Uncertain population forecasts Nico Keilman Department of Economics, University of Oslo
Main points Uncertainty in forecasts of certain population variables surprisingly large Forecasts for the young and the old age groups are the least reliable Forecast errors increase as forecast interval lengthens European forecasts have not become more accurate during the past 2-3 decades Traditional forecasts with their high and low scenarios do not give a correct impression of uncertainty probabilistic forecasts
Focus National forecasts in industrialized countries (to a large extent)
Where does uncertainty manifest itself? Forecasts of: Total population Age structure Fertility Mortality Migration
Measuring uncertainty Empirical findings – historical forecasts evaluated against actual population numbers (ex post facto)
Total population size fairly accurate Forecasts of population size -all countries of the world -made by the UN, the World Bank, and the US Census Bureau between 1972 and 1994 were too high by, on average, -0.8 %, 5 years ahead -2.4 %, 15 years ahead -3.5 %, 25 years ahead
Characteristic age pattern
Errors in age structure forecasts Europe
United Kingdom - men
United Kingdom - women
Young age groups fertility Old age groups mortality
Uncertain Population of Europe (UPE) Joint work with Juha Alho, Harri Cruijsen, Maarten Alders, Timo Nikander, Din Quang Pham Evaluated historical accuracy of population forecasts -national agencies in 14 European countries One (of several) source of information for probabilistic forecasts
European forecasters have under-predicted gains in life expectancy: - by 2.3 years of life for forecasts 15 years ahead - by 4.5 years of life for forecasts 25 years ahead
European forecasters have predicted too high fertility: - by 0.2 children per woman for 15 years ahead - by 0.4 children per woman for 25 years ahead
European forecasters have predicted too low levels of migration: - by 1 per thousand of population for 6-8 years ahead - by 3 per thousand of population for years ahead
Why uncertain? Data quality Social science predictions No accurate behavioural theory Rely on observed regularities instead Problems when sudden trend shifts occur assumption drag
Assumption drag: fertility
Assumption drag: mortality
Forecast accuracy has NOT improved over the last 25 years
Error indicator for TFR forecasts, 14 countries The graph shows estimated forecast effects in a model that also controls for period, duration, country, and forecast variant. Log of absolute error in TFR is dependent variable. Estimates in black, 95% confidence intervals in red. Launch years are reference category for the forecast effects. R 2 = 0.704, N = 4847
Error indicator for e0 forecasts, 14 countries The graph shows estimated forecast effects in a model that predicts the log of absolute error in e0. The model controls for period, duration, country, sex, and forecast variant. Estimates in black, 95% confidence intervals in red. Launch years are reference category for the forecast effects. R 2 = 0.722, N = NB No data for launch years
Three problems related to deterministic population forecasts 1. Wide margins for some variables, narrow margins for others
Example: Old Age Dependency Ratio (OADR) for Norway in 2060 Source: 2005-based forecast of Statistics Norway HighMiddle Low|H-L|/M millions % POP POP OADR (!)
Problems … (cntd) 2. Too narrow margins in the short run, too wide margins in the long run
Problems … 3. A limited number of variants, without probability statements, leave room for politically motivated choices.
Views about the demographic future have changed over time
TFR assumptions for 18 EEA+ countries, Averages across countries
Life expectancy assumptions for 18 EEA+ countries, Men Averages across countries
Net migration assumptions for 18 EEA+ countries, Averages across 18 EEA+ countries (UN, UPE), across 15 EU-15 countries (Eurostat)
Implications Forecast users should be informed about the reliability of the future population numbers Historical errors just a first step Expected errors for the current forecast probabilistic forecasts UPE: probabilistic forecasts for 18 European countries. See
Forecast users should be prepared for the unexpected - use buffers? - flexibility? - risk aversion?
Users should check whether overpredictions are more costly, or less costly, than underpredictions Loss function Forecasters should educate the users, cf. -weather forecasts: EPS (Ensemble Prediction System) Meteograms: series of Box plots -inflation and interest rate forecasts: uncertainty fans
Bank of Norway’s forecast of future interest rate (%) with uncertainty fan 30%50%70%90% Source: Norges Bank