Conformal Prediction Credibility Intervals

Abstract

In the predictive modeling context, the credibility estimator is a point predictor; it is easy to calculate and avoids the model misspecification risk asymptotically, but it provides no quantification of inferential uncertainty. A Bayesian prediction interval quantifies uncertainty of prediction, but it often requires expensive computation and is subject to model misspecification risk even asymptotically. Is there a way to get the best of both worlds? Based on a powerful machine learning strategy called conformal prediction, this article proposes a method that converts the credibility estimator into a conformal prediction credibility interval. This conformal prediction credibility interval contains the credibility estimator, has computational simplicity, and guarantees finite-sample validity at a pre-assigned coverage level.

ACKNOWLEDGMENTS

Thanks are due to the co-editor and the two anonymous reviewers for many helpful comments and suggestions.

Notes

1 Note that ${\mathbb{R}}_{+}$ is open in ${\mathbb{R}}_{+},$ though it is not open in $\mathbb{R}.$

Additional information

Funding

This work is supported jointly by the Casualty Actuarial Society and the Society of Actuaries through the 2021 Individual Grant.