Pricing models: validation must follow the insurance decision
A lower validation error is useful only if the prediction targets the right insurance quantity and uses information available at the time of the decision.
Set the target and exposure first
For a claim-count model, policy counts alone can be misleading when observation lengths differ. An exposure offset can model expected count as exposure × frequency rate. A severity model fitted only on positive claims answers a different question from a pure-premium model across all policies.
A leakage example
A field populated after a claim investigation predicts claims very well in a random split. It is unavailable when the policy is quoted. Remove it from a prospective pricing model, even if accuracy falls. Split and preprocess data in a way that respects the intended deployment timing.
The validation pack
Compare benchmark and candidate models; examine calibration and stability by meaningful segment; inspect out-of-time performance; test rating-plan implementation and document limitations. Review the effect of sparse classes, changed exposure mix and extreme predictions.
A deployable model needs data definitions, version control, implementation tests, an owner and monitoring triggers. A model that fits well but cannot be reconciled with the production rating engine is unfinished.
Routine monitoring asks whether actual experience and input distributions have changed. A new model or recalibration is a project requiring actuarial, business and applicable regulatory review. No metric alone authorises use of a rating variable.
Reading sources
- CAS · Generalized Linear Models for Insurance Rating
- FSRA · Active Ontario major auto filing guidance
ActNet editorial guide · October 1, 2026 · Original illustrative examples.