Exploring Forecast Models for Actor-Centric Conflict Onset and Escalation

• In Event: Forecasting Conflict

Abstract

Attempts at predicting the behaviour of rebel armed actors is in its infancy, in part due to a lack of empirical understanding of the processes determining onset of violence at rebel group level. This in turn has been (at least partially) caused by a lack of data which is especially visible when it comes to those armed groups that have had the capacity of conducting armed conflict but have chosen not to escalate violence. Using a novel dataset that includes violent actors that have initiated violence, but not escalated into full-blown armed conflict and leveraging the large empirical corpus of the Uppsala Conflict Data Program, this paper introduces a large, theoretically-driven, ensemble of actor-centric forecast models that explores the predictors of violent escalation of armed conflict. The paper trains this ensemble using a heterogenous pool of supervised machine learning algorithms, each covering an individual, theoretically-informed, them, and evaluates them through an evaluation framework derived from the Violence Early Warning System (ViEWS) programme. The ensemble shows good capacity to predict escalation of armed actors, as well as improvements in the predictive accuracy of classical conflict forecast models when informed by the actor-centric ensemble.

Author

• Mihai Croicu, Uppsala University, Dept. of Peace and Conflict Research