Hierarchical semantic extraction and heterogeneous graph neural networks for event-driven time series forecasting.

Event-driven time series forecasting constitutes a fundamental challenge across multiple scientific domains. Existing deep learning approaches rely on the assumption of historical pattern continuity and thus exhibit inadequate responsiveness to exogenous shocks, while mainstream text-enhanced methods encode documents as singular dense vectors, causing the compression of fine-grained semantic information such as entity stances, event types, and causal propagation relationships. To address these limitations, this study proposes a tripartite methodological framework building upon multimodal graph neural networks and the Temporal Fusion Transformer architecture. The first component employs large language models to perform hierarchical semantic extraction that decomposes unstructured text into four structured layers comprising entity stance quantification, event ontology mapping, inter-event causal chain reasoning, and aggregate sentiment indicators, thereby preserving prediction-relevant information that would otherwise be lost through flattened vectorization. The second component constructs heterogeneous information graphs containing multiple node types and relation types to explicitly model cumulative effects, propagation effects, and synergistic effects among events through relational graph convolutional networks with relation-aware attention mechanisms. The third component utilizes semantic analysis to achieve rapid environmental state identification, enabling dynamic recalibration of feature importance weights according to prevailing conditions and thereby addressing the non-stationarity inherent in event-driven prediction tasks. Experimental validation using crude oil price forecasting as a representative scenario demonstrates that the proposed framework achieves 17.7% improvement over baseline methods in overall prediction accuracy, with improvement reaching 34.7% during event-intensive periods, while ablation studies confirm independent contributions from each component. The proposed methodology possesses domain generality and can be transferred to other event-driven prediction tasks including epidemic spread forecasting, supply chain risk assessment, and financial market analysis.