Close Calls of Conflict: How False Positives Can Improve Conflict Onset Prediction

In conflict forecasting, the general aim is to correctly predict the actual conflict events. Current state-of-the-art research in this field uses advanced machine learning to get at the processes driving political violence. Yet, while the field has advanced significantly, predicting which new outbreaks of violence will escalate into conflict onset – and which will not – remains a challenge in conflict prediction. This study aims to address this limitation through a methodological innovation intended to capture the underlying processes of averted onsets, where new outbreaks of violence did not lead to conflict onset. In this paper, I ask why countries at high predicted risk of conflict onset in some cases experience no, or only a few, incidences of violence. I propose that the averted onsets stem from other processes than those driving violence, and our models fail to capture whether we here see structural risk but no trigger, or high risk met with restraint or prevention. I suggest that over-predicted onsets – false positives – are signals from such stalled escalatory processes that took our model by surprise. Empirically, I test how information on false positives from onset models can be incorporated into forecasting models to improve their predictive performance. Using data from countries with low-intensity violence in the UCDP Candidate Events Dataset, I predict conflict onsets and apply a cascaded neural network to extract signals from the over-predicted onsets. The results show that we can use knowledge from the onsets that never were, to improve our ability to correctly predict true onsets. The approach offers a solution to separate model-based uncertainty from the stochastic uncertainty inherent in prediction errors. The findings also highlight that false positives open a gateway to explore, theoretically and empirically, what makes high-risk cases resilient against conflict onset.