Insights from Inside Neural Networks

Abstract

We provide a tutorial that illuminates the aspects which need to be considered when fitting neural network regression models to claims frequency data in insurance. We discuss feature pre-processing, choice of loss function, choice of neural network architecture, class imbalance problem, balance property and bias regularization as well as over-fitting. This discussion is based on a publicly available real car insurance data set.

Keywords: Neural Networks, Architecture, Over-Fitting, Loss Function, Dropout, Regularization, LASSO, Ridge, Gradient Descent, Class Imbalance, Car Insurance, Claims Frequency, Poisson Regression Model, Machine Learning, Deep Learning

JEL Classification: G22, C10, C13, C14, C67