Disability insurance : predicting accurate inflow probabilities in an imbalanced data setting

A core business of insurance companies is to set a premium for the policy holders. To set an accurate premium, one needs to model the probability of the insured to produce a claim. In the disability insurance branch for high income employees in the Netherlands, this poses a problem due to the imbalance regarding claims and non-claims. The current pricing model applies two features and is a relatively simple model. In this research we use more advanced machine learning techniques to predict the probability of a disability with more accuracy. We add external and internal data to improve the predictions. We conclude that using a gradient boosting model improves the ability to seperate claim from non-claim data while retaining the accuracy of the probabilities. Furthermore, the addition of features such as geographical data can help increase the performance of machine learning models.