Mortality forecasting using factor models: Time-varying or time-invariant factor loadings?

Many existing mortality models follow the framework of classical factor models, such as Lee–Carter model and its variants. Latent common factors in are defined time-related indices (such κt model). Factor loadings, which capture linear relationship between age variables latent βx model), assumed to be time-invariant framework. This assumption is usually too restrictive reality datasets typically span a long period time. Driving forces medical improvement certain diseases, environmental changes technological progress may significantly influence different variables. In this paper, we first develop with time-varying loadings (time-varying model) an extension for modelling. Two forecasting methods extrapolate local regression method naive method, proposed model. From empirical data analysis, find that new can feature improve over horizons countries. Further, propose novel approach based on change point analysis estimate optimal ‘boundary’ short-term long-term forecasting, favoured by respectively. Additionally, simulation studies provided show performance under various scenarios.