Endogenous conflict and the limits of predictive optimization

Forecasting models in political violence research increasingly rely on high-dimensional covariates and machine learning. Yet in practice, the most reliable conflict forecasts often come from much simpler systems: autoregressive models that predict future events based solely on recent past outcomes. This paper argues that such models are not merely convenient baselines but theoretically appropriate tools for sparse, dynamic environments like armed conflict. We show that autoregressive models consistently outperform or match more complex alternatives across multiple countries and specifications, while structural covariates frequently add little or degrade performance. We explain this pattern both theoretically and empirically: conflict is driven by internal feedback, burstiness, and short-term adaptation-not by slow-changing structural conditions. By foregrounding the limits of causal modeling in high-entropy settings, we make a broader case for epistemic modesty in prediction. Autoregression, we argue, is not a shortcut, but a principled strategy in systems that resist control.