Stijn de Boer

+1 This would be very useful

> @AuguB, the example I've included in our discussion (DelayedSaturatedMMM) is an Hierarchical model using coordinates. Just in my opinion defining a property and its dedicated setter isn't justifiable when...

> Also not that much related to the topic, but pymc-marketing is already integrated with this API, so any change done here would result in quite an effort there to...

>About the privacy issues. That is interesting, but not really unique to ModelBuilder? When you do pm.sample by default the observed data (and any Constant/MutableData) are stored in the arviz...

> I don't follow what hierarchical models have to do with coords. You can (and should) use coords for any kind of model You're right. The point is that the...

> I may be wrong again (mostly dealing with implementation, not the actual usage) but you shouldn't build model again if you want to make predictions. _data_setter method should be...

Okay, great. One related thing that has not been adressed, and which guided my thinking in this issue, is the following: I want to apply the same preprocessing on the...

> in that case please put the preprocessing logic in _data_setter, it should contain everything needed to successfully replace the data in your model for predictions. I understand your reasoning,...

> Somewhat related, some methods like `sample_posterior_predictive` (used in `predict`) will behave differently if the model has `MutableData` variables that are missing or have changed from the `idata` (it will...

> I would make it raise `NotImplementedError` not abstract (so far we haven't needed it, so I wouldn't require every subclass to implement it) Wouldn't this then raise the error...