Just thought of another question w.r.t hierarchical forecast network. If I have exogenous variables that are specific to the lowest level, how will that work with HINT? E.g. I have a hierarchy of supermarket sales by region made up of supermarket sales by individual stores. Individual stores have categorical exogenous variables that describe location (rural, urban). I'm not sure how that will aggregate at the higher level, does it allow for exogenous variables to be used only for the lowest level in the hierarchy?

Hey @Rachel Yee, Thanks for checking out HINT, in the DPMN paper (main HINT inspiration) what I did was to carefully craft aggregate level exogenous. For example I used static dummies from the location hierarchy to capture geographic effects. It is possible to give that information to HINT through the `static_exog_list`parameter and the

static_df

inputs. In your specific example, you can transform rural/urban dummy into an average of urban population as an aggregate measure. Hope these ideas help.

Complementing Kin's answer, note that you still need to define values for all exogenous variables for all time series. The base model of the

HINT

class operates as any other

nf

model. If there is no obvious aggregation rule for the exogenous variables, it might also work to simply set missing/not defined variables to 0 (or other value which does not collide with true values of the lower level variables), which is equivalent to "masking out" the missing data.