Neural Prophet is a time-series forecasting model inspired by Facebook Prophet and AR-Net (for autocorrelation modeling). It is built on PyTorch and combines Neural Networks and traditional time-series algorithms.
Neural Prophet includes all the components from the original Prophet model: trend, seasonality, recurring events, and regressors. Further, Neural Prophet now also provides support for auto-regression and lagged covariates. That's particularly relevant in the kinds of applications in which the near-term future depends on the current state of the system.
It also provides an automatic selection for the hyperparameters related to model training. In addition, the model is adaptable to different forecast horizons (greater than 1) and its components can be interpreted with the corresponding visualizations.
Bricks → Machine Learning → Neural Prophet
Datetime
Dataset column that specifies the datetime of the observation.
Target Variable
A column that contains the values to be predicted.
Frequency (’Auto’ by default)
Data step sizes (frequency of data recording).
Available time units (‘Custom’ frequency):
Regression / Auto-Regression
Future Regressors (optional)
Regressors as lagged covariates with order 1 or as known in advance.
Lags (0 by default)
Previous time series steps to include in auto-regression i.e. AR-order.
Lag-Regressors (optional)
Covariate or list of covariate time series added as additional lagged regressors to be used for fitting and predicting.
Forecasts (1 by default)
The number of steps ahead of prediction time step to forecast.
Events (optional, if optional events input is connected)
Datetime
A column from the optional events input that contains the dates for the user-specific events.
Events column
A column from the optional events input that specifies the names of the user-specific events.