Time series forecasting with NeuralProphet
In Chapter 7, Machine Learning-Based Approaches to Time Series Forecasting, we covered the Prophet algorithm created by Meta (formerly Facebook). In this recipe, we will look into an extension of that algorithm—NeuralProphet.
As a brief refresher, the authors of Prophet highlighted good performance, interpretability, and ease of use as the model’s key advantages. The authors of NeuralProphet also had this in mind for their approach. They retained all the advantages of Prophet while adding new components that lead to improved accuracy and scalability.
The critique of the original Prophet algorithm included its rigid parametric structure (based on a generalized linear model) and the fact that it was a sort of “curve-fitter” that was not adaptive enough to fit the local patterns.
Traditionally, time series models used lagged values of the time series to predict the future value. Prophet’...
