I think that we could add note mentioning that we can achieve a balanced random forest classifier by setting max_features='auto' of the decision tree. I don't think that we should implement a new estimator since scikit-learn is going to do it.