Reproduce “Unified Zero-Shot Time Series Forecasting: A Darts Foundation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhihao Dai, Dennis Bader, Alain Gysi
Darts is extended with a common `FoundationModel` interface that brings Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM into a shared time-series forecasting workflow. The wrappers preserve zero-shot inference while adding access to Darts features such as covariates, probabilistic prediction, fine-tuning, batching, backtesting, and explainability; on a 20-dataset benchmark, three integrations closely match their original implementations, with skill-score differences of at most 1.3 percentage points.
Key insights
- The main contribution is interoperability rather than a new forecasting architecture or pretraining objective: users can switch among supported models with limited changes to application code.
- A unified interface can standardize evaluation and deployment workflows, but it does not by itself resolve data leakage, domain shift, calibration, computational cost, or benchmark-design concerns.
- Darts’ `fit` call can represent either parameter training or zero-shot model initialization, giving frozen and trainable foundation models a consistent programming model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.