NeuralProphet

NeuralProphet is a time-series forecasting package that extends Facebook’s Prophet with neural-network components. It keeps Prophet’s interpretable, decomposable design — trend, seasonality and events — and adds auto-regression and lagged regressors powered by PyTorch, so forecasts can react to the recent past instead of only following a fixed curve. The project has about 4,300 stars on GitHub and is actively maintained in 2026.

Overview

Built for practitioners, NeuralProphet fits a model in a few lines: give it a dataframe with dates and values and it decomposes the series into trend, yearly/weekly/daily seasonality, holiday and event effects, then layers neural components on top where they help. Auto-regressive terms let the model learn local dynamics, lagged external regressors connect the target to drivers like weather or promotions, and quantile regression produces uncertainty intervals. It handles multiple time series at once with global modelling, supports missing data gracefully, and plots every component so analysts can explain a forecast rather than just ship a number. Typical uses are demand and sales forecasting, energy load prediction, web-traffic planning and financial series. Development continues actively, with recent releases in 2026.

Under the hood it is a modular PyTorch model, so the neural components train with modern gradient methods while the statistical components keep their closed-form interpretability. It supports growth modes, changepoint control, sparsity-inducing priors for event effects and automatic handling of multiple seasonalities, and it can export forecasts with full component attribution for stakeholder reports. Because the API mirrors Prophet’s, teams already using Prophet can migrate by changing an import and gain the auto-regressive upgrades incrementally.

Key Features

• Prophet-style interpretable components plus neural auto-regression
• Trend, multi-seasonality, holidays and custom events built in
• Lagged regressors for external drivers such as weather or price
• Uncertainty intervals via quantile regression
• Global models that learn across many related time series
• Component plots that make every forecast explainable

💳 Pricing

NeuralProphet is free and open source under the MIT license; install it from PyPI and run it anywhere PyTorch runs.

Who Is It For

NeuralProphet is for data scientists, analysts and operations teams who forecast demand, traffic, energy or sales and want Prophet’s simplicity with more expressive power.