Bias Regularization in Neural Network Models for General Insurance Pricing

Generalized linear models have the important property of providing unbiased estimates on a portfolio level. This implies that generalized linear models manage to provide accurate prices on a portfolio level. On the other hand, neural networks may provide very accurate prices on an individual policy level, but state-of-the-art use of neural networks does not pay attention to unbiasedness on a portfolio level. In fact, this is an implicit consequence of using early stopping rules in gradient descent methods for model fitting. In the present paper we discuss this deficiency and we provide two different techniques that remove this drawback of neural network model fitting.