Forecasting mortality rates using population composition data

In an environment where human life expectancy continues to improve, it has become increasingly challenging to produce accurate mortality forecasts. Most of the existing methods extrapolate future mortality rates from historical patterns in some way. One difficulty in mortality forecasting is the potential time-varying age effects of mortality development. This paper handles this issue by introducing an additional population composition factor into the LC model via the locally connected neural (LCN) network approach. To reduce dimensionality, population composition data are modelled as a bilinear structure of age and time effects. The population composition factor serves as an indicator of phases of demographic transition, which helps to explain the evolution of age patterns of mortality development. Our analysis indicates that with the incorporation of population composition information, the proposed mortality model produces more reasonable and accurate mortality forecasts for different age groups than the original LC model.