{
  "name": "back-cast, life expectancy: from data up to 1999, life expectancy at birth in 2019, by sex, in every country of 1M+ people",
  "registered": "2026-10-03",
  "target": "docs/LIVES.md target 2",
  "data": {
    "source": "UN World Population Prospects 2024, WPP2024_Demographic_Indicators_Medium.csv.gz (population.un.org), columns LExFemale, LExMale, TPopulation1July",
    "countries": "every location with LocTypeID 4 (a country or area) and TPopulation1July in 2000 of at least 1,000 (thousands)",
    "fit": "years 1950-1999 only",
    "held_out": "2019 (before COVID-19); 2009 also reported as a 10-year horizon",
    "caveat": "WPP 2024 re-estimates the past with censuses taken after 1999. Every method sees the same revised past, so the comparison between methods is fair; absolute errors are kinder than a forecast made in 1999 would have been."
  },
  "methods": {
    "no-change": "e0(2019) = e0(1999)",
    "own-trend": "a straight line fitted by least squares to the country's own 1980-1999 values, extended",
    "pooled-gain": "the one being tested. The yearly gain in e0 as a function of the level of e0, learned across every country's own series of that sex in 1950-1999: the median yearly gain in each 1-year bin of e0 level (bins with fewer than 30 country-years merged with their neighbour), linearly interpolated. A country's own effect: the mean of its own residuals (actual gain minus the pooled gain) over 1985-1999, shrunk by n/(n + 15) with n = 15 years, and decaying by half every 10 years ahead. Projected year by year: e0(t+1) = e0(t) + g(e0(t)) + own effect. Fitted per sex.",
    "intervals": "pooled-gain only: 2,000 paths per country, each year's gain plus a residual drawn from the pooled 1950-1999 residuals of the same sex and e0 bin, in 5-year blocks from one country chosen at random (to keep the persistence of real shocks); the 10th and 90th percentiles of 2019 are the 80% interval"
  },
  "outcomes": {
    "primary": "mean absolute error of e0 in 2019 across the countries, per sex, per method",
    "also": "the same in 2009; root mean square error; the worst 10 countries per method; 80% interval coverage of pooled-gain in 2019",
    "decision": "target 2 is met only if pooled-gain's 2019 MAE is at most 1.0 year for BOTH sexes AND its 80% intervals cover between 70% and 90% of countries. Whatever happens, all methods are reported."
  },
  "seed": 20261003
}