Darts Unifies Zero-Shot Time Series Models with a One-Line Swap

Swapping in any of four foundation models for zero-shot time series forecasting now takes a single name-string change in a Darts pipeline. Darts, the open-source Python library that has been around since 2020, just shipped a unified FoundationModel class collection that wraps Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM. No more stitching together isolated packages with incompatible APIs.

Four Foundation Models, One Interface

Each of those models was previously released as its own standalone package with its own predict/forecast interface, its own input format quirks, and its own dependencies. Darts' new abstraction provides a standardized, full-cycle forecasting interface for all four. Minimal external dependencies—just the model weights and a thin wrapper. The library handles data preprocessing, evaluation tooling, and backtesting integrally.

What That Means for Existing Pipelines

If you already have a Darts pipeline, you can point it at any of these foundation models by changing the model name in your code. Same predict(), same backtest(), same uncertainty estimation calls. New pipelines get the same zero-shot or fine-tuned forecasting capabilities without writing glue code. Darts' ecosystem already includes data loading, scaling, metrics, and plotting; now foundation models slot in as first-class citizens.

Zero-Shot Forecasting Gets Practical

The claimed accuracy improvements from these foundation models have been real—but the friction of integrating them into production pipelines has kept many teams on custom models. Darts' move eliminates that friction. One library, one import, one call. For anyone building time series pipelines in Python, this is the difference between theoretical access and actual use.

The next step is obvious: more models get the same treatment, and the Darts team can focus on benchmarking and standardization rather than reinventing wrappers for each new foundation model release.

Source: Unified Zero-Shot Time Series Forecasting: A Darts Foundation
Domain: arxiv.org