Grounding Time Series in Language: Interpretable Reasoning with Large Language Models

Abstract

Can large language models (LLMs) classify time-series data by reasoning like a domain expert—if given the right language? We propose a method that expresses statistical time-series features in natural language, enabling LLMs to perform classification with structured, interpretable reasoning. By grounding low-level signal descriptors in semantic context, our approach reframes time-series classification as a language-based reasoning task. We evaluate this method across 23 diverse univariate datasets spanning biomedical, sensor, and human activity domains. Despite requiring no fine-tuning, it achieves competitive accuracy compared to traditional and foundation model baselines. Our method also enables models to generate expert-style justifications, providing interpretable insights into their decision-making process. We present one of the first large-scale analyses of LLM reasoning over statistical time-series features, examining calibration, explanation structure, and reasoning behavior. This work highlights the potential of language native interfaces for interpretable and trustworthy time-series classification.