When making final predictions for a stacking ensemble, all submodels need to be refitted on full dataset before calling modeltime_fit_resamples, otherwise the stacking ensemble predictions obtained with modeltime_refit will be problematic, especially if there is a trend in the time series.
When making final predictions for a stacking ensemble, all submodels need to be refitted on full dataset before calling modeltime_fit_resamples, otherwise the stacking ensemble predictions obtained with modeltime_refit will be problematic, especially if there is a trend in the time series.
I don't know if this commit is the best solution to the problem, but it seems to be a simple imediate solution. I have made this gist to illustrate the problem: https://gist.github.com/regisely/435d4a2d444c44bd83bb8669d163f4f9