An 80-year history input reduced squared log-share error by 0.52% in ten-year ridge forecasts of US names for 2007–2017

We tested whether old name popularity improves ten-year forecasts beyond the previous 40 years. We used pinned US Social Security applicant name records and matching totals from babynames 1.0.1. The main test had 34,951 name, recorded sex, and forecast start cases, with at least 100 applicants at forecast start. Forecast start years were 1997–2007. Each ridge model used only training outcomes and inputs known by its start year. Adding a popularity input from 80–84 years earlier reduced mean squared log-share error from 0.52504 to 0.52231, or 0.52%. A 95% name-cluster resampling interval was -0.35% to 1.40%. It held the 11 tested target years fixed. The reduction was -1.10% when the threshold was 50 applicants and 1.53% when it was 200. Inputs from 40–44 and 60–64 years earlier gave reductions of 0.34% and 0.60%. The tested 80-year input thus added little mean forecast value, with no clear advantage over the other history windows. These reconstructed forecasts use a later source snapshot. They do not identify family memory or exclude other revival processes. The package preserves source files, code, all predictions, validation scores, fitted parameters, resampling draws, figures, and reproduction instructions.

name choice · cultural change · forecasting · historical memory · Social Security · open data