neural_prophet

Built on PyTorch, it maintains Prophet's human-interpretable model structure where each component is separately visualizable and auditable, while adding auto-regression via an AR-Net component, lagged covariate effects, and configurable neural network layers for residual nonlinearity.

The library is designed for practitioners who value both accuracy and interpretability use cases in retail demand forecasting, energy consumption prediction, web traffic modeling, and financial planning where explaining a forecast to stakeholders matters as much as minimizing error metrics.

NeuralProphet decomposes forecasts into additive components that can be plotted separately: the trend component showing long-run direction, seasonal components showing weekly and yearly patterns, holiday effects quantified as additive offsets, and regressor contributions from external variables.

NeuralProphet is open-source under the MIT license and compatible with the broader PyTorch and scikit-learn ecosystems. It provides a Prophet-compatible API that makes migration from Prophet straightforward for teams that want improved accuracy without completely rewriting their forecasting pipelines.

The library includes hyperparameter tuning utilities, cross-validation with temporal train-test splits, and residual diagnostic plots, providing the infrastructure for rigorous forecasting model development and validation beyond the initial fit.