A parametric distributional reinforcement learning framework for conditional systemic risk estimation

We extend the distributional reinforcement learning framework by incorporating parametric tail modeling via the generalized extreme value (GEV) distribution into the value estimation process, augmented with superior learning enhancements. This allows for principled modeling of systemic risk measures such as the System Conditional Value-at-Risk (S-CoVaR) and enables risk-sensitive policy learning in environments associated with heavy-tailed reward distributions, such as asset class returns. Additionally, we explore nonparametric empirical distribution modeling to provide a flexible alternative and evaluate the agents’ estimated results across bull and bear market conditions. Our research findings highlight the vital role that distributional assumptions play in frameworks geared towards risk-sensitive decision-making under economic and financial market uncertainty.