We do our best to avoid the introduction of breaking changes, but cannot always guarantee backwards compatibility. Changes that may break code which uses a previous release of Darts are marked with a “🔴”.


Full Changelog

For users of the library:




For developers of the library:

0.29.0 (2024-04-17)

For users of the library:


  • 🚀🚀 New forecasting model: TSMixerModel as proposed in this paper. An MLP based model that combines temporal, static and cross-sectional feature information using stacked mixing layers. #2293, by Dennis Bader and Cristof Rojas.

  • 🚀🚀 Improvements to metrics, historical forecasts, backtest, and residuals through major refactor. The refactor includes optimization of multiple process and improvements to consistency, reliability, and the documentation. Some of these necessary changes come at the cost of breaking changes. #2284 by Dennis Bader.

    • Metrics:

      • Optimized all metrics, which now run > n * 20 times faster than before for series with n components/columns. This boosts direct metric computations as well as backtest and residuals computation!

      • Added new metrics:

        • Time aggregated metric merr() (Mean Error)

        • Time aggregated scaled metrics rmsse(), and msse() : The (Root) Mean Squared Scaled Error.

        • “Per time step” metrics that return a metric score per time step: err() (Error), ae() (Absolute Error), se() (Squared Error), sle() (Squared Log Error), ase() (Absolute Scaled Error), sse (Squared Scaled Error), ape() (Absolute Percentage Error), sape() (symmetric Absolute Percentage Error), arre() (Absolute Ranged Relative Error), ql (Quantile Loss)

      • All scaled metrics (mase(), …) now accept insample series that can be overlapping into pred_series (before they had to end exactly one step before pred_series). Darts will handle the correct time extraction for you.

      • Improvements to the documentation:

        • Added a summary list of all metrics to the metrics documentation page

        • Standardized the documentation of each metric (added formula, improved return documentation, …)

      • 🔴 Breaking changes:

        • Improved metric output consistency based on the type of input series, and the applied reductions. For some scenarios, the output type changed compared to previous Darts versions. You can find a detailed description in the metric API documentation.

        • Renamed metric parameter reduction to component_reduction.

        • Renamed metric parameter inter_reduction to series_reduction.

        • quantile_loss() :

          • Renamed to mql() (Mean Quantile Loss).

          • Renamed quantile parameter tau to q.

          • The metric is now multiplied by a factor 2 to make the loss more interpretable (e.g. for q=0.5 it is identical to the MAE)

        • rho_risk() :

          • Renamed to qr() (Quantile Risk).

            • Renamed quantile parameter rho to q.

        • Scaled metrics do not allow seasonality inference anymore with m=None.

        • Custom metrics using decorators multi_ts_support and multivariate_support must now act on multivariate series (possibly containing missing values) instead of univariate series.

    • Historical Forecasts:

      • 🔴 Improved historical forecasts output consistency based on the type of input series : If series is a sequence, historical forecasts will now always return a sequence/list of the same length (instead of trying to reduce to a TimeSeries object). You can find a detailed description in the historical forecasts API documentation.

    • Backtest:

      • Metrics are now computed only once on all series and historical_forecasts, significantly speeding things up when using a large number of series.

      • Added support for scaled metrics as metric (such as ase, mase, …). No extra code required, backtest extracts the correct insample series for you.

      • Added support for passing additional metric (-specific) arguments with parameter metric_kwargs. This allows for example to parallelize the metric computation with n_jobs, customize the metric reduction with *_reduction, specify seasonality m for scaled metrics, etc.

      • 🔴 Breaking changes:

        • Improved backtest output consistency based on the type of input series, historical_forecast, and the applied backtest reduction. For some scenarios, the output type changed compared to previous Darts versions. You can find a detailed description in the backtest API documentation.

        • reduction callable now acts on axis=1 rather than axis=0 to aggregate the metrics per series.

        • Backtest will now raise an error when user supplied historical_forecasts don’t have the expected format based on input series and the last_points_only value.

    • Residuals: While the default behavior of residuals() remains identical, the method is now very similar to backtest() but that it computes any “per time step” metric on historical_forecasts :

      • Added support for multivariate series.

      • Added support for all historical_forecasts() parameters to generate the historical forecasts for the residuals computation.

      • Added support for pre-computed historical forecasts with parameter historical_forecasts.

      • Added support for computing the residuals with any of Darts’ “per time step” metric with parameter metric (e.g. err(), ae(), ape(), …). By default, uses err() (Error).

      • Added support for passing additional metric arguments with parameter metric_kwargs. This allows for example to parallelize the metric computation with n_jobs, specify seasonality m for scaled metrics, etc.

      • 🔴 Improved residuals output and consistency based on the type of input series and historical_forecast. For some scenarios, the output type changed compared to previous Darts versions. You can find a detailed description in the residuals API documentation.

  • Improvements to TimeSeries :

    • from_group_dataframe() now supports parallelized creation over the pandas.DataFrame groups. This can be enabled with parameter n_jobs. #2292 by Bohdan Bilonoha.

    • New method slice_intersect_values(), which returns the sliced values of a series, where the time index has been intersected with another series. #2284 by Dennis Bader.

    • Performance boost for methods: slice_intersect(), has_same_time_as(). #2284 by Dennis Bader.

  • Improvements to forecasting models:

    • Improvements to RNNModel, #2329 by Dennis Bader:

      • 🔴 Enforce training_length>input_chunk_length since otherwise, during training the model is never run for as many iterations as it will during prediction.

      • Historical forecasts now correctly infer all possible prediction start points for untrained and pre-trained RNNModel.

    • Added a progress bar to RegressionModel when performing optimized historical forecasts (retrain=False and no autoregression) to display the series-level progress. #2320 by Dennis Bader.

    • Renamed private ForecastingModel._is_probabilistic property to public supports_probabilistic_prediction. #2269 by Felix Divo.

  • Other improvements:

    • All InvertibleDataTransformer now supports parallelized inverse transformation for series being a list of lists of TimeSeries (Sequence[Sequence[TimeSeries]]). This type represents the output of historical_forecasts() when using multiple series with last_points_only=False. #2267 by Alicja Krzeminska-Sciga.

    • Added release notes to the Darts Documentation. #2333 by Dennis Bader.

    • 🔴 Moved around utils functions to clearly separate Darts-specific from non-Darts-specific logic, #2284 by Dennis Bader:

      • Moved function generate_index() from darts.utils.timeseries_generation to darts.utils.utils

      • Moved functions retain_period_common_to_all(), series2seq(), seq2series(), get_single_series() from darts.utils.utils to darts.utils.ts_utils.


  • Fixed the order of the features when using component-specific lags so that they are grouped by values, then by components (before, they were grouped by components, then by values). #2272 by Antoine Madrona.

  • Fixed a bug when using a dropout with a TorchForecastingModel and pytorch lightning versions >= 2.2.0, where the dropout was not properly activated during training. #2312 by Dennis Bader.

  • Fixed a bug when performing historical forecasts with an untrained TorchForecastingModel and using covariates, where the historical forecastable time index generation did not take the covariates into account. #2329 by Dennis Bader.

  • Fixed a bug in quantile_loss, where the loss was computed on all samples rather than only on the predicted quantiles. #2284 by Dennis Bader.

  • Fixed a segmentation fault that some users were facing when importing a LightGBMModel. #2304 by Dennis Bader.

  • Fixed type hint warning “Unexpected argument” when calling historical_forecasts() caused by the _with_sanity_checks decorator. The type hinting is now properly configured to expect any input arguments and return the output type of the method for which the sanity checks are performed for. #2286 by Dennis Bader.

For developers of the library:

  • Fixed failing docs build by adding new dependency lxml_html_clean for nbsphinx. #2303 by Dennis Bader.

  • Bumped black from 24.1.1 to 24.3.0. #2308 by Dennis Bader.

  • Bumped codecov-action from v2 to v4 and added codecov token as repository secret for codecov upload authentication in CI pipelines. #2309 and #2312 by Dennis Bader.

  • Improvements to linting, switch from flake8 to Ruff. #2323 by Jirka Borovec.

0.28.0 (2024-03-05)

For users of the library:


  • Improvements to GlobalForecastingModel :

    • 🚀🚀🚀 All global models (regression and torch models) now support shifted predictions with model creation parameter output_chunk_shift. This will shift the output chunk for training and prediction by output_chunk_shift steps into the future. #2176 by Dennis Bader.

