Lloyd's 2026 capital briefing notes that historical cycles can be insufficient for validation and that catastrophe back-tests can become too focused on averages. The actuarial implication is straightforward: a model may pass a mean-loss test while remaining weak on dependence, accumulation, severity or regime change.
Average performance is not tail validation
Average loss ratios and mean catastrophe frequency are useful diagnostics, but they do not test whether the model recognises the shape of the tail. Validation should also examine exceedance probabilities, event clustering, spatial accumulation, exposure drift and the effect of policy terms.
Define pass and fail before looking at the result
Validation becomes difficult to govern when thresholds are selected after results are known. A control framework should specify the test, threshold, materiality, escalation and remediation path in advance. Where a formal statistical threshold is not appropriate, the rationale for expert judgement should be recorded.
Use scenarios to test what history cannot
Historical data cannot contain every plausible future event. Severe-but-plausible scenarios therefore complement back-testing. They should be linked to the syndicate view of risk and should cover second-order effects, model exclusions and changes in portfolio structure. Scenario results should be compared with capital appetite and management action capacity.
My recommendation as CEO of Quantica Risk Modelling is to make validation a chain of evidence: data quality, model design, parameter uncertainty, scenario performance, governance and remediation closure. Source: Lloyd's Capital Briefing 2026.