Point and interval forecasts of age-specific life expectancies: A model averaging approach

BACKGROUND Any improvement in the forecast accuracy of life expectancy would be beneficial for policy decision regarding the allocation of current and future resources. In this paper, I revisit some methods for forecasting age-specific life expectancies. OBJECTIVE This paper proposes a model averaging approach to produce accurate point forecasts of age-specific life expectancies. METHODS Illustrated by data from fourteen developed countries, we compare point and interval forecasts among ten principal component methods, two random walk methods, and two univariate time-series methods. RESULTS Based on averaged one-step-ahead and ten-step-ahead forecast errors, random walk with drift and Lee-Miller methods are the two most accurate methods for producing point forecasts. By combining their forecasts, point forecast accuracy is improved. As measured by averaged coverage probability deviance, the Hyndman-Ullah methods generally provide more accurate interval forecasts than the Lee-Carter methods. However, the Hyndman-Ullah methods produce wider half-widths of prediction interval than the Lee-Carter methods. CONCLUSIONS Model averaging approach should be considered to produce more accurate point forecasts. COMMENTS This study is a sequel to another Demographic Research paper by Shang, Booth and Hyndman (2011), in which the authors compared the principal component methods for forecasting age-specific mortality rates and life expectancy at birth.