Predicting motor policy loss – a ZAIG model or a two stage neural network approach?

Abstract

Artificial neural networks have increasingly being applied to solve problems which traditionally would have fallen under the domain of more classical statistical methodology, and the latter has long been a staple of popular actuarial methodology. We aim to compare a two-stage artificial neural network approach with the zero-adjusted inverse Gaussian model for predicting the claim of a motor insurance policy, which is a popular method with actuaries. The performance of both approaches is analysed by means of K-fold cross-validation. The conclusion reached is that our approach provides a comparable, if not superior, overall performance in predicting policy loss which is more robust to extreme observations.