Using machine learning to forecast conflict events for use in forced migration models

Forecasting the movement of populations during conflict outbreaks remains a significant challenge in contemporary humanitarian efforts. Accurate predictions of displacement patterns are crucial for improving the delivery of aid to refugees and other forcibly displaced individuals. Over the past decade, generalized modeling approaches have demonstrated their ability to effectively predict such movements, provided that accurate estimations of conflict dynamics during the forecasting period are available. However, deriving precise conflict forecasts remains difficult, as many existing methods for conflict prediction are overly coarse in their spatial and temporal resolution, rendering them inadequate for integration with displacement models. In this paper, we propose a hybrid methodology to enhance the accuracy of conflict-driven population displacement forecasts by combining machine learning-based conflict prediction with agent-based modeling (ABM). Our approach uses a coupled model that combines a Random Forest classifier for conflict forecasting with the Flee ABM-a model of the movements of refugees and internally displaced persons (IDPs). The coupled model is validated using case studies from historical conflicts in Mali, Burundi, South Sudan, and the Central African Republic. Our results demonstrate comparable predictive accuracy over traditional methods without the need for manual conflict estimations in advance, thus reducing the effort and expertise needed for humanitarian professionals to provide urgent displacement forecasts.