Rolling average: when is it possible to consider it?

You can! But the implications can be huge.

If you choose a big number, you automatically shorten the dataset from which you can train your time-series or forecast model. Also, if things start to "change" in recent time, your model might take too long to react if it weights those rolling averages heavily.

If you go too short, it might overreact to recent changes, or you're just missing that information from which your model might be able to learn on. I usually make that decision based on how I want to use the forecast on unseen data. This is not an easy decision for which there is a solid rule-of-thumb.

Welcome to the real-world complexities of a time-series problem!