Forecasting Mortality with Autoencoders: An Application to Italian Mortality Data

Predictions of human survival probabilities are an extremely relevant topic in many fields of human activities and interests, including in particular the insurance field. The model considered the most reliable, and, for this reason, most widely used both in the literature and in practical applications, is the Lee–Carter model. In this paper, we propose to project survival probabilities making use of Autoencoders. Empirical evidence on real data shows that the ability of autoencoders to model highly nonlinear structures leads to significant improvements in prediction accuracy over the widely used Lee–Carter model.