Error when try to use the direct multi-step approach with ARIMA models

[NunopRolo]

Hi,
I'm trying to forecast a time series using an ARIMA model with a direct multi-step approach, but when running the code found in the documentation, I always get the same error, which is: TypeError: n_periods must be an int

Here I put some example code that I'm using:

forecaster = ForecasterAutoregDirect(
                 regressor     = ARIMA(order=(1,0,0)),
                 steps         = 24,
                 lags          = 24
             )

forecaster.fit(y=data_train)

Does anyone know what the problem is?

Thanks!

[JoaquinAmatRodrigo]

Hi @NunopRolo,
Thanks for using skforecast.
The current implementation of ForecasterAutoreg and ForecasterAutoregDirect does not allow to use ARIMA models as regressors. The reason is that the underlying matrices for fitting the model are not compatible. For ARIMA-SARIMAX models, we suggest using ForecasterSarimax.
Hope this helps.

[NunopRolo]

Hi @JoaquinAmatRodrigo

Thanks for the quick response.
Ok, but with ForecasterSarimax it is not possible to perform the direct multi-step approach.
Do you know of any way, or another library, to perform this direct multi-step approach, in ARIMA models?

Thank you!

[JoaquinAmatRodrigo]

Unfortunately, there is currently no way to do this in skforecast, and I am not aware of any other library that does it.
Still, I am not sure if it makes sense to combine a direct strategy with an arima model. These types of models are forced to use a recursive strategy because they need the previous errors in their MA component. Therefore, for an arima model to predict step +3, it needs to predict step +1 and step +2 before. They cannot point directly to step +3.

[NunopRolo]

Ok @JoaquinAmatRodrigo, thank you.
I closed the issue then.