NeuralForecast is an open-source deep learning library from Nixtla for time series forecasting, published on GitHub where it has earned about 4,291 stars and was still actively updated in 2026. It packages state-of-the-art neural forecasting architectures โ the same family of models behind well-known research results โ behind a clean, user-friendly API, so data scientists can train, evaluate, and deploy accurate forecasters without wrestling with low-level PyTorch code.
The library includes more than thirty models, from classics such as DeepAR and ESRNN to Nixtla’s own NBEATS, NBEATSx, and NHITS, plus Transformers like the Temporal Fusion Transformer and PatchTST. It supports probabilistic forecasting with prediction intervals, exogenous variables, and hierarchical forecasting through its HINT framework, which keeps forecasts coherent across aggregation levels. A familiar fit-and-predict interface, cross-validation utilities, and GPU acceleration make it suitable for both experimentation and production pipelines with thousands of series. The project is part of Nixtla’s wider open-source ecosystem, with extensive documentation, tutorials, and example notebooks covering common problems from retail demand planning to financial series.
NeuralForecast is free and open source under the Apache-2.0 license; you can install it from PyPI and use it in commercial or research projects at no cost. Nixtla separately offers hosted forecasting services, but the library itself carries no fees.
NeuralForecast is for data scientists and ML engineers who forecast demand, sales, energy, traffic, or any other time series and want modern neural architectures with research-grade implementations rather than hand-rolled model code.