Darts: User-Friendly Modern Machine Learning for Time Series

Julien HerzenFrancesco LässigSamuele Giuliano PiazzettaThomas NeuerLéo TaftiGuillaume RailleTomas Van PottelberghMarek PasiekaAndrzej SkrodzkiNicolas Huguenin

article2022JMLR328 citations

Presents Darts, an open-source Python library that unifies classical forecasting methods and modern deep learning architectures under a scikit-learn-like interface supporting multivariate series, multi-series training, and probabilistic forecasting.

Listen

Time series forecasting is critical across multiple industries, including energy, logistics, manufacturing, retail, and healthcare. Although traditional statistical models remain widely used, modern machine learning and deep learning approaches offer greater capacity to handle complex, high-dimensional datasets and multiple time series simultaneously. However, existing software ecosystems often lack unified interfaces, forcing organizations to navigate fragmented tools with steep technical barriers.

The article introduces and evaluates Darts, an open-source Python library designed to democratize modern time series forecasting by bridging classical statistical techniques and advanced deep learning methods under a single, consistent programming interface.

The authors developed a unified software framework centered on an immutable, three-dimensional time series data container. The framework integrates both classical statistical models (such as ARIMA and Exponential Smoothing) and deep learning architectures (including recurrent networks, temporal convolutional networks, and temporal fusion transformers). The evaluation demonstrates the library's ability to train a single model across multiple separate time series, manage external past and future covariates, and execute probabilistic forecasting using Monte Carlo simulations and 17 built-in probability distributions.

Key findings show that Darts standardizes the training and prediction workflow across fundamentally different model families using a single high-level interface. The library natively supports training deep neural networks on collections of multiple time series, significantly reducing data management complexity. Additionally, it provides built-in mechanisms for handling external explanatory data, automated model backtesting, ensembling, and memory-efficient data loading for large-scale datasets.

For organizations, these capabilities mean lower operational costs, shorter development cycles, and reduced implementation risk when deploying advanced forecasting pipelines. Data science teams can seamlessly benchmark classical methods against state-of-the-art neural networks within the same pipeline, accelerating the transition from prototyping to production.

Decision-makers should consider adopting the library as a standardized foundation for enterprise forecasting projects. Technical teams should run pilot evaluations on their own operational datasets to assess performance gains. At the same time, practitioners should note current boundary conditions, as support for static covariates and pre-trained foundation models remains in ongoing development.

Unit8 SA libraries python deep learning Darts time series forecasting probabilistic forecasting

arXiv: 2110.03224 unit8co/darts

Paper: API design for machine learning software: experiences from the scikit-learn project, Lars Buitinck et al. (2013). It introduces the standardized scikit-learn API design principles that Darts adopts to unify diverse time series models under a consistent interface.
Paper: DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks, David Salinas et al. (2020). It establishes the DeepAR architecture, which is one of the core probabilistic deep-learning forecasting models directly supported in Darts.
Paper: N-BEATS: Neural basis expansion analysis for interpretable time series forecasting, Boris N. Oreshkin et al. (2020). It introduces the N-BEATS neural architecture for interpretable and deep time series forecasting, a foundational neural model provided by Darts.
Paper: Temporal Fusion Transformers for interpretable multi-horizon time series forecasting, Bryan Lim et al. (2021). It details the Temporal Fusion Transformer architecture for multi-horizon forecasting with static and dynamic covariates, forming a core neural baseline in the library.
Paper: Time-series forecasting with deep learning: a survey, Bryan Lim et al. (2020). It provides a comprehensive survey of modern deep learning architectures and probabilistic methods in time series forecasting that contextualizes Darts' model offerings.
Paper: Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks, Guokun Lai et al. (2017). It develops the LSTNet model for multivariate forecasting with recurrent and convolutional components, which is implemented within Darts' suite of neural models.
Paper: Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting, Haixu Wu et al. (2021). It introduces the Autoformer decomposition transformer architecture, representing the class of modern transformer-based models supported in Darts.
Paper: A state space framework for automatic forecasting using exponential smoothing methods, Rob J. Hyndman et al. (2002). It formulates the automatic exponential smoothing state-space framework that underpins the classical statistical forecasting baselines in Darts.

