The lay abstract featured today (for Conformal Prediction Inference in Regularized Insurance Models by Alokesh Manna, Aditya Vikram Sett, Dipak K. Dey, Yuwen Gu, Elizabeth D. Schifano, Jichao He) is from Applied Stochastic Models in Business and Industry with the full article now available to read here. This is a Data Science in Business and Industry article.
How to cite
, , , , , and , “ Conformal Prediction Inference in Regularized Insurance Models,” Applied Stochastic Models in Business and Industry 41, no. 5 (2025): e70045, https://doi.org/10.1002/asmb.70045.
Lay Abstract
Accurately predicting uncertain outcomes is one of the biggest challenges in both science and business. In the insurance industry, this challenge is especially important: when estimating how much drivers might cost an insurer in claims, companies need not only a best guess but also a realistic range of possible costs. These ranges help insurers set fairer premiums and prepare for unexpected losses. Traditional approaches such as Generalized Linear Models (GLMs) have long been used to model insurance claims, while more modern machine learning methods such as Gradient Boosting (GBMs) have become popular for their flexibility and predictive power. Yet, providing reliable measures of uncertainty for these predictions remains a difficult task. Standard statistical practice often uses a 95% confidence interval based on the ‘3-sigma rule’, but this can produce overly wide intervals, limiting their usefulness in practice.
This research introduces new ways to apply a statistical framework known as conformal prediction to both GLMs and GBMs. Conformal prediction provides a distribution-free method to quantify uncertainty, meaning it does not rely on strong assumptions about the data. We demonstrate that by using specialized residual-based measures (Pearson, Anscombe, Deviance) within this framework, it is possible to construct prediction intervals that are both accurate and informative. In particular, when applied to real insurance claims data, a locally weighted approach with LightGBM produced narrower intervals while still capturing the true outcomes as often as expected. Beyond insurance, these ideas offer a general strategy for combining modern machine learning with principled uncertainty quantification, with potential benefits for finance, healthcare, and many other fields where decision-making under uncertainty is critical.
Supplementary Material
Conformal-Prediction-Inference-in-Regularized-Insurance-Models.pdf