How AI Predicts Refugee Movements: A Technical Deep Dive Into Forecasting Models

Why displacement forecasting is now a first-class discipline

For most of the post-Cold War period, refugee forecasting meant a UNHCR planning scenario refreshed twice a year. By 2026 it is a daily nowcasting product. The shift was forced by Syria 2015, Venezuela 2017-2022, Ukraine 2022, Sudan 2023, and a steady drumbeak of climate-driven movement across the Sahel and Central America. Donor agencies and operational responders both wanted lead time, and the data needed to provide it finally arrived: globally consistent conflict event feeds from ACLED, sub-daily satellite revisit, and standardised flow monitoring from the IOM DTM. The result is a small but mature field, sitting at the intersection of demography, econometrics, and machine learning.

The three families of models in production

Almost every operational system in 2026 belongs to one of three families. Classical time-series models (ARIMA, Prophet, state-space) still dominate short-horizon arrival forecasts at the border-post level because they are cheap, transparent, and easy to retrain weekly. Gravity and radiation models, borrowed from transport planning, estimate flows between origin and destination using pairwise pull and push variables; UNHCR's earliest Jetson pilots in the Horn of Africa used a gravity backbone. Machine learning ensembles (gradient-boosted trees, temporal fusion transformers, graph neural networks over country-pair graphs) handle the messier signal: conflict shocks, weather anomalies, market price spikes, and social-media velocity. The Danish Refugee Council Foresight model and the JRC Dynamic Data Hub sit in this third group.

What data actually moves the needle

Across published evaluations, four feature classes consistently improve forecast skill: lagged ACLED fatalities and event counts at admin-1 resolution; food price indices from WFP VAM; cross-border registration counts from UNHCR proGres and IOM DTM with explicit reporting-lag adjustment; and weather anomalies from ERA5. Mobile network metadata helps where regulators permit aggregation, but it is rarely in the global pipelines because of access constraints. Social-media velocity is useful for detecting novel shocks (Ukraine February 2022, Sudan April 2023) but is a poor steady-state predictor.

Reading uncertainty without being misled

Point forecasts are the wrong unit of analysis. Every credible 2026 system publishes prediction intervals. A useful rule: if a model reports a 7-day forecast of 12,000 arrivals with an 80% interval of 6,000-22,000, that is not a failure of the model, that is honesty about a fundamentally noisy process. Treat any displacement forecast presented as a single number with suspicion. The IASC operational guidance on data responsibility explicitly recommends publishing intervals and methodology with every estimate.

Where the models break

Three failure modes recur. First, regime-change events: every major model underweighted the run-up to Ukraine, Sudan, and Gaza because the training distribution did not contain the shock. Second, slow-onset climate displacement, which moves at a pace the models conflate with seasonal noise. Third, urban and undocumented IDPs, who never enter the registration data the models are calibrated against. The honest framing in operational meetings is that current systems are good at amplifying signals already present in the data and bad at anticipating phase transitions.

What a defensible 2026 workflow looks like

Mature teams combine a transparent baseline (classical time series), a richer ML ensemble, and a structured expert-elicitation layer where country analysts adjust the machine output before publication. They version every model, log every input, and rerun back-tests at least quarterly. Public outputs cite primary sources, publish intervals, and never disclose protection-sensitive origin-destination pairs at named-locality granularity. For a reporter, the cleanest sourcing is: 'According to the UNHCR-IOM joint forecast of [date], with an 80% interval of X-Y...' rather than a single headline number.

Further reading and primary sources

• UNHCR Project Jetson and Innovation Service: https://www.unhcr.org/innovation/
• IOM Displacement Tracking Matrix: https://dtm.iom.int/
• ACLED data and methodology: https://acleddata.com/
• DRC Foresight: https://pro.drc.ngo/resources/news/foresight/
• JRC Dynamic Data Hub: https://drmkc.jrc.ec.europa.eu/
• WFP VAM food prices: https://dataviz.vam.wfp.org/
• ECMWF ERA5 reanalysis: https://cds.climate.copernicus.eu/
• IASC Operational Guidance on Data Responsibility: https://interagencystandingcommittee.org/