Agentic Reasoning for Social Event Extrapolation: Integrating Knowledge Graphs and Language Models

Accurate prediction of socio-political events is a longstanding challenge with profound implications for risk management, policy planning, and international relations. Traditional machine learning approaches, such as graph neural networks and recurrent neural networks, have achieved notable progress but often struggle to integrate rich textual context and provide interpretable reasoning. Recent advances in large language models (LLMs) have demonstrated unprecedented capabilities across diverse tasks, including text generation, code synthesis, and complex multimodal reasoning, making them promising candidates for event prediction in dynamic, data-rich environments. This research presents an agentic reasoning framework combining temporal knowledge graphs, large language models (LLMs), and a modular tool-based architecture to address event extrapolation in complex, real-world settings. The methodology integrates agent-based reasoning, iterative tool invocation, and explicit validation mechanisms to ensure logical consistency and transparency in predictions. Experiments on country-specific subsets of the POLECAT dataset employ multiple LLM architectures and multiple evaluation metrics, including Hit@k, MRR, F1-scores and Prediction Entropy. Comparative analyses demonstrate that the agentic framework achieves robust predictive performance and interpretability, complementing fine-tuned LLM baselines. Furthermore, the ethical implications of deploying AI in social computing are addressed, including bias, transparency, and accountability. This study advances event prediction by demonstrating how agentic LLMs, equipped with explicit reasoning and validation, provide scalable and ethically grounded solutions for complex social event extrapolation.