Assessing the accuracy of ARIMA models to forecast population at the intraurban level in Greater Mexico City

Population forecasts at the intraurban level are crucial for governments and businesses to implement spatial planning and identify potential markets. However, population forecasts are often produced at coarse scales, namely national and regional, and state of the art small area forecasting methods are scarce and adjusted to work on Global North countries. We aim to explore the accuracy of autoregressive integrated moving average (ARIMA) models to forecast population totals and age-group populations at the intraurban level for the municipalities of Greater Mexico City. We use WorldPop 2000-10 population estimates as input data and perform retrospective forecasts for 2011-20. Comparing our forecasts with actual population counts, we found that ARIMA models return a median error of 1.5% when forecasting population totals in a 10-year horizon, and 5.7% for population forecasts by age-group. The age-groups 0-14, 15-39 and 40-64 performed particularly well, with error levels below 5% in most municipalities, while the age-group 65+ displays higher errors, typically over 20%. We conclude that ARIMA models outperform accuracy levels of small area population forecasts from previous studies which use different methods across the Global North. Yet, ARIMA models have limitations. They are powerful when rates of population change are linear or unchanged but cannot predict sudden changes in demographic components or ageing processes in cohorts with different population size.