A Practitioner-Oriented Evaluation of Mortality Forecasting Methods: The Case of Australia

Abstract

Practitioners seeking a suitable mortality model for forecasting population by age and sex are presented with many possible choices from the large and growing academic literature on mortality forecasting. Despite this abundance, there is relatively little practical guidance on selecting the most appropriate models for their needs. This study evaluates the accuracy of mortality forecasting methods and provides guidance on model selection. The evaluation includes eight methods from the StMoMo and demography R packages, and a benchmark extrapolative method based on the Ediev (2008) model. We also consider the accuracy of simple combinations of individual methods. We evaluate models by preparing mortality ‘forecasts’ for Australia for past periods using data obtained from the Human Mortality Database. For each method, we created five sets of 30-year retrospective forecasts and evaluated the accuracy of the forecast mortality rates, life expectancies at birth, and life expectancy at age 65. We also evaluated the accuracy of mortality forecasts in terms of projected total deaths calculated using a pseudo-projection method. The Age-Period-Cohort model from the StMoMo R package, based on the Cairns et al. (2009) implementation, was the standout performer in our evaluation, followed by the benchmark extrapolative method. This study presents a comprehensive evaluation of mortality forecasting methods using a variety of metrics, including a new way to evaluate mortality forecasts using a pseudo-projection method. We hope that this evaluation proves useful for practitioners looking to select a mortality forecasting method.