Unified Zero-Shot Time Series Forecasting: A Dar ts Foundation
Zhihao Dai, Dennis Bader, Alain Gysi · 2026
Read Unified Zero-Shot Time Series Forecasting: A Dar ts Foundation by Zhihao Dai, Dennis Bader, Alain Gysi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-27438.txt.
Introduction
Dai, Bader, and Gysi fold the new breed of time series foundation models into Darts, the open-source forecasting library they help maintain. Foundation models promise zero-shot forecasting: pre-trained once, applied to unseen series without retraining. In practice each arrives as an isolated package with its own interface, blocking joint evaluation and pipeline integration. Their unified FoundationModel class collection standardises several leading models behind one full-cycle interface. Existing Darts pipelines can swap one in by changing little more than a model name in this software paper from 2026.
Zero-shot and fine-tuned forecasting, uncertainty estimation, and backtesting all flow through the same framework. Data processing and evaluation tooling from the library come along for free. External dependencies stay minimal, easing installation pain. The work targets the fragmentation that kept foundation models as demos rather than daily tools.
Catalog record
| CREATOR | Zhihao Dai, Dennis Bader, Alain Gysi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.27438
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Unified Zero-Shot Time Series Forecasting: A Dar ts Foundation Zhihao Dai, Dennis Bader, Alain Gysi Abstract Since its initial release in 2020, Dar ts has become a widely used open-source Python library for time series analysis. A series of foundation models have recently claimed accuracy improvements in zero-shot forecasting, promising a paradigm shift from training custom models to harnessing pre-trained general-purpose forecasters. Foundation models, however, are often released as isolated packages with fragmented interfaces and limited interoperability with common tooling, making joint evaluation and integration within complete pipelines difficult. In Dar ts , we developed a unified FoundationModel class collection (Chronos-2, TimesFM 2.5, TiRex, PatchTST-FM) that provides standardized, full-cycle forecasting interfaces with minimal external dependencies for integrating foundation models into the ecosystem. Existing Dar ts pipelines can now use foundation models with only a name change; new pipelines can use them for zero-shot or fine-tuned forecasting, uncertainty estimation, and backtesting, combined with data processing and evaluation tooling, all within a unified framework. ∗ Equal contribution https://github.com/unit8co/darts Quickstart ⬇ # Install via ‘ pip install darts [ torch ]‘ from darts . datasets import WineDataset from darts . models import TimesFM2p5Model series = WineDataset (). load (). astype (" f ") model = TimesFM2p5Model ( input_chunk_length =36, output_chunk_length =12 ). fit ( series ) forecast = model . predict ( n =12, series = series ) 1 Introduction Pre-trained on massive real-world and synthetic data ( 19 ; 1 ; 20 ) , time series foundation models (TSFMs) ( 17 ; 20 ; 43 ; 4 ; 11 ; 5 ; 3 ; 21 ; 41 ; 16 ) capture diverse temporal patterns and are emerging as general-purpose forecasters applicable to unseen time series without further training, i.e., zero-shot forecasting. All of them have successively claimed improvements over series- and task-specific models on public benchmarks ( 44 ; 1 ; 37 ) and envisioned a future where TSFMs replace specialized models for most, if not all, forecasting tasks. Successful applications include forecasting for respiratory disease incidence ( 13 ) , retail sales ( 9 ; 35 ) , and defect detection ( 45 ) . Yet that vision will remain a vision should TSFMs continue to live in isolated packages with neither unified APIs nor interoperability with common time series tooling. Existing forecasting pipelines desiring TSFMs must be retrofitted for idiosyncratic interfaces of each TSFM. Dependency conflicts and divergent behaviors (e.g., probabilistic sampling) often ensue with multiple TSFMs. Further, new pipelines using TSFMs would still demand common processing (e.g., resampling) and evaluation (e.g., metrics) tooling, which is either unavailable or has been rewritten in TSFM packages, leading to fragmentation and duplicated efforts. Since its initial release in 2020, Dar ts ( 23 ) has aspired to be a user-friendly, versatile machine learning library for time series data, providing unified APIs for forecasting, classification, and anomaly detection and encompassing data transformation and evaluation tooling. While TSFMs are still in their infancy and face concerns about generalizability ( 32 ) and benchmarking practices ( 7 ) , we believe in the value of a common TSFM foundation facilitating user adoption and fair comparison, and that Dar ts is well-positioned to provide it. Table 1 : Foundation model support in open-source Python libraries, including model availability, forecasting capabilities (covariates, probabilistic, and multiple series), fine-tuning capability, and explainability, as of June 2026. 