TimeID: LLM-Centric Framework for Source-Free Time Series Forecasting

A new research paper introduces TimeID, a framework for source-free time series forecasting that leverages large language models (LLMs) for proxy denoising. The study addresses the challenge of adapting pretrained models to sparse target data without accessing source data, ensuring data protection. TimeID comprises three key components: dual-branch invariant disentangled feature learning, which enforces representation- and gradient-wise invariance; and two other components not fully detailed in the abstract. The approach is inspired by the generalization capabilities of LLMs and aims to maximize value from sparse data in real-world applications, particularly in mobile device contexts. The paper is available on arXiv under identifier 2510.05589, with an announcement type of replace-cross.