Coherent mortality forecasting via the aggregate measure of life expectancy

Tiziana Torri, Università di Roma "La Sapienza" and Max Planck Institute for Demographic Research

Modeling age-specific death rates provides insights into mortality dynamics over age and time; such trends are less evident when directly working on aggregate measures such as life expectancy. On the other hand, working with disaggregated data does not guarantee the coherence of the outcomes obtained in terms of future life expectancy. As consequence, forecasting mortality rates frequently leads to underestimation of future life expectancy. In this paper we intend to model directly the time series of life expectancy. We first apply the classic univariate ARIMA model and look at the different future results obtained changing the time window under analysis. Given the high variability of the results, we test for the presence of structural breaks in the series, which would return biased results. The use of a more dynamic model as the Structural Time Series Models seems also advisable. The performance of our approaches is assessed through back-testing on the data.