TLDR: TimeCopilot is the first open-source agentic framework that unifies multiple Time Series Foundation Models (TSFMs) with Large Language Models (LLMs) through a single API. It automates the entire forecasting pipeline, from data analysis to forecast generation, provides natural language explanations, and achieves state-of-the-art performance on benchmarks like GIFT-Eval at low cost, making advanced forecasting more accessible and reproducible.
In the rapidly evolving landscape of artificial intelligence, time series forecasting has seen a proliferation of sophisticated models. However, this growth has also introduced significant complexity, with each model often having its own unique API, training pipeline, and data requirements. This fragmentation makes it challenging for forecasters to compare, integrate, and deploy these powerful tools effectively. Addressing this challenge, researchers Azul Garza and Reneé Rosillo have introduced TimeCopilot, an innovative open-source agentic framework designed to unify diverse forecasting models with Large Language Models (LLMs) through a single, accessible API.
TimeCopilot stands out as the first open-source solution that combines multiple Time Series Foundation Models (TSFMs) with LLMs, creating a powerful and intuitive system for automated forecasting. The framework streamlines the entire forecasting pipeline, encompassing crucial steps such as feature analysis, model selection, cross-validation, and the generation of forecasts. Beyond automation, TimeCopilot provides natural language explanations for its decisions and supports direct queries about future predictions, making the complex world of forecasting more transparent and user-friendly.
The Core Design: Agentic Intelligence Meets Forecasting
The design of TimeCopilot is built on two key principles. Firstly, it leverages LLMs to orchestrate actions at every stage of the forecasting process. LLMs act as intelligent controllers, reasoning over various signals and deploying specialized tools as needed. Secondly, LLMs are used to explain both the model selection process and the resulting forecasts in natural language, enhancing trust and accessibility for users.
The framework is divided into two main components: the TimeCopilot Agent and the TimeCopilot Forecaster.
-
TimeCopilot Agent: This component manages the forecasting workflow in three structured steps. It begins with Time Series Feature Analysis, where it computes diagnostics like trend, seasonality, and stationarity to inform model choice. Next, during Model Selection and Evaluation, it proposes candidate models, evaluates them through cross-validation, and escalates to more complex models only when necessary. Finally, in Final Model Selection and Forecasting, it selects the best model, generates forecasts, and interprets patterns, uncertainty, and reliability. A crucial aspect of the Agent is its commitment to explainability, allowing users to query the reasoning behind each decision.
-
TimeCopilot Forecaster: This is the execution engine, providing a unified hub for a wide array of forecasting models. It integrates state-of-the-art TSFMs such as Chronos, Moirai, Sundial, TabPFN, TiRex, TimeGPT, TimesFM, and Toto. In addition, it supports classical statistical baselines like ADIDA, AutoARIMA, AutoETS, and Prophet, as well as machine learning methods like AutoLGBM, and neural networks including AutoNHITS and AutoTFT. The Forecaster also supports ensemble techniques, like MedianEnsemble, to combine forecasts from heterogeneous models for improved robustness.
State-of-the-Art Performance and Accessibility
TimeCopilot’s capabilities have been rigorously tested on the large-scale GIFT-Eval benchmark, which includes over 24 datasets, 144,000 time series, and 177 million data points. The framework demonstrated state-of-the-art probabilistic forecasting performance (CRPS) and achieved the second-best point forecast performance (MASE) among non-leaking models. Importantly, these impressive results were achieved at a low computational cost, approximately $24 for GPU-distributed inference, highlighting its efficiency and reproducibility.
The framework’s API is designed for simplicity, allowing users to seamlessly switch between agent-driven automation and manual benchmarking of specific models. This flexibility, combined with its open-source nature, makes advanced forecasting techniques more accessible to a wider audience, from seasoned forecasters to those new to the field.
Also Read:
- Achieving Balance: BALM-TSF’s Strategy for LLM-Enhanced Time Series Prediction
- Evaluating Large Language Models for Multi-Step Time Series Analysis
Looking Ahead
The introduction of TimeCopilot marks a significant step towards more reproducible, explainable, and accessible agentic forecasting systems. The authors envision future enhancements, including integration with the Model Context Protocol (MCP) for seamless interaction with external tools, expansion into diverse real-world domains such as energy, climate, finance, and supply chain forecasting, and extended capabilities for hierarchical and multivariate forecasting with detailed explanations.
For more in-depth information, you can read the full research paper here.
