Why does the area where a policyholder lives tell us so much about their risk? In motor insurance, the link is intuitive, our driving habits and accident risks are shaped by the environments we navigate daily.
Existing research, such as Burdett et al. (2017), has highlighted the “close to home” effect, noting that a significant portion of crashes occur near a driver’s residence, often at a higher rate than driving volume alone would suggest. This implies that the residential environment (local road design, traffic density, and land use) carries a “spatial context” that is highly relevant to risk.
The Challenge: Modeling with Limited Data. While this geographic context is clearly important, researchers often face a major difficulty: public datasets frequently lack granular location data. In our latest study, we set out to see if we could still capture these vital geographic signals under these constraints by using a zone-level modeling framework.
Our Approach: We investigated whether the environment surrounding the center of a policyholder’s municipality postcode could provide the necessary predictive information. We integrated several alternative data sources and use the BeMTPL97 CAS data set:
* Environmental Indicators: Extracted from OpenStreetMap and CORINE Land Cover.
* Aerial Imagery: High-resolution images from the Belgian National Geographic Institute for academic purposes.
Our Results:
- While both linear and tree-based models benefited on average the most from environmental features extracted at a 5 km scale from the municipality center, we also found that smaller neighborhoods also improve baseline specifications. This confirms that geographic signals are present across multiple spatial resolutions.
- 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.
- These performance gains are not accidental. We conducted robustness checks using different data splits, confirming that the predictive behavior remain stable and reliable across unseen postcodes.
The Takeaway The most important lesson from this work is that the predictive value of geography depends less on the complexity of the AI model and more on how geography is represented. And also, we have shown that even limited spatial representations can significantly outperform traditional insurance variables alone.
Read the preprint in this ArXiV link. The paper is still a preprint, so please take it as a work in progress as it will evolve through the peer-review process. We welcome any feedback!
You can also hear the machine-podcast generated below
