Improving Crude Oil Price Prediction Using LLM-Based Hierarchical Semantic Extraction and Heterogeneous Graph Neural Networks
IEEE Access · 2026 · 14 · 52911-52926
DOI: 10.1109/access.2026.3681371 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Shang Cao und 4 weiteren Autor:innen. Die Arbeit nennt 35 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Guangxi Key Laboratory of Machine Vision and Intelligent Control, 2024 Guangxi University Young and Middle-Aged Teachers’ Research Basic Ability Enhancement Project, Wuzhou Science and Technology Plan Project.
Abstract (Originalsprache)
Crude oil price forecasting serves energy policy formulation and financial risk management. Existing deep temporal models rely on the assumption of historical pattern continuity and respond insufficiently to exogenous events such as geopolitical conflicts. Mainstream large language model approaches encode news as single vectors, and fine-grained semantic information is compressed and lost in this process. Causal propagation relationships among events and heterogeneous temporal lag characteristics have not yet been incorporated into modeling. To address these limitations, this paper proposes three improvements building upon multimodal graph neural network and Temporal Fusion Transformer architectures. At the semantic extraction level, a four-layer pyramid architecture based on large language models is designed to parse news texts into structured representations covering entity stances, event types, causal chains, and market sentiment. At the graph modeling level, a heterogeneous information graph containing multiple node types and edge types is constructed, introducing causal and temporal edges beyond existing static entity associations to characterize cumulative, propagation, and synergistic effects among events. At the feature fusion level, a semantic-driven market state identification module is designed to replace numerical-statistics-based lagging detection, dynamically adjusting feature weights according to the identified state. Using Brent crude oil price forecasting as a validation case, the proposed method achieves a 17.7% RMSE improvement over the Temporal Fusion Transformer baseline, with improvement reaching 34.7% during event-intensive periods and the coefficient of determination $R^{2}$ reaching 0.903.
Förderung
- Guangxi Key Laboratory of Machine Vision and Intelligent Control
- 2024 Guangxi University Young and Middle-Aged Teachers’ Research Basic Ability Enhancement Project
- Wuzhou Science and Technology Plan Project
- Wuzhou University Research Foundation for Advanced Talents
- National Natural Science Foundation of China
Zitieren
APA
Cao, S., Qu, Z., Cen, C., Xu, Z., & Zou, Y. (2026). Improving Crude Oil Price Prediction Using LLM-Based Hierarchical Semantic Extraction and Heterogeneous Graph Neural Networks. IEEE Access, 14, 52911-52926. https://doi.org/10.1109/access.2026.3681371
Vancouver
Cao S, Qu Z, Cen C, Xu Z, Zou Y. Improving Crude Oil Price Prediction Using LLM-Based Hierarchical Semantic Extraction and Heterogeneous Graph Neural Networks. IEEE Access. 2026;14:52911-52926. doi:10.1109/access.2026.3681371
BibTeX
@article{Cao2026,
author = {Cao, Shang and Qu, Ziqi and Cen, Caichun and Xu, Zeling and Zou, Yuntao},
title = {Improving Crude Oil Price Prediction Using LLM-Based Hierarchical Semantic Extraction and Heterogeneous Graph Neural Networks},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {52911--52926},
doi = {10.1109/access.2026.3681371}
}
EUVIMED