Revealing Geography-Driven Signals in Zone-Level Claim Frequency Models: An Empirical Study using Environmental and Visual Predictors
2026-04-23 • Machine Learning
Machine Learning
AI summaryⓘ
The authors looked at how to use geographic information to better predict car insurance claims when exact location data is limited. They combined zone-level data with environmental features from maps and aerial images, testing different types of models. They found that adding geographic features generally improved prediction accuracy, especially when combining map coordinates and environmental data at a 5 km scale. Image data helped only when environmental features were missing. Their work shows that it’s not just model complexity but how geography is represented that matters for predicting insurance claims.
Motor Third Party Liability (MTPL)claim-frequency modelsactuarial modelingOpenStreetMapCORINE Land Coverzone-level modelinggeneralized linear models (GLMs)gradient-boosted treesconvolutional neural networksvision-transformer embeddings
Authors
Sherly Alfonso-Sánchez, Cristián Bravo, Kristina G. Stankova
Abstract
Geographic context is often consider relevant to motor insurance risk, yet public actuarial datasets provide limited location identifiers, constraining how this 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 such constraints. Using the BeMTPL97 dataset, we adopt a zone-level modeling framework and evaluate predictive performance on unseen postcodes. Geographic information is introduced through two channels: environmental indicators from OpenStreetMap and CORINE Land Cover, and orthoimagery released by the Belgian National Geographic Institute for academic use. We evaluate the predictive contribution of coordinates, environmental features, and image embeddings across three baseline models: generalized linear models (GLMs), regularized GLMs, and gradient-boosted trees, while raw imagery is modeled using convolutional neural networks. Our results show that augmenting actuarial variables with constructed geographic information improves accuracy. Across experiments, both linear and tree-based models benefit most from combining coordinates with environmental features extracted at 5 km scale, while smaller neighborhoods also improve baseline specifications. Generally, image embeddings do not improve performance when environmental features are available; however, when such features are absent, pretrained vision-transformer embeddings enhance accuracy and stability for regularized GLMs. Our results show that the predictive value of geographic information in zone-level MTPL frequency models depends less on model complexity than on how geography is represented, and illustrate that geographic context can be incorporated despite limited individual-level spatial information.