, such as algorithmic trading, have greatly benefited from their unison with cutting-edge machine learning research in recent years. However, despite significant research efforts directed towards leveraging supervised learning methods alone for designing superior algorithmic trading strategies, existing studies continue to confront significant hurdles like striking the optimum balance of risk and return, incorporating real-world complexities, and minimizing max drawdown periods. This research work proposes a modified deep reinforcement learning (DRL) approach to automated stock trading with the inclusion of short selling, a new thresholding framework, and employs turbulence as a safety switch. The DRL agents’ performance is evaluated on the U.S. stock market’s DJIA index constituents. The modified DRL agents are shown to outperform previous DRL approaches and the DJIA index, in terms of absolute returns, risk-adjusted returns, and lower max drawdowns, while giving insights into the effects of short selling inclusion and proposed thresholding.