Bayesian mortality forecasting with overdispersion

The ability to produce accurate mortality forecasts, accompanied by a set of representative uncertainty bands, is crucial in the planning of public retirement funds and various life-related businesses. In this paper, we focus on one of the drawbacks of the Poisson Lee–Carter model (Brouhns et al., 2002) that imposes mean–variance equality, restricting mortality variations across individuals. Specifically, we present two models to potentially account for overdispersion. We propose to fit these models within the Bayesian framework for various advantages, but primarily for coherency. Markov Chain Monte Carlo (MCMC) methods are implemented to carry out parameter estimation. Several comparisons are made with the Bayesian Poisson Lee–Carter model (Czado et al., 2005) to highlight the importance of accounting for overdispersion. We demonstrate that the methodology we developed prevents over-fitting and yields better calibrated prediction intervals for the purpose of mortality projections. Bridge sampling is used to approximate the marginal likelihood of each candidate model to compare the models quantitatively.