Selection of variables for automobile insurance rating

Abstract

This paper uses generalized linear models to assess risk in an automobile insurance portfolio. The effect on the model fitting process of the aggregation of individual policies into homogeneous classes is examined. Problems of scale are discussed with respect to (a) the size of the data set, (b) the size of a model involving many factors, (c) the number of interaction parameters where rating factors have large numbers of levels. The methods are illustrated using a data set from a UK automobile insurer.