Darts Library Unifies Zero-Shot Time Series Forecasting with Foundation Models

The Darts open-source Python library, a widely used tool for time series analysis since 2020, has introduced a significant update. Recent advancements in foundation models have promised a paradigm shift in forecasting, moving from custom model training to leveraging pre-trained, general-purpose forecasters. However, these foundation models often come as isolated packages with fragmented interfaces, making their joint evaluation and integration into existing pipelines challenging. To address this, Darts has developed a unified `FoundationModel` class collection. This collection provides standardized, full-cycle forecasting interfaces for several prominent foundation models, including Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM, with minimal external dependencies. This integration means that existing Darts pipelines can now incorporate these powerful foundation models with a simple name change. Furthermore, new pipelines can utilize them for zero-shot or fine-tuned forecasting, uncertainty estimation, and backtesting, all within Darts' cohesive framework, alongside its data processing and evaluation tools.