🤖 machine learning

Deep learning four decades of human migration

This paper introduces an open-source deep learning framework that utilizes a recurrent neural network trained on 18 covariates to generate detailed, annual, origin-destination migration flow estimates with uncertainty bounds for 230 countries from 1990 to the present, significantly outperforming traditional methods in both accuracy and temporal resolution.

Original authors: Thomas Gaskin, Guy J. Abel

Published 2026-07-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Thomas Gaskin, Guy J. Abel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to track a global game of "musical chairs" where the music is the economy, the politics, and the weather, and the chairs are countries. For decades, trying to count exactly who moved where, when, and why has been a messy, frustrating game. Some countries keep strict lists at the door (like Germany), others just guess based on who bought a ticket (like the UK), and many places in the world have no list at all. The result? The numbers often don't match. If Germany says 160,000 people came from Poland, but Poland only says 12,300 left, that's a lot of missing people in the math.

Enter a new team of digital detectives: Thomas Gaskin and Guy J. Abel. They didn't just count heads; they built a time-traveling robot brain (a deep recurrent neural network) to figure out the story of human movement from 1990 to 2023.

The Magic Brain with a Memory

Most old ways of guessing migration were like taking a snapshot every five years and trying to guess what happened in between. They assumed people only moved based on what was happening right now. But humans aren't like that. We remember. A crisis from ten years ago might still be pushing someone to leave today.

The authors' robot brain is special because it has a memory. It's like a detective who doesn't just look at the crime scene today, but remembers every clue from the last 30 years. By feeding the brain 18 different types of clues—like how much money a country makes, how far apart they are, whether they speak the same language, if they share a religion, and even how many people are dying in wars—the brain learned to predict the flow of people.

Crucially, this brain doesn't just guess a single number. It also tells you how unsure it is. If the data is fuzzy (like in many parts of Africa), the brain says, "I'm not 100% sure here, you might need to check your records." This is a huge deal because it highlights exactly where we need to collect better data.

What They Found (The Plot Twist)

The brain's story reveals some big changes in the world's movement:

The Great Surge: Since the year 2000, the number of people moving across borders every year has jumped from 13 million to over 36 million in 2023. This isn't just because there are more people on Earth; the percentage of people moving has nearly doubled, from 0.21% to 0.45%.
The Big Hiccups: The only times this number dropped were during the Great Recession (2008–2009) and the Covid-19 pandemic (2020).
The Biggest Moves: The single biggest yearly jump the brain spotted was in 2019, when about 850,000 people fled Venezuela for Colombia. Other massive waves included people moving from Mexico to the US in 2022 and from Ukraine to Russia in 2022.
The Middle East Magnet: The Middle East has been the biggest magnet for new arrivals, mostly from South Asia. Since 2010, an average of 1.3 million people a year (totaling 18.4 million) have moved from India, Pakistan, and Bangladesh to Saudi Arabia, Qatar, Bahrain, and the UAE. That's more than the total number of people who moved from Mexico to the US since 1990 (which was 14.8 million).
Europe's Shuffle: Before the pandemic, about 3 million people moved around inside Europe every year. After the Soviet Union fell in 1991, there was a massive shuffle, with over 800,000 people born in Poland, Russia, Ukraine, and Romania moving within Europe that single year.

What They Argue Against (The "Don't Do This" List)

The authors are very clear about what doesn't work well:

The "Five-Year Snapshot" is too slow: Old methods only gave data every five or ten years. The authors argue this is too slow to catch sudden disasters or quick changes. You can't understand a sudden war or a quick economic crash if you only look at the data every five years.
The "Current State Only" model is wrong: They argue against models that treat humans as if they have no memory (Markovian models). They say it's silly to think a person decides to leave a country based only on today's weather or economy, ignoring the crises of the past few years.
The "Residual" Math is shaky: Many official UN numbers for "net migration" (people arriving minus people leaving) are calculated by taking the total population change and subtracting births and deaths. The authors show this is often wrong because population counts themselves can be messy. For example, their model suggests Russia's net migration actually turned negative around 2005, contradicting the UN's suggestion of a positive flow since 1995.

How Sure Are They?

The authors didn't just guess; they tested their robot brain rigorously.

The Test: They hid 20% of the known migration routes from the brain during training and then asked it to guess those hidden routes.
The Score: The brain got a 95% correlation on the data it studied and an 80% correlation on the hidden test data. That's a strong score, especially considering migration data is notoriously noisy.
The Comparison: When they compared their brain to the old "stock-differencing" methods (the five-year snapshots) and other standard techniques, their deep learning model significantly outperformed them all.
The Caveat: They admit that for some regions, like Sub-Saharan Africa, the uncertainty is still high because the input data is missing. They explicitly state that for countries like Nigeria, the uncertainty on their estimates is among the highest in the world, signaling that more data collection is needed there.

The Bottom Line

This isn't just a list of numbers; it's a new way of seeing the world. By giving the computer a memory and teaching it to learn from 18 different types of clues, the authors have created the most detailed map of human movement ever made, covering 230 countries and regions from 1990 to 2023. They've shown us that migration is rising, that it's driven by complex memories of the past, and that while we have a great new map, there are still some foggy spots where we need to shine a brighter light.

