How to use covariates at different level for hierarchicalforecast?

[blackhole-canada]

Hello everyone, is there a way to incorporate covariates but at different levels of hierarchy? For example I have the following hierarchy:

spec = [
['Zone'],
['Zone', 'Market'],
['Zone', 'Market', 'Country'],
['Zone', 'Market', 'Country', 'Category'],
]

I have different covariates available at different level, for example price of a category, holidays of a country, inflation of a zone etc. Is there a way to incorporate them at different levels of the hierarchy?

[theforecastnoob]

I've only got 2 levels, and only one type of exogenous- variable for the different levels, but here's how I do it.

1. Prepare the exogenous DataFrame
   It needs at least:

   • ds (date)
   • exog_name (your covariate)
   • One or more columns specifying the hierarchy level(s).

   # Example columns:
   df_exog.columns  # ['ds', 'exog_name', 'category']

Here, exog_name (e.g. “Price”) repeats for each category

Now lets focus on the target, Y_df

Y_df, S_df, tags = aggregate(...)
Here, Y_df has all information I need, from the 'unique_id' column, where my unique_id's are for example
(1)

unique_id
item_total
item_total
...
item_total/item_categoryA
item_total/item_categoryA
item_total/item_categoryA

I can obtain the categories with some code and create a new column
Y_df['category'] = Y_df['unique_id'].astype(str).str.rsplit('_', n=1).str[-1]
(2)

unique_id	category
item_total	total
item_total	total
...	...
item_total/item_categoryA	categoryA
item_total/item_categoryA	categoryA
item_total/item_categoryA	categoryA

From this, i simply merge, because now Y_df and df_exog has similar a column: category
Y_df = Y_df.merge(df_exog[['ds', 'exog_name', 'category']], on=['ds', 'category'], how='left')

All that's left is the exog forecast that should also have the same unique_id's as Y_df as I've written above (1).

I guess in your case, since you have different covariates, you could create a mapping between the unique_id from Y_df to your covariates and then proceed to merge.

Hope this helps you somewhat, though we have different data.