Geographic segmentation of a rate, built by spatially smoothing a model's residual claims experience to separate the risk signal from the noise of sparsely populated areas.
Place of residence is among the most predictive criteria in motor and homeowners lines, but it raises a problem the others do not: it holds tens of thousands of levels, many carrying three policies and no claims. Estimating a coefficient per municipality would return noise. Territorial rating solves this in three steps. First fit a model on every other criterion, which isolates a residual per municipality, the share of claims experience geography explains. Then smooth that residual across space, by a weighted average of neighboring municipalities or by a spatial model, with weights given by exposure volume so that a sparsely populated area borrows information from its neighbors. Finally group the smoothed municipalities into a small number of homogeneous zones by clustering, to obtain a communicable table. Two cautions apply: territorial rating also captures socioeconomic variables correlated with geography, which exposes it to disparate impact challenges, and it dates quickly when urbanization shifts.
French homeowners territorial rating, 2026 calibration, 34,900 municipalities reduced to twelve zones. Before smoothing, 61% of municipalities carry fewer than fifty policies and their raw residual ranges from minus 70% to plus 180%. After exposure-weighted spatial smoothing and clustering, the spread between the cheapest and the most expensive zone is 2.3, and validation on the following year shows predicted against actual claims experience within 4% per zone.
zonage tarifaire, territorial rating, zonier tarifaire, lissage spatial du risque