Assessing the accuracy of probabilistic population forecasts

Abstract

Stochastic demographic forecasts indicate to their users the ex ante level of uncertainty to be expected. There are approximately 25 years’ worth of data for population stocks and flows on the performance of such forecasts. The demographic context has numerous special features. The data come in the form of counts. The vital processes of births and deaths, and migration, combine to produce the future counts, so it is not obvious how their separate effects might be assessed. Finally, the ex post assessment is hampered by quality problems in the official population data. These issues are tackled in the context of a stochastic forecast of Norway for the years 2003–2023. A novel approach infers predictions of births, deaths, and net migration from changes in predicted population stocks in carefully chosen age segments. We use the deviance as a scoring rule to measure the lack of fit between the predictive distribution and the realized value. For a 20-year forecast horizon, predictive distributions for deaths in ages 80 years and over are less accurate than the birth distributions. The situation is the reverse for ages below 75 years. The accuracy of the predictive distribution for births appears to deteriorate much more rapidly than that of age-specific deaths for forecast horizons beyond 20 years. Forecasts for net migration are systematically less accurate than those for births or deaths.