This paper introduces a robust method for evaluating Conditional Value‑at‑Risk (CVaR) when data distribution can't be simulated. Using rolling data windows as proxies for independent samples, the approach effectively assesses worst‑case risk. Applied to Danish fire insurance data, it outperformed traditional DRO (distributional risk optimization) methods—achieving accurate, less conservative estimates in 87% of cases. This advancement enables reliable risk management even with limited tail data. Future research will focus on refining robustness guarantees and integrating extreme value theory into decision‑making models involving rare but impactful events.