Predicting (de-)Escalation of Sub-National Violence Using Gradient Boosting: Does It Work?

This article presents a prediction model of (de-)escalation of sub-national violence using gradient boosting. The prediction model builds on updated data from the PRIO-GRID data aggregator, contributing to the ViEWS prediction competition by predicting changes in violence levels, operationalized using monthly fatalities at the 0.5 × 0.5-degree grid (pgm) level. Our model's predictive performance in terms of mean square error (MSE) is marginally worse than the ViEWS baseline model and inferior to most other submissions, including our own supervised random forest model. However, while we knew that the model was comparatively worse than our random forest model in terms of MSE, we propose the gradient boosting model because it performed better where it matters—in predicting when (de-)escalation happens. This choice means that we question the usefulness of using MSE for evaluating model performance and instead propose alternative performance measurements that are needed to understand the usefulness of predictive models. We argue that future endeavors using this outcome should measure their performance using the Concordance Correlation, which takes both the trueness and the precision elements of accuracy into account, and, unlike MSE, seems to be robust to the issues caused by zero inflation.