Paper: N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting, Cristian Challu et al. (2023). It introduces N-HiTS as a hierarchical interpolation advance over N-BEATS for long-horizon forecasting, which subsequently expanded modern Python forecasting toolkits.
Paper: A Time Series is Worth 64 Words: Long-term Forecasting with Transformers, Yuqi Nie et al. (2023). It introduces PatchTST, advancing transformer-based long-term time series forecasting beyond earlier architectures supported in initial deep learning toolkits.
Paper: TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis, Haixu Wu et al. (2023). It proposes TimesNet, generalizing 1D temporal forecasting models into a 2D variation framework for multi-task time series analysis.
Paper: Unified Training of Universal Time Series Forecasting Transformers, Gerald Woo et al. (2024). It introduces MOIRAI, shifting time series modeling toward unified, zero-shot universal forecasting transformers across diverse datasets.
Paper: Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu 0007 et al. (2025). It presents Sundial, a foundation model family using continuous generative flow-matching to extend probabilistic time series forecasting beyond conventional deep models.
Paper: Timer: Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu et al. (2024). It develops Timer, extending generative pre-trained transformer modeling to general-purpose time series analysis at scale.
Paper: Transformers in Time Series: A Survey, Qingsong Wen et al. (2022). It synthesizes subsequent progress, structural modifications, and emerging challenges in adapting transformer architectures for time series modeling.

Cover for Darts: User-Friendly Modern Machine Learning for Time Series

Abstract

We present Darts¹, a Python machine learning library for time series, with a focus on forecasting. Darts offers a variety of models, from classics such as ARIMA to state-of-the-art deep neural networks. The emphasis of the library is on offering modern machine learning functionalities, such as supporting multidimensional series, fitting models on multiple series, training on large datasets, incorporating external data, ensembling models, and providing a rich support for probabilistic forecasting. At the same time, great care goes into the API design to make it user-friendly and easy to use. For instance, all models can be used using fit()/predict(), similar to scikit-learn (Pedregosa et al., 2011).

Topics

libraries python deep learning

Keywords

Darts time series forecasting probabilistic forecasting recurrent neural networks future covariates past covariates joint distributions

1. Introduction
2. Design Principles and Main Features of Darts
2.1 Time Series Representation
2.2 Unified High-Level Forecasting API
2.3 Training Models on Collections of Time Series
2.4 Support for Past and Future Covariates
2.5 Probabilistic Forecasting
2.6 Other Features
3. Usage Example
4. Conclusions
References

Knowls

Knowl 1 — TimeSeries Data Container Representation
model/method
In the Darts library, the core abstraction representing time series data is the immutable TimeSeries container class. It encapsulates a 3-dimensional xarray.DataArray structured with dimensions (time, component, sample):
- time: Indexed by a monotonically sorted time index, supporting either continuous timestamps (pandas.DatetimeIndex) or integer indices (pandas.RangeIndex).
- component: Represents the dimensions/channels of multivariate series.
- sample: Represents Monte Carlo stochastic realizations for probabilistic forecasting.
To ensure data integrity, TimeSeries guarantees well-formed shapes, valid data types, and sorted temporal order upon construction. To optimize memory efficiency and prevent costly deep copies during model training and batching, TimeSeries utilizes NumPy array views while maintaining an immutable external interface. The container provides built-in utilities for conversions (to/from Pandas DataFrames and NumPy arrays), indexing, temporal slicing, differencing, interpolation, applying mapping functions, timestamp feature embedding, and computing marginal empirical quantiles.
Knowl 2 — Unified High-Level Forecasting Interface
model/method
Darts establishes a standardized, high-level API inspired by scikit-learn that unifies both classical statistical forecasting algorithms and modern deep neural networks under a single interface.

