Model selection for forecasting mortality rates

The increasing life expectancy, driven mainly by improvements in sanitation, housing and education, is a positive change for individuals and a substantial social achievement but, at the same time, the implications for public spending need to be properly evaluated. Forecasting mortality appears indeed a key issue in different fields of insurance and financial markets as pricing annuity, evaluating mortality-linked securities and quantifying longevity risk. Lee-Carter proposed in 1992 a model widely used in order to forecast mortality rates. Actuarial literature has subsequently provided several extensions with the aim to correct some weaknesses of the original model. The purpose of this paper is to compare a range of both new and existing methodologies proposed over years. The study pays considerable attention to both consistency with historical data and robustness relative to the range of data employed in order to find a good balance between the goodness of fit, the simplicity of the model and the robustness of projections. At this regard, several nested methods are here compared by an application to five different countries and by analyzing several data-sets. Results confirm that some models perform better than others, but no single model can be defined as the best method. Furthermore the original version of Lee-Carter provides a good fit when a cohort effect is not significant leading to the strictest confidence interval.