Tail-aware reinforcement learning via copula-based scenario generation: computational algorithms and empirical evidence

Traditional risk models often fail during market crises due to extreme volatility and fat-tail characteristics. To address this, this study proposes a “Tail-Aware RL” architecture that integrates Extreme Value Theory (EVT), Student-t copula, and the Soft Actor-Critic (SAC) algorithm. The framework models marginal tails via POT-EVT and dependence via t-copula, employing a hybrid training strategy (70% synthetic, 30% real data) to mitigate overfitting. Empirical analyses on S&P 100 and BIST 30 indices (2012–2025), covering the COVID-19 and 2023 Banking crises, demonstrate that the model significantly outperforms baseline RL approaches. Specifically, Maximum Drawdown rates improved from 34% to 19% for BIST 30 and from 29% to 14% for S&P100. Validated by Jobson-Korkie and Block-Bootstrap tests, these results indicate that the proposed architecture enhances capital protection and offers a robust alternative for dynamic portfolio management during periods of systemic risk.