Is there any structure similar to 'Transformation' that doesn't work on a stream?

[VictorJouault]

Hi, just looking for the most 'GluonTS-y' method to create Transformation of a dataset that does multiple pass on the data. Ideally, it could have a 'fit' and a 'transform' method like in sklearn.

My specific use-case is about conditioning a a time series forecast based on other time series in the dataset (sort of univariate output but multivariate intput).
It goes like this:

For each time series i in dataset:
    for each time series j in dataset:
        If time series i and time series j are in the same 'group' (all groups don't have the same size, those groups can be defined by the same static cat feature):
            Add time series j as dynamic real feature to time series i

We can see that there are 2 'for' loops inside. Any suggestion here?