TLDR: DualSG is a new framework for multivariate time series forecasting that uses Large Language Models (LLMs) as “semantic guides” rather than direct forecasters. It combines a numerical prediction stream with a textual reasoning stream, using natural language “Time Series Captions” to explicitly summarize trends. This dual-stream approach, along with semantic-guided fusion modules, improves forecasting accuracy and interpretability by addressing issues of numerical imprecision and modality alignment common in other LLM-based methods.
Multivariate Time Series Forecasting, which involves predicting future values based on multiple related data streams, is crucial in many fields like health monitoring, weather prediction, and finance. Traditionally, these predictions relied on single data types or statistical methods. More recently, Large Language Models (LLMs) have been explored for this task, leveraging their powerful reasoning abilities.
However, current approaches using LLMs for time series forecasting face significant challenges. Some methods treat LLMs as direct forecasters by converting numerical data into text, which often leads to a loss of precision and forces LLMs to handle patterns they weren’t designed for. Other methods try to align textual and time series data in a hidden “latent space,” but this often results in alignment difficulties that distort important time series properties.
To overcome these limitations, researchers have introduced a new framework called DualSG: A Dual-Stream Explicit Semantic-Guided Multivariate Time Series Forecasting Framework. This innovative approach redefines the role of LLMs, using them not as standalone forecasters, but as “semantic guidance modules” within a dual-stream system. DualSG combines a numerical forecasting stream, which focuses on precise, fine-grained temporal patterns, with a textual reasoning stream that provides high-level semantic correction.
A key innovation within DualSG is the concept of “Time Series Caption” (TSC). Instead of relying on implicit alignment, TSCs are explicit, natural language summaries that describe trend patterns in the time series data. These captions provide interpretable context for the LLMs, allowing them to refine predictions based on clear semantic information rather than trying to implicitly understand numerical patterns. This explicit guidance helps to mitigate issues like numerical imprecision and the mismatch between LLM design and time series data.
DualSG also incorporates two semantic-aware fusion modules. The first, called SemFuse, uses the generated captions to build sparse and interpretable connections between different variables (channels) in the time series. For instance, if two variables show similar trends described by their captions, SemFuse helps them exchange relevant information, reducing noise and computation compared to traditional methods. The second module, Spatial & Temporal Attention Matrix (STAM), dynamically adjusts the contributions of the numerical and textual streams. This means DualSG can emphasize either fine-grained numerical signals or broader semantic corrections as needed, improving accuracy for both short-term and long-term predictions, especially in complex or changing data environments.
Experiments conducted on various real-world datasets demonstrate that DualSG consistently outperforms 15 state-of-the-art baseline models, including other LLM-based, Transformer-based, CNN-based, and MLP-based models. For example, it significantly reduced Mean Squared Error (MSE) and Mean Absolute Error (MAE) compared to leading LLM baselines. The ablation studies confirmed the critical role of each component, particularly the Time Series Caption Generation (TSCG) module and Multi-scale Adaptive Patching (MAP), in enhancing performance. The research highlights that explicit semantic guidance, when properly integrated, can effectively bridge the gap between LLMs and numerical forecasting tasks. For more technical details, you can refer to the full research paper.
Also Read:
- Advancing Time Series Prediction with Adaptive Fuzzy Logic and Asymmetric Convolution
- Bridging Data Gaps in Time Series with Representation Decomposition
The code for DualSG is made available at https://github.com/BenchCouncil/DualSG.
