Introduction
Time series forecasting has recently pivoted toward leveraging Large Language Models (LLMs), typically following the LLM-as-Predictor paradigm. In this setup, a pretrained LLM (such as GPT) acts as the primary engine for numerical prediction. However, this approach faces two significant hurdles: the "semantic gap" between numerical patches and natural language, and the high computational cost of running full forward passes through massive models for every prediction.
FiCoTS addresses these challenges by proposing the LLM-as-Enhancer paradigm. Instead of using the LLM for direct prediction, it leverages the model's semantic strengths to support a specialized time series encoder through a hierarchical "Fine-to-Coarse" interaction framework.
The FiCoTS Framework
The core of FiCoTS is its hierarchical alignment of text and time series data across three distinct levels: token, feature, and decision. This approach ensures that local temporal fluctuations are grounded in linguistic context before being aggregated into global representations.
- Token-Level Modality Alignment: The model segments time series into patches and text into tokens, constructing a dynamic heterogeneous graph. Using GraphSAGE, it passes messages between semantically similar patches and tokens, effectively grounding numerical data in linguistic context at a granular level. A dynamic filtering mechanism prunes weak connections to reduce noise.
- Feature-Level Modality Interaction: A global cross-attention mechanism follows, where time series representations act as the "query" and text representations provide the "key" and "value." This allows the model to weigh global textual context (like dataset descriptions) against specific temporal features.
- Decision-Level Modality Fusion: Finally, an adaptive gated network fuses the multimodal "enhanced" features with raw temporal patterns. This ensures the final forecast remains grounded in the actual data while benefiting from semantic insights.
Experimental Setup
The framework was evaluated on seven widely recognized real-world benchmarks, including the Electricity Transformer Temperature (ETT) series, Weather, Electricity, and Traffic datasets. These benchmarks cover various temporal resolutions and domain-specific characteristics, providing a robust testbed for cross-modality interaction.
Key Results and Insights
FiCoTS achieves state-of-the-art performance, outperforming both specialized temporal models (like PatchTST and iTransformer) and other LLM-based approaches (like Time-LLM and GPT4TS). Notably, it achieves an average 2.0% improvement in Mean Squared Error (MSE) over TimeVLM, the previous leading LLM-as-Enhancer model.
Few-Shot Mastery and Latent Space Clarity
The model demonstrates remarkable robustness in low-data scenarios. When trained on only 5% of the training data, FiCoTS achieved improvements of 22.5% to 25.8% over other recent LLM-based forecasting methods.
Furthermore, visualization of the latent space confirms the efficacy of the alignment process. As shown in the t-SNE plots, the token-level alignment successfully organizes numerical data into more coherent and distinct clusters.
This clustering proves that semantic text information provides a valuable structural prior, helping the model "understand" the underlying patterns in the numerical data more effectively than numerical analysis alone.
Conclusion
FiCoTS demonstrates that LLMs are most effective in time series forecasting when used as semantic assistants rather than direct predictors. By aligning modalities hierarchically from fine to coarse levels, the framework bridges the semantic gap while maintaining computational efficiency and superior accuracy. Future work may explore more sophisticated prompt engineering or alternative graph neural network architectures to further refine this cross-modal synergy.
Summary updated