A Governance Framework for Explainable Artificial Intelligence in Catastrophe Modelling

Catastrophe models are among the highest-stakes artificial intelligence systems in financial services, yet they lack a dedicated explainability governance framework. This paper supplies one. Anchored in the Bermuda Monetary Authority's July 2025 Discussion Paper on the responsible use of artificial intelligence and its February 2026 stakeholder letter, we develop a technical governance framework for explainable AI (XAI) in catastrophe modelling. The framework translates the Authority's principles-led expectations into testable controls using SHapley Additive exPlanations (SHAP): component-level AI inventories with behavioural baselines, risk-tiered explainability requirements, quantitative attribution-shift triggers for mandatory human review, proxy-variable bias diagnostics, explanation-based validation protocols, and an application pathway for third-party vendor models. An illustrative machine-learning hurricane loss model, built on NOAA HURDAT2 hazard parameters for sixteen Florida landfalls, a synthetic exposure portfolio, and a stylised vulnerability function, demonstrates the framework end to end. The model achieves R 2 = 0.96 on held-out data; SHAP recovers the known loss drivers; and a simulated vulnerability update shifts roof-age attribution by +82%, firing the proposed oversight trigger exactly as designed. The framework embeds within existing enterprise risk management, consistent with the BMA's stated supervisory direction, and is transferable to any jurisdiction adopting principles-based AI governance.