Boosted Regression Trees for Small-Area Population Forecasting

Small-area population forecasting, such as the forecasting of age/gender groupings at the level of US Census Tracts, is challenged by thorny issues including (1) small population sizes, (2) frequent and sometimes directionally opposing shifts in population dynamics between censuses, (3) data availability, and (4) the ongoing evolution of the US census geographies. It is, therefore, not surprising that evaluation studies suggest wide-ranging forecast errors. Estimates vary between lows between 10% and 20% and highs sometimes exceeding 100% within any given age/gender interval. Despite its successes, only recently have population forecasters begun to explore the possibilities presented by machine learning. Using 1990 and 2000 census data, we develop 10-year age/gender-structured 2010 population forecasts for 50,965 census tracts in the U.S. using a well-known machine learning technique: boosted regression trees. Using standard ex post facto measures of forecast error (MAPE, MALPE, and MAPE-R), we demonstrate that forecasts based on “out-of-the-box” boosted regression trees have greater accuracy and produce fewer and less extreme outliers than comparison forecasts produced by the Hamilton-Perry method (reported in Baker et al. in Population Res Policy Rev 40:1341–1354, 2021. https://doi.org/10.1007/s11113-020-09601-y ).