Reserving Methods9 min read

Bootstrap and Over-Dispersed Poisson Models for Reserve Variability

By Jonas Mohamed Osman Abdelghafour

Published 16 September 2026

How the ODP bootstrap actually works, where it breaks down, and the diagnostics that should accompany every bootstrap reserve distribution.

Executive answer

The over-dispersed Poisson bootstrap reproduces the chain-ladder best estimate while generating a distribution around it by resampling scaled residuals. It is popular because it is fast, reproducible and well understood, but it inherits every assumption of the chain ladder and adds assumptions of its own about residual behaviour.

Where it works and where it fails

The method performs acceptably for stable, high-volume, short-to-medium-tail triangles with consistent case-reserving practice. It performs poorly where development patterns have shifted, where a few large claims dominate a cell, where negative incrementals are common, or where the triangle is too small for residuals to carry information. In those cases the resulting distribution looks precise and is not.

Diagnostics that should always accompany the output

Residual plots by origin period, development period and calendar period are the minimum. A calendar-period trend in residuals almost always signals an inflation or case-strength effect that the model has absorbed incorrectly. Backtesting against held-out diagonals is the strongest available evidence and is still too rarely produced.

Governance considerations

A bootstrap result presented without residual diagnostics and a statement of scope should not pass validation. Where the diagnostics fail, the honest response is a wider judgemental range rather than a different resampling scheme that hides the same problem.

Conclusion

The bootstrap is a good tool held to a low evidentiary standard in much of the market. Raising that standard costs little and materially improves the credibility of reserve uncertainty reporting.

Primary sources

bootstrapODPreserve variability