Predicting Risk: Trumping Standard Practice in Auto Insurance
This thesis presents an approach for identifying extremes in auto insurance by means of risk profiles. A standard generalised linear model serves as benchmark whereas gradient boosting and backpropagation neural networks are considered as alternatives. To handle class skew, the latter methods are combined with Multi-Minority SMOTE, an alteration of standard SMOTE presented in this thesis, as well as cost-sensitivity. Considering the Area under the ROC curve, the ability to identify the high risk class and the depth of insights provided, gradient boosting combined with MM SMOTE is deemed to be the most appropriate method for predicting risk.