Forecasting Cohort Fertility with a Simple Neural Network
Year: 2026
Type: preprint
Abstract: Forecasting completed cohort fertility (CCF) helps resolve ambiguities in total fertility rate (TFR) trends, yet conventional cohort forecasting methods perform little better than a freeze-rate baseline (Bohk-Ewald et al., 2018). Inspired by deep learning advances in mortality forecasting, this paper tests whether an untuned ensemble of Multi-Layer Perceptrons (MLPs), named CCF-Net, can improve cohort fertility forecasts. Using data from the Human Fertility Database and the models and evaluation criteria of Bohk-Ewald et al. (2018), CCF-Net ranks first among established methods. On average across jump-off years 1985–2013, CCF-Net forecasts higher CCF than other methods; it appears to better capture rises in fertility among some cohorts born in the late 1970s and early 1980s. However, its accuracy varies considerably across jump-off years, and it falls behind other methods in 2006–2011. When trained on data through 2019, CCF-Net forecasts substantially and perhaps unrealistically lower CCF than all benchmarks, which raises doubts about whether its historical performance will continue.
Source: SocArXiv (OSF Preprints)
Author: Paige Park
Institution: University of California, Berkeley
Language: English
Subfield: Demography
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225875853
