Those are more analytic than BCI problems. More and more research in BCIs (and classification of EEG i general) is focused on getting models to work with raw, unprocessed data. In my opinion, way too little attention is paid to model evaluation and to final intended purpose. I think you should pay attention to when overfitting (and evaluation with n-fold crossvalidation) is beneficial (when creating a speller for a disabled person, for example meant to be used by that specific person and noone else) or detrimental (when a model is meant to be used by anyone in a plug-and-play fasion).