Darts — User-Friendly Time Series Forecasting and Anomaly Detection

Introduction

Darts is an open-source Python library developed by Unit8 that makes time series forecasting accessible to both beginners and experts. It wraps a wide range of models — from ARIMA and exponential smoothing to N-BEATS and Temporal Fusion Transformer — behind a consistent, easy-to-use API centered around its own TimeSeries data structure.

What Darts Does

Provides a unified fit()/predict() interface across 30+ forecasting models
Supports probabilistic forecasting with confidence intervals and quantile predictions
Handles multivariate, multi-step, and hierarchical time series natively
Includes built-in anomaly detection and scoring modules
Offers data preprocessing utilities for scaling, missing value handling, and feature engineering

Architecture Overview

Darts is built around the TimeSeries class, an immutable wrapper over xarray that stores time-indexed data with optional static covariates. Models inherit from base classes (LocalForecastingModel, GlobalForecastingModel) that define the training interface. Deep learning models use PyTorch Lightning under the hood, enabling GPU training, early stopping, and learning rate scheduling out of the box.

Self-Hosting & Configuration

Install via pip install darts or pip install "u8darts[all]" for all optional dependencies
Use darts[torch] to include PyTorch-based models only
Configure deep learning models with pl_trainer_kwargs for PyTorch Lightning options
Set num_loader_workers for parallel data loading during training
Export trained models with model.save() and reload with Model.load()

Key Features

Consistent API across statistical, machine learning, and deep learning models
First-class support for covariates (past and future external variables)
Backtesting and evaluation with historical_forecasts() and multiple metrics
Ensemble models and pipeline composition for stacking approaches
Rich plotting utilities built on matplotlib

Comparison with Similar Tools

sktime — Broader ML task coverage; Darts focuses on making forecasting simple
statsmodels — Lower-level statistical API; Darts wraps statsmodels with a friendlier interface
GluonTS — AWS probabilistic forecasting; Darts is framework-agnostic and easier to start
Prophet — Single model optimized for business series; Darts offers 30+ model choices
NeuralForecast — Pure neural focus; Darts mixes classical and deep models

FAQ

Q: Can Darts handle multiple time series at once? A: Yes. Global models in Darts can be trained on multiple series simultaneously and transfer learned patterns across them.

Q: Does Darts support GPU training? A: Yes. All deep learning models leverage PyTorch Lightning and can be trained on GPUs by passing appropriate trainer kwargs.

Q: What data formats does Darts accept? A: Darts accepts pandas DataFrames, NumPy arrays, or its native TimeSeries objects.

Q: Is Darts suitable for production use? A: Yes. Models can be serialized, and the library supports reproducible pipelines suitable for deployment.