Boosting Regression Models With Neural Networks, Because We CANN

The purpose of this thesis is to gain an understanding of neural networks and how they can be used to boost regression models. A theoretical foundation is presented, covering both generalized linear models (GLMs) and feed-forward neural networks (FNNs). This is followed by application in a non-life actuarial environment using the well-studied dataset freMTPL2freq, which contains French insurance data. GLMs are used as cornerstone models and are compared with feed-forward neural networks FNNs. Subsequently, the two are combined into a combined actuarial neural network (CANN) model in an attempt to boost the initial GLM model via skip connection, with successful results.