Evaluating Reinforcement Learning Agents for Portfolio Management

This study examines the efficacy of cutting-edge model-free Reinforcement Learning (RL) agents in trading, juxtaposing their performance against that of equally weighted portfolios and mean-variance optimization benchmarks. The RL agents undergo training using price data and features, and their performance is evaluated across various factor environments using a Carhart four-factor model. The findings reveal that the RL approach surpasses mean-variance specifications in terms of risk-adjusted and probabilistic portfolio performance, aligning closely with equally weighted portfolios. The RL agents are capable of partially deciphering the nonlinear structure of the Stochastic Discount Factor (SDF) and pinpointing crucial performance determinants. This investigation suggests that RL models hold practical utility and contribute significantly to the burgeoning integration of AI and big data within the finance domain. Ultimately, the study underscores the potential of RL agents in enhancing trading performance while offering valuable insights into their application in finance.