All models implement standard methods:
- fit(series, ...): Trains the model. The input series can be a single TimeSeries object or a Sequence[TimeSeries] representing a collection of multiple distinct series.
- predict(n, series=None, ...): Generates an out-of-sample forecast for $n$ future time steps after the end of the specified target series.
This abstraction covers diverse underlying model mechanics—including autoregressive, recurrent (e.g., DeepAR-like RNNs), sequence-to-sequence, convolutional (e.g., Temporal Convolutional Networks - TCN), transformer-based (e.g., Temporal Fusion Transformers - TFT), basis expansion (e.g., N-BEATS), tabular regression wrappers (e.g., scikit-learn regressors), and classical methods (e.g., ARIMA, VARIMA, AutoARIMA, Exponential Smoothing, Theta, Prophet, FFT)—allowing seamless model comparison, ensembling, and automated backtesting without altering client code.
Knowl 3 — Global Model Training on Multiple Time Series Collections
model/method
Unlike classical single-series forecasting methods that require fitting an individual model per time series, Darts supports global model training where a single model is fitted across a dataset comprising many distinct time series.

Key mechanisms include:
1. Data Slicing: The darts.utils.data module provides sample generation logic that extracts input and output windows (sub-series chunks) along with associated covariates from a collection of series. Each model architecture provides a default slicing policy that can be customized.
2. PyTorch Integration: Neural network models are implemented in PyTorch, enabling stochastic gradient descent (SGD) optimization and GPU hardware acceleration.
3. Lazy Data Loading: For datasets too large to fit in RAM, Darts supports lazy sample extraction through custom PyTorch Sequence and Dataset implementations, streaming slices into memory on-the-fly during mini-batch construction.
Knowl 4 — Past and Future Covariate Differentiation and Alignment
model/method
Darts explicitly separates external auxiliary series (covariates) into two distinct semantic categories in model interfaces:
- Past Covariates (past_covariates): Auxiliary time series whose values are strictly known only up to the time of prediction (e.g., historical sensor readings, measured demand). These cannot be assumed known in the forecast window at inference time.
- Future Covariates (future_covariates): Auxiliary time series whose future values are known in advance across the entire forecasting horizon $n$ (e.g., promotional calendars, holidays, weather forecasts).
Models accept past_covariates and/or future_covariates parameters in their fit() and predict() methods. The library's internal slicing engine automatically aligns the temporal indices of the target series and covariate series based on their respective time axes, eliminating the requirement that input series share identical start or end dates.
Knowl 5 — Monte Carlo Sample Representation for Probabilistic Forecasting
model/method
Darts implements probabilistic forecasting without requiring a restrictive parametric form for the joint distribution over time and multivariate components. Instead, distributions are represented empirically via Monte Carlo samples stored along the sample dimension of the TimeSeries container.

Probabilistic deep neural network architectures optimize parameterized distribution heads via maximum likelihood estimation by minimizing differentiable negative log-likelihood (NLL) loss functions. Features include:
- Support for continuous, discrete, univariate, and multivariate likelihood distributions (17 distributions available out-of-the-box).
- Support for specifying time-independent prior distributions over distribution parameters to inject domain beliefs.
- Efficient vectorized Monte Carlo sampling during inference using batched operations, which produces stochastic TimeSeries objects containing $S$ realizations that directly support arbitrary quantile extraction (e.g., $10\%$ , $50\%$ , $90\%$ ).
Knowl 6 — Time Series Ensembling, Filtering, and Backtesting Utilities
model/method
Beyond individual forecasting models, Darts provides a comprehensive ecosystem of modular utilities:
- Model Ensembling: Allows combining predictions from diverse individual forecasting models (e.g., statistical, tabular, neural), including meta-models that learn ensembling weights using tabular regression models.
- Filtering Models: Implements probabilistic state-space and filtering techniques, such as Kalman filters and Gaussian Processes, for noise reduction and probabilistic state estimation.
- Automated Backtesting: A standardized backtest() method on forecasting models to evaluate historical roll-forward predictive accuracy across user-specified metrics.
- Data Processing and Metrics: Pipelines and scikit-learn-compatible transformers (e.g., Scaler), dynamic time warping (DTW) similarity measures, evaluation metrics, and built-in benchmark datasets via darts.datasets.

Coverage note — None was omitted; all core components, architectural abstractions, API features, and capabilities described in the software paper are fully covered.

