Improving the Early Warning Function of Civil War Onset Models Using Automated Event Data

Abstract

This research note explains our approach to political violence forecasting employing a combination of structural and automated event data. We have created four civil war onset models using data from 1979-1999 (in sample), which we then use to predict the onset of civil war from 2000-2015 (out-of-sample). Our time horizon is one month into the future. The predictive performance of our best model as measured by its area under the receiver operating characteristic (ROC) curve is 0.881. All 24 conflict onsets between 2000-2015 occurred within countries in the top half of our risk ranking, with 79% of these onsets falling in the top 20% category, and 58% in the top 10% category. Our findings suggest that models incorporating both structural and automated event data have significantly stronger predictive power than those that rely exclusively on structural data. Conclusions and guidelines for future research are provided on the basis of our results.