From features to tariff: a short, defensible path.
Match distribution/link to the task (Poisson for frequency, Gamma/Tweedie for severity/premium; often a log link). Set exposure/offset and keep loadings outside the model.
For features, prefer sensible grouping (categoricals), binning/transformations (numerics), and only explainable interactions. Validate with hold-out/rolling, add regularisation, and track coefficient stability.
When communicating, translate effects into tariff, LR, and net impact, not just betas. Before shipping, freeze model version and feature mappings, define caps/floors, monitor drift (e.g., PSI), and keep a full audit trail.
Takeaway: GLM is craft. Simple features + clear validation + explicit model→tariff mapping = production-proof decisions.