raise ValueError(

`"There are missing combinations of ids and times in

futr_df

.\n"` `f"You can run the

{expected_cmd}

method to get the expected combinations or "` `f"the

{missing_cmd}

method to get the missing combinations."`

)

Would this happen if you don't reindex, since all the combos are in the future dataframe, other than the earlier training dates?

Hey. You can run any of the suggested commands to know for sure

@José Morales can you explain this? Training Frame: unique_id ds 0 Bensalem (6046) 08 2022-02-25 1 Bensalem (6046) 08 2022-02-26 2 Bensalem (6046) 08 2022-02-27 3 Bensalem (6046) 08 2022-02-28 4 Bensalem (6046) 08 2022-03-01 ... ... ... 101004 Willow Grove (6427) 20 2024-03-28 101005 Willow Grove (6427) 20 2024-03-29 101006 Willow Grove (6427) 20 2024-03-30 101007 Willow Grove (6427) 20 2024-03-31 101008 Willow Grove (6427) 20 2024-04-01 Future Frame: unique_id ds 763 Bensalem (6046) 08 2024-04-02 764 Bensalem (6046) 08 2024-04-03 765 Bensalem (6046) 08 2024-04-04 766 Bensalem (6046) 08 2024-04-05 767 Bensalem (6046) 08 2024-04-06 ... ... ... 101018 Willow Grove (6427) 20 2024-04-11 101019 Willow Grove (6427) 20 2024-04-12 101020 Willow Grove (6427) 20 2024-04-13 101021 Willow Grove (6427) 20 2024-04-14 101022 Willow Grove (6427) 20 2024-04-15 Here is the result of get_missing_future(futr_df): unique_id ds 0 Bensalem (6046) 08 2024-03-31 1 Bensalem (6046) 08 2024-04-01 2 Bensalem (6046) 09 2024-03-31 3 Bensalem (6046) 09 2024-04-01 4 Bensalem (6046) 10 2024-03-31 .. ... ... 229 Royersford (6317) 18 2024-04-01 230 Royersford (6317) 19 2024-03-31 231 Royersford (6317) 19 2024-04-01 232 Royersford (6317) 20 2024-03-31 233 Royersford (6317) 20 2024-04-01 It's working for the univariate and regression models, but not neural networks.

Do you have those dates in your training set? Seems like you're providing the future values starting at April 2nd but for some series the last value was march 30th

Yes, the first dataframe is the tail of the training frame, which goes through 4-1. Then the future dataframe starts at 4-2. So why does get_missing_future_df return 3-31 and 4-1?

Those are computed per serie, so for example maybe you don't have march 31st in your training df for the first id. statsforecast currently only checks the shape, so it wouldn't catch this

But why would it work for univariate and regression models with the exact same dataframes?

What do you mean by univariate and regression? If you don't use exogenous features then you don't need to provide the futr_df argument

I do for mlforecast but understand your overarching point. I think what you said was possible and we’ll need to investigate that. Thanks for your time Jose. You see every angle, and sometimes it’s hard to pin down the exact problem. Truly appreciate that.