This message was deleted.

Which mlforecast version are you using? I believe this should issue a warning

I'm on 0.10, I'll see if upgrading helps

I think 0.11.1 changed this to a warning. Although I'm not entirely sure if the intervals will work, you may get a different error. The most reliable solution would be to drop the small series for the intervals computation

This error has indeed been changed to a warning. However, now I am getting an error on the predict step: model_rec.predict(horizon,new_df=test_set,X_df=X_df, level = [90]) ValueError: cannot reshape array of size 2168 into shape (5,271,4)

That's what I feared, since the warning drops the series from the CV we don't have a score for each serie in each window, so there are missing entries in that array. It's hard to keep track of which series were dropped to kind of pad them. My suggestion is to drop the series that have less than h * n_windows + 1 samples or try using less windows if possible

I didn't realize each series got its own score in training that was then used in prediction; does this mean that the prediction will not work if new_df contains unique_ids not present in the training set?

Yes, that won't work. We compute the conformity scores during fit (performing CV) and store them for each id and window, we then use these to create the intervals during predict. It wouldn't be reliable to use these scores during transfer learning because they're associated to each serie and horizon

I see... I misunderstood the science, I thought multiple trees were trained on each CV window and then in prediction the results for each tree were compared to get a distribution. I guess I would have to make a secondary system that uses conformity scores as the target variable if I wanted to continue down this path. Probably easier to just use quantile regression on a few levels and backsolve probability distributions from that. Thank you for the help!