ABSTRACT The accelerating infusion of advanced computational methods into geopolitical analysis has created new opportunities to anticipate unrest, economic shocks and diplomatic shifts. Traditional machine learning pipelines can extract statistical patterns from large event corpora, but they often struggle to incorporate real‐time contextual information or explain their predictions in language accessible to decision‐makers. This study proposes a comprehensive framework, LLM4Geopolitics , that couples a domain‐adapted large language model with a retrieval‐augmented generation mechanism grounded in a structured knowledge graph. The forecasting component employs a transformer architecture tailored to sparse, irregular event streams, while the generative component translates model outputs into dialogue‐ready assessments enriched with up‐to‐date economic and peace‐index indicators. Experiments conducted on the Gdelt dataset demonstrate that the integrated approach improves event‐severity prediction and generates fact‐consistent narratives compared with baseline time series and text‐only models. These findings highlight the potential of combining specialised sequence models, on‐demand knowledge retrieval and generative reasoning to deliver timely and interpretable insights for geopolitical forecasting.
Mouakher et al. (Fri,) studied this question.