Conformal prediction for frequency-severity modeling

Helton Graziadei, Paulo C. Marques F., Eduardo F. L. de Melo and Rodrigo S. Targino

https://arxiv.org/abs/2307.13124

• synthetic/ – §4.1
  Scripts to generate data and run the two-stage frequency–severity pipeline with conformalization (split-conformal baseline; GLM/RF severity variants). Produces summary metrics such as empirical coverage and average width.

• mtpl/ – §4.2
  Real-data application to Motor Third-Party Liability (Belgium), implementing the same pipeline and evaluation metrics as in the synthetic study.

• crop/ – §4.3
  Real-data application to Brazilian crop insurance (municipality-level aggregation as described in the paper), with the same reporting of coverage and width.

• Split conformal prediction (two-stage) – main procedure across §4.1–4.3.
• Out-of-bag (OOB) extension – §5: when the severity model is a Random Forest, OOB residuals can replace a held-out calibration set. OOB variants live alongside split-conformal scripts within each folder.