Dynamic factor-informed reinforcement learning for enhancing portfolio optimization

Abstract Portfolio optimization is essential for investors seeking to manage risk, diversify assets, and maximize returns. Although recent studies have focused primarily on enhancing technical aspects such as model architecture through the application of deep learning or reinforcement learning, knowledge of factor portfolios, grounded in modern portfolio theory, remains paramount. Therefore, to effectively utilize the knowledge of factor investment strategies, we propose a novel hybrid portfolio investment method that integrates reinforcement learning with dynamic factors, called the dynamic factor portfolio model. The dynamic factors encompass five fundamental factors: size, value, beta, investment, and quality. The proposed model comprises two modules: a dynamic factor module that calculates a score on the basis of factors reflecting the macro market and a price score module that calculates a score on the basis of prices expressing the relationship between assets and their future value. With dynamic factor-informed knowledge, the proposed model can make portfolio decisions adaptively on the basis of market conditions. Through comprehensive experiments, we validate the effectiveness of the dynamic factor module and demonstrate that the proposed model outperforms both traditional portfolio investment strategies and existing reinforcement learning-based strategies. Moreover, the proposed model offers interpretability by identifying critical factors across varying market scenarios, thereby enhancing portfolio management practices.