Embedding layer and LocalGLMnet feature selection in actuarial setting

Details

Supervisors

Hainaut, Donatien

Faculty

Faculté des sciences

Degree label

Master [120] en science des données, orientation statistique, à finalité spécialisée

Abstract

Inspired by the structure of generalized linear models, Wuthirch M. and Richman R. (2021) propose a new network architecture that shares similar features as generalized linear models, but provides a superior predictive power benefiting from the representation learning. This architecture allows for variable selection of tabular data and for interpretation of the calibrated deep learning model. The purpose of this master thesis is to develop and apply this network to predict the default probability of personnal loans but also to generalize its behaviour in an actuarial setting with response features from the exponential dispersion family. Extensions such as the integration of embedding layers to manage text features and SHAP surrogate model will be discussed and applied. We will also take the liberty of criticising certain components of the architecture and making modifications, in particular for the variable selection part which is one of the cornerstones of this modern prediction tool.