pytorch-forecasting
Deep learning time series forecasting library for PyTorch using PyTorch Lightning for scalable training.
Build, train, and interpret advanced neural network architectures on multi-horizon time series data.
PyTorch Forecasting is a Python package designed for high-level deep learning time series prediction. It helps data scientists quickly implement complex architectures like Temporal Fusion Transformers and N-BEATS using PyTorch Lightning for hardware acceleration. You should use it if you need multi-horizon probabilistic forecasting on tabular time series data.
What it is
The package abstracts data preparation using a specialized dataset class that handles missing values, variable transformations, and subsampling automatically. It provides built-in neural architectures optimized for real-world deployments alongside TensorBoard logging and dependency visualizations. Training scales across CPU and multi-GPU environments through its PyTorch Lightning integration.
Who it's for
It suits data scientists and machine learning engineers building production-grade predictive models for complex temporal data. It is not suitable for non-programmers or teams looking for code-free tools, as it requires proficiency in Python and PyTorch.
Key features
TimeSeriesDataSet abstraction class
Handles feature transformations, missing values, variable history lengths, and subsampling automatically.
Advanced neural architectures
Provides implementations of Temporal Fusion Transformers, N-BEATS, N-HiTS, PatchTST, and DeepAR.
Automated logging and visuals
Integrates with TensorBoard to output training metrics, actual versus predicted plots, and dependency charts.
Optuna hyperparameter tuning
Includes integration with Optuna to tune model parameters across complex forecasting architectures.
Up and running in 3 steps
Our honest verdict
What's great
- Includes ready-to-use implementations of Temporal Fusion Transformers, N-HiTS, and PatchTST.
- Uses PyTorch Lightning to scale training across multi-GPU setups without manual distributed code.
- TimeSeriesDataSet wrapper automates complex feature engineering and windowing tasks.
- Provides built-in interpretability features and multi-horizon prediction evaluation metrics.
Where it falls short
- High computational overhead and hardware costs for training large transformer models.
- Steep learning curve for practitioners who lack PyTorch or PyTorch Lightning experience.
- Requires extra manual installation steps for specific loss functions like MQF2.
- Windows installations require extra manual steps to select the correct PyTorch binaries.
Pricing & license
PyTorch Forecasting is free and open-source under the MIT License. The primary expense involved is the server and GPU infrastructure required to train these deep learning models.
Open remote AI-training roles
874 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
What deep learning architectures are included?
It includes Temporal Fusion Transformers, N-BEATS, N-HiTS, PatchTST, DeepAR, standard LSTM/GRU networks, and simple baseline models.
Does PyTorch Forecasting support hyperparameter tuning?
Yes, it integrates directly with Optuna for automated hyperparameter optimization.
Can models be trained on multi-GPU setups?
Yes, PyTorch Lightning provides out-of-the-box support for CPU, single GPU, and multi-GPU hardware acceleration.
