{
 "name": "back-cast, men's life expectancy as women's minus a gap that narrows: does forecasting the gap between the sexes remove the centre bias for men?",
 "registered": "2026-10-04",
 "honesty": "The idea comes from results-backcast-e0-shocks.json, which showed men's 2019 values above their intervals (50 of 54 misses), so 2019 has been seen in that sense. The version was chosen on earlier origins only (develop-gap.ts: fit to 1984 scored 2003, fit to 1979 scored 1998; nothing after 2003 read). Men on their own (e0.ts robust) against this version, MAE all countries / without the generalised-HIV list: 1984->2003 3.243 / 2.265 -> 3.146 / 2.180; 1979->1998 3.760 / 2.880 -> 3.458 / 2.560. Of the eight gap variants tried (hinge 70 or 75, trimmed refit or not, own effect or not), this one was best or within 0.04 of best in all four. Frozen before the 2019 values are read.",
 "data": "WPP 2024 life expectancy at birth by sex, the 155 countries of 1M+ people in 2000 (wpp.ts), years up to 1999 for fitting, 2019 for scoring",
 "method": "gap.ts with hinge 75, trim true, own true: the next year's gap (women minus men) is fitted by least squares on women's level, the current gap and women's level above 75, pooled over every country and year up to 1999, refitted once without residuals beyond 3 median absolute deviations; plus the median of the country's last 15 residuals within 3 MAD, shrunk by 15/30 and halving every 10 years. Iterated along women's frozen crisis-robust forecast (e0.ts robust: true, as backcast-e0-robust.json). Men's forecast is women's minus the gap.",
 "comparator": "men's own crisis-robust forecast (e0.ts robust: true), the frozen centre of backcast-e0-robust.json and backcast-e0-shocks.json",
 "scored_sets": "all 155; the 135 outside bayesLife's generalised-HIV list; of those, the ones not on the shock list (both lists read from results-backcast-e0-bayes.json, fixed before this registration)",
 "outcomes": {
  "primary": "men's 2019 absolute error, gap version against the comparator, paired by country over all 155: the gap version has the lower mean absolute error AND wins more countries than it loses, one-sided sign test p < 0.05",
  "also": "the same on the 135 and on the non-shock subset; 2009; mean signed error (the centre bias) in each set; the 10 worst countries; whether men's MAE, with women's frozen centre unchanged, reaches target 2 as revised (at most 1.2 without HIV or shocks, at most 1.8 overall on the 135)",
  "decision": "the claim 'men as women minus a gap removes the centre bias' is supported only if the primary holds; reported whatever happens"
 },
 "seed": 20261004
}