Darts unifies zero-shot forecasting with Chronos-2, TimesFM 2.5, and more
Swap one line of code to tap pre-trained forecasters like Chronos-2 or TimesFM 2.5.
Since its 2020 debut, Darts has been a staple open-source Python library for time series analysis. Foundation models – large pre-trained forecasters – have recently promised a paradigm shift: instead of training custom models, users can apply a single general-purpose forecaster out of the box. Yet these models were often released as isolated packages with fragmented interfaces, making joint evaluation and pipeline integration difficult.
Darts’ new `FoundationModel` class collection solves this by standardizing interfaces for Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM. Users get full-cycle forecasting – zero-shot or fine-tuned – plus uncertainty estimation and backtesting, all within a unified framework. Existing Darts pipelines work with just a name change, while new projects can combine these models with data processing and evaluation tooling from day one.
- Standardized interface for Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM foundation models
- Zero-shot, fine-tuned, uncertainty estimation, and backtesting capabilities fully integrated
- Existing Darts pipelines require only a single name change to use any supported foundation model
Why It Matters
Makes powerful zero-shot time series forecasting plug-and-play, accelerating adoption across data science teams.