Abstract
Within non-life insurance pricing, an accurate evaluation of claim frequency, also known in theory as
count data, represents an essential part in determining an insurance premium according to the
policyholder’s degree of risk. Count regression analysis allows the identification of the risk factors
and the prediction of the expected frequency of claims given the characteristics of policyholders. The
aim of this paper is to verify several hypothesis related to the methodology of count data models and
also to the risk factors used to explain the frequency of claims. In addition to the standard Poisson
regression, Negative Binomial models are applied to a French auto insurance portfolio. The best
model was chosen by means of the log-likelihood ratio and the information criteria. Based on this
model, the profile of the policyholders with the highest degree of risk is determined.