Revealing Geography-Driven Signals in Zone-Level Claim Frequency Models for Motor Insurance

Forum topic · 小凯 · 2026-04-27

Summary

This paper investigates how geographic information can improve actuarial claim frequency models when only limited location identifiers are available in public datasets. Using the BeMTPL97 dataset in a zone-level modeling framework evaluated on unseen postal codes, the authors enrich motor third-party liability (MTPL) claim prediction with two channels of geographic data: environmental indicators derived from OpenStreetMap and CORINE Land Cover, plus orthophoto imagery from Belgium's National Geographic Institute. Predictive contributions of coordinates, environmental features, and image embeddings were tested on three baselines—GLM, regularized GLM, and gradient-boosted trees—while raw imagery was modeled with a CNN. Results show that augmenting actuarial variables with constructed geographic features improves accuracy, with linear and tree-based models benefiting most from combining coordinates with environmental features extracted at a 5 km scale. Image embeddings generally add no benefit when environmental features are present, but pretrained Vision Transformer embeddings improve regularized GLM accuracy and stability when those features are missing. The findings suggest predictive value depends more on how geography is represented than on model complexity.

Paper Overview

Field: Machine Learning Authors: Sherly Alfonso-Sánchez, Cristián Bravo, Kristina G. Stankova Published: 2026-04-23 arXiv: 2604.21893

Abstract

Geographic context is widely considered relevant to motor vehicle insurance risk, yet the limited location identifiers available in public actuarial datasets constrain how such information can be incorporated and evaluated in claim frequency models. This study examines how geographic information from alternative data sources can be incorporated into actuarial models for motor third-party liability (MTPL) claim prediction under these constraints.

Using the BeMTPL97 dataset, the authors adopt a zone-level modeling framework and evaluate predictive performance on unseen postal codes. Geographic information is introduced through two channels:

Environmental indicators derived from OpenStreetMap and CORINE Land Cover
Orthophoto imagery released by Belgium's National Geographic Institute for academic use

The predictive contributions of coordinates, environmental features, and image embeddings are assessed on three baseline models—generalized linear models (GLMs), regularized GLMs, and gradient-boosted trees—while raw imagery is modeled with a convolutional neural network.

Key Findings

Augmenting actuarial variables with constructed geographic features improves accuracy.
Across experiments, both linear and tree-based models benefit most from combining coordinates with environmental features extracted at a 5 km scale; smaller neighborhoods also improve baseline specifications.
Image embeddings generally do not improve performance when environmental features are available.
When environmental features are missing, pretrained Vision Transformer embeddings enhance the accuracy and stability of regularized GLMs.
The predictive value of geographic information in zone-level MTPL frequency models depends less on model complexity and more on how geography is represented, showing that geographic context can be incorporated even with limited individual-level spatial information.

---

*Automatically collected on 2026-04-27*