TBH I'm taking the conclusions with a grain of salt because (1) They only looked at a single Kaggle dataset; (2) The dataset was not really very imbalanced, with a 20% churn rate; (3) they don't address the calibration consequences of using an oversampling technique like SMOTE.