Model uncertainty approach in mortality projection with model assembling methodologies

Forecasting mortality rates has become a key task for all who are concerned with payments for non-active people, such as Social Security or life insurance firms managers. The non-ending process of reduction in the mortality rates is forcing to continuously improve the models used to project these variables. Traditionally, actuaries have selected just one model, supposing that this model were able to generate the observed data. Most times the results have driven to a set of questionable decisions linked to those projections. This way to act does not consider the model uncertainty when selecting a specific one. This drawback can be reduced through model assembling. This technique is based on using the results of a set of models in order to get better results. In this paper we introduce two approaches to ensemble models: a classical one, based on the Akaike information criterion (AIC), and a Bayesian model averaging method. The data are referred to a Spanish male population and they have been obtained from the Human Mortality Database. We have used four of the most widespread models to forecast mortality rates (Lee-Carter, Renshaw-Haberman, Cairns-Blake-Dowd and its generalization for including cohort effects) together with their respective Bayesian specifications. The results suggest that using assembling models techniques gets more accurate predictions than those with the individual models.