References

1.A. Alexandrov, K. Benidis, M. Bohlke-Schneider, V. Flunkert, J. Gasthaus, T. Januschowski, D. C. Maddix, S. Rangapuram, D. Salinas, J. Schulz, L. Stella, A. C. Turkmen, and Y. Wang. Gluonts: Probabilistic and neural time series modeling in python. Journal of Machine Learning Research, 21(116):1–6, 2020. URL http://jmlr.org/papers/v21/19-820.html.
2.V. Assimakopoulos and K. Nikolopoulos. The theta model: a decomposition approach to forecasting. International Journal of Forecasting, 16(4):521–530, 2000. ISSN 0169-2070. doi: https://doi.org/10.1016/S0169-2070(00)00066-2. URL https://www.sciencedirect.com/science/article/pii/S0169207000000662. The M3- Competition.
3.S. Bai, J. Z. Kolter, and V. Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. CoRR, abs/1803.01271, 2018. URL http://arxiv.org/abs/1803.01271.
4.Beitner et al. Pytorch forecasting. https://github.com/jdb78/pytorch-forecasting. [Online; accessed 27-September-2021].
5.D. J. Berndt and J. Clifford. Using dynamic time warping to find patterns in time series. In KDD workshop, volume 10, pages 359–370. Seattle, WA, USA:, 1994.
6.A. Bhatnagar, P. Kassianik, C. Liu, T. Lan, W. Yang, R. Cassius, D. Sahoo, D. Arpit, S. Subramanian, G. Woo, A. Saha, A. K. Jagota, G. Gopalakrishnan, M. Singh, K. C. Krithika, S. Maddineni, D. Cho, B. Zong, Y. Zhou, C. Xiong, S. Savarese, S. Hoi, and H. Wang. Merlion: A machine learning library for time series, 2021.
7.V. Flunkert, D. Salinas, and J. Gasthaus. Deepar: Probabilistic forecasting with autoregressive recurrent networks. CoRR, abs/1704.04110, 2017. URL http://arxiv.org/abs/1704.04110.
8.C. R. Harris, K. J. Millman, S. J. van der Walt, R. Gommers, P. Virtanen, D. Cournapeau, E. Wieser, J. Taylor, S. Berg, N. J. Smith, R. Kern, M. Picus, S. Hoyer, M. H. van Kerkwijk, M. Brett, A. Haldane, J. F. del R´ıo, M. Wiebe, P. Peterson, P. G´erard-Marchant, K. Sheppard, T. Reddy, W. Weckesser, H. Abbasi, C. Gohlke, and T. E. Oliphant. Array programming with NumPy. Nature, 585(7825):357–362, Sept. 2020. doi: 10.1038/s41586-020-2649-2. URL https://doi.org/10.1038/s41586-020-2649-2.
9.Hosseini et al. Linkedin greykite. https://github.com/linkedin/greykite, 2021. [Online; accessed 27-September-2021].
10.S. Hoyer and J. Hamman. xarray: N-D labeled arrays and datasets in Python. Journal of Open Research Software, 5(1), 2017. doi: 10.5334/jors.148. URL http://doi.org/10.5334/jors.148.
11.R. Hyndman and G. Athanasopoulos. Forecasting: Principles and Practice. OTexts, Australia, 2nd edition, 2018.
12.R. J. Hyndman and Y. Khandakar. Automatic time series forecasting: the forecast package for R. Journal of Statistical Software, 26(3):1–22, 2008. URL https://www.jstatsoft.org/article/view/v027i03.
13.Jiang et al. Facebook kats. https://github.com/facebookresearch/Kats, 2021. [Online; accessed 27-September-2021].
14.B. Lim, S. O. Arik, N. Loeff, and T. Pfister. Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4):1748–1764, 2021. ISSN 0169-2070. doi: https://doi.org/10.1016/j.ijforecast.2021.03.012. URL https://www.sciencedirect.com/science/article/pii/S0169207021000637.
15.M. L¨oning, A. J. Bagnall, S. Ganesh, V. Kazakov, J. Lines, and F. J. Kir´aly. sktime: A unified interface for machine learning with time series. CoRR, abs/1909.07872, 2019. URL http://arxiv.org/abs/1909.07872.
16.B. N. Oreshkin, D. Carpov, N. Chapados, and Y. Bengio. N-BEATS: neural basis expansion analysis for interpretable time series forecasting. CoRR, abs/1905.10437, 2019. URL http://arxiv.org/abs/1905.10437.
17.B. N. Oreshkin, D. Carpov, N. Chapados, and Y. Bengio. Meta-learning framework with applications to zero-shot time-series forecasting. Proceedings of the AAAI Conference on Artificial Intelligence, 35(10):9242–9250, May 2021. URL https://ojs.aaai.org/index.php/AAAI/article/view/17115.
18.A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024–8035. Curran Associates, Inc., 2019.
19.F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.
20.T. G. Smith et al. pmdarima: Arima estimators for Python, 2017. URL http://www.alkaline-ml.com/pmdarima. [Online; accessed 27-September-2021].
21.S. J. Taylor and B. Letham. Forecasting at scale. The American Statistician, 72(1):37–45, 2018.
22.Wes McKinney. Data Structures for Statistical Computing in Python. In St´efan van der Walt and Jarrod Millman, editors, Proceedings of the 9th Python in Science Conference, pages 56 – 61, 2010. doi: 10.25080/Majora-92bf1922-00a.

