A1255
Title: Unstable gains: Evaluating the reproducibility of deep reinforcement learning in trading and portfolio management
Authors: Przemyslaw Gradzki - University of Warsaw (Poland) [presenting]
Abstract: The aim is to investigate the reproducibility and robustness of deep reinforcement learning (DRL) in financial applications, focusing on algorithmic trading and portfolio management across two asset classes: Stocks and cryptocurrencies. While DRL has gained popularity in these domains, most studies rely on single-run evaluations and overlook the high variance inherent to these methods. Influential DRL-based strategies are reproduced under identical hyperparameters but across multiple independent random seeds, and it is shown that both performance and learned policies vary widely under fixed configurations. These experiments highlight the fragility of commonly reported results. Even the best-performing algorithms display substantial variability across runs. To improve reliability, a checkpointing strategy is introduced, and uncertainty is quantified using bootstrapping and permutation tests. Findings reveal that prevailing evaluation practices risk misleading conclusions about strategy efficacy and also conceal the true risk profile of DRL-based financial models. This underscores the need for more rigorous and reproducible protocols to ensure dependable advancements and foster genuine risk assessment in financial DRL research
