Feature configuration effects in DRL portfolio management: a risk-focused evaluation under market stress

This study investigates whether deep reinforcement learning (DRL) agents can effectively use financial feature information for risk-aware portfolio management. We design a controlled experimental framework that compares four feature configurations: BARRA-derived systematic risk information, technical indicators, their combination, and a no-feature baseline. Using the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm across 800 independently trained agents on randomly sampled 30-stock S&P 500 portfolios, we evaluate out-of-sample performance during the volatile 2022 market and assess statistical significance via paired permutation tests. The BARRA-derived information provides significant downside protection, improving maximum drawdown by 0.71% (p = 0.02) relative to the baseline, while technical indicators do not offer significant benefit alone or in combination. These results indicate that DRL agents can leverage systematic risk information to manage tail risk, and that targeted feature selection based on financial theory may be more effective than indiscriminate feature augmentation.