Conflict Forecasting With Automated Machine Learning - Vito D'Orazio, November 2|3, 2021
Posted on Saturday, Oct 30, 2021
Conflict forecasting models provide information about potential future violence in a country or region. Toward the goal of improving sub-national and country level forecasts, I run a series of experiments devised around three broad criteria informed by the literatures on political violence and predictive modeling: (1) flexibility to capture nonlinear relationships, (2) an emphasis on the endogenous nature of violence, and (3) avoid overfitting, which is especially problematic with flexible algorithms and the types of highly disaggregate data used here. To meet these criteria, I experimented with combinations of automated machine learning (autoML) systems and limited datasets that emphasized the endogenous features of conflict. Two core findings emerged. One, autoML improves predictive performance, and the stacked ensemble method performs best. Two, the emphasis on endogenous factors improves predictive performance, but particularly at the sub-national level. The best performing model from these experiments was entered into a conflict forecasting competition hosted by the Violence Early Warning System (ViEWS) project. It won for both predictive accuracy and originality. Beyond conflict forecasting, this research has implications for modeling other types of violence and social events at highly disaggregate levels.