Pretrained models that forecast time series zero-shot

What it does The Chronos library is a Python interface to Amazon’s family of pretrained time series forecasting models. It ingests historical series—optionally paired with future covariates—and emits probabilistic forecasts without task-specific training. The package supports three distinct lineages: the original T5-based Chronos that quantizes values into tokens, the patch-based Chronos-Bolt built for direct multi-step quantile regression, and the newer Chronos-2 that targets universal zero-shot forecasting.

The interesting bit The original Chronos models literally treat a time series as a sentence: they scale and quantize observations into a vocabulary, train a T5 transformer with cross-entropy loss, and sample future trajectories the way a language model generates text. Chronos-Bolt ditches tokenization in favor of patching contiguous observations, a design the README says yields up to 250× faster inference and 20× lower memory use than same-sized originals while shaving 5% off error.

Key highlights

• Chronos-2 supports zero-shot univariate, multivariate, and covariate-informed forecasting; the project claims it ranks best among pretrained models on fev-bench, GIFT-Eval, and Chronos Benchmark II, and wins over 90% of head-to-head comparisons against Chronos-Bolt.
• Chronos-Bolt models range from 9M to 205M parameters; the original Chronos family spans 8M to 710M.
• Chronos-Bolt uses patch-based encoding that the README says yields up to 250× faster inference and 20× lower memory use than same-sized original models, plus 5% lower error.
• All model weights are hosted on HuggingFace and the codebase is Apache-2.0 licensed.

Verdict Worth testing if you need quick zero-shot forecasts—especially with exogenous covariates—and are comfortable relying on pretrained transformer outputs. Pass if you require bespoke model architectures or cannot pull weights from HuggingFace.