Pushing the boundaries of anticipatory action using machine learning

Summary

Open access peer-reviewed article describing the AHEAD model in detail. Documents the Bayesian state-space model using gradient boosted trees and linear models for sub-district displacement forecasting in Liptako-Gourma (Burkina Faso, Mali, Niger). Reports 15% mean absolute percentage error across 315 hindcasts in Burkina Faso. Details the Akobo County, South Sudan anticipatory action case with community-level triggers via peace committees. Discusses combining ML predictions with Protection Monitoring System (PMS/P21) community data. Includes confusion matrix analysis showing ~70% correct classification at 1,200-person threshold. Replication data on GitHub.