GLMs Versus Gradient Boosting in General Insurance Pricing
By Jonas Mohamed Osman Abdelghafour
How pricing teams are combining generalised linear models with gradient boosting, and the governance that makes the hybrid defensible.
Executive answer
The practical settlement in most pricing teams is not one or the other. Gradient boosting is used to discover structure, interactions and residual signal; a GLM or constrained model is used for the deployed rating structure where interpretability, stability and regulatory explanation are required.
What boosting finds that GLMs miss
Interactions and non-linearities that the modeller did not hypothesise. Reviewing boosted-model diagnostics against the incumbent GLM is often the fastest way to find missing factors, mis-specified banding and segments where the current rating structure is systematically wrong.
Why GLMs persist in deployment
Stability across refreshes, transparency to underwriters and regulators, controllability of individual relativities, and straightforward implementation in rating engines. Where a black-box model is deployed directly, the firm takes on explanation, monitoring and fairness obligations that should be planned for deliberately.
Governance considerations
Pricing models face conduct as well as financial scrutiny: proxy discrimination, price-walking rules and outcome testing all apply regardless of model family. Documenting factor rationale and testing outcomes across customer groups is now a baseline expectation rather than a refinement.
Conclusion
Use machine learning to learn and interpretable models to decide, unless there is a specific, governed reason to do otherwise.
Primary sources
More in Pricing & Underwriting
- Large Loss and Catastrophe Loadings in Technical Pricing
By Jonas Mohamed Osman Abdelghafour · 8 min read