imbalanced-data

Accuracy is a bad comfort metric when the positive class is rare. I care more about precision, recall, PR AUC, calibration, and how thresholding changes operational workload. The right metric depends on the cost of false negatives versus false positiv