darts

Darts is a Python 3.10+ library for training, evaluating, and using models on time-series data. Its TimeSeries abstraction supports univariate and multivariate series and can exchange data with pandas, Polars, NumPy, PyArrow, xarray, and other backends. Forecasting models include ARIMA, VARIMA, ExponentialSmoothing, Prophet, N-BEATS, and scikit-learn-compatible regressors. Models follow consistent fit() and predict() workflows and may support multiple series, sample weights, shifted output windows, past-observed or future-known covariates, and static covariates. Many models handle multivariate data and probabilistic forecasts through samples, distribution parameters, or quantiles. Conformal prediction models can produce calibrated quantile intervals for pretrained global forecasting models. The library also provides moving-window backtesting, hierarchical reconciliation transformers, time-series metrics including R² and mean absolute scaled error, SHAP-based explanation for some models, and data transformations such as scaling, imputation, differencing, and Box-Cox. Filtering models include KalmanFilter, GaussianProcessFilter, and MovingAverageFilter. The darts.ad module combines anomaly scorers, detectors, and aggregators. It includes examples such as KMeansScorer and QuantileDetector, can wrap forecasting or filtering models by comparing predictions with actuals, and can apply PyOD detectors through PyODScorer. Deep-learning models use PyTorch Lightning, with support for custom callbacks, GPUs or TPUs, and custom trainers. MLflow integration supports experiment tracking, comparison, and persistence. Installation is available through pip, and the README links to a quickstart, user guide, API reference, examples, and topic articles.