Y_df argument in reconcile

[ngupta23]

I was wondering what the use of the Y_df argument in reconcile is. In the following doc, it is not used

reconciled_forecasts = hrec.reconcile(Y_hat_df=forecasts_df, S=S, tags=tags)

In this doc, it is used, but if I do not specify it, the results do not change. So it is not clear to me what the use of this argument is and when it is needed.

Y_rec_df = hrec.reconcile(Y_hat_df=Y_hat_df, Y_df=Y_fitted_df, S=S_df, tags=tags)

Any further details would be helpful. Thanks!

[ngupta23]

So, it looks like Y_df is only needed for certain reconciliation methods. Is this understanding correct?
If so, should this contain the actual Y values or the in-sample predictions?

[kdgutier]

Hi @ngupta23,

As you mentioned some mean/probabilistic reconciliation methods require Y_df insample predictionsy_hat and target y.
Here is the reconciliation methods catalog: https://nixtla.github.io/hierarchicalforecast/methods.html

Here is a usage example of such methods, where we use StatsForecast's fitted values:
https://github.com/Nixtla/hierarchicalforecast/blob/main/nbs/examples/TourismLarge-Evaluation.ipynb

[ngupta23]

Going through this, but have some questions.

For Middle out, why are the in-sample predictions required for forecast_proportions (they are not needed for top-down)?

https://nixtla.github.io/hierarchicalforecast/methods.html#middleout.fit_predict

[ngupta23]

Maybe it is a typo - "Only" should be changed to "Optional" in the Middle-Out section

[kdgutier]

Thanks for the comment, we will update the documentation of the TopDown method, to mention it.
The documentation needs an overall revision. It is currently in its 0.0 version
:)

Please let us know if you find other things.
One thing we definitely need to improve is the introduction to the hierarchical forecast structures and the S_df inputs.

[ngupta23]

we will update the documentation of the TopDown method, to mention it

So just to confirm, is the typo in "Top Down" or "Middle Out"? I would think "Middle Out" since you should not need in-sample when using forecast proportions.