An Open Source AI Framework for Forecasting Armed Conflict.
Abstract – Armed conflicts are increasing in frequency and severity worldwide, resulting in devastating humanitarian and economic consequences. Despite advances in conflict forecasting, leading academic early warning models often lack complete openness, limiting reproducibility and the potential for advancements in the field. Addressing this critical issue, this thesis presents an entirely open-source framework designed to reproduce and extend state-of-the-art early
conflict forecasting models using publicly available datasets.
Specifically, we systematically evaluate three advanced machine learning approaches—XGBoost, AutoGluon, and TabPFN—across existing datasets from ViEWS and Conflict Forecast, as well as our own extended dataset based on the Uppsala Conflict Data Program (UCDP), covering a period from 1989 to 2024. Our results demonstrate that fully open-source models can match, and even slightly surpass, the predictive performance of current closed-source benchmarks, as
measured by Precision-Recall AUC. By openly sharing all developed pipelines, datasets, models, and evaluation frameworks developed in this thesis, we aim to foster transparency, reproducibility, and further collaborative
innovation in early conflict forecasting research.
Key words: Early Conflict Forecasting, Civil War, AI for Good.
Master thesis VU Amsterdam, August 2025
Download the full thesis.