ConceptTS Offers Interpretable Multivariate Time-Series Forecasting with LLMs.

This research introduces ConceptTS, an innovative framework designed to bring interpretability to multivariate time-series forecasting. While state-of-the-art forecasters can model complex dependencies, their opaque nature often hinders adoption in critical settings where understanding the "why" behind a prediction is essential. ConceptTS addresses this by structuring its predictions around named, human-readable concepts. The framework leverages a large language model (LLM) to automatically propose relevant concepts for a given task and generate executable rules for labeling them, bypassing the need for costly manual annotation. These concepts are then organized into three complementary "bottlenecks" that capture historical context, local forecast intervals, and the full forecast horizon. A shared decoder combines representations from these concept activations to produce the final forecast, making the decision process transparent and allowing for direct intervention at the concept level. Experiments on the Beijing Multi-Site Air Quality dataset show ConceptTS achieves accuracy comparable to strong black-box models while providing semantically meaningful concept explanations.