TimeCopilot: Framework for Forecasting combining Time Series Models with LLMs (arxiv.org)

TimeCopilot is an open-source, agentic forecasting framework that unifies Time Series Foundation Models (TSFMs) and Large Language Models (LLMs) behind a single API to automate the end-to-end forecasting pipeline. It handles feature analysis, model selection, cross-validation, ensemble construction and probabilistic forecast generation, while enabling natural-language explanations and direct conversational queries about future trajectories. The system is LLM‑agnostic (compatible with both commercial and open-source LLMs) and supports ensembles across diverse forecasting families, making it practical for both research and production workflows. For the AI/ML community, TimeCopilot is significant because it operationalizes reproducible, explainable, and accessible agentic forecasting: researchers can benchmark and compare TSFMs more easily, and practitioners get automated model orchestration plus human-friendly explanations. On the technical side, TimeCopilot attains state-of-the-art probabilistic forecasting on the large-scale GIFT-Eval benchmark at low cost, demonstrating that combining TSFMs with LLM-driven orchestration and ensemble strategies yields strong, efficient performance. The open-source code, data, and demos provide a foundation for building transparent, conversational forecasting tools and for extending model-agnostic pipelines in time-series ML.