Sezibra/conflict-forecasting-ml

Sub-national conflict forecasting for Ethiopia using Random Forest, temporal cross-validation, and SHAP interpretation. Built on UCDP GED v25.1 data. # Conflict Forecasting with Machine Learning ## Research Question Which features best predict escalation of violence at the sub-national level, and how do different machine learning models compare when evaluated with temporal cross-validation that prevents information leakage? ## Navigation | Section | Description | |---------|-------------| | Motivation | Why conflict forecasting matters and how this project relates to VIEWS | | Key Findings | Main results across all models and analytical steps | | Data | UCDP GED panel construction for 11 Ethiopian regions | | Methods | Feature engineering, temporal CV, model comparison, SHAP, calibration | | Results in Detail | Figures and interpretation for each analytical step | | Limitations | Honest assessment of constraints | | Notebooks | Three-notebook analytical progression | | How to Reproduce | Setup and replication instructions | | References | Academic sources | ## Motivation Conflict forecasting is the most policy-relevant application of computational social science in the conflict domain. The VIEWS (Violence and Impacts Early-Warning System), directed by Håvard Hegre at Uppsala University and PRIO, generates monthly predictions of armed conflict fatalities up to 36 months ahead using ensembles of ML models trained on 535 million data points and 250+ conflict variables. VIEWS received the Kluz Prize for PeaceTech Special Distinction in September 2024. Cederman and Weidmann (2017) identified critical pitfalls in conflict forecasting: overfitting to rare events, ignoring spatial dependence, poor out-of-sample evaluation, lack of theoretical grounding, and insufficient attention to calibration. This project addresses each pitfall explicitly. This project replicates the VIEWS analytical logic at a smaller scale: panel construction, event-history feature engineering, temporal cross-validation, model comparison, SHAP interpretation, and calibration analysis. It focuses on Ethiopia at admin1 resolution with 30 event …