ourownstory/neural_prophet

What it solves

NeuralProphet provides an easy-to-learn framework for interpretable time series forecasting. It addresses the challenge of building forecasting models that are not only accurate but also understandable, allowing users to iteratively refine their models through a human-in-the-loop process.

How it works

Built on PyTorch, the framework combines traditional time-series algorithms with neural networks. It allows users to define, customize, and evaluate models using a few lines of code. The system supports various model components including autoregression (linear or NN), piecewise linear trends with automatic changepoint detection, Fourier terms for seasonality (yearly, daily, weekly, hourly), lagged and future regressors, and event handling for holidays and custom events.

Who it’s for

It is best suited for users working with higher-frequency (sub-daily) time series data over longer durations (at least two full periods or years).

Highlights

• Interpretable Components: Includes tools to plot forecast components and model coefficients for better understanding.
• Global Modeling: Supports local, global, or 'glocal' (global + regularized local) parameters across multiple time series.
• Quantile Regression: Ability to estimate specific quantiles for uncertainty estimation.
• Integrated Tooling: Features built-in plotting functions, time series cross-validation utilities, and model checkpointing.