Citation

MLA

Herzen, J., et al. “Darts: User-Friendly Modern Machine Learning for Time Series”. Journal of Machine Learning Research, vol. 23, no. 124, 2022, pp. 1–6, https://www.jmlr.org/papers/v23/21-1177.html.

APA

Herzen, J., Lässig, F., Piazzetta, S. G., Neuer, T., Tafti, L., Raille, G., Pottelbergh, T. V., Pasieka, M., Skrodzki, A., Huguenin, N., Dumonal, M., Kościsz, J., Bader, D., Gusset, F., Benheddi, M., Williamson, C., Kosinski, M., Petrik, M., & Grosch, G. (2022). Darts: User-Friendly Modern Machine Learning for Time Series. Journal of Machine Learning Research, 23(124), 1–6. https://www.jmlr.org/papers/v23/21-1177.html

Chicago

Herzen, J., F. Lässig, S. G. Piazzetta, et al. 2022. “Darts: User-Friendly Modern Machine Learning for Time Series”. Journal of Machine Learning Research 23 (124): 1–6. https://www.jmlr.org/papers/v23/21-1177.html.

Harvard

Herzen, J. et al. (2022) “Darts: User-Friendly Modern Machine Learning for Time Series”, Journal of Machine Learning Research, 23(124), pp. 1–6. Available at: https://www.jmlr.org/papers/v23/21-1177.html.

Vancouver

1. Herzen J, Lässig F, Piazzetta SG, et al (2022) Darts: User-Friendly Modern Machine Learning for Time Series. Journal of Machine Learning Research 23:1–6

BibTeX

@article{JMLR:v23:21-1177,
  author  = {Julien Herzen and Francesco LÃ¤ssig and Samuele Giuliano Piazzetta and Thomas Neuer and LÃ©o Tafti and Guillaume Raille and Tomas Van Pottelbergh and Marek Pasieka and Andrzej Skrodzki and Nicolas Huguenin and Maxime Dumonal and Jan KoÅcisz and Dennis Bader and FrÃ©dÃ©rick Gusset and Mounir Benheddi and Camila Williamson and Michal Kosinski and Matej Petrik and GaÃ«l Grosch},
  title   = {Darts: User-Friendly Modern Machine Learning for Time Series},
  journal = {Journal of Machine Learning Research},
  year    = {2022},
  volume  = {23},
  number  = {124},
  pages   = {1--6},
  url     = {http://jmlr.org/papers/v23/21-1177.html}
}

Metadata:DOI registry

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

Darts: User-Friendly Modern Machine Learning for Time Series

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — TimeSeries Data Container Representation

Knowl 2 — Unified High-Level Forecasting Interface

Knowl 3 — Global Model Training on Multiple Time Series Collections

Knowl 4 — Past and Future Covariate Differentiation and Alignment

Knowl 5 — Monte Carlo Sample Representation for Probabilistic Forecasting

Knowl 6 — Time Series Ensembling, Filtering, and Backtesting Utilities

References

Citation

BibTeX

Source Code

Access the Paper