Artificial neural networks as an alternative to a traditional statistical modelling approach : a comparison through an insurance application

Abstract

Artificial neural networks are increasingly being applied to solve problems which traditionally would have fallen under the domain of more classical statistical methods such as regression analysis. This project aims to take a look at the theoretical foundations of such neural networks and compare them to the statistical techniques we are familiar with. Such a comparative analysis is strengthened through the application of both neural network and actuarial methodologies to the problem of predicting the loss of a motor insurance policy. Both a novel neural network approach and a conventional zero-adjusted inverse Gaussian regression approach were applied in an attempt to model the loss of policies from a local motor insurance dataset. The performance of both models was analysed by means of k-fold cross validation and the conclusion was reached that in this scenario the neural network approach provided an equivalent if not superior overall performance.