Ten prior birth years gave lower Brier loss than all prior years since 1950 in listed US name counts for 2000 to 2017

We tested how the training period changes name-based probabilities for recorded sex in later US birth cohorts. We used the pinned babynames 1.0.1 table. Each test year from 2000 to 2017 used only earlier birth years. A rule based on the previous ten years gave a Brier loss of 0.020992 on 67,619,496 listed applicants. A rule based on all earlier years since 1950 gave 0.022552. Loss fell by 6.92%, with a decrease in every test year. The gain was concentrated. Loss fell for 5,769 spellings and rose for 36,823; the improving group held 69.0% of test applicants. A fixed common-name check covered 85.65% of listed applicants. Under every allowed addition of zero to four applicants to absent training cells, its loss reduction was at least 0.001351. The conservative bound for all listed names included zero. Hidden test applicants remained outside the test. The data describe one historical release and a binary recorded sex field. They do not measure gender identity or validate an individual decision. The package includes source data, probabilities, code, figures, bounds, independent checks, and repeat instructions.

names · birth cohorts · temporal transfer · Brier loss · count suppression · human behavior