TLDR: BALM-TSF is a new lightweight framework for time series forecasting that uses large language models (LLMs) by addressing the “modality imbalance” between textual and temporal data. It processes time series and descriptive statistics (as text) in separate branches, then uses a balanced alignment strategy involving scaling and contrastive learning to integrate them effectively. This approach achieves state-of-the-art performance in both long-term and few-shot forecasting with significantly fewer trainable parameters and high efficiency.
Time series forecasting, which involves predicting future values based on historical data, is a critical task across many industries, from predicting energy consumption to analyzing financial markets. Traditionally, this field has relied on specialized time series models. However, with the rapid advancements in large language models (LLMs), there’s a growing trend to incorporate textual information to improve forecasting accuracy.
While LLMs offer powerful capabilities for extracting meaning from text, integrating them with time series data presents a significant challenge: modality imbalance. This imbalance manifests in two main ways. First, LLMs are primarily trained on text, making them less adept at understanding the precise numerical patterns and dynamic trends inherent in time series data. This can lead to “text-dominant” interpretations where the LLM overemphasizes linguistic patterns and overlooks crucial temporal dynamics. Second, there’s a distributional mismatch; time series data is often normalized to a narrow range, while textual embeddings from LLMs can have a much wider value range, causing one modality to dominate during fusion.
To tackle this fundamental problem, researchers have introduced BALM-TSF (Balanced Multimodal Alignment for LLM-Based Time Series Forecasting). This innovative framework is designed to maintain a crucial balance between textual and temporal modalities, ensuring that both contribute effectively to the forecasting process.
How BALM-TSF Works
BALM-TSF employs a lightweight, dual-branch architecture that processes time series and textual data separately before carefully aligning them. The core idea is to leverage the strengths of each modality without letting one overpower the other.
-
Time Series Branch: This branch focuses on the raw time series data. It first normalizes the input to handle distribution shifts and then uses a patch encoder to extract local patterns, converting the time series into a compact numerical representation.
-
Text Branch: Instead of feeding raw time series directly into an LLM, which can lead to semantic imbalance, BALM-TSF takes a different approach. It summarizes each time series using descriptive statistics (like minimum, maximum, median values, trend, and lag features). These statistics are then formatted into a compact textual prompt. This prompt is augmented with a small set of learnable prompt tokens, which are updated during training. A frozen pre-trained LLM (specifically, GPT-2) then processes this combined prompt to generate rich textual embeddings, capturing high-level semantic information without ever seeing the raw numerical sequence.
-
Balanced Multimodal Alignment: This is the key module that addresses the modality imbalance. It involves a two-step strategy:
-
Scaling: The textual embeddings, which might have a higher variance, are rescaled to match the numerical range of the time series embeddings. This step ensures distributional consistency. Additionally, the textual embeddings are adaptively truncated based on the prediction horizon; shorter horizons retain fewer tokens (as temporal patterns are more dominant), while longer horizons incorporate more textual context to supplement weaker temporal signals.
-
Semantic Alignment: After scaling, a contrastive learning objective (InfoNCE loss) is used. This objective ensures that the textual and temporal representations of the same time series instance are pulled closer together in a shared latent space, while representations of different instances are pushed apart. This promotes effective semantic alignment between the two modalities.
-
Finally, the aligned textual and time series embeddings are concatenated and fed into a lightweight forecasting head to produce the final predictions.
Impressive Results and Efficiency
Extensive experiments on standard benchmarks demonstrate that BALM-TSF achieves state-of-the-art performance in both long-term and few-shot forecasting scenarios. It consistently outperforms existing LLM-based forecasters and strong non-LLM baselines, showing significant reductions in Mean Squared Error (MSE) and Mean Absolute Error (MAE).
For instance, in long-term forecasting, BALM-TSF achieved average reductions of 11.1% in MSE and 8.6% in MAE compared to DLinear, a strong non-LLM baseline. Against Time-LLM, one of the strongest existing LLM-based models, BALM-TSF reduced average MSE by 8.9% and MAE by 5.8%. In few-shot forecasting, where models are trained with limited data, BALM-TSF also showed superior performance, reducing the average MSE of GPT4TS by 10.7% and MAE by 5.8%.
Beyond performance, BALM-TSF is remarkably efficient. It requires significantly fewer trainable parameters (only 0.97 million, or 1.17% of the total parameters) compared to other LLM-based models like Time-LLM (53.44 million trainable parameters). This lightweight design also translates to lower GPU memory usage during inference, making it a highly practical solution.
The ablation studies confirmed the importance of each component, with the learnable prompt and the balanced alignment module being crucial for performance and stability.
Also Read:
- Language Models Learn to Predict and Explain Connections in Dynamic Networks
- Unlocking Deeper Understanding: How Multi-Agent LLMs Are Revolutionizing Causal AI
Conclusion
BALM-TSF offers a compelling solution to the modality imbalance challenge in LLM-based time series forecasting. By intelligently decoupling and then harmoniously aligning textual and temporal information, it unlocks the full potential of LLMs to enrich forecasting inputs. Its lightweight design, coupled with state-of-the-art accuracy, positions BALM-TSF as an effective and efficient benchmark for multimodal time series forecasting. You can find the full research paper here.
