In this question, it was asked for a way to repeat a 5-folds cross-validation 10 times in a
GridSearchCV. The accepted answer suggests a nested cross-validation using 10-fold in the outer loop and 5-fold in the inner loop.
I believe that both have similar applicability in terms of model evaluation, but it feels like those two things are different. In the original question (and a later response given using
RepeatedKFold), the grid search is able to find the model that performs best considering all 50 splits. In the nested cross-validation answer, it isn't even possible to retrieve a best performing model because we end up with several of those (one for each outer split).
In the end, what are the most notable differences between the two approaches? Why should I prefer one for the other, if at all?