Nixtla HierarchicalForecast: Extracting reconciled y_hat_insample after reconciliation (MintShrink)

[hfc25]

Dear Nixtla community,

I think this feature is not available yet but is there any chance to extract the reconciled y_hat_insample df after fitting the reconciliation model based on the MintShrink approach? The origin of this idea is that I'm aiming to combine several reconciliation stages for business-related reasons and would like to extract the reconciled y_hat_insample df from the first stage so I can later use it as an input in a later reconciliation stage.

Thank you in advance for any hints!

[elephaint]

Hi!

I think you can create that, but it doesn't really make sense? y_hat_insample is used to estimate the covariance for the reconciliation, and then you'd be using it for its own reconciliation, which is kinda... circular?

[hfc25]

Hi, maybe I can provide a bit more context on my idea. In my specific use case, I'm planning to first apply MintShrink reconciliation from an organisational perspective (starting with this to avoid major accuracy drops in the third step due to potential biases after top down reconciliation), then apply top down reconciliation for temporal purposes, and eventually make sure all time series are aligned on an organisational level again by applying MintShrink once again.

So I would like to extract the reconciled y_hat_insample from stage 1, then apply the historical proportions from the top down reconciliation to it, and use that resulting dataframe as an input for the final reconciliation.

Do you have any recommendations on how to approach this? Thank you in advance