Increasing computing power is opening up a wide range of predictive modelling techniques to actuaries, from the fields of traditional statistics and data mining. These techniques are often quite different and difficult to compare directly.
This paper looks at some of the main classes of model available, and compares their performance on a number of real general insurance modelling problems. Performance is measured both in terms of predictiveness, using objective measures – and more subjectively on other qualities that might collectively be described as “elegance”. The models investigated include table-based analysis, generalised linear models, generalised additive models, decision trees, and neural nets.