Geographic ratemaking with spatial embeddings

Our paper on geographic ratemaking was recently published online in the ASTIN Bulletin. We present a method to leverage external and unstructured spatial data to improve spatial ratemaking.

We decompose the ratemaking process into two steps. First, we train spatial embeddings that capture spatial effects from the external data. The embeddings capture the spatial nature of the data. Then, we use the spatial embeddings as input variables to a generalized linear model.

The great advantage of our approach is that the regression model does not require a spatial component, since the spatial embeddings exhibit all desirable attributes of spatial models.

Link to paper

Here are the slides I prepared for the annual meeting of the Statistical Society of Canada, which earned the Actuarial Science Student Research Presentation Award.