Predicting war, coups and riots has been a goal for a generation of social studies researchers. We reviewed a wide array of literature from several ongoing research projects across academia, the security sector and the humanitarian sector. Advances in machine learning and newly available historical datasets and predictors have given momentum to the field. Nevertheless, the problem of conflict prediction remains complex and hard to solve.
Our research focused on the feasibility of predicting conflict for anticipatory action. In our paper, we define and evaluate three types of conflict prediction models – classification, risk prediction, and continuous prediction – and conclude with a set of recommendations and next steps for the humanitarian sector.
In brief, we found insufficient justification for exclusively relying on conflict prediction models to drive anticipatory action due to several factors:
\- Poor performance in predicting the onset of new conflicts.
\- The lack of clear connection between predicted conflict and resulting humanitarian impact.
\- The dominance of ongoing conflict as a predictor of future conflict.
To make use of conflict prediction for anticipatory action in the humanitarian sector, we recommend that future work:
\- Utilize flexible models that do not pre-suppose a theoretical framework of conflict causality.
\- Focus models on predicting shifts in conflicts, such as an increase in intensity or onset.
\- Explore the use of human inputs through superforecasters or prediction markets and use local data to improve model performance in specific contexts.
\- Improve predictions on the humanitarian impact of conflict as opposed to conflict itself.
\- Ensure that model development and evaluation is done in a reproducible and transparent way that highlights the model performance in all relevant metrics.
\- Learn from the state-of-the-art research underway in the academic field and ensure that applied research is relevant for humanitarian decision making.