Specificity

Specificity measures how well a model avoids false alarms among the cases that are truly negative. It is calculated as TN divided by all actual negatives (TN + FP), so it answers: "If the true class is negative, how often does the model correctly say negative?" This works because the denominator counts every actual negative case, while the numerator counts only the ones the model got right.

Use specificity when false positives matter, such as in medical screening, fraud detection, spam filtering, or any test where incorrectly flagging a negative case has a cost. It is especially useful when you want to know how reliably a model leaves normal or non-target cases alone.

• Using true positives instead of true negatives in the numerator
• Dividing by all predictions instead of only actual negatives
• Confusing specificity with sensitivity, which measures how well positives are detected