Lower Risks, Better Choices: Stock Correlation Based Portfolio Selection in Stock Markets

Di Luo, Weiheng Liao, Rui Yan

Abstract: Over the past few years, we’ve seen a huge interest in applying AI techniques to develop investment strategies both in academia and the finance industry. However, we note that generating returns is not always the sole investment objective. Take large pension funds for example, they are considerably more risk-averse as opposed to profit-seeking. With this observation, we propose a Risk-balanced Deep Portfolio Constructor (RDPC) that takes risk into explicit consideration. RDPC is an end-to-end reinforcement learning-based transformer trained to optimize both returns and risk, with a hard attention mechanism that learns the relationship between asset pairs, imitating the powerful pairs trading strategy widely adopted by many investors. Experiments on real-world data show that RDPC achieves state-of-the-art performance not just on risk metrics such as maximum drawdown, but also on risk-adjusted returns metrics including Sharpe ratio and Calmar ratio.