Time series forecasting with a scikit-learn accent

What it does

Darts is a Python library for forecasting and anomaly detection on time series. It provides a common fit() and predict() interface across a wide range of models, from classical statistical methods like ARIMA to deep learning architectures built on PyTorch Lightning. The library also handles multivariate series, external covariates, and probabilistic forecasting, while offering utilities for backtesting and model evaluation.

The interesting bit

The real convenience is the uniformity. You can train a global model on multiple heterogeneous series, plug in any scikit-learn regressor as a forecaster, or turn a forecasting model into an anomaly detector by comparing predictions against actuals. It also supports conformal prediction and hierarchical reconciliation, features often missing from higher-level wrappers.

Key highlights

• Single API for statistical, machine-learning, and deep-learning forecasters
• Native multivariate series support and training on multiple time series simultaneously
• Built-in anomaly detection via PyOD integration or by wrapping forecasting/filtering models
• Probabilistic forecasting with quantile estimates and conformal prediction calibration
• Extras like SHAP explainability, hierarchical reconciliation, and data backend compatibility (pandas, polars, xarray)

Verdict

Worth a look if you juggle multiple time series methodologies and want one consistent interface. Probably overkill if you only need a single, well-understood model and prefer to avoid abstraction layers.