source: arxiv machine learning: unified zero-shot time series forecasting: a darts foundation
level: technical
darts, a popular open-source python library for time series analysis, now includes a unified foundationmodel class. this class supports several recent foundation models: chronos-2, timesfm 2.5, tirex, and patchtst-fm. these models offer zero-shot forecasting, meaning they can predict future values without training on specific data. previously, each model came as a separate package with different interfaces, making them hard to compare and use together.
the new class provides a standard way to load, fit, and predict with these models. users can switch between models by changing a single name in their code. the interface handles the full forecasting cycle, from input preparation to output formatting. it also keeps external dependencies minimal, so adding a foundation model does not require installing many extra packages. this design lets existing darts pipelines adopt foundation models with almost no code changes.
this update addresses a key problem in the field: the fragmentation of foundation model releases. by bringing these models under one roof, darts makes it easier to evaluate them side by side. researchers and practitioners can now test which model works best for their data without learning new apis each time. the integration also means that darts' other features, like backtesting and anomaly detection, can be used directly with foundation models.
why it matters: it simplifies testing and deploying zero-shot forecasters, saving time and reducing errors in time series projects.
source: arxiv machine learning: unified zero-shot time series forecasting: a darts foundation