Parametric and XGBoost Hurdle Model for estimating accident frequency

Dafnis Krasniqi (1) , Jean-Marc Bardet (1) , Joseph Rynkiewicz (1)

1 SAMM - Statistique, Analyse et Modélisation Multidisciplinaire (SAmos-Marin Mersenne)

A well known and used method for modeling count data from exogenous variables is based on a Poisson model. Nevertheless, it has many shortcomings when faced with the problem of counting data with an excess of zeros. In this article, we focus on the Hurdle model as an alternative model. We will analyze its properties and build it as well with parametric and non-parametric estimates, notably with an XGBoost method. These new models will then be applied to a car insurance portfolio of a French insurance company, in which the number of annual accidents per driver presents a significant excess of zero accident. We show that the performance of the Hurdle model improved by XGBoost estimation is superior to that of several alternative models.

Mots clés: Insurance pricing; XGBoost algorithm; GLM; Hurdle Model

Domaines: Statistiques [math.ST]

Dafnis Krasniqi, Jean-Marc Bardet, Joseph Rynkiewicz. Parametric and XGBoost Hurdle Model for estimating accident frequency. 2022. ⟨hal-03739838v1⟩