  • Improvements to TimeSeries, #2196 by Dennis Bader:

    • 🚀🚀🚀 Significant performance boosts for several TimeSeries methods resulting increased efficiency across the entire Darts library. Up to 2x faster creation times for series indexed with “regular” frequencies (e.g. Daily, hourly, …), and >100x for series indexed with “special” frequencies (e.g. “W-MON”, …). Affects:

      • All TimeSeries creation methods

      • Additional boosts for slicing with integers and Timestamps

      • Additional boosts for from_group_dataframe() by performing some of the heavy-duty computations on the entire DataFrame, rather than iteratively on the group level.

    • Added option to exclude some group_cols from being added as static covariates when using TimeSeries.from_group_dataframe() with parameter drop_group_cols.

  • 🚀 New global baseline models that use fixed input and output chunks for prediction. This offers support for univariate, multivariate, single and multiple target series prediction, one-shot- or autoregressive/moving forecasts, optimized historical forecasts, batch prediction, prediction from datasets, and more. #2261 by Dennis Bader.

    • GlobalNaiveAggregate : Computes an aggregate (using a custom or built-in torch function) for each target component over the last input_chunk_length points, and repeats the values output_chunk_length times for prediction. Depending on the parameters, this model can be equivalent to NaiveMean and NaiveMovingAverage.

    • GlobalNaiveDrift : Takes the slope of each target component over the last input_chunk_length points and projects the trend over the next output_chunk_length points for prediction. Depending on the parameters, this model can be equivalent to NaiveDrift.

    • GlobalNaiveSeasonal : Takes the target component value at the input_chunk_lengthth point before the end of the target series, and repeats the values output_chunk_length times for prediction. Depending on the parameters, this model can be equivalent to NaiveSeasonal.

  • Improvements to TorchForecastingModel :

    • Added support for additional lr scheduler configuration parameters for more control (“interval”, “frequency”, “monitor”, “strict”, “name”). #2218 by Dennis Bader.

  • Improvements to RegressionModel, #2246 by Antoine Madrona:

    • Added a get_estimator() method to access the underlying estimator

    • Added attribute lagged_label_names to identify the forecasted step and component of each estimator

    • Updated the docstring of get_multioutout_estimator()

  • Other improvements:

    • Added argument keep_names to WindowTransformer and window_transform to indicate whether the original component names should be kept. #2207 by Antoine Madrona.

    • Added new helper function darts.utils.utils.n_steps_between() to efficiently compute the number of time steps (periods) between two points with a given frequency. Improves efficiency for regression model tabularization by avoiding pd.date_range(). #2176 by Dennis Bader.

    • 🔴 Changed the default start value in ForecastingModel.gridsearch() from 0.5 to None, to make it consistent with historical_forecasts and other methods. #2243 by Thomas Kientz.

    • Improvements to ARIMA documentation: Specified possible p, d, P, D, trend advanced options that are available in statsmodels. More explanations on the behaviour of the parameters were added. #2142 by MarcBresson.


  • Fixed a bug when using RegressionModel with lags=None, some lags_*covariates, and the covariates starting after or at the same time as the first predictable time step; the lags were not extracted from the correct indices. #2176 by Dennis Bader.

  • Fixed a bug when calling window_transform on a TimeSeries with a hierarchy. The hierarchy is now only preseved for single transformations applied to all components, or removed otherwise. #2207 by Antoine Madrona.

  • Fixed a bug in probabilistic, where the model attribute was not pointing to all underlying estimators. #2205 by Antoine Madrona.

  • Raise an error in RegressionEsembleModel when the regression_model was created with multi_models=False (not supported). #2205 by Antoine Madrona.

  • Fixed a bug in coefficient_of_variation() with intersect=True, where the coefficient was not computed on the intersection. #2202 by Antoine Madrona.

  • Fixed a bug in gridsearch() with use_fitted_values=True, where the model was not propely instantiated for sanity checks. #2222 by Antoine Madrona.

  • Fixed a bug in TimeSeries.append/prepend_values(), where the component names and the hierarchy were dropped. #2237 by Antoine Madrona.

  • Fixed a bug in get_multioutput_estimator(), where the index of the estimator was incorrectly calculated. #2246 by Antoine Madrona.

  • 🔴 Fixed a bug in datetime_attribute_timeseries(), where 1-indexed attributes were not properly handled. Also, 0-indexing is now enforced for all the generated encodings. #2242 by Antoine Madrona.


For developers of the library:

  • Updated pre-commit hooks to the latest version using pre-commit autoupdate. Change pyupgrade pre-commit hook argument to --py38-plus. #2228 by MarcBresson.

  • Bumped dev dependencies to newest versions, #2248 by Dennis Bader:

    • black[jupyter]: from 22.3.0 to 24.1.1

    • flake8: from 4.0.1 to 7.0.0

    • isort: from 5.11.5 to 5.13.2

    • pyupgrade: 2.31.0 from to v3.15.0

0.27.2 (2024-01-21)

For users of the library:


  • Added darts.utils.statistics.plot_ccf that can be used to plot the cross correlation between a time series (e.g. target series) and the lagged values of another time series (e.g. covariates series). #2122 by Dennis Bader.

  • Improvements to TimeSeries : Improved the time series frequency inference when using slices or pandas DatetimeIndex as keys for __getitem__. #2152 by DavidKleindienst.


  • Fixed a bug when using a TorchForecastingModel with use_reversible_instance_norm=True and predicting with n > output_chunk_length. The input normalized multiple times. #2160 by FourierMourier.

For developers of the library:

0.27.1 (2023-12-10)

For users of the library:


  • 🔴 Added CustomRNNModule and CustomBlockRNNModule for defining custom RNN modules that can be used with RNNModel and BlockRNNModel. The custom model must now be a subclass of the custom modules. #2088 by Dennis Bader.


  • Fixed a bug in historical forecasts, where some fit/predict_kwargs were not passed to the underlying model’s fit/predict methods. #2103 by Dennis Bader.

  • Fixed an import error when trying to create a TorchForecastingModel with PyTorch Lightning v<2.0.0. #2087 by Eschibli.

  • Fixed a bug when creating a RNNModel with a custom model. #2088 by Dennis Bader.

For developers of the library:

  • Added a folder docs/generated_api to define custom .rst files for generating the documentation. #2115 by Dennis Bader.

0.27.0 (2023-11-18)

For users of the library:


  • Improvements to TorchForecastingModel :

    • 🚀🚀 We optimized historical_forecasts() for pre-trained TorchForecastingModel running up to 20 times faster than before (and even more when tuning the batch size)!. #2013 by Dennis Bader.

    • Added callback darts.utils.callbacks.TFMProgressBar to customize at which model stages to display the progress bar. #2020 by Dennis Bader.

    • All InferenceDatasets now support strided forecasts with parameters stride, bounds. These datasets can be used with TorchForecastingModel.predict_from_dataset(). #2013 by Dennis Bader.

  • Improvements to RegressionModel :

    • New example notebook for the RegressionModels explaining features such as (component-specific) lags, output_chunk_length in relation with multi_models, multivariate support, and more. #2039 by Antoine Madrona.

    • XGBModel now leverages XGBoost’s native Quantile Regression support that was released in version 2.0.0 for improved probabilistic forecasts. #2051 by Dennis Bader.

  • Improvements to LocalForecastingModel

    • Added optional keyword arguments dict kwargs to ExponentialSmoothing that will be passed to the constructor of the underlying statsmodels.tsa.holtwinters.ExponentialSmoothing model. #2059 by Antoine Madrona.

  • General model improvements:

    • Added new arguments fit_kwargs and predict_kwargs to historical_forecasts(), backtest() and gridsearch() that will be passed to the model’s fit() and / or predict methods. E.g., you can now set a batch size, static validation series, … depending on the model support. #2050 by Antoine Madrona

    • For transparency, we issue a (removable) warning when performing auto-regressive forecasts with past covariates (with n >= output_chunk_length) to inform users that future values of past covariates will be accessed. #2049 by Antoine Madrona

  • Other improvements:

    • Added support for time index time zone conversion with parameter tz before generating/computing holidays and datetime attributes. Support was added to all Time Axis Encoders, standalone encoders and forecasting models’ add_encoders, time series generation utils functions holidays_timeseries() and datetime_attribute_timeseries(), and TimeSeries methods add_datetime_attribute() and add_holidays(). #2054 by Dennis Bader.

    • Added new data transformer: MIDAS, which uses mixed-data sampling to convert TimeSeries from high frequency to low frequency (and back). #1820 by Boyd Biersteker, Antoine Madrona and Dennis Bader.

    • Added new dataset ElectricityConsumptionZurichDataset : The dataset contains the electricity consumption of households in Zurich, Switzerland from 2015-2022 on different grid levels. We also added weather measurements for Zurich which can be used as covariates for modelling. #2039 by Antoine Madrona and Dennis Bader.

