[Paper Review] Darts: User-Friendly Modern Machine Learning for Time Series

Julia Herzen, Francesco Lässig+17 more|arXiv (Cornell University)|Oct 7, 2021

Forecasting Techniques and Applications16 references57 citations

TL;DR

Darts provides a user-friendly Python library unifying classical and ML-based time series forecasting under a common fit/predict API, with support for multi-series training, covariates, probabilistic forecasting, and ensemble methods.

ABSTRACT

We present Darts, 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, meta-learning 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.

Motivation & Objective

• Democratize modern ML-based time series forecasting by unifying classical and ML approaches under a single, easy-to-use API.
• Enable training on collections of time series and handling of high-dimensional data.
• Support probabilistic forecasting and covariates (past and future) to improve forecast quality.
• Provide tools for backtesting, hyperparameter search, and ensemble modeling to aid practitioners.

Proposed method

• Introduce a TimeSeries container that is immutable and guarantees well-formed shapes and time indices.
• Provide a unified high-level API with fit(series) and predict(n) applicable to diverse models.
• Allow training on collections of time series and lazy loading for large datasets.
• Support past and future covariates to condition forecasts.
• Offer probabilistic forecasting by storing Monte Carlo samples within TimeSeries and supporting multiple distributions.
• Incorporate a range of models (classical ARIMA, Exponential Smoothing, Theta, Prophet, FFT-based, RNNs like DeepAR, N-BEATS, TCN, TFT, and regression wrappers).
• Include additional features such as data preprocessing pipelines, backtesting, hyperparameter search, dynamic time warping, ensembles, Kalman filters, Gaussian Processes, and a datasets module.

Experimental results

Research questions

• RQ1 Can a single unified API support both classical and ML-based time series models effectively?
• RQ2 How well can a model be trained on large collections of time series and still perform accurate forecasts for individual series?
• RQ3 What is the impact of incorporating covariates (past and future) on forecasting accuracy?
• RQ4 How can probabilistic forecasting be implemented in a flexible, non-parametric way for multi-dimensional time series?
• RQ5 What are the benefits and limitations of a user-friendly API in facilitating adoption of ML-based forecasting?

Key findings

• Darts enables training and forecasting with a consistent fit/predict interface across diverse models.
• It supports training on multiple time series and large datasets via lazy data loading and sequence slicing utilities.
• The library provides probabilistic forecasting by storing Monte Carlo samples and supports 17 distributions out-of-the-box.
• It differentiates past and future covariates and handles their alignment automatically during slicing.
• A broad set of models (including ARIMA, Exponential Smoothing, Theta, Prophet, FFT, RNNs, N-BEATS, TCN, TFT) can be used under a unified API.
• It includes additional tooling such as backtesting, hyperparameter search, pipelines, preprocessing, and ensemble modeling.