Why this model
What the data supports.
Most of the people alive in 2050 are alive now and their ages are known, so a model that ages a real population forward is right about most of the future before any assumption is made. The assumptions it does need are few and each has a meaning: how many children women have, how long people live, and who moves.
The UN's estimates give every nation's population by single year of age and sex for every year from 1950 to 2023, with death rates and fertility rates in the same detail. That supports single years of age and one-year steps, which avoids the approximations of five-year groups, and still runs all 230 nations to 2100 in under a second.
Net migration is the exception: the estimates hold it as one number a nation a year, with no ages and no origins or destinations. So migration here is the simplest part of the model, and the weakest.
Data
What the model reads, and what it is only compared with.
- Inputs. The base population is 1 January 2024.
- Before 1950. It is shown as history and is not a model input.
- Benchmarks. The UN's projections (every variant) and the Wittgenstein Centre's (every SSP) are read only by the comparison and the calibration. The projection code does not open them.
One year
The accounting, from 1 January to 1 January.
- Half the year's net migrants are added. Arrivals follow a standard age profile (children with their parents, a peak in the twenties); departures follow that profile times the people there are to leave.
- Each cohort is survived one year by the survival ratios of that year's life table, by sex. Those aged 99 and 100 and over join the open group.
- Births are the fertility rates of ages 15 to 49 applied to the mean of the women present at the start and the end of the year, split by the nation's sex ratio at birth, and survived to the end of the year.
- The other half of the migrants are added.
Mid-year populations, which the benchmarks publish, are the mean of two 1 Januaries. World net migration is balanced to zero every year.
What moves the rates
Three levers. Each reads a function off the pooled record of all nations, 1950 to 2023, so no nation's own history decides its future alone.
Fertility. Births per woman fall at the pace the record shows for their level, once a nation's fall has begun: fastest around 4.5 to 5, slower above and below. Near the nation's long-run level the fall gives way to convergence; from below, fertility recovers towards that level by a set share of the gap each year. The long-run level is one number for the world, pulled towards a nation's own level where that is already lower. The trend of the last five years carries on and fades. The age pattern moves towards the late schedule of today's low-fertility nations as fertility falls.
Mortality. Life expectancy, by sex, gains each year what the record shows for its level (about half a year a year at 50, a fifth at 80), times a pace. Death rates by age then fall in the pattern the record shows for that level of life expectancy (among children where it is low, among the old where it is high), scaled so the life table gives the target. Where life expectancy in 2023 is more than a year below the best of the last five years (war, epidemic), that best year's death rates are the starting point.
Migration. Each nation's net flow is the median of its last twenty years, which a few years of flight or return do not move. It is then spread over corridors between nations: the UN's migrant stocks by origin say who moves where, and the matrix of flows is fitted so that every nation's arrivals less departures equal its net. In the projection each origin sends along its corridors in proportion to its people of migrating age, so a nation that grows sends more and one that ages sends fewer, and every arrival is a departure somewhere. A scenario scales what a destination admits.
Scenarios. A scenario is a plain block of numbers for these levers. It may set a value for the nations at one demographic stage (high, mid or low fertility in 2023) and may carry dated shocks to death rates, fertility or migration. The five scenarios
The central parameters
The values Continuity runs on. Those in bold were tuned; the rest are set by judgement.
Only numbers that apply to every nation alike are tuned, so the model does not lean on the benchmarks nation by nation and stands when they are taken away. The calibration in full
Known limits
What the model cannot support.
- Migration is the past carried forward. The corridors keep their recent shape and respond only to the size and age of each origin: not to wages, war or policy, and a destination does not turn to new sources as its old ones age. The size of the two-way flows is a judgement, because stocks do not record moves. Small nations with heavy migration are where the model and the benchmarks differ most.
- Contraception is thinly measured. Prevalence is known only where and when a survey was taken, and not at all for some nations; the link to fertility is an association in the record, not a mechanism.
- The base is the UN's estimate of the past. The model does not read the UN's projections, but its starting population and rates are the UN's reading of censuses and surveys. Where those are weak, so is the start.
- One pace for every nation. Fertility and life expectancy follow the pooled record. Nations with a cause of their own (HIV, conflict, a pronatalist state) are not treated differently unless a scenario says so.
- No education. The Wittgenstein Centre projects people by schooling, which drives its fertility and mortality. Ours has age and sex only; schooling enters through a scenario's settings, not through the model.
- No uncertainty bands. Each scenario is one path. The spread between the five is a spread of stories, not a probability interval.
- History before 1950 is totals only. Age structures, fertility and life expectancy before 1950 exist for a few nations and are not reconstructed for the rest.
Running it
The model is one module with scenario parameters as plain data.
From the repository root: python -m pipeline.population build runs the five scenarios and writes the stores these pages read; calibrate re-tunes the central parameters; fit re-reads the functions off the record; extract refreshes the inputs from the metrics repository. The projection itself is pipeline/population/model.py: NumPy only, no file access, one call a scenario.