    • Adapted the example notebooks to properly apply data transformers and avoid look-ahead bias. #2020 by Samriddhi Singh.


  • Fixed a bug when calling historical_forecasts() and overlap_end=False that did not generate the last possible forecast. #2013 by Dennis Bader.

  • Fixed a bug when calling optimized historical_forecasts() for a RegressionModel trained with varying component-specific lags. #2040 by Antoine Madrona.

  • Fixed a bug when using encoders with RegressionModel and series with a non-evenly spaced frequency (e.g. Month Begin). This raised an error during lagged data creation when trying to divide a pd.Timedelta by the ambiguous frequency. #2034 by Antoine Madrona.

  • Fixed a bug when loading the weights of a TorchForecastingModel that was trained with a precision other than float64. #2046 by Freddie Hsin-Fu Huang.

  • Fixed broken links in the Transfer learning example notebook with publicly hosted version of the three datasets. #2067 by Antoine Madrona.

  • Fixed a bug when using NLinearModel on multivariate series with covariates and normalize=True. #2072 by Antoine Madrona.

  • Fixed a bug when using DLinearModel and NLinearModel on multivariate series with static covariates shared across components and use_static_covariates=True. #2070 by Antoine Madrona.

For developers of the library:

No changes.

0.26.0 (2023-09-16)

For users of the library:


  • Improvements to RegressionModel, #1962 by Antoine Madrona:

    • 🚀🚀 All models now support component/column-specific lags for target, past, and future covariates series.

  • Improvements to TorchForecastingModel :

    • 🚀 Added RINorm (Reversible Instance Norm) as an input normalization option for all models except RNNModel. Activate it with model creation parameter use_reversible_instance_norm. #1969 by Dennis Bader.

    • 🔴 Added past covariates feature projection to TiDEModel with parameter temporal_width_past following the advice of the model architect. Parameter temporal_width was renamed to temporal_width_future. Additionally, added the option to bypass the feature projection with temporal_width_past/future=0. #1993 by Dennis Bader.

  • Improvements to EnsembleModel, #1815 by Antoine Madrona and Dennis Bader:

    • 🔴 Renamed model constructor argument models to forecasting_models.

    • 🚀🚀 Added support for pre-trained GlobalForecastingModel as forecasting_models to avoid re-training when ensembling. This requires all models to be pre-trained global models.

    • 🚀 Added support for generating the forecasting_model forecasts (used to train the ensemble model) with historical forecasts rather than direct (auto-regressive) predictions. Enable it with train_using_historical_forecasts=True at model creation.

    • Added an example notebook for ensemble models.

  • Improvements to historical forecasts, backtest and gridsearch, #1866 by Antoine Madrona:

    • Added support for negative start values to start historical forecasts relative to the end of the target series.

    • Added a new argument start_format that allows to use an integer start either as the index position or index value/label for series indexed with a pd.RangeIndex.

    • Added support for TimeSeries with a RangeIndex starting at a negative integer.

  • Other improvements:

    • Reduced the size of the Darts docker image unit8/darts:latest, and included all optional models as well as dev requirements. #1878 by Alex Colpitts.

    • Added short examples in the docstring of all the models, including covariates usage and some model-specific parameters. #1956 by Antoine Madrona.

    • Added method TimeSeries.cumsum() to get the cumulative sum of the time series along the time axis. #1988 by Eliot Zubkoff.


  • Fixed a bug in TimeSeries.from_dataframe() when using a pandas.DataFrame with != None. #1938 by Antoine Madrona.

  • Fixed a bug in RegressionEnsembleModel.extreme_lags when the forecasting models have only covariates lags. #1942 by Antoine Madrona.

  • Fixed a bug when using TFTExplainer with a TFTModel running on GPU. #1949 by Dennis Bader.

  • Fixed a bug in TorchForecastingModel.load_weights() that raised an error when loading the weights from a valid architecture. #1952 by Antoine Madrona.

  • Fixed a bug in NLinearModel where normalize=True and past covariates could not be used at the same time. #1873 by Eliot Zubkoff.

  • Raise an error when an EnsembleModel containing at least one LocalForecastingModel is calling historical_forecasts with retrain=False. #1815 by Antoine Madrona.

  • 🔴 Dropped support for lambda functions in add_encoders’s “custom” encoder in favor of named functions to ensure that models can be exported. #1957 by Antoine Madrona.

For developers of the library:


0.25.0 (2023-08-04)

For users of the library:


  • 🔴 Removed Prophet, LightGBM, and CatBoost dependencies from PyPI packages (darts, u8darts, u8darts[torch]), and conda-forge packages (u8darts, u8darts-torch) to avoid installation issues that some users were facing (installation on Apple M1/M2 devices, …). #1589 by Julien Herzen and Dennis Bader.

    • The models are still supported by installing the required packages as described in our installation guide.

    • The Darts package including all dependencies can still be installed with PyPI package u8darts[all] or conda-forge package u8darts-all.

    • Added new PyPI flavor u8darts[notorch], and conda-forge flavor u8darts-notorch which are equivalent to the old u8darts installation (all dependencies except neural networks).

  • 🔴 Removed support for Python 3.7 #1864 by Dennis Bader.


  • General model improvements:

    • 🚀🚀 Optimized historical_forecasts() for RegressionModel when retrain=False and forecast_horizon <= output_chunk_length by vectorizing the prediction. This can run up to 700 times faster than before! #1885 by Antoine Madrona.

    • Improved efficiency of historical_forecasts() and backtest() for all models giving significant process time reduction for larger number of predict iterations and series. #1801 by Dennis Bader.

    • 🚀🚀 Added support for direct prediction of the likelihood parameters to probabilistic models using a likelihood (regression and torch models). Set predict_likelihood_parameters=True when calling predict(). #1811 by Antoine Madrona.

    • 🚀🚀 New forecasting model: TiDEModel as proposed in this paper. An MLP based encoder-decoder model that is said to outperform many Transformer-based architectures. #1727 by Alex Colpitts.

    • Prophet now supports conditional seasonalities, and properly handles all parameters passed to Prophet.add_seasonality() and model creation parameter add_seasonalities #1829 by Idan Shilon.

    • Added method generate_fit_predict_encodings() to generate the encodings (from add_encoders at model creation) required for training and prediction. #1925 by Dennis Bader.

    • Added support for PathLike to the save() and load() functions of all non-deep learning based models. #1754 by Simon Sudrich.

    • Added model property ForecastingModel.supports_multivariate to indicate whether the model supports multivariate forecasting. #1848 by Felix Divo.

  • Improvements to EnsembleModel :

    • Model creation parameter forecasting_models now supports a mix of LocalForecastingModel and GlobalForecastingModel (single TimeSeries training/inference only, due to the local models). #1745 by Antoine Madrona.

    • Future and past covariates can now be used even if forecasting_models have different covariates support. The covariates passed to fit()/predict() are used only by models that support it. #1745 by Antoine Madrona.

    • RegressionEnsembleModel and NaiveEnsembleModel can generate probabilistic forecasts, probabilistics forecasting_models can be sampled to train the regression_model, updated the documentation (stacking technique). #1692 by Antoine Madrona.

  • Improvements to Explainability module:

  • Improvements to documentation #1904 by Dennis Bader:

    • made model sections in, covariates user guide and forecasting model API Reference more user friendly by adding model links and reorganizing them into model categories.

    • added the Dynamic Time Warping (DTW) module and improved its appearance.

  • Other improvements:

    • Improved static covariates column naming when using StaticCovariatesTransformer with a sklearn.preprocessing.OneHotEncoder. #1863 by Anne de Vries.

    • Added MSTL (Season-Trend decomposition using LOESS for multiple seasonalities) as a method option for extract_trend_and_seasonality(). #1879 by Alex Colpitts.

    • Added RINorm (Reversible Instance Norm) as a new input normalization option for TorchForecastingModel. So far only TiDEModel supports it with model creation parameter use_reversible_instance_norm. #1865 by Alex Colpitts.

    • Improvements to TimeSeries.plot() : custom axes are now properly supported with parameter ax. Axis is now returned for downstream tasks. #1916 by Dennis Bader.


  • Fixed an issue not considering original component names for TimeSeries.plot() when providing a label prefix. #1783 by Simon Sudrich.

  • Fixed an issue with the string representation of ForecastingModel when using array-likes at model creation. #1749 by Antoine Madrona.

  • Fixed an issue with TorchForecastingModel.load_from_checkpoint() not properly loading the loss function and metrics. #1759 by Antoine Madrona.

