Achieving an exceptionally high AUC ROC value of 99% on training data with little effort is often indicative of data leakage, where future information inadvertently influences the training process, leading to overfitting and diminished performance on new data. Applying nested cross-validation is a robust strategy to detect and address data leakage by ensuring that the model is evaluated on unseen data within each fold, thereby providing a more accurate assessment of its performance. Additionally, validating the model's performance on a completely unseen dataset is crucial to confirm that the high AUC ROC is not a result of data leakage. While simplifying the model or tweaking hyperparameters can sometimes alleviate overfitting, these measures may not be as effective if the underlying issue is data leakage. Similarly, removing highly correlated features might help, but it's essential to weigh the potential loss of informative features against the benefits of their removal.