Endogenous conflict and the limits of predictive optimization.

Area of Science:

• Political Science
• Computational Social Science
• Conflict Studies

Background:

• Forecasting political violence often uses complex machine learning with high-dimensional data.
• However, simpler autoregressive models, using only past outcomes, frequently yield more reliable conflict predictions.
• This highlights a discrepancy between common practice and effective forecasting in political violence research.

Purpose of the Study:

• To argue that autoregressive models are theoretically appropriate for forecasting in sparse, dynamic environments like armed conflict.
• To demonstrate the empirical performance of autoregressive models against more complex alternatives.
• To provide a theoretical explanation for why simpler models excel in predicting political violence.

Main Methods:

• Comparative analysis of forecasting model performance across multiple countries and specifications.
• Empirical testing of autoregressive models against complex machine learning and structural covariate models.
• Theoretical modeling to explain the dynamics of political violence and its predictability.

Main Results:

• Autoregressive models consistently matched or outperformed more complex forecasting models.
• Inclusion of structural covariates often added little predictive value or degraded model performance.
• Empirical evidence suggests conflict dynamics are driven by internal feedback, burstiness, and short-term adaptation.

Conclusions:

• Autoregression is a theoretically sound and empirically validated strategy for forecasting political violence in high-entropy systems.
• Complex models may be less effective due to the inherent nature of conflict dynamics, which are not solely driven by slow-changing structural conditions.
• The study advocates for epistemic modesty in prediction, favoring principled, simpler strategies like autoregression in unpredictable environments.