  • Fixed a bug when loading the weights of a TorchForecastingModel trained with encoders or a Likelihood. #1744 by Antoine Madrona.

  • Fixed a bug when using selected target_components with ShapExplainer. #1803 by Dennis Bader.

  • Fixed TimeSeries.__getitem__() for series with a RangeIndex with start != 0 and freq != 1. #1868 by Dennis Bader.

  • Fixed an issue where DTWAlignment.plot_alignment() was not plotting the alignment plot of series with a RangeIndex correctly. #1880 by Ahmet Zamanis and Dennis Bader.

  • Fixed an issue when calling ARIMA.predict() and num_samples > 1 (probabilistic forecasting), where the start point of the simulation was not anchored to the end of the target series. #1893 by Dennis Bader.

  • Fixed an issue when using TFTModel.predict() with full_attention=True where the attention mask was not applied properly. #1392 by Dennis Bader.

For developers of the library:


  • Refactored the ForecastingModelExplainer and ExplainabilityResult to simplify implementation of new explainers. #1392 by Dennis Bader.

  • Adapted all unit tests to run successfully on M1 devices. #1933 by Dennis Bader.

0.24.0 (2023-04-12)

For users of the library:


  • General model improvements:

    • New baseline forecasting model NaiveMovingAverage. #1557 by Janek Fidor.

    • New models StatsForecastAutoCES, and StatsForecastAutoTheta from Nixtla’s statsforecasts library as local forecasting models without covariates support. AutoTheta supports probabilistic forecasts. #1476 by Boyd Biersteker.

    • Added support for future covariates, and probabilistic forecasts to StatsForecastAutoETS. #1476 by Boyd Biersteker.

    • Added support for logistic growth to Prophet with parameters growth, cap, floor. #1419 by David Kleindienst.

    • Improved the model string / object representation style similar to scikit-learn models. #1590 by Janek Fidor.

    • 🔴 Renamed MovingAverage to MovingAverageFilter to avoid confusion with new NaiveMovingAverage model. #1557 by Janek Fidor.

  • Improvements to RegressionModel :

    • Optimized lagged data creation for fit/predict sets achieving a drastic speed-up. #1399 by Matt Bilton.

    • Added support for categorical past/future/static covariates to LightGBMModel with model creation parameters categorical_*_covariates. #1585 by Rijk van der Meulen.

    • Added lagged feature names for better interpretability; accessible with model property lagged_feature_names. #1679 by Antoine Madrona.

    • 🔴 New use_static_covariates option for all models: When True (default), models use static covariates if available at fitting time and enforce identical static covariate shapes across all target series used for training or prediction; when False, models ignore static covariates. #1700 by Dennis Bader.

  • Improvements to TorchForecastingModel :

    • New methods load_weights() and load_weights_from_checkpoint() for loading only the weights from a manually saved model or checkpoint. This allows to fine-tune the pre-trained models with different optimizers or learning rate schedulers. #1501 by Antoine Madrona.

    • New method lr_find() that helps to find a good initial learning rate for your forecasting problem. #1609 by Levente Szabados and Dennis Bader.

    • Improved the user guide and added new sections about saving/loading (checkpoints, manual save/load, loading weights only), and callbacks. #1661 by Antoine Madrona.

    • 🔴 Replaced ":" in save file names with "_" to avoid issues on some operating systems. For loading models saved on earlier Darts versions, try to rename the file names by replacing ":" with "_". #1501 by Antoine Madrona.

    • 🔴 New use_static_covariates option for TFTModel, DLinearModel and NLinearModel : When True (default), models use static covariates if available at fitting time and enforce identical static covariate shapes across all target series used for training or prediction; when False, models ignore static covariates. #1700 by Dennis Bader.

  • Improvements to TimeSeries :

    • Added support for integer indexed input to from_* factory methods, if index can be converted to a pandas.RangeIndex. #1527 by Dennis Bader.

    • Added support for integer indexed input with step sizes (freq) other than 1. #1527 by Dennis Bader.

    • Optimized time series creation with fill_missing_dates=True achieving a drastic speed-up . #1527 by Dennis Bader.

    • from_group_dataframe() now warns the user if there is suspicion of a “bad” time index (monotonically increasing). #1628 by Dennis Bader.

  • Added a parameter to give a custom function name to the transformed output of WindowTransformer; improved the explanation of the window parameter. #1676 and #1666 by Jing Qiang Goh.

  • Added historical_forecasts parameter to backtest() that allows to use precomputed historical forecasts from historical_forecasts(). #1597 by Janek Fidor.

  • Added feature values and SHAP object to ShapExplainabilityResult, giving easy user access to all SHAP-specific explainability results. #1545 by Rijk van der Meulen.

  • New quantile_loss() (pinball loss) metric for probabilistic forecasts. #1559 by Janek Fidor.


  • Fixed an issue in BottomUp/TopDownReconciliator where the order of the series components was not taken into account. #1592 by David Kleindienst.

  • Fixed an issue with DLinearModel not supporting even numbered kernel_size. #1695 by Antoine Madrona.

  • Fixed an issue with RegressionEnsembleModel not using future covariates during training. #1660 by Rajesh Balakrishnan.

  • Fixed an issue where NaiveEnsembleModel prediction did not transfer the series’ component name. #1602 by David Kleindienst.

  • Fixed an issue in TorchForecastingModel that prevented from using multi GPU training. #1509 by Levente Szabados.

  • Fixed a bug when saving a FFT model with trend=None. #1594 by Antoine Madrona.

  • Fixed some issues with PyTorch-Lightning version 2.0.0. #1651 by Dennis Bader.

  • Fixed a bug in QuantileDetector which raised an error when low and high quantiles had identical values. #1553 by Julien Adda.

  • Fixed an issue preventing TimeSeries from being empty. #1359 by Antoine Madrona.

  • Fixed an issue when using backtest() on multiple series. #1517 by Julien Herzen.

  • General fixes to historical_forecasts()

    • Fixed issue where retrain functions were not handled properly; Improved handling of start, and train_length parameters; better interpretability with warnings and improved error messages (warnings can be turned of with show_warnings=False). By #1675 by Antoine Madrona and Dennis Bader.

    • Fixed an issue for several models (mainly ensemble and local models) where automatic start did not respect the minimum required training lengths. #1616 by Janek Fidor and Dennis Bader.

    • Fixed an issue when using a RegressionModel with future covariates lags only. #1685 by Maxime Dumonal.

For developers of the library:


  • Option to skip slow tests locally with pytest . --no-cov -m "not slow". #1625 by Blazej Nowicki.

  • Major refactor of data transformers which simplifies implementation of new transformers. #1409 by Matt Bilton.

0.23.1 (2023-01-12)

Patch release


  • Fix an issue in TimeSeries which made it incompatible with Python 3.7. #1449 by Dennis Bader.

  • Fix an issue with static covariates when series have variable lengths with RegressionModels. #1469 by Eliane Maalouf.

  • Fix an issue with PyTorch Lightning trainer handling. #1459 by Dennis Bader.

  • Fix an issue with historical_forecasts() retraining PyTorch models iteratively instead of from scratch. #1465 by Dennis Bader.

  • Fix an issue with historical_forecasts() not working in some cases when future_covariates are provided and start is not specified. #1481 by Maxime Dumonal.

  • Fix an issue with slice_n_points functions on integer indexes. #1482 by Julien Herzen.

0.23.0 (2022-12-23)

For users of the library:


  • 🚀🚀🚀 Brand new Darts module dedicated to anomaly detection on time series: More info on the API doc page: #1256 by Julien Adda and Julien Herzen.

  • New forecasting models: DLinearModel and NLinearModel as proposed in this paper. #1139 by Julien Herzen and Greg DeVos.

  • New forecasting model: XGBModel implementing XGBoost. #1405 by Julien Herzen.

  • New multi_models option for all RegressionModels: when set to False, uses only a single underlying estimator for multi-step forecasting, which can drastically increase computational efficiency. #1291 by Eliane Maalouf.

  • All RegressionModels (incl. LightGBM, Catboost, XGBoost, Random Forest, …) now support static covariates. #1412 by Eliane Maalouf.

  • historical_forecasts() and backtest() now work on multiple series, too. #1318 by Maxime Dumonal.

  • New window transformation capabilities: TimeSeries.window_transform() and a new WindowTransformer which allow to easily create window features. #1269 by Eliane Maalouf.

  • 🔴 Improvements to TorchForecastingModels : Load models directly to CPU that were trained on GPU. Save file size reduced. Improved PyTorch Lightning Trainer handling fixing several minor issues. Removed deprecated methods load_model and save_model #1371 by Dennis Bader.

