POINT AND INTERVAL FORECASTS OF DEATH RATES USING NEURAL NETWORKS

Abstract The Lee–Carter model has become a benchmark in stochastic mortality modeling. However, its forecasting performance can be significantly improved upon by modern machine learning techniques. We propose convolutional neural network (NN) architecture for rate forecasting, empirically compare this as well other NN models to the and find that lower forecast errors are achievable many countries Human Mortality Database. provide details on forecasts of our make it more understandable and, thus, trustworthy. As default only yield point estimates, previous works applying them modeling have not investigated prediction uncertainty. address gap literature implementing bootstrapping-based technique demonstrate yields highly reliable intervals model.