SuperGLM

Penalised GLMs and GAM-style pricing models for insurance, with exact and discrete REML, solver-backed monotone splines, out-of-fold validation, and deployable fitted estimators.

Recommended Workflows Monotone Splines API Reference

Start Here

If you are new to the package, the intended path is:

1. build an explicit feature spec
2. fit with fit_reml() and selection_penalty=0
3. use cross_validate(..., fit_mode="fit_reml") for model comparison
4. validate challengers with Lorenz and double-lift charts
5. serialize the fitted estimator for scoring

What It Covers

Workflow-first pricing models

Exact REML for standard GAM-style pricing, discrete REML for large data, and clear guidance on when to use sparse selection instead.

Feature engineering in-model

P-splines, cubic regression splines, natural splines, ordered categoricals, grouped categoricals, and interaction terms.

Monotone solvers

Solver-backed monotone spline fitting through QP and SCOP paths, with post-fit repair retained only as a fallback.

Validation and deployment

Cross-validation with out-of-fold predictions, Lorenz and double-lift charts, diagnostics, and deployable fitted estimators.

Example Plots

Poisson frequency model on French MTPL2 (678k policies), fitted with REML smoothness selection. Plots show pointwise confidence bands, weighted density, and interior knot positions.

Vehicle Age (quantile_rows knots)	Bonus-Malus (quantile_tempered, alpha=0.2)