  • Improvements to encoders: Added support for encoders to all models with covariate support through add_encoders at model creation. Encoders now generate the correct minimum required covariate time spans for all models. #1338 by Dennis Bader.

  • New datasets available in darts.datasets (ILINetDataset, ExchangeRateDataset, TrafficDataset, WeatherDataset) #1298 by Kamil Wierciak. #1291 by Eliane Maalouf.

  • New Diff transformer, which can difference and “undifference” series #1380 by Matt Bilton.

  • Improvements to KalmanForecaster: The model now accepts different TimeSeries for prediction than the ones used to fit the model. #1338 by Dennis Bader.

  • Backtest functions can now accept a list of metric functions #1333 by Antoine Madrona.

  • Extension of baseline models to work on multivariate series #1373 by Błażej Nowicki.

  • Improvement to TimeSeries.gaps() #1265 by Antoine Madrona.

  • Speedup of TimeSeries.quantile_timeseries() method #1351 by @tranquilitysmile.

  • Some dependencies which can be hard to install (LightGBM, Catboost, XGBoost, Prophet, Statsforecast) are not required anymore (if not installed the corresponding models will not be available) #1360 by Antoine Madrona.

  • Removed IPython as a dependency. #1331 by Erik Hasse

  • Allow the creation of empty TimeSeries #1359 by Antoine Madrona.


0.22.0 (2022-10-04)

For users of the library:


  • New explainability feature. The class ShapExplainer in darts.explainability can provide Shap-values explanations of the importance of each lag and each dimension in producing each forecasting lag for RegressionModels. #909 by Maxime Dumonal.

  • New model: StatsForecastsETS. Similarly to StatsForecastsAutoARIMA, this model offers the ETS model from Nixtla’s statsforecasts library as a local forecasting model supporting future covariates. #1171 by Julien Herzen.

  • Added support for past and future covariates to residuals() function. #1223 by Eliane Maalouf.

  • Added support for retraining model(s) every n iteration and on custom conditions in historical_forecasts method of ForecastingModels. #1139 by Francesco Bruzzesi.

  • Added support for beta-NLL in GaussianLikelihoods, as proposed in this paper. #1162 by Julien Herzen.

  • New LayerNorm alternatives, RMSNorm and LayerNormNoBias #1113 by Greg DeVos.

  • 🔴 Improvements to encoders: improve fitting behavior of encoders’ transformers and solve a couple of issues. Remove support for absolute index encoding. #1257 by Dennis Bader.

  • Overwrite min_train_series_length for Catboost and LightGBM #1214 by Anne de Vries.

  • New example notebook showcasing and end-to-end example of hyperparameter optimization with Optuna #1242 by Julien Herzen.

  • New user guide section on hyperparameter optimization with Optuna and Ray Tune #1242 by Julien Herzen.

  • Documentation on model saving and loading. #1210 by Amadej Kocbek.

  • 🔴 torch_device_str has been removed from all torch models in favor of Pytorch Lightning’s pl_trainer_kwargs method #1244 by Greg DeVos.


0.21.0 (2022-08-12)

For users of the library:


  • New model: Catboost, incl quantile, poisson and gaussian likelihoods support. #1007, #1044 by Jonas Racine.

  • Extension of the add_encoders option to RegressionModels. It is now straightforward to add calendar based or custom past or future covariates to these models, similar to torch models. #1093 by Dennis Bader.

  • Introduction of StaticCovariatesTransformer, categorical static covariate support for TFTModel, example and user-guide updates on static covariates. #1081 by Dennis Bader.

  • ARIMA and VARIMA models now support being applied to a new series, different than the one used for training. #1036 by Samuele Giuliano Piazzetta.

  • All Darts forecasting models now have unified save() and load() methods. #1070 by Dustin Brunner.

  • Improvements in logging. #1034 by Dustin Brunner.

  • Re-integrating Prophet >= 1.1 in core dependencies (as it does not depend on PyStan anymore). #1054 by Julien Herzen.

  • Added a new AustralianTourismDataset. #1141 by Julien Herzen.

  • Added a new notebook demonstrating hierarchical reconciliation. #1147 by Julien Herzen.

  • Added drop_columns() method to TimeSeries. #1040 by @shaido987

  • Speedup static covariates when no casting is needed. #1053 by Julien Herzen.

  • Implemented the min_train_series_length method for the FourTheta and Theta models that overwrites the minimum default of 3 training samples by 2*seasonal_period when appropriate. #1101 by Rijk van der Meulen.

  • Make default formatting optional in plots. #1056 by Colin Delahunty

  • Introduce retrain option in residuals() method. #1066 by Julien Herzen.

  • Improved error messages. #1066 by Julien Herzen.

  • Small readability improvements to user guide. #1039, #1046 by Ryan Russell


0.20.0 (2022-06-22)

For users of the library:



  • A bug with tail() and head() #942 by Julien Herzen.

  • An issue with arguments being reverted for the metric function of gridsearch and backtest #989 by Clara Grotehans.

  • An error checking whether fit() has been called in global models #944 by Julien Herzen.

  • An error in Gaussian Process filter happening with newer versions of sklearn #963 by Julien Herzen.

For developers of the library:


0.19.0 (2022-04-13)

For users of the library:



0.18.0 (2022-03-22)

For users of the library:


  • LinearRegressionModel and LightGBMModel can now be probabilistic, supporting quantile and poisson regression. #831, #853 by Gian Wiher.

  • New models: BATS and TBATS, based on tbats. #816 by Julien Herzen.

  • Handling of stochastic inputs in PyTorch based models. #833 by Julien Herzen.

  • GPU and TPU user guide. #826 by @gsamaras.

  • Added train and validation loss to PyTorch Lightning progress bar. #825 by Dennis Bader.

  • More losses available in darts.utils.losses for PyTorch-based models: SmapeLoss, MapeLoss and MAELoss. #845 by Julien Herzen.

  • Improvement to the seasonal decomposition #862. by Gian Wiher.

  • The gridsearch() method can now return best metric score. #822 by @nlhkh.

  • Removed needless checkpoint loading when predicting. #821 by Dennis Bader.

  • Changed default number of epochs for validation from 10 to 1. #825 by Dennis Bader.


  • Fixed some issues with encoders in fit_from_dataset(). #829 by Julien Herzen.

  • Fixed an issue with covariates slicing for DualCovariatesForecastingModels. #858 by Dennis Bader.

0.17.1 (2022-02-17)

Patch release

For users of the library:


  • Fixed issues with (now deprecated) torch_device_str parameter, and improved documentation related to using devices with PyTorch Lightning. #806 by Dennis Bader.

  • Fixed an issue with ReduceLROnPlateau. #806 by Dennis Bader.

  • Fixed an issue with the periodic basis functions of N-BEATS. #804 by Vladimir Chernykh.

  • Relaxed requirements for pandas; from pandas>=1.1.0 to pandas>=1.0.5. #800 by @adelnick.

0.17.0 (2022-02-15)

For users of the library:


  • 🚀 Support for PyTorch Lightning: All deep learning models are now implemented using PyTorch Lightning. This means that many more features are now available via PyTorch Lightning trainers functionalities; such as tailored callbacks, or multi-GPU training. #702 by Dennis Bader.

  • The RegressionModels now accept an output_chunk_length parameter; meaning that they can be trained to predict more than one time step in advance (and used auto-regressively to predict on longer horizons). #761 by Dustin Brunner.

  • 🔴 TimeSeries “simple statistics” methods (such as mean(), max(), min() etc, …) have been refactored to work natively on stochastic TimeSeries, and over configurable axes. #773 by Gian Wiher.

  • 🔴 TimeSeries now support only pandas RangeIndex as an integer index, and does not support Int64Index anymore, as it became deprecated with pandas 1.4.0. This also now brings the guarantee that TimeSeries do not have missing “dates” even when indexed with integers. #777 by Julien Herzen.

  • New model: KalmanForecaster is a new probabilistic model, working on multivariate series, accepting future covariates, and which works by running the state-space model of a given Kalman filter into the future. The fit() function uses the N4SID algorithm for system identification. #743 by Julien Herzen.

  • The KalmanFilter now also works on TimeSeries containing missing values. #743 by Julien Herzen.

  • The estimators (forecasting and filtering models) now also return their own instance when calling fit(), which allows chaining calls. #741 by Julien Herzen.


For developers of the library:

0.16.1 (2022-01-24)

Patch release

For users of the library:

For developers of the library:

0.16.0 (2022-01-13)

For users of the library:


  • The documentation page has been revamped and now contains a brand new Quickstart guide, as well as a User Guide section, which will be populated over time.

