The role of theory in conflict prediction

7 April 2026

Over the past decade, quantitative conflict research has experienced a marked shift from explanation toward prediction. Enabled by advances in data availability and computational power, machine-learning models have become the dominant approach for forecasting violent conflict. These models are well suited to capturing the non-linear relationships and complex interactions observed in conflicts. The paper explains how such prediction models work, and why their focus on predictive performance has led to a reduced role of causal theories regarding the origins of violence.

Using predictions of communal violence in West Africa, the paper highlights that “naïve” attempts to incorporate theory through comparing models with different theory-based feature sets provide limited and potentially misleading insights. It shows that differences between theoretical may only appear under artificially constrained specifications. Instead, the paper highlights the potential of interpretability methods to reconnect theory and prediction after model training. By decomposing predictions into contributions from theoretically meaningful feature groups, predictive models can generate insights into underlying conflict processes and produce theory-informed risk assessments while retaining their predictive realism.

The article can be downloaded here.

More information on the Working Paper Series of CISS Munich can be found here.

Picture: CISS