Machine Learning Prediction of Conflict-Driven Refugee Migration: Evidence from Syria, Afghanistan, and Ukraine

Abstract

Refugee migration is a complex global phenomenon influenced by geopolitical instability, economic conditions, and historical migration networks. Accurate prediction of refugee movement patterns is essential for humanitarian planning and policy development. In this study, we develop a machine learning framework to predict refugee migration flows using historical refugee statistics, conflict indicators, and macroeconomic variables. Using datasets spanning 2000–2023, including refugee statistics from the United Nations High Commissioner for Refugees (UNHCR), conflict event data from the Uppsala Conflict Data Program (UCDP), and socioeconomic indicators from the World Bank, we construct predictive models to estimate migration flow intensity between origin and destination countries. The analysis focuses on three large-scale conflicts: the Syrian Civil War, the War in Afghanistan, and the Russian invasion of Ukraine. Migration flows between origin and destination countries from 2000–2023 were analyzed to quantify displacement surges associated with conflict onset. Multiple machine learning algorithms were evaluated, including Random Forest, Gradient Boosting, Long Short-Term Memory networks, Graph Neural Networks, and Transformer models. Results indicate that classical ensemble models outperform deep learning approaches in this dataset, with Random Forest achieving the highest area under the receiver operating characteristic curve (AUC = 0.56). Feature importance analysis suggests that historical migration patterns and economic indicators are stronger predictors of refugee flows than conflict intensity alone. These findings highlight the importance of structural migration networks in shaping refugee movement and demonstrate the potential of machine learning methods to support humanitarian forecasting and migration policy planning.