Chronos-2 is AWS's latest time series foundation model (TSFM), released October 2025, featuring 120M parameters and an encoder-only Transformer architecture with patch-based embeddings, time attention, and group attention. It supports four forecasting modes via a 'group ID' mechanism: univariate, multivariate, covariate-informed, and cross-learning for cold-start scenarios. A hands-on case study with synthetic building electricity demand data demonstrates each mode using the chronos-forecasting Python package. Results show zero-shot WAPE of 8.6% (univariate), 5.4% (multivariate), and 4% (covariate-informed), with cross-learning reducing cold-start WAPE from 22.2% to 16.7%. The post also outlines four signals indicating when zero-shot is insufficient and fine-tuning is needed.
Table of contents
1. What is a time series foundation model, and how does it change the analytics workflow?2. Why would a foundation model even work for time series?3. What is Chronos-2, specifically?4. What new things can we actually do with Chronos-2?5. Where does zero-shot stop being enough?ReferencesSort: