NeuralForecast is a Python library that provides scalable neural network-based time series forecasting algorithms.
The library addresses the challenge of applying deep learning to forecasting tasks by offering a collection of neural architectures designed specifically for time series prediction. It implements several established models including N-BEATS, N-HiTS, and other neural forecasting approaches, alongside classical baselines. The tool is built on PyTorch and designed to handle both univariate and multivariate forecasting problems, with support for hierarchical forecasting scenarios where predictions must respect structural constraints across grouped time series.
Developers should choose NeuralForecast when building production forecasting systems that need to scale across large numbers of time series or when exploring deep learning approaches to forecasting problems. The library suits projects requiring probabilistic forecasts, where uncertainty quantification matters alongside point predictions. It is particularly valuable for teams already working in the Python ecosystem who want neural methods without building custom training pipelines from scratch. The tool emphasizes user-friendliness, aiming to make neural forecasting accessible without requiring deep expertise in deep learning frameworks.
The project maintains active development with regular updates to its model implementations and API. The codebase shows consistent refinement of existing functionality rather than rapid feature expansion. Documentation and examples receive ongoing attention, supporting the stated goal of accessibility. The project integrates with the broader Nixtla ecosystem, positioning itself as part of a coordinated suite of forecasting tools rather than as an isolated library.
