Open breadwall opened 1 month ago
I've found that if I go back to neuralforecast version 1.6.1
Then re-add the changes to hint.py from this commit: https://github.com/Nixtla/neuralforecast/pull/964/files
I get properly reconciled forecasts using fit then predict with the HINT column produced.
What happened + What you expected to happen
I walked through the tutorial and it doesn't produce columns that are labeled as HINT when fitting then predicting. Therefore, when you get to the step to plot it errors out.
Additionally, since they're not HINT predictions but the model chosen (NHITS or TFT), they are not reconciled together. https://nixtlaverse.nixtla.io/neuralforecast/docs/tutorials/hierarchical_forecasting.html or https://github.com/Nixtla/neuralforecast/blob/main/nbs/docs/tutorials/09_hierarchical_forecasting.ipynb
However, when I follow the two test cases in the source code which uses nf.cross_validation() it does produce HINT columns. These HINT predictions are reconciled with sums that are the same across hierarchy levels and combinations. https://github.com/Nixtla/neuralforecast/blob/main/nbs/models.hint.ipynb
Versions / Dependencies
neuralforecast==1.7.4 python 3.10.7
Reproduction script
Using Code from https://github.com/Nixtla/neuralforecast/blob/main/nbs/models.hint.ipynb
Issue Severity
Medium: