The Geometry of Instability: Patterns in Conflict and Migration Flow Forecasting

Abstract

Forecasting in social sciences has progressed rapidly, driven by increased demand from decision-makers. Yet, phenomena like conflict and migration remain difficult to predict due to their complexity of mechanisms, context dependency, and reliance on multi-individual human decision-making. These dynamics often vary in speed, contain time lags, and pose challenges to rigid classical modeling approaches. This thesis introduces ShapeFinder, a shape-based forecasting model that focuses on temporal dynamics rather than static values. It uses flexible distance measures, Dynamic Time Warping (DTW) and Earth Mover's Distance (EMD), to identify similar historical patterns, even when sequences differ in timing, scale, or context. The model generates scenario-based forecasts with probabilities and returns matched historical cases, providing transparency and interpretability crucial for practical use, unlike black-box models such as deep learning. Across country and regional scales, ShapeFinder outperforms classical statistical methods, as well as machine and deep learning models, in forecasting conflict fatalities and migration flows. In conflict settings, the model versions without covariates outperform complex models with hundreds of variables. In migration flows settings, the model outperforms classical statistical and deep learning models, with covariates showing little added value unless used dynamically. Its lightweight, flexible, and transparent method makes ShapeFinder suited for real-world early warning systems in conflict and migration flow forecasting, but also offers strong potential for academic use.