Training-Free Time Series Classification via In-Context Reasoning with LLM Agents

Researchers introduce FETA, a multi-agent framework that enables large language models to classify time series data without any training or fine-tuning. The system decomposes multivariate time series into individual channels, retrieves similar labeled examples, and uses LLM reasoning to make predictions with confidence scores, achieving competitive accuracy on benchmark datasets.

FETA addresses a fundamental challenge in machine learning: the scarcity of labeled training data for specialized tasks. Traditional time series classification requires task-specific model training, which demands substantial labeled datasets and computational resources. This research demonstrates that reasoning-capable LLMs can function as zero-shot classifiers by leveraging in-context learning with exemplar retrieval, eliminating the need for parameter optimization entirely.

The framework's innovation lies in its decomposition strategy and confidence-weighted aggregation. By treating each channel of a multivariate time series as an independent subproblem, FETA reduces complexity and enables targeted exemplar retrieval. The self-assessed confidence scores provide interpretability—a critical advantage in domains like healthcare or finance where understanding prediction rationale matters as much as accuracy. This approach builds on the broader trend of prompt engineering and retrieval-augmented generation becoming viable alternatives to traditional supervised learning.

For practitioners, FETA offers significant operational advantages. It eliminates preprocessing overhead, model validation cycles, and hardware requirements for fine-tuning. Organizations can deploy it immediately to new time series classification tasks without maintaining separate trained models. The plug-and-play nature reduces technical debt and enables rapid experimentation across domains.

The results on UEA datasets—surpassing trained baselines—suggest that LLM reasoning capability combined with structural pattern matching can compete with purpose-built neural architectures. This validates a broader hypothesis: that foundation models with sufficient reasoning capacity and proper prompting architecture may become universal tools for temporal analysis, reducing the need for specialized model development across industries.

• →FETA enables time series classification without training by decomposing problems into channel-level reasoning tasks with exemplar grounding.
• →The framework achieves competitive accuracy on benchmark datasets by outperforming several trained baseline models despite using zero-shot inference.
• →Confidence-weighted aggregation provides interpretability and reliability estimates, addressing a critical gap in black-box machine learning systems.
• →Elimination of training requirements dramatically reduces deployment complexity and enables immediate application to new domains.
• →This work demonstrates that LLM reasoning capabilities combined with retrieval-augmented design can rival specialized neural architectures for temporal pattern recognition.