Catastrophic Risk-Aware Methods in Deep Reinforcement Learning

This study extends the Extreme Value Theory (EVT) method for Conditional Value-at-Risk (CVaR) optimization to high-dimensional neural network policies. It obtains a differentiable tail risk objective through peak-over-threshold representation and moment estimation, allowing direct calculation of finite-sample gradients via automatic differentiation. In deep hedging problem tests involving rare extreme market fluctuations, EVT-trained strategies exhibit lower hedging losses under extreme scenario evaluation, providing a feasible alternative for training neural network policies when observations related to extreme risks are scarce.