Neural networks vs GLMs in pricing general insurance

Document description

Neural Networks are often referred to, with awe, as some mysterious representation of the human brain that can solve problems. They have also been referred to in previous GISG papers as having potential applications to general insurance pricing or reserving. The purpose of this paper is to remove some of this awe by explaining what Neural Networks are, how they compare with traditional statistical models, and consider what scope there is for their use in general insurance. The paper is in three main sections. The first section describes what Neural Networks are. The second section briefly describes GLMs, and makes a few observations on the practical nitty gritty of using such models. The third section compares the two approaches from a theoretical perspective and with some practical examples based on real general insurance data. Finally, some references to Neural Network software and publications are given for anyone interested in pursuing the topic further. The practical examples presented in this paper are only half finished. The University of Edinburgh's Artificial Intelligence department will be training some Neural Networks with the data during the summer. The results will be compared with some PC-based Neural Network models and the results of traditional modelling techniques in a further paper to be presented at the conference.