Standard insurance models produce point estimates. A Tweedie GBM gives you expected loss cost — not how confident the model is in that estimate. That distinction matters. A standard motor risk with ten thousand similar policies behind it is not the same as a borderline fleet risk in a sparse corner of feature space. The model outputs a number in both cases. Without uncertainty quantification you cannot tell them apart, and you will misprice the risks you are least sure about most consistently.

The classical response is parametric confidence intervals: bootstrap the GLM coefficients, or propagate uncertainty through the Tweedie dispersion. Both depend on distributional assumptions. If the model is misspecified — and it is always at least partially misspecified — the intervals are wrong in ways that are difficult to characterise.

Conformal prediction offers a different kind of guarantee. It makes no assumptions about the error distribution and produces intervals with a finite-sample coverage guarantee: the true value falls within the interval at least 90% of the time (or whatever coverage level you set), unconditionally. The insurance-conformal library applies this to insurance models, handling the exposure weighting, heteroscedasticity, and non-exchangeable claims series that standard implementations ignore.

Library: insurance-conformal on GitHub · pip install insurance-conformal


Tutorials and introductions


Techniques and extensions


Benchmarks and validation


Library comparisons