Deep reinforcement learning for stock portfolio optimization by connecting with modern portfolio theory

With artificial intelligence and data quality development, portfolio optimization has improved rapidly. Traditionally, researchers in the financial market have utilized the modern portfolio theory for portfolio optimization; however, with the recent development of artificial intelligence, attempts to optimize portfolios with reinforcement learning are increasing. Many studies have developed reinforcement learning and deep learning algorithms and conducted portfolio optimization research. However, in reality, thus far, the securities industry thus has used the modern portfolio theory, which is sufficiently valuable. Nevertheless, to the best of our knowledge, there has yet to be an attempt to combine modern portfolio theory and reinforcement learning. To bridge this gap in the literature, we propose a novel deep reinforcement learning approach that combines the modern portfolio theory and a deep learning approach. As far as we know, we are the first to combine recent deep learning technology and traditional financial theory. Specifically, we solved the multimodal problem through the Tucker decomposition of a model with the input of technical analysis and stock return covariates. The results show that the proposed method outperforms state-of-the-art algorithms regarding the Sharpe ratio, annualized return, and maximum drawdown. In addition, the proposed method dynamically changes the weight according to the market trend, unlike other state-of-the-art algorithms.