  • The API documentation has been revamped and improved, notably using numpydoc.

  • The datasets building procedure has been improved in RegressionModel, which yields dramatic speed improvements.


  • The KalmanFilter can now do system identification using fit() (using nfoursid).


For developers of the library:

  • We have switched to black for code formatting (this is checked by the CI pipeline).

0.15.0 (2021-12-24)

For users of the library:


  • On-the-fly encoding of position and calendar information in Torch-based models. Torch-based models now accept an option add_encoders parameter, specifying how to use certain calendar and position information as past and/or future covariates on the-fly.


    from darts.dataprocessing.transformers import Scaler
        'cyclic': {'future': ['month']},
        'datetime_attribute': {'past': ['hour', 'dayofweek']},
        'position': {'past': ['absolute'], 'future': ['relative']},
        'custom': {'past': [lambda idx: (idx.year - 1950) / 50]},
        'transformer': Scaler()

    This will add a cyclic encoding of the month as future covariates, add some datetime attributes as past and future covariates, an absolute/relative position (index), and even some custom mapping of the index (such as a function of the year). A Scaler will be applied to fit/transform all of these covariates both during training and inference.

  • The scalers can now also be applied on stochastic TimeSeries.

  • There is now a new argument max_samples_per_ts to the fit() method of Torch-based models, which can be used to limit the number of samples contained in the underlying training dataset, by taking (at most) the most recent max_samples_per_ts training samples per time series.

  • All local forecasting models that support covariates (Prophet, ARIMA, VARIMA, AutoARIMA) now handle covariate slicing themselves; this means that you don’t need to make sure your covariates have the exact right time span. As long as they contain the right time span, the models will slice them for you.

  • and mappers data transformers now work on stochastic TimeSeries.

  • Granger causality function: utils.statistics.granger_causality_tests can test if one univariate TimeSeries “granger causes” another.

  • New stationarity tests for univariate TimeSeries : darts.utils.statistics.stationarity_tests, darts.utils.statistics.stationarity_test_adf and darts.utils.statistics.stationarity_test_kpss.

  • New test coverage badge 🦄


  • Fixed various issues in different notebooks.

  • Fixed a bug handling frequencies in Prophet model.

  • Fixed an issue causing PastCovariatesTorchModels (such as NBEATSModel) prediction to fail when n > output_chunk_length AND n not being a multiple of output_chunk_length.

  • Fixed an issue in backtesting which was causing untrained models not to be trained on the initial window when retrain=False.

  • Fixed an issue causing residuals() to fail for Torch-based models.

For developers of the library:

  • Updated the contribution guidelines

  • The unit tests have been re-organised with submodules following that of the library.

  • All relative import paths have been removed and replaced by absolute paths.

  • pytest and pytest-cov are now used to run tests and compute coverage.

0.14.0 (2021-11-28)

For users of the library:


  • Probabilistic N-BEATS: The NBEATSModel can now produce probabilistic forecasts, in a similar way as all the other deep learning models in Darts (specifying a likelihood and predicting with num_samples >> 1).

  • We have improved the speed of the data loaing functionalities for PyTorch-based models. This should speedup training, typically by a few percents.

  • Added num_loader_workers parameters to fit() and predict() methods of PyTorch-based models, in order to control the num_workers of PyTorch DataLoaders. This can sometimes result in drastic speedups.

  • New method TimeSeries.astype() which allows to easily case (e.g. between np.float64 and np.float32).

  • Added dtype as an option to the time series generation modules.

  • Added a small performance guide for PyTorch-based models.

  • Possibility to specify a (relative) time index to be used as future covariates in the TFT Model. Future covariates don’t have to be specified when this is used.

  • New TFT example notebook.

  • Less strict dependencies: we have loosened the required dependencies versions.


  • A small fix on the Temporal Fusion Transformer TFTModel, which should improve performance.

  • A small fix in the random state of some unit tests.

  • Fixed a typo in Transformer example notebook.

0.13.1 (2021-11-08)

For users of the library:


  • Factory methods in TimeSeries are now classmethods, which makes inheritance of TimeSeries more convenient.


  • An issue which was causing some of the flavours installations not to work

0.13.0 (2021-11-07)

For users of the library:


  • New forecasting model, Temporal Fusion Transformer (TFTModel). A new deep learning model supporting both past and future covariates.

  • Improved support for Facebook Prophet model (Prophet):

    • Added support for fit & predict with future covariates. For instance:, future_covariates=train_covariates) and model.predict(n=len(test), num_sample=1, future_covariates=test_covariates)

    • Added stochastic forecasting, for instance: model.predict(n=len(test), num_samples=200)

    • Added user-defined seasonalities either at model creation with kwarg add_seasonality (Prophet(add_seasonality=kwargs_dict)) or pre-fit with model.add_seasonality(kwargs). For more information on how to add seasonalities,

      see the Prophet docs.

    • Added possibility to predict and return the base model’s raw output with model.predict_raw(). Note that this returns a pd.DataFrame pred_df, which will not be supported for further processing with the Darts API. But it is possible to access Prophet’s methods such as plots with model.model.plot_compenents(pred_df).

  • New n_random_samples in gridsearch() method, which allows to specify a number of (random) hyper parameters combinations to be tried, in order mainly to limit the gridsearch time.

  • Improvements in the checkpointing and saving of Torch models.

    • Now models don’t save checkpoints by default anymore. Set save_checkpoints=True to enable them.

    • Models can be manually saved with YourTorchModel.save_model(file_path) (file_path pointing to the .pth.tar file).

    • Models can be manually loaded with YourTorchModel.load_model(file_path) or the original method YourTorchModel.load_from_checkpoint().

  • New QuantileRegression Likelihood class in darts.utils.likelihood_models. Allows to apply quantile regression loss, and get probabilistic forecasts on all deep learning models supporting likelihoods. Used by default in the Temporal Fusion Transformer.


  • Some issues with darts.concatenate().

  • Fixed some bugs with RegressionModels applied on multivariate series.

  • An issue with the confidence bounds computation in ACF plot.

  • Added a check for some models that do not support retrain=False for historical_forecasts().

  • Small fixes in install instructions.

  • Some rendering issues with bullet points lists in examples.

0.12.0 (2021-09-25)

For users of the library:


  • Improved probabilistic forecasting with neural networks

    • Now all neural networks based forecasting models (except NBEATSModel) support probabilistic forecasting, by providing the likelihood parameter to the model’s constructor method.

    • darts.utils.likelihood_models now contains many more distributions. The complete list of likelihoods available to train neural networks based models is available here:

    • Many of the available likelihood models now offer the possibility to specify “priors” on the distribution’s parameters. Specifying such priors will regularize the training loss to make the output distribution more like the one specified by the prior parameters values.

  • Performance improvements on TimeSeries creation. creating TimeSeries is now be significantly faster, especially for large series, and filling missing dates has also been significantly sped up.

  • New rho-risk metric for probabilistic forecasts.

  • New method darts.utils.statistics.plot_hist() to plot histograms of time series data (e.g. backtest errors).

  • New argument fillna_value to TimeSeries factory methods, allowing to specify a value to fill missing dates (instead of np.nan).

  • Synthetic TimeSeries generated with darts.utils.timeseries_generation methods can now be integer-index (just pass an integer instead of a timestamp for the start argument).

  • Removed some deprecation warnings

  • Updated conda installation instructions


  • Removed extra 1x1 convolutions in TCN Model.

  • Fixed an issue with linewidth parameter when plotting TimeSeries.

  • Fixed a column name issue in datetime attribute time series.

For developers of the library:

  • We have removed the develop branch.

  • We force sklearn<1.0 has we have observed issues with pmdarima and sklearn==1.0

0.11.0 (2021-09-04)

For users of the library:


  • New model: LightGBMModel is a new regression model. Regression models allow to predict future values of the target, given arbitrary lags of the target as well as past and/or future covariates. RegressionModel already works with any scikit-learn regression model, and now LightGBMModel does the same with LightGBM. If you want to activate LightGBM support in Darts, please read the detailed install notes on the README carefully.

  • Added stride support to gridsearch


  • A bug which was causing issues when training on a GPU with a validation set

  • Some issues with custom-provided RNN modules in RNNModel.

  • Properly handle kwargs in the fit function of RegressionModels.

  • Fixed an issue which was causing problems with latest versions of Matplotlib.

  • An issue causing errors in the FFT notebook

0.10.1 (2021-08-19)

For users of the library:


  • A bug with memory pinning that was causing issues with training models on GPUs.