The best part? The whole map, the robot brain, and the code are free for anyone to use. It's like they handed the keys to the time machine to the rest of the world.

Technical Summary: Deep Learning Four Decades of Human Migration

Problem Statement
Human migration is a critical driver of demographic change, labor markets, and social policy, yet it remains notoriously difficult to define and track. Existing data suffers from three primary limitations:

Inconsistency and Discrepancy: Different countries use varying definitions (e.g., 12-month thresholds) and collection methods (registration vs. border statistics vs. surveys), leading to significant divergences between reported emigration and immigration figures.
Data Scarcity: Reliable migration statistics are largely absent for the Global South, and sudden shocks (conflict, disasters) often render data collection impossible.
Low Temporal Resolution: Traditional estimation methods rely on migrant stock data published by the UN Department of Economic and Social Affairs (UN DESA) at five-year intervals. These methods, such as stock differencing or demographic accounting, assume migration is a Markovian process (dependent only on the current state) and fail to capture long-term temporal correlations or sudden, short-term shifts in migration dynamics. Furthermore, they often treat stock data as ground truth without accounting for the uncertainty inherent in demographic residuals.

Methodology
The authors propose a novel framework combining a mechanistic flow model with a deep recurrent neural network (RNN) to estimate annual bilateral migration flows ( $F_{ij}$ ) and stocks ( $S_{ij}$ ) for 230 countries and regions from 1990 to 2023.

Architecture: The core of the model is a deep recurrent neural network ( $u_\theta$ ) with a latent state ( $z_{ijk}$ ). This architecture introduces "memory" to the system, allowing current flow estimates to be conditioned on the entire history of covariates and past flows, thereby capturing long-range temporal correlations that traditional gravity models miss.
Input Covariates: The network is trained on 18 covariates for the country of birth ( $i$ $i$ ), origin ( $j$ $j$ ), and destination ( $k$ $k$ ). These include:
- Demographic: Population, life expectancy, birth/death rates.
- Economic: Real GDP per capita, GDP growth, bilateral trade flows.
- Geographic/Cultural/Political: Distance, religious and linguistic similarity, colonial ties, EU membership, conflict deaths, and refugee stocks.
Training Objective: The model is trained to minimize a loss function ( $J$ $J$ ) that aligns predicted outputs with target data. The targets include:
- UN DESA migrant stocks (disaggregated by birth).
- Facebook location data (2019–2022).
- Harmonized QuantMig flows (Europe, 2009–2019).
- National net migration statistics for ~30 countries.
- Note: The loss function operates on the differences in stocks over five-year intervals rather than absolute stock values to avoid conditioning on potentially erroneous initial baselines.
Data Transformation: To handle the heavy skew and long tails of migration data, a symmetrized Yeo-Johnson transformation is applied to inputs and targets to approximate normal distributions before training.
Uncertainty Quantification: The authors employ two strategies to estimate confidence bounds:
1. Ensemble Learning: Training an ensemble of 15 neural networks to capture parameter uncertainty.
2. Input Propagation: "Pushing" uncertainty from input covariates (specifically initial stock estimates) through the trained network to generate a distribution of flow estimates.
Recursive Training: The network operates recursively; the estimated flow table for year $t$ is used to update the migrant stock $S(t+1)$ , which is then fed back as an input covariate for the next time step.

Key Results

Global Trends: The dataset reveals that global annual migration flows rose from 13 million in 2000 to over 36 million in 2023. Per-capita migration increased steadily from 0.21% to 0.45% over the same period.
Event Detection: The model successfully captures major historical events, including the 2019 Venezuela-to-Colombia exodus (~850k), the 2022 Ukraine-to-Russia movement, and the Rwandan civil war. It identifies the Middle East as the region with the highest total inflow, particularly from South Asia.
Validation: On unseen test sets, the model achieves an 80% correlation on flow corridors, significantly outperforming traditional stock-differencing and demographic accounting methods. It matches the uncertainty estimates of the QuantMig project while providing annual resolution.
Net Migration Corrections: The model provides corrected net migration figures for various countries, often diverging from UN World Population Prospects (WPP) estimates which rely on demographic residuals. For instance, the model suggests Russian net migration turned negative around 2005, contrary to UN figures suggesting positive inflow until 2022.
Uncertainty Mapping: The study identifies that uncertainty is highest in Sub-Saharan Africa, highlighting specific regions where data collection efforts are most needed.

Significance and Claims
The authors claim this work represents a "paradigm shift" in computational migration modeling. By moving beyond the static, five-year resolution of UN DESA data and the Markovian assumptions of gravity models, the deep learning approach:

Increases Resolution: Delivers annual, bilateral flow estimates disaggregated by country of birth for the entire globe.
Captures Memory: Explicitly models the influence of past economic, political, and social shocks on current migration decisions.
Quantifies Uncertainty: Provides confidence bounds that allow researchers to pinpoint geographic regions with inconsistent data, guiding future data collection.
Open Source: The authors release the full dataset, training code, and neural network weights, aiming to serve as a foundational resource for economists, epidemiologists, and demographers.

The paper concludes that while the current framework operates at the country level, the recurrent architecture is well-suited for future transitions to finer spatial grids (e.g., 10km² resolution) to model internal migration, and that deep learning is essential for capturing the non-linear, threshold-based responses of human systems to unforeseen global events.