Models Versus Rankings: Forecasting Political Violence

We compare the predictive performance in forecasting the onset of large scale political violence worldwide of five statistical models and three commonly used fragility/instability indices using PITF and UCDP data for the period 2000-2015. We find that the models typically outperform the rankings and that a ‘consensus’ model performs better than the individual models. We highlight problems with measurement of the dependent conflict variable, reflect on problems associated with forecasting political violence, and we outline ways forward for future research.