Mortality improvement neural-network models with autoregressive effects

Abstract

In this paper, we propose a neural network (NN) architecture of mortality improvement model with cohort effect. We then extend the mortality improvement NN model to consider autoregressive effects, which allows mortality improvement to depend on the lagged mortality rates. The advantage of our NN model setup is that the parameters of period and cohort effects are implicitly estimated by the NN models, and hence, the mortality projection can be obtained without taking the extra steps of selecting and estimating the suitable time-series model for period and cohort effects. Our empirical results suggests that, based on 48 populations in the Human Mortality Database with complete sets of observations from 1950 with the age span of 55–90, the NN models with cohort and autoregressive effects improve the forecast accuracy of mortality rate projections and provide better prediction performance.