Geographically Weighted Regression as a Tool in Estimating Auto Insurance Loss Cost
by Mazdak Moini © 2007

The Canadian property and casualty industry has not seen heavy application of spatial statistics. For this reason, there is an underlying need for case studies that serve as pedagogical material to business executives. The basis for a spatial model is that areas that are closer together are more likely to be similar than areas that are farther apart. In insurance, this means that we expect adjacent areas to have similar levels of underlying risk. We have applied the modified version of a widely used statistical method, regression analysis, to a central business problem. We propose an exploratory method of spatial data analysis that allows spatial variation to be examined, and serves as a method for predicting automobile claim loss cost at a finer geographic level than previously contemplated. Geographically weighted regression and a family of predictive variables are used to predict auto insurance loss cost at the Forward Sortation Area level of geography in southern Ontario.