Implement/demo a restricted VAR model

[brookslogan]

See description of use case / idea here. This particular case actually sounds like it might be AR plus an ARX model (or is it VAR plus VARX?), so this may be an instance of composing models (a generalization of iterative 1-ahead models; see connected issue here).

[dajmcdon]

My understanding from the slack thread is that this is essentially just arx_forecaster(). Though perhaps what they really want is to not have the training data become further out-of-date as the horizon increases resulting in an iterative approach? I would have thought that fable combined with cases as exogenous would have done this, though it's not immediately clear how it would forecast cases.