The Hard Problem of Conflict Forecasting: Predicting Conflict Onset with Early Signs of Violence and Prevention

When civil strife turns violent, the risk of armed conflict onset increases, especially in societies with a history of civil war. Yet, only in some cases do we see low-intensity violence lead to armed conflict outbreak or recurrence. While conflict research has disentangled processes that drive and prevent the escalation of violence into armed conflict, predicting which contentious situations will lead to onset - and which will not - remains a challenge in conflict forecasts. This paper focuses on why some high-risk cases see armed conflict onset while others do not, and proposes that we can improve onset prediction by drawing on theory regarding processes that stall escalation. Building on a theoretical framework of conflict prevention, I develop a forecasting model including both risk factors and variables of societal resilience, quelled dissent and compromises made through agreements, which represent ways to limit violence before conflict breaks out. For a case universe of potential conflict onsets, I combine georeferenced conflict event data from the UCDP with raw data from the UCDP records on cases of low-intensity violence that never reached its threshold of active armed conflict. I train a risk model and the theorized prevention model on global data 1990-2021 to make out-of-sample predictions of conflict onset within one year, for January 2022 to January 2026. Preliminary results indicate that the risk model without preventive factors improve onset forecasts, likely benefiting from the new data on cases with only low-intensity violence. While the prevention model does not improve predictions, a closer analysis shows that an ensemble balancing the risk and prevention models yields the fewest incorrect predictions of onset, especially when considering false alarms that later proved to correctly anticipate onset.