fabletools: Using forecast function on stretching window using external regressors

Question

I am using fable to train and test multiple timeseries.

I have created a stretching window using tsibble_stretch to prepare the time series cross validation. Then I use ARIMA to find the best model for each fold and forecast to test each timeseries.

I use the following two lines of code for this and it works great:

ARIMA_model <- stretching_window %>%
  model(ARIMA(pend_app))

ARIMA_forecast <- ARIMA_model %>% 
  forecast(h = "2 months")

However, when I want to do this using an external regressor, I get a lot of errors using the forecast function. I train my model like this:

 ARIMA_model_ext <- stretching_window %>% 
   model(arima_regr = ARIMA(pend_app ~ new_app + rec_rate))

when I simply try to forecast two months into the future it says it is missing external regressors. Although these are in the stretching_window dataset. When I try to use the new_data function:

 test <- new_data(stretching_window, 2) %>% 
  group_by_key() %>% 
  mutate( pend_app = last(stretching_window$pend_app), 
          new_app = last(stretching_window$new_app),
          rec_rate = last(stretching_window$rec_rate))

ARIMA_model_ext %>% 
  fabletools::forecast( new_data = test)

I get the following error:

Forecasts from an ARIMA model must start one step beyond the end of the trained data.

Ideally, I would not want to use the new_data() function but simply use my stretching_window dataset. Because from my understanding it should already split the data in test and train folds with all the required data ( incl. external regressors) to forecast.