Forecasting mortality rates of Visegrád group countries by neural networks

Abstract

In our paper, we apply recurrent neural networks, which have been used with great success in many fields in recent years, to predict age-specific mortality rates for ages 18 to 99 years for the Czech Republic, Poland, Hungary and Slovakia between 1970 and 2019, separately for women and men. We investigate the popular Recurrent Neural Network, Long-Short Term Memory and Gated Recurrent Unit architectures, and put a special emphasis on the optimization of the hyperparameters of the networks by applying cross-validation and splitting the base period into learning and testing subsets. We compare our predictions with those of both the classical Lee--Carter and the coherent Li--Lee multipopulation models in terms of accuracy, and determine which procedures are able to generate the most reliable projections for each country, gender and age group. Our models can be applied in practice by life, pension and health actuaries as well as demographers.