Blog · Research · Oct 21, 2026 · 6 min
Living Worlds' life expectancy forecasts, frozen twenty years back and scored once on years no run had seen: met for women and for men on the latest national figures, not met against 2024 national figures or across 155 countries, with honest intervals for women and a stacked UN model that did not help.
How long will a child born today live, and how long will the people of a country live twenty years from now? Life expectancy at birth is one of the most watched numbers in public life. Pensions, health systems and housing plans all lean on a forecast of it, and the forecast is usually made decades before anyone can check it.
Living Worlds asks a narrower question that can be checked: if a forecast had been frozen twenty years ago, how close would it have come to the figure a national statistics office publishes today? Every population in Living Worlds lives through simulated time on stated birth and death rates, so the length of a life is something the world has to get right before anything built on it can be trusted.
Each test here follows the same discipline. The forecast is fitted on history up to a cut-off year, the rule for success is registered, and only then is the target year read. It is scored once. "A year nobody had read" means exactly that: no run, and no person tuning the engine, had looked at the figures it was scored on.
This matters because a forecast checked against data its author has already seen measures fit, and fit is cheap. A forecast checked once, on a year that was sealed when the rule was written, measures what a real forecaster would have experienced.
The first large test, Life expectancy, predicted twenty years ahead, was fitted on 1950 to 1999 and scored once on 2019 across 155 countries. It came within 2.06 years for women and 2.42 for men, against a registered target of 1.0. Not met. Most of what remained was shocks: wars, state collapse and HIV. The history it was fitted on is the UN's 2024 revision, which re-estimates the past with later censuses, so those errors are kinder than a forecast truly made in 1999 would have been.
The second, Life expectancy twenty years on, against 2024 national figures, froze the forecast in 2004 and scored it on 2024 figures from Eurostat and the OECD for 37 countries. Each forecast was a twenty-year gain added to the country's own 2004 figure. It came within 1.75 years for women and 2.13 for men, against a bar of 1.2. Not met. It expected the longest-lived countries to keep gaining as they had; Japan, Germany, Italy, France, Canada and Australia came in 2.3 to 5 years below it, while the Baltic countries did better. The protocol was amended once, before anything was read, when the OECD refused a request.
That pattern is visible in the latest European figures too. Eurostat's preliminary estimate puts EU life expectancy at birth at 81.7 years in 2024, with the largest gains since 2019 in Lithuania, Czechia, Latvia and Romania, and the Netherlands slightly below its 2019 level. The biggest recent gains came in countries that started lower.
The two misses changed what was tested next. Women's life expectancy twenty years on, a year nobody had read was registered after the 2024 test missed and before any 2025 value was read. Frozen on UN data up to 2005, it was scored on 2025 figures from Norway, Sweden, Finland and the Netherlands, the four national offices that publish a single-year figure openly. It came within 0.92 years on average, inside the bar of 1.2 and the 1.0 stretch goal. Recorded as Met (women only). All four came out high: Sweden by 0.15 years, Finland 0.87, Norway 1.25 and the Netherlands 1.39.
Men's life expectancy twenty years on, a year nobody had read followed, registered before any 2025 value for men was read and scored on six offices: Norway, Sweden, Finland, the Netherlands, Estonia and Latvia. It came within 1.09 years on average, inside the bar of 1.2, with no lean either way: 0.15 years low on average, inside a bar of 0.5. Its 80% intervals held all six countries, against three for the earlier method, with a better interval score (4.56 against 7.13). Met. The same frozen forecast is registered against the UN's 2026 revision for every country.
Two earlier steps fed into it. Men as women minus a gap forecast men through their gap to women and lowered the error from 2.42 to 2.40 years, better in 92 countries and worse in 63 (p 0.012), recorded as Primary met; bias remains, since outside HIV countries men still came out about 0.9 years low.
A forecast twenty years out needs a range that means what it says. Shocks as events treated wars, collapses and epidemics as events with their own chances. Women's 80% intervals covered 79% of countries, better in 93 of 155 (p 0.008). Men's covered 65%, centred too low. Met for women, not for men.
Honest intervals twenty years ahead, a fresh year repeated the test on a year nobody had read: fitted on 1950 to 2003, scored on 2023. Women's intervals held, with 81% of countries inside the 80% intervals and a lower interval score than the baseline (8.87 against 12.01, p below 0.00000001). Men's were too cautious for a calmer period: 89% coverage and a worse score than the baseline. Met for women, not for men. The next step is for men's width to follow each period's risk of shocks.
Countries hide a great deal. Life expectancy in European regions, ten years ahead was fitted on countries only and scored on Eurostat NUTS 2 regions. Men came within 0.57 years across 180 regions, against 1.38 for each region's own trend. Women came in at 1.10 years across 235 regions against a bar of 1.0, almost all over-forecast (+1.06), as gains stalled across Western Europe in the 2010s. Both sexes were needed, so the test stands as not met, recorded as Met for men, not for women.
The follow-up, European regions on a window nobody had read, forecast from 2004-2006 to 2014-2016. The 2014-2016 values had never been read; the starting values had. It came within 0.64 years for women and 0.33 for men, against 0.66 and 0.64 for each region's own trend, with 84% of regions within a year for women and 98% for men. Met. Women's forecasts still ran 0.51 years high on average.
The UN projects life expectancy with a Bayesian hierarchical model. The UN's Bayesian model, stacked asked whether combining it with Living Worlds' own pooled-gain model would do better than either. Across 135 countries, Living Worlds' model did better alone than stacked; neither reached the target. Not met.
Read together, these results draw a clear line. On the most recent national figures, where offices publish openly and nobody had read them, the forecasts for both women and men came inside the registered bar. Across 155 countries over twenty years, they did not: shocks dominate what is left, and the longest-lived countries stopped gaining at the pace their past suggested. Women's intervals are now honest on fresh years; men's are still being set.
The met results rest on four and six countries, the ones whose offices publish a single-year figure openly. They say nothing yet about the 155-country picture, where the registered target of 1.0 was missed. The regional window that met had its starting values read before registration. And women's forecasts lean high in every test where the direction is reported, which is a bias still to be removed.
Living Worlds is a parallel dimension where populations live through every age of human life, on each year's official rates, month by month. How long those lives last is the first thing that has to hold. Every forecast above is registered before the year it is scored on is read, scored once, and published as Met or Not met, so anyone can see which claims the world has earned and which it has yet to.
Blog
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.
Language models answering as real people, scored once on survey answers nobody had read: a frontier model met 0 of 6, prompted voices overstated births threefold, and carrying the last shares forward beat a small model. Lives fitted to records came closer, a model's estimate counted as a prior meant fewer people to ask, and subgroup splits landed inside the survey's own noise in the United States and in 30 of 31 African countries.
A population lived month by month in the United States from 1950 matched how many children Americans have had and how many brothers and sisters they report, closer than a frontier model answering from what it has read. The father's side, first counted through partnerships, missed, and became the next test.