1 1 footnotemark: 1 Library Models 1 1 footnotemark: 1 Forecasting Fine- Tuning XAI Cov. Prob. Multi. Dar ts 4 \checkmark \checkmark \checkmark \checkmark \checkmark TimeCopilot 11 \times \checkmark \checkmark \checkmark \times nixtla 1 \checkmark \checkmark \checkmark \checkmark \checkmark sktime 11 \checkmark \checkmark \checkmark \checkmark \times AutoGluon 2 \checkmark \checkmark \checkmark \checkmark \times skforecast 4 \checkmark \checkmark \checkmark \times \times Transformers 1 \times \checkmark \checkmark \checkmark \times aeon 0 - - - - - NF 0 - - - - - PTF 0 - - - - - Models from the same family are counted as one. In Dar ts , we developed a FoundationModel class collection with unified APIs and minimal external dependencies, initially comprising Chronos-2 ( 3 ) , TimesFM 2.5 ( 11 ) , TiRex ( 5 ) , and PatchTST-FM ( 41 ) . A three-liner Quickstart is shown on the cover, and darts[torch] is the only prerequisite installation. Within the Dar ts ecosystem, end users can now co-deploy TSFMs with specialized models and common tooling, ensuring consistent forecasting protocols and minimizing dependency conflicts. Under unified APIs, switching between models is as easy as a name change, and joint evaluation or ensembling becomes straightforward. On the other end, TSFM providers can ship their models in Dar ts to reach a wide user base and reduce the effort of reinventing common tooling, knowing that we offer a level playing field for all TSFMs and a short path to end users. Dar ts is not the only framework providing zero-shot forecasting; Table 1 lists contemporary open-source libraries and their TSFM support to the best of our knowledge. This includes nixtla ( 17 ) , sktime ( 29 ) , AutoGluon ( 38 ) , skforecast ( 2 ) , aeon ( 31 ) , NeuralForecast ( 33 ) , PyTorch Forecasting ( 6 ) , and Transformers ( 42 ) . A notable mention is TimeCopilot ( 18 ) , an agentic toolkit that uses Language Models to orchestrate a forecasting workflow. We do not claim that Dar ts is superior or more complete than other libraries, since each appeals to its target users in its own way. Our contributions in this paper are: • A unified FoundationModel API integrating four TSFMs with minimal external dependencies. • Native support for covariates, probabilistic forecasting, and fine-tuning under a single interface. • Model-agnostic explainability via SHAP values for all PyTorch-based models. • Full interoperability with existing Dar ts tooling for evaluation, backtesting, and data processing. 2 Design Principles and Key Features 2.1 FoundationModel Unified APIs The new FoundationModel class minimally extends TorchForecastingModel (no extra methods) and is the base class for all PyTorch-native TSFMs. Like all Dar ts models, it provides standardized fit(series: TimeSeries) and predict(n: int) methods and can consume and produce TimeSeries objects, ensuring framework-wide interoperability. TimeSeries is an immutable container ( 23 ) housing 3D time-indexed data, (time, components, samples) , alongside optional fields such as metadata. It can be converted to and from common formats such as pandas DataFrames ( 34 ) and NumPy arrays ( 22 ) . As shown in the Quickstart, fit must be called before predict , even for training-free zero-shot forecasting, to enforce a uniform workflow and allow for model creation and sanity checks. Under the hood, FoundationModel uses the PyTorch Lightning backend ( 14 ) to dispatch training and inference loops. We expose Lightning’s Trainer keyword arguments to users for fine-grained control, such as early stopping and hardware acceleration. All TorchForecastingModel functionalities are supported, including optimized historical forecasting, error backtesting, residual computation, and model loading and saving, providing consistent behavior across models. For TSFM providers, implementing a model requires only two steps: (1) subclass a PLForecastingModule and define its forward pass, and (2) subclass FoundationModel and implement the _create_model method to return a module instance. The rest is handled by Dar ts ,