https://github.com/nixtla logo
n

Nikola Shahpazov

07/21/2023, 8:11 AM
Hi guys, I'm testing some reconciliation algorithms from the
hierarchicalforecast
package and I want to test using a
lgbm
mlforecast
model with some additional features. I couldn't find example of that, the one in the documentation is not using any additional features. Is this possible any way?
n

Nikola Shahpazov

07/21/2023, 1:30 PM
Thanks Kin, but I still can't see how I can use exogenous covariates in the reconciliation forecasting process. Is there some place where this is decsribed?
k

Kin Gtz. Olivares

07/21/2023, 2:08 PM
I don’t believe anyone has estimated reconciliation matrices using exogenous features yet. That is a very novel idea.
There might be some work where different reconciliation matrices are computed for different calendar events. But it is a guess. Usually one shares the matrix across the horizon.
If you have very different residual distributions based on seasonality, the idea makes a lot of sense.
n

Nikola Shahpazov

07/23/2023, 12:44 PM
Thanks Kin. Let me try to rephrase my question. I have a time series which I've aggregated with the aggregate function. It produces a time dataframe with a single unique_id which is <upper_level_id> when it's the upper level and it's <upper_level_id>/<lower_level_id. Now, I have some additional features like
is_holiday
,
is_special_day
, etc that I can use in the forecasting model. I'm curious if there's a quick way to integrate those in the model, i.e. do a left join or something.