Hi Flávio! Thanks, I'm glad you enjoy this space.

The point of purging is to avoid the inherent correlation between the final part of the training data and the start of validation. In principle, this will result in a better estimation concerning the performance of the model in the long-term.

But, as far as I known, purging is a CV-only thing. After CV, I would re-train the model using the whole data set. Does this answer your question?

Regarding the second point. Yes, you should try to use the information from other products to forecast the sales of another. A form of leakage can occur can occur if you have the same entity (product) in both training and validation.

I think this can be situational. Here's a more comprehensive read on this problem: https://www.kaggle.com/code/jorijnsmit/found-the-holy-grail-grouptimeseriessplit.

Hope this helps!