Large collection of classical and state-of-the-art forecasting models: from statistical models (such as ARIMA) to deep learning models (such as N-BEATS). We are constantly scouting research and bringing the best models to Darts. Some of Darts’ models have extra features not present in the original versions.

Detect and classify anomalies in your series: It’s trivial to apply PyOD, or Darts’ forecasting and Filtering models to obtain fully fledged anomaly detection models.

Handling multiple series: All of Darts machine learning models support being trained on multiple series, as well as large datasets.

Multivariate series: Each of the time series can have multiple dimensions, to more easily model complex systems

Taking static, past, and future external data into account to improve your forecasts: Use some static information about your target series and/or past-observed as well as future-known external time series as inputs for producing forecasts.

Probabilistic Support: Usually point forecasts are not sufficient. Darts’ models can output whole probability distributions or just the distribution parameters, for example to obtain confidence intervals or take more optimal decisions.

Regression Models: You can also easily use your favorite scikit-learn regression model for forecasting your series

A variety of metrics for evaluating time series’ goodness of fit; from R2-scores to Mean Absolute Scaled Error.

Evaluating forecasts correctly is very important: Darts provides the tools to simulate and evaluate historical forecasts, using moving or expanding time windows.

Understand what your model learned: from feature importances and shap explanations for regression models, to self attention for transformer models.

Data processing: Easily manipulate and/or apply (and potentially revert) common transformations on time series data (window transformations, missing values handling, normalization, etc …)

De-noise your series: Darts provides several filtering methods from Gaussian Process Regressor to Moving Average filters..

All deep learning models are implemented using PyTorch Lightning: supporting among other things custom callbacks, GPUs/TPUs training, checkpointing, weights loading for fine-tuning, etc.