NeuralProphet is a time series forecasting framework built on PyTorch that combines neural networks with traditional forecasting algorithms.
The tool addresses the challenge of building interpretable forecasting models through a human-in-the-loop approach. It merges neural network capabilities with classical time series methods inspired by Facebook Prophet and AR-Net, allowing users to quickly define, customize, visualize, and evaluate forecasting models with minimal code. The framework emphasizes interpretability and iterative refinement, enabling practitioners to build an initial model, examine results, make adjustments, and repeat until satisfied with performance.
NeuralProphet is best suited for higher-frequency time series data spanning longer durations, such as sub-daily observations collected over at least two full years. The tool prioritizes ease of learning and customization over out-of-the-box accuracy, making it appropriate for practitioners who want to actively shape their models rather than rely on automatic optimization. The README acknowledges that users may need to adjust and iterate to achieve desired results, positioning it as a tool for exploratory and collaborative model development rather than a black-box solution.
The project remains in beta phase and actively solicits issue reports and contributions from users. Documentation is acknowledged as incomplete, with the README directing users to docstrings as the more reliable reference while improvements are underway. The maintainers maintain active community engagement through GitHub discussions and a Slack channel, and they provide example notebooks and datasets to help users get started. The project welcomes contributions and has compiled practical guidance for potential contributors.