  • Clarified conda support on the README

0.10.0 (2021-08-13)

For users of the library:


  • 🔴 Improvement of the covariates support. Before, some models were accepting a covariates (or exog) argument, but it wasn’t always clear whether this represented “past-observed” or “future-known” covariates. We have made this clearer. Now all covariate-aware models support past_covariates and/or future_covariates argument in their fit() and predict() methods, which makes it clear what series is used as a past or future covariate. We recommend this article for more information and examples.

  • 🔴 Significant improvement of RegressionModel (incl. LinearRegressionModel and RandomForest). These models now support training on multiple (possibly multivariate) time series. They also support both past_covariates and future_covariates. It makes it easier than ever to fit arbitrary regression models (e.g. from scikit-learn) on multiple series, to predict the future of a target series based on arbitrary lags of the target and the past/future covariates. The signature of these models changed: It’s not using “exog” keyword arguments, but past_covariates and future_covariates instead.

  • Dynamic Time Warping. There is a brand new darts.dataprocessing.dtw submodule that implements Dynamic Time Warping between two TimeSeries. It’s also coming with a new dtw metric in darts.metrics. We recommend going over the new DTW example notebook for a good overview of the new functionalities

  • Conda forge installation support (fully supported with Python 3.7 only for now). You can now conda install u8darts-all.

  • TimeSeries.from_csv() allows to obtain a TimeSeries from a CSV file directly.

  • Optional cyclic encoding of the datetime attributes future covariates; for instance it’s now possible to call my_series.add_datetime_attribute('weekday', cyclic=True), which will add two columns containing a sin/cos encoding of the weekday.

  • Default seasonality inference in ExponentialSmoothing. If left to None, the seasonal_periods is inferred from the freq of the provided series.

  • Various documentation improvements.


  • Now transformations and forecasting maintain the columns’ names of the TimeSeries. The generation module darts.utils.timeseries_generation also comes with better default columns names.

  • Some issues with our Docker build process

  • A bug with GPU usage


  • For probabilistic PyTorch based models, the generation of multiple samples (and series) at prediction time is now vectorized, which improves inference performance.

0.9.1 (2021-07-17)

For users of the library:


  • Improved GaussianProcessFilter, now handling missing values, and better handling time series indexed by datetimes.

  • Improved Gaussian Process notebook.


  • TimeSeries now supports indexing using pandas.Int64Index and not just pandas.RangeIndex, which solves some indexing issues.

  • We have changed all factory methods of TimeSeries to have fill_missing_dates=False by default. This is because in some cases inferring the frequency for missing dates and resampling the series is causing significant performance overhead.

  • Fixed backtesting to make it work with integer-indexed series.

  • Fixed a bug that was causing inference to crash on GPUs for some models.

  • Fixed the default folder name, which was causing issues on Windows systems.

  • We have slightly improved the documentation rendering and fixed the titles of the documentation pages for RNNModel and BlockRNNModel to distinguish them.


  • The dependencies are not pinned to some exact versions anymore.

For developers of the library:

  • We have fixed the building process.

0.9.0 (2021-07-09)

For users of the library:


  • Multiple forecasting models can now produce probabilistic forecasts by specifying a num_samples parameter when calling predict(). Stochastic forecasts are stored by utilizing the new samples dimension in the refactored TimeSeries class (see ‘Changed’ section). Models supporting probabilistic predictions so far are ARIMA, ExponentialSmoothing, RNNModel and TCNModel.

  • Introduced LikelihoodModel class which is used by probabilistic TorchForecastingModel classes in order to make predictions in the form of parametrized distributions of different types.

  • Added new abstract class TorchParametricProbabilisticForecastingModel to serve as parent class for probabilistic models.

  • Introduced new FilteringModel abstract class alongside MovingAverage, KalmanFilter and GaussianProcessFilter as concrete implementations.

  • Future covariates are now utilized by TorchForecastingModels when the forecasting horizon exceeds the output_chunk_length of the model. Before, TorchForecastingModel instances could only predict beyond their output_chunk_length if they were not trained on covariates, i.e. if they predicted all the data they need as input. This restriction has now been lifted by letting a model not only consume its own output when producing long predictions, but also utilizing the covariates known in the future, if available.

  • Added a new RNNModel class which utilizes and rnn module as both encoder and decoder. This new class natively supports the use of the most recent future covariates when making a forecast. See documentation for more details.

  • Introduced optional epochs parameter to the TorchForecastingModel.predict() method which, if provided, overrides the n_epochs attribute in that particular model instance and training session.

  • Added support for TimeSeries with a pandas.RangeIndex instead of just allowing pandas.DatetimeIndex.

  • ForecastingModel.gridsearch now makes use of parallel computation.

  • Introduced a new force_reset parameter to TorchForecastingModel.__init__() which, if left to False, will prevent the user from overriding model data with the same name and directory.


  • Solved bug occurring when training NBEATSModel on a GPU.

  • Fixed crash when running NBEATSModel with log_tensorboard=True

  • Solved bug occurring when training a TorchForecastingModel instance with a batch_size bigger than the available number of training samples.

  • Some fixes in the documentation, including adding more details

  • Other minor bug fixes


  • 🔴 The TimeSeries class has been refactored to support stochastic time series representation by adding an additional dimension to a time series, namely samples. A time series is now based on a 3-dimensional xarray.DataArray with shape (n_timesteps, n_components, n_samples). This overhaul also includes a change of the constructor which is incompatible with the old one. However, factory methods have been added to create a TimeSeries instance from a variety of data types, including pd.DataFrame. Please refer to the documentation of TimeSeries for more information.

  • 🔴 The old version of RNNModel has been renamed to BlockRNNModel.

  • The historical_forecast() and backtest() methods of ForecastingModel have been reorganized a bit by making use of new wrapper methods to fit and predict models.

  • Updated to reflect the new additions to the library.

0.8.1 (2021-05-22)


  • Some fixes in the documentation


  • The way to instantiate Dataset classes; datasets should now be used like this .. code-block:

    from darts.datasets import AirPassengers
    ts: TimeSeries = AirPassengers().load()

0.8.0 (2021-05-21)

For users of the library:


  • RandomForest algorithm implemented. Uses the scikit-learn RandomForestRegressor to predict future values from (lagged) exogenous variables and lagged values of the target.

  • darts.datasets is a new submodule allowing to easily download, cache and import some commonly used time series.

  • Better support for processing sequences of TimeSeries.

    • The Transformers, Pipelines and metrics have been adapted to be used on sequences of TimeSeries (rather than isolated series).

    • The inference of neural networks on sequences of series has been improved

  • There is a new utils function darts.utils.model_selection.train_test_split which allows to split a TimeSeries or a sequence of TimeSeries into train and test sets; either along the sample axis or along the time axis. It also optionally allows to do “model-aware” splitting, where the split reclaims as much data as possible for the training set.

  • Our implementation of N-BEATS, NBEATSModel, now supports multivariate time series, as well as covariates.


  • RegressionModel is now a user exposed class. It acts as a wrapper around any regression model with a fit() and predict() method. It enables the flexible usage of lagged values of the target variable as well as lagged values of multiple exogenous variables. Allowed values for the lags argument are positive integers or a list of positive integers indicating which lags should be used during training and prediction, e.g. lags=12 translates to training with the last 12 lagged values of the target variable. lags=[1, 4, 8, 12] translates to training with the previous value, the value at lag 4, lag 8 and lag 12.

  • 🔴 StandardRegressionModel is now called LinearRegressionModel. It implements a linear regression model from sklearn.linear_model.LinearRegression. Users who still need to use the former StandardRegressionModel with another sklearn model should use the RegressionModel now.


  • We have fixed a bug arising when multiple scalers were used.

  • We have fixed a small issue in the TCN architecture, which makes our implementation follow the original paper more closely.

For developers of the library:


0.7.0 (2021-04-14)

Full Changelog

For users of the library:


  • darts Pypi package. It is now possible to pip install darts. The older name u8darts is still maintained and provides the different flavours for lighter installs.

  • New forecasting model available: VARIMA (Vector Autoregressive moving average).

  • Support for exogeneous variables in ARIMA, AutoARIMA and VARIMA (optional exog parameter in fit() and predict() methods).

  • New argument dummy_index for TimeSeries creation. If a series is just composed of a sequence of numbers without timestamps, setting this flag will allow to create a TimeSeries which uses a “dummy time index” behind the scenes. This simplifies the creation of TimeSeries in such cases, and makes it possible to use all forecasting models, except those that explicitly rely on dates.

  • New method TimeSeries.diff() returning differenced TimeSeries.

  • Added an example of RegressionEnsembleModel in intro notebook.


  • Improved N-BEATS example notebook.

