A case study in the application of machine learning methods in non-life insurance

Specialeforsvar ved Philip Harmuth

Titel: A case study in the application of machine learning methods in non-life insurance

Abstract: This thesis has its focus on machine learning models, methods, and concept and how we apply them to insurance data. The case we are studying is data for comprehensive car insurance from the Danish insurance company Tryg A/S for the years 2013-2017. The current model, as well as most other risks premium models in Tryg, is a Tweedie Glm. The overall goal of the case study is to examine if we can create better performing models than the current GLM. Methods for reducing the dimensionality of data are represented, as well as tools for exploring interactions between features. We then introduce the Deep Tweedie neural network, a tool to model risk premiums. The results of the case study show that the Deep Tweedie neural network performs better than both Tryg’s current GLM and better than a GBM model. We then apply an out-of-time stacking scheme to improve the out-of-time generalization of the different models created. The results show that the scheme does provide better performance, but we encounter problems with the training of the base learners and provide solutions to how one should train the base learners.

Vejledere: Jostein Paulsen & Martin Rose
Censor: Mette Magdalene Havning