Hi, Geo-,
generalities are risky because there are usually data that will prove a principle wrong. That said, people's buying patterns change with time. Therefore, the last six months of data would be more representative of buying patterns in the near future than the last year's data. So, if there are enough data in the last six months, use those. Perhaps you can explore how long buying patterns look stable.
With regard to (3) in which someone may be positive last month but negative this year, the negative examples should be every customer who is not positive. (I assume it is the customer who is either positive or negative, not an individual transaction.) Retraining the model every month or so is prudent models because models get stale (because people change their buying patterns). When retraining, a customer who was positive when training the earlier model could be negative when training the new model.
I might not understand your question correctly, and I certainly do not know the data or concerns as well as you do, so I might have missed the mark here.
Good luck, Padraic
... View more