Researchers have published a new global dataset estimating bilateral migration flows between all countries from 1960 through 2024, providing what they describe as the first comprehensive, sex-disaggregated record spanning more than six decades. The dataset, released in *Nature Scientific Data*, combines historical migrant stock data with demographic accounting and machine learning to estimate directional migration flows—who moved from where to where—for 234 countries and territories across 65 years.
The dataset covers migration flows at intervals matching available stock data: every ten years from 1960 to 1990, every five years from 1990 to 2020, and a final snapshot in 2024, with all estimates eventually downscaled to annual resolution. According to the report, the harmonized time series reveals that the global number of international migrants grew continuously from 1960 to 2024, with separate tracking for female and male populations. The authors trained an ensemble of 100 neural networks on migration flow observations from 2019–2022 (derived from Facebook data published by Chi et al.) to estimate gross migration—the total bidirectional movement between country pairs—which they then combined with stock-based net migration calculations to determine directional flows. The model achieved an R-squared of 0.83 when comparing predicted versus observed gross flows in logarithmic space for the 2019–2022 training period.
The researchers applied a two-step "Delta method" to reconcile inconsistencies between UN DESA migrant stock data (available from 1990 onward at five-year intervals) and older World Bank estimates (covering 1960–2000 at ten-year intervals). The authors write that because the two datasets used different collection and processing methods, reported migrant stocks for the same country pairs in overlapping years (1990 and 2000) frequently disagreed. They bias-corrected the World Bank data to align with UN DESA figures, operating on migrant shares relative to destination-country populations rather than absolute numbers to prevent corrected stocks from exceeding total populations in small countries. They then used closed demographic accounting—adjusting for births and deaths—to estimate bilateral net migration flows from successive stock observations, ensuring that derived flows remained consistent with both the stock data and country-level net migration totals reported by the UN.
The methodology addresses a long-standing problem in migration research: most driver-based flow estimates incorporate economic, social, or political variables as predictors, which means any subsequent analysis linking migration to those same factors risks circular reasoning. This dataset avoids that trap, the authors explain, by estimating flows exclusively from migration and demographic variables—migrant stocks, births, deaths, and population sizes—without using GDP, conflict indicators, climate data, or other non-migration drivers. The model's four key predictors were all stock-based: the lower bound of gross migration derived from net flows by country of birth, the bilateral net migration flow, the total bilateral migrant stock, and a measure of imbalance between the two migrant populations. The stock-based lower bound showed the strongest influence on model output at 69 percent, followed by total bilateral stock at 18 percent, with the two directional variables contributing 8 percent and 5 percent respectively.
The dataset is designed to enable researchers to analyze relationships between migration flows and potential drivers—economic shocks, policy changes, environmental disasters—without the risk of merely recovering correlations embedded in the flow estimates themselves. The authors note that while earlier stock-based methods provided multi-year flow estimates, this approach uniquely combines historical depth back to 1960, global geographic coverage, annual resolution, and sex disaggregation. The full dataset and code are publicly available, allowing researchers to examine migration trends across specific corridors, test hypotheses about what drives migration timing and volume, and build forecasts grounded in observed patterns rather than assumed relationships.