Forecasting Ukrainian Refugee Flows With Organic Data Sources

Abstract

Although European countries seek to understand the volume and destinations of forced migrant flows out of Ukraine, it is difficult to collect timely data for many reasons including dangerous conditions for on-the-ground survey data collection. This article combines different organic data to predict forced migration from Ukraine to five neighboring countries receiving refugees: Poland, Romania, Slovakia, Moldova, and Hungary. We pair online Ukrainian-language Twitter conversation with event and fatality data from Armed Conflict Location and Event Data Project and to develop predictive models of forced displacement, and assess the quality of our predictions using United Nations High Commissioner for Refugees (UNHCR) border crossing data. Using a Bayesian hierarchical approach that accounts for heterogeneity in the forced migration process and fine temporal granularity of the data, results suggest that, after an initial rise in out-migration at the start of the conflict, migrant flows persist albeit at lower rates. In addition, countries with the highest initial volume of migrant arrivals have higher rates of prolonged flows. Finally, in terms of prediction quality, Twitter variables were more important predictors in the first phase of the conflict while event-based predictors were more important in the second phase.