Abstract

To date, in the insurance industry, the premium for a given risk is based on the expected claim amount since the models used are only meant to calculate a mean response. However, getting more information about the distribution of each single risk would be useful for risk assessment. A method in Neural Networks (NN) called Tractable Approximate Gaussian Inference (TAGI) by Goulet et al. (2020) allows to study each response individually since each output follows its own Normal distribution. The main contributions of this thesis are to make this technique available through an open source package, to apply TAGI in insurance and compare it to other techniques and to study risk measures with it.