A description of the modelling process and results for the risk analysis of the motor portfolio for a major Australian insurer. Components of the process include: 1. The use of Treenet for variable filtering to reduce 2,500 potential predictors to around 100; 2. Sampling and stratification strategies to manage data size without sacrificing signal; 3. The filtering of correlated variables; 4. Overall measures of model performance; 5. The fitting of a main effects GLM with key interactions; 6. Benchmarking the effect of interactions; 7. A GLM brute force search for interactions followed by iterative Treenet searches; 8. The need (or not) for geographic and vehicle category overlays. Although the need for confidentiality limits our ability to show the fine detail of the process, we illustrate key results using a model of claim frequency for at-fault collision claims.

Moderator:
Peter J. Mulquiney, Actuary, Taylor Fry Consulting Actuaries

Panelist:
Richard Brookes, Actuary, Taylor Fry Consulting Actuaries