Deep learning four decades of human migration

Main

The movement of people—within countries and between them—is an important topic across multiple domains. Migration drives demographic change, shaping the size and composition of populations; it can influence labour markets 1 , inform social policy 2 and is a popular topic for public debate 3 . Although migration often follows long-term trends driven by development 4 , it can be dramatically altered by short-term shocks—armed conflict, famine, natural disasters, political instability, changes in national borders, peace agreements or independence movements 5 .

Human migration, however, is notoriously difficult to define and track 6 . Current analyses of global migration systems rely heavily on migrant population data published at five-year intervals by the United Nations (UN) and at ten-year intervals by the World Bank. These datasets provide counts of migrants in each country by country of birth, typically referred to as stock data. Although relatively straightforward to collect, they offer only a snapshot at a fixed point in time and provide limited insight into the temporal dynamics of migration: migrants may have arrived immediately before the observation point or several decades earlier. To better capture migration dynamics, researchers have developed methods that estimate migration flows over multi-year periods by comparing changes in migrant stocks at the beginning and end of each interval 9 . However, as these estimates are tied to gaps in the underlying stock data, the resulting five- or ten-year estimates inevitably smooth or completely miss movements that occur in the intervening years. What researchers on global migration ideally need are annual flow data for all countries. Such data would allow them to track the tempo of migration systems with far greater precision, integrate migration patterns with other annually reported datasets on drivers such as economic change, conflict, climate or policy reforms, feed into annual population projection models, and facilitate both causal and comparative analyses across countries and regions. Yet existing annual migration flow data are predominantly available only from high-income Western countries with the statistical infrastructure to monitor migration. Such data only cover a small share of the global migration system 7 , 8 (Fig. 1a ) and reinforce a receiving-country bias in global migration research 15 .

Fig. 1: Availability of flow data across global migration corridors. The alternative text for this image may have been generated using AI.

Full size image

a , Fraction of corridors that have reported flow values in the 1990–2020 period by any of the validation datasets used in this work 7 , 66 , 67 , 68 , 69 . Statistics for both origin- and birth-destination corridors are shown; these are further disaggregated by corridors for which neither, only one of, or both the sending and the receiving country has reported figures. b , Migration flow estimates based on domicile registration (usually with a local authority) are available for a small number of countries, but the discrepancies can be large: estimates of flows—based on registrations of people arriving from Poland as reported by German authorities (red) and de-registrations of people leaving Poland for Germany, as reported by Polish authorities (blue) 7 —are shown. The harmonized QuantMig estimates (orange; error bands show the 97.5% quantile) and the recent digital-trace estimates based on Facebook data are also shown. c , Various estimates of the net migration for France, such as those from the UN Population Division’s World Population Prospects (2024 Revision), the US Census Bureau (USCB)’s International Dataset 70 , and the French National Institute of Statistics and Economic Studies (INSEE) 71 . d , UN DESA estimates of the migrant stock of Somalians in Ethiopia, which do not agree across revisions. In some cases, they are based on refugee data figures from the UNHCR 72 .

In countries in which migration flow statistics are published, the definitions of what constitutes a migration event are determined by criteria designed to meet domestic policy needs 16 , 17 , which can bias comparative analyses. Although the UN recommends a twelve-month threshold 18 , where anyone relocating for the majority of a year or more qualifies as a migrant, this definition is not applied consistently. Some countries such as Germany mandate residential registration, requiring migrants to report their country of origin upon arrival. Others, such as the UK, rely on visa records, administrative data and, until recently, passenger surveys. A third common approach uses border entry statistics collected by immigration authorities. Each method has limitations: registration systems typically undercount emigration, as few individuals de-register when leaving; passenger surveys and border data are not comprehensive and may conflate short- and long-term travellers. As a result, estimates from sending and receiving countries of…