// the find
ourownstory/neural_prophet
NeuralProphet: A simple forecasting package
NeuralProphet is a PyTorch-based time series forecasting library that extends Facebook Prophet with neural network components: AR-Net for autoregression, optional NN layers for lagged/future regressors, and global/glocal modeling across multiple series. It targets practitioners who want Prophet's interpretability but need to handle sub-daily data, multiple series, or more expressive seasonality patterns.
The 'glocal' modeling approach — global parameters with regularized local adjustments — is genuinely useful for heterogeneous panel data where full sharing fails but full independence wastes data. Component-wise plotting (trend, seasonality, AR coefficients) is first-class, not an afterthought, which matters when you need to explain a forecast to a stakeholder. Quantile regression for uncertainty is built in and doesn't require wrapping the model or post-hoc calibration. The migration path from Prophet is documented and the API surface is deliberately close, so porting an existing Prophet workflow takes hours not days.
The README itself admits it may not be the most accurate model out-of-the-box, and that's an understatement — on standard benchmarks NeuralProphet frequently loses to simpler statistical baselines like ARIMA or ETS, especially on shorter series. The 'Coming soon' section has items that have been listed as coming soon for years (logistic growth for trend, static metadata regression), which is a signal about the maintenance cadence. Documentation is acknowledged to be out of date in the README itself, meaning you'll hit undocumented API changes and need to read source. Last meaningful push was January 2025; the project has the feel of research software that peaked at the arxiv paper and is now in slow decline.