Analysis · Finance
The Simple Baseline That Most Fertility Forecasts Failed to Beat
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
Holding fertility rates constant sounds like an easy forecast to beat. Yet in a 2018 comparison by Christina Bohk-Ewald, Peng Li and Mikko Myrskylä, only four of 20 major forecasting methods consistently outperformed that baseline across the overall accuracy assessment. Two were simple extrapolation methods; two used Bayesian approaches.
The researchers evaluated methods for completing cohort fertility: estimating the average number of children women born in the same year would have. They examined 162 variants, retaining the best-performing variant of each method for the final comparison. Their main dataset covered 1,096 birth cohorts across 29 countries in the Human Fertility Database.
This was a retrospective forecasting evaluation. The team withheld later fertility observations, supplied data available up to ages ranging from 20 to 39, and predicted the remaining fertility through age 40. They compared predictions with recorded outcomes using several error measures. The baseline kept each age-specific fertility rate at its latest observed level.
The amount of history available mattered. Simple extrapolation often improved on the baseline when observations extended beyond age 30, but generally offered little advantage below age 25. Uncertainty estimates posed another problem: only eight methods supplied prediction intervals, and six substantially understated uncertainty. Of the other two, one substantially overstated it.
The main data mostly represented developed countries. A broader check using reconstructed UN data for 201 countries supported the general finding, although data quality was less consistent. The evaluation stopped at age 40 and did not establish performance for long-term fertility forecasts. Overfitting was a proposed explanation for some complex methods’ weaker results, not a demonstrated cause.
For readers assessing market forecasts, the transferable insight is a testing question: does the model beat a simple baseline on comparable historical predictions, and do its uncertainty ranges match observed errors? This fertility study supplies a useful evaluation approach, not evidence that simple models will win in every market.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Forecast accuracy hardly improves with method complexity when completing cohort fertility Christina Bohk-Ewald, Peng Li, Mikko Myrskylä
