DeepAR: do the lags generated by get_lags_for_frequency() apply *just* to the target when there are also dynamic features?

[Serendipity31]

Hello! I've been trying to find the answer in the DeepAR code, but haven't been successful, so I'm hoping someone can help me.

In my situation, I will have multiple dynamic real value features (exogenous real value covariates that influence the target variable of interest). I'm trying to figure out how the internal, automatic lag functionality in DeepAR actually works in practice in this case.

I know that if I don't supply any lags (in lags_seq) that DeepAR will automatically determine the lags to use based on the frequency of the data (using 'get_lags_for_frequency').

In the case of monthly data, for example, get_lags_for_frequency returns the following lags by default: [1, 2, 3, 4, 5, 6, 7, 11, 12, 13, 23, 24, 25, 35, 36, 37]

To what are these lags applied? Are they only applied to the target series? Or also they also applied to all dynamic real value feature series in the dataset?

If the lags are only applied to the target series, and if there was a reason to think the past values of an exogenous covariate influences the target in a given month, then would it be appropriate to create the lags of the coviarates in a data preprocessing step and include them as additional dynamic real value features?