What has Changed from Prophet#
NeuralProphet has a number of added features with respect to original Prophet. They are as follows.
Gradient Descent for optimisation via using PyTorch as the backend.
Modelling Auto-Regression of time series using AR-Net
Modelling lagged regressors using a separate linear or Feed-Forward Neural Network.
Directly predict specific forecast horizons.
Train a single model on many related time-series (global modelling).
Due to the modularity of the code and the extensibility supported by PyTorch, any component trainable by gradient descent can be added as a module to NeuralProphet. Using PyTorch as the backend, makes the modelling process much faster compared to original Prophet which uses Stan as the backend.