Locally-coherent multi-population mortality modelling via neural networks

This manuscript proposes an approach for large-scale mortality modelling and forecasting with the assumption of locally-coherence of the mortality forecasts. In general, the coherence prevents diverging long-term mortality forecasts between two or more populations. Despite being considered a desirable property in a multi-population modelling framework, it could be perceived as a strong assumption when a large collection of countries is considered. We propose a neural network model which requires the coherence of the mortality forecasts only within sub-groups of similar populations. The architecture is designed to be easily interpretable and induces the creation of some clusters of countries with similar mortality patterns. This aspect also makes the model an interesting tool for analysing similarities and differences between different countries’ mortality dynamics and identifying opportunities for longevity risk diversification and mitigation. An extensive set of numerical experiments performed using all the available data from the Human Mortality Database shows that our model produces more accurate mortality forecasts with respect to some well-known stochastic mortality models. Furthermore, a massive reduction of the parameters to optimise is achieved with respect to the benchmark mortality models.