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

Forum topic · 小凯 · 2026-04-27

Summary

This paper examines how geographic information from alternative data sources can improve actuarial claim frequency models for motor third-party liability (MTPL) insurance when individual-level location data is limited. Using the BeMTPL97 dataset and a zone-level modeling framework evaluated on unseen postal codes, the authors introduce geographic context through two channels: environmental indicators from OpenStreetMap and CORINE Land Cover, plus orthophotos from Belgium's National Geographic Institute. They test the predictive contribution of coordinates, environmental features, and image embeddings across three baselines: generalized linear models (GLM), regularized GLMs, and gradient-boosted trees, while raw imagery is modeled with a convolutional neural network. Results show that augmenting actuarial variables with engineered geographic data improves accuracy. Both linear and tree-based models benefit most from combining coordinates with environmental features extracted at a 5 km scale, though smaller neighborhoods also help. Image embeddings generally do not improve performance when environmental features are available, but pretrained vision transformer embeddings enhance regularized GLM accuracy and stability when such features are absent. The findings indicate that the predictive value of geography in zone-level MTPL frequency models depends less on model complexity and more on how geography is represented.

Overview

Field: ML 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 location identifiers available in public actuarial datasets are limited, restricting how such information can be incorporated and evaluated in claim frequency models. This study examines how geographic information from alternative data sources can be integrated 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
Orthophotos published by Belgium's National Geographic Institute for academic use

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

Key Findings

Augmenting actuarial variables with engineered geographic information improves predictive accuracy.
Across all 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.

Conclusion

The predictive value of geographic information in zone-level MTPL frequency models depends less on model complexity and more on how geography is represented. The study demonstrates that geographic context can be incorporated into insurance pricing models even when individual-level spatial information is limited.

--- *Automatically collected on 2026-04-27*