Anticipating humanitarian emergencies with a high risk of conflict-induced displacement

This exploratory study assesses the risk of future onset of large-scale, conflict-related internal displacement in countries facing humanitarian emergencies. We train a variety of machine learning models on near-real-time data, which we compare against a simple baseline model, to assess the risk, one and three months into the future, of whether at least 1,000 people per month will flee their homes due to conflict. Measures of past displacement, conflict, risk of humanitarian crises, humanitarian access, the severity of humanitarian crises, and free elections improve forecasting performance. Limitations include the fact that displacement onsets are rare and hard to predict, and limited data availability and quality. Still, the best random forest model flagged 24 of 26 cases of displacement onset three months into the future and identified a high-risk group of country-months with a 33 times higher probability of displacement onset than a low-risk group. Providing such monthly forecasts to humanitarian practitioners could help them prepare better for new displacement or even mitigate the human suffering caused by conflict.