Modelling Lagged Regressors

In the current state of NeuralProphet development, Lagged Regressor support is only available when the AR-Net is enabled. This is because they are both handled in a similar way internally using Feed-Forward Neural Networks and need to specify the n_lags value. For simplicity, at the moment we use the same n_lags value for both the AR-Net and the Lagged Regressors. Therefore, with Lagged Regressors, the NeuralProphet object is instantiated similar with AR-Net like below.

m = NeuralProphet(
    n_forecasts=3,
    n_lags=5,
    yearly_seasonality=False,
    weekly_seasonality=False,
    daily_seasonality=False,
)

When fitting the model, the dataframe provided to the fit function should have additional columns for your lagged regressors like below.

ds

y

A

0

2007-12-10 00:00:00

9.59076

9.59076

1

2007-12-11 00:00:00

8.51959

9.05518

2

2007-12-12 00:00:00

8.18368

8.76468

3

2007-12-13 00:00:00

8.07247

8.59162

4

2007-12-14 00:00:00

7.89357

8.45201

In this example, we have a Lagged Regressor named A. You also need to register these Lagged Regressors with the NeuralProphet object by calling the add_lagged_regressor function and giving the necessary configs.

m = m.add_lagged_regressor(names='A')

By setting the only_last_value argument of the add_lagged_regressor function, the user can specify either to use only the last known value of the regressor within the input window or else use the same number of lags as auto-regression. Now you can perform the model fitting and forecasting as usual. The plotted components should look like below.

plot-comp-1{: style=”height:500px”}

You can see the components corresponding to both auto-regression and the Lagged Regressor A. The coefficients plot looks like below.

plot-param-1{: style=”height:700px”}

It shows both the AR and Lagged Regressor relevance at the 5 lags corresponding to the input window.