Forecasting bug when using lagged predictor and multiple training sets

Hi all,

I am trying to do cross-validation on a model with lagged predictors. I had some unexpected forecasting results. Below, you'll find an example.

library(fpp3)

insurance_train <- insurance %>%
  stretch_tsibble(.init = 39, .step = 1, .id=".id")

# Commenting out this rule leads to a different forecast for May 2005
# insurance_train <- filter(insurance_train, .id == 2)

fit <- insurance_train %>%
  model(
    lag = ARIMA(Quotes ~ pdq(d = 0) + lag(TVadverts))
  )

insurance_test <- new_data(insurance_train, 1) %>%
  left_join(select(insurance, TVadverts), by='Month')

fit %>%
  forecast(insurance_test)

So far, I noticed that:

• The issue does not occur if the predictor is not lagged (Quotes ~ pdq(d = 0) + TVadverts).
• The issue does not occur if the lagged predictor is created manually before stretching the tsibble, thus by adding a variable TVadverts_lag = lag(TVadverts) and using Quotes ~ pdq(d = 0) + TVadverts_lag.

Thanks for creating the fable package and having a look at this issue 🙏.
Timo