  • Methods TimeSeries.split_before() and split_after() now also accept integer or float arguments (in addition to timestamp) for the breaking point (e.g. specify 0.8 in order to obtain a 80%/20% split).

  • Argument value_cols no longer has to be provided if not necessary when creating a TimeSeries from a DataFrame.

  • Update of dependency requirements to more recent versions.


  • Fix issue with MAX_TORCH_SEED_VALUE on 32-bit architectures (

  • Corrected a bug in TCN inference, which should improve accuracy.

  • Fix historical forecasts not returning last point.

  • Fixed bug when calling the TimeSeries.gaps() function for non-regular time frequencies.

  • Many small bug fixes.

0.6.0 (2021-02-02)

Full Changelog

For users of the library:


  • Pipeline.invertible() a getter which returns whether the pipeline is invertible or not.

  • TimeSeries.to_json() and TimeSeries.from_json() methods to convert TimeSeries to/from a JSON string.

  • New base class GlobalForecastingModel for all models supporting training on multiple time series, as well as covariates. All PyTorch models are now GlobalForecastingModels.

  • As a consequence of the above, the fit() function of PyTorch models (all neural networks) can optionally be called with a sequence of time series (instead of a single time series).

  • Similarly, the predict() function of these models also accepts a specification of which series should be forecasted

  • A new TrainingDataset base class.

  • Some implementations of TrainingDataset containing some slicing logic for the training of neural networks on several time series.

  • A new TimeSeriesInferenceDataset base class.

  • An implementation SimpleInferenceDataset of TimeSeriesInferenceDataset.

  • All PyTorch models have a new fit_from_dataset() method which allows to directly fit the model from a specified TrainingDataset instance (instead of using a default instance when going via the fit() method).

  • A new explanatory notebooks for global models:


  • 🔴 removed the arguments training_series and target_series in ForecastingModels. Please consult the API documentation of forecasting models to see the new signatures.

  • 🔴 removed UnivariateForecastingModel and MultivariateForecastingModel base classes. This distinction does not exist anymore. Instead, now some models are “global” (can be trained on multiple series) or “local” (they cannot). All implementations of GlobalForecastingModels support multivariate time series out of the box, except N-BEATS.

  • Improved the documentation and README.

  • Re-ordered the example notebooks to improve the flow of examples.


  • Many small bug fixes.

  • Unit test speedup by about 15x.

0.5.0 (2020-11-09)

Full Changelog

For users of the library:


  • Ensemble models, a new kind of ForecastingModel which allows to ensemble multiple models to make predictions:

    • EnsembleModel is the abstract base class for ensemble models. Classes deriving from EnsembleModel must implement the ensemble() method, which takes in a List[TimeSeries] of predictions from the constituent models, and returns the ensembled prediction (a single TimeSeries object)

    • RegressionEnsembleModel, a concrete implementation of EnsembleModelwhich allows to specify any regression model (providing fit() and predict() methods) to use to ensemble the constituent models’ predictions.

  • A new method to TorchForecastingModel : untrained_model() returns the model as it was initially created, allowing to retrain the exact same model from scratch. Works both when specifying a random_state or not.

  • New ForecastingModel.backtest() and RegressionModel.backtest() functions which by default compute a single error score from the historical forecasts the model would have produced.

    • A new reduction parameter allows to specify whether to compute the mean/median/… of errors or (when reduction is set to None) to return a list of historical errors.

    • The previous backtest() functionality still exists but has been renamed historical_forecasts()

  • Added a new last_points_only parameter to historical_forecasts(), backtest() and gridsearch()


  • 🔴 Renamed backtest() into historical_forecasts()

  • fill_missing_values() and MissingValuesFiller used to remove the variable names when used with fill='auto' – not anymore.

  • Modified the default plotting style to increase contrast and make plots lighter.


  • Small mistake in the NaiveDrift model implementation which caused the first predicted value to repeat the last training value.

For developers of the library:


  • @random_method decorator now always assigns a _random_instance field to decorated methods (seeded with a random seed). This doesn’t change the observed behavior, but allows to deterministically “reset” TorchForecastingModel by saving _random_instance along with the other parameters of the model upon creation.

0.4.0 (2020-10-28)

Full Changelog

For users of the library:


  • Data (pre) processing abilities using DataTransformer, Pipeline :

    • DataTransformer provide a unified interface to apply transformations on TimeSeries, using their transform() method

    • Pipeline :

      • allow chaining of DataTransformers

      • provide fit(), transform(), fit_transform() and inverse_transform() methods.

    • Implementing your own data transformers:

      • Data transformers which need to be fitted first should derive from the FittableDataTransformer base class and implement a fit() method. Fittable transformers also provide a fit_transform() method, which fits the transformer and then transforms the data with a single call.

      • Data transformers which perform an invertible transformation should derive from the InvertibleDataTransformer base class and implement a inverse_transform() method.

      • Data transformers which are neither fittable nor invertible should derive from the BaseDataTransformer base class

      • All data transformers must implement a transform() method.

  • Concrete DataTransformer implementations:

    • MissingValuesFiller wraps around fill_missing_value() and allows to fill missing values using either a constant value or the pd.interpolate() method.

    • Mapper and InvertibleMapper allow to easily perform the equivalent of a map() function on a TimeSeries, and can be made part of a Pipeline

    • BoxCox allows to apply a BoxCox transformation to the data

  • Extended map() on TimeSeries to accept functions which use both a value and its timestamp to compute a new value e.g.f(timestamp, datapoint) = new_datapoint

  • Two new forecasting models:


  • 🔴 Removed cols parameter from map(). Using indexing on TimeSeries is preferred.

    # Assuming a multivariate TimeSeries named series with 3 columns or variables.
    # To apply fn to columns with names '0' and '2':
    #old syntax, cols=['0', '2']) # returned a time series with 3 columns
    #new syntax
    series[['0', '2']].map(fn) # returns a time series with only 2 columns
  • 🔴 Renamed ScalerWrapper into Scaler

  • 🔴 Renamed the preprocessing module into dataprocessing

  • 🔴 Unified auto_fillna() and fillna() into a single fill_missing_value() function

    #old syntax
    fillna(series, fill=0)
    #new syntax
    fill_missing_values(series, fill=0)
    #old syntax
    auto_fillna(series, **interpolate_kwargs)
    #new syntax
    fill_missing_values(series, fill='auto', **interpolate_kwargs)
    fill_missing_values(series, **interpolate_kwargs) # fill='auto' by default

For developers of the library


  • GitHub release workflow is now triggered manually from the GitHub “Actions” tab in the repository, providing a #major, #minor, or #patch argument. #211

  • (A limited number of) notebook examples are now run as part of the GitHub PR workflow.

0.3.0 (2020-10-05)

Full Changelog

For users of the library:


  • Better indexing on TimeSeries (support for column/component indexing) #150

  • New FourTheta forecasting model #123, #156

  • map() method for TimeSeries #121, #166

  • Further improved the backtesting functions #111:

    • Added support for multivariate TimeSeries and models

    • Added retrain and stride parameters

  • Custom style for matplotlib plots #191

  • sMAPE metric #129

  • Option to specify a random_state at model creation using the @random_method decorator on models using neural networks to allow reproducibility of results #118


  • 🔴 Refactored backtesting #184

    • Moved backtesting functionalities inside ForecastingModel and RegressionModel

      # old syntax:
      backtest_forecasting(forecasting_model, *args, **kwargs)
      # new syntax:
      forecasting_model.backtest(*args, **kwargs)
      # old syntax:
      backtest_regression(regression_model, *args, **kwargs)
      # new syntax:
      regression_model.backtest(*args, **kwargs)
    • Consequently removed the backtesting module

  • 🔴 ForecastingModel fit() method syntax using TimeSeries indexing instead of additional parameters #161

    # old syntax:, target_indices=[0, 1])
    # new syntax:, multivariate_series[["0", "1"]])
    # old syntax:, component_index=2)
    # new syntax:["2"])


  • Solved issue of TorchForecastingModel.predict(n) throwing an error at n=1. #108

  • Fixed MASE metrics #129

  • [BUG] ForecastingModel.backtest: Can bypass sanity checks #188

  • ForecastingModel.backtest() fails if forecast_horizon isn’t provided #186

For developers of the library


  • Gradle to build docs, docker image, run tests, … #112, #127, #159

  • M4 competition benchmark and notebook to the examples #138

  • Check of test coverage #141


  • Dependencies’ versions are now fixed #173

  • Workflow: tests trigger on Pull Request #165


  • Passed the freq parameter to the TimeSeries constructor in all TimeSeries generating functions #157

Older releases

Full Changelog