Portfolio construction through reinforcement learning: an empirical study on the Korean stock market via interpretable AI

Abstract

Attempts have been made to incorporate advanced machine learning techniques to handle financial data due to their non-linearity, high-dimensionality, and non-stationarity. However, even with wider applications in finance, such techniques have given limited insights to investors due to their black-box nature. In addition, typical portfolio optimization through machine learning takes two steps, return prediction and weight assignment, which magnify optimization error. To overcome aforementioned limitation, this paper implements deep reinforcement learning-based portfolio optimization model introduced by Cong, Tang, Wang, and Zhang (2020) to the Korean stock market. The model consisted of Transformer Encoder and Cross-Asset Attention Network optimizes Sharpe ratio of directly constructed portfolio. The portfolio exhibits outstanding out of sample performance even under various restrictions, including transaction costs, and short sale constraints. Lastly, the interpretation of the model is achieved by replicating policy network of the model into second degree polynomial. Through dominant feature ranking and feature dynamic analysis, it is concluded that the model depicts SG&A to sales ratio as one of the most dominant features, and it tends to put more weights on profitability characteristics during earlier months in the test period from 2006 to 2020 and value characteristics during the later months.