Deep Reinforcement Learning for Portfolio Management: A Simulation Study

Abstract

We propose a comprehensive Monte Carlo simulation study of deep reinforcement learning methods applied to the portfolio management task when returns are predictable and transactions costly. We discover significant discrepancies in performance depending on the underlying algorithm. Our findings emphasize the necessity of restricting the asset sample for the adequate performance of DRL approaches and show that some algorithms achieve robust outstanding cumulative returns under the right circumstances. Moreover, we contribute to the ecosystem of the Julia programming language by providing all necessary codes to replicate and extend our work.