Explainable Forecasting of International Crude Oil Price Driven by Geopolitical News Decision-Making with Large Language Models

Abstract International crude oil price forecasting plays a pivotal role in market forecasting and energy trading. Currently, forecasting systems are confronted with two major challenges: the inability to rapidly respond to geopolitical changes and the lack of result interpretability. This paper introduces an end-to-end, weekly-updated international crude oil price forecasting system that integrates geopolitical decision extraction, multi-scale model prediction, and result attribution analysis. Firstly, we design a geopolitical decision extraction algorithm driven by LLM (Large Language Model) summarization and reasoning, establishing a comprehensive workflow from geopolitical news collection to decision-making. Secondly, we propose a multi-scale prediction model that facilitates concurrent forecasting at weekly, monthly, and quarterly levels using a unified deep learning architecture. Finally, we construct an attribution analysis engine powered by LLM reasoning, which offers robust and interpretable explanations for forecasting results. The entire system guarantees the timeliness of geopolitical analysis through its weekly update mechanism. Experimental results demonstrate that, driven by geopolitical decisions, the multi-scale prediction model has achieved accuracy improvements of 5%, 3.5%, and 1% at weekly, monthly, and quarterly forecasting levels, respectively. Additionally, the consistency rate between the attribution analysis engine and expert analyses exceeds 80%.