What is specificity?
Specificity is the share of actual negative cases that a model correctly labels as negative. If 1,000 tax returns are legitimate and an audit-selection model leaves 950 of them alone, its specificity is 0.95. The formula is true negatives divided by the sum of true negatives and false positives. Specificity and the false positive rate always add up to one, so a specificity of 0.95 means a false positive rate of 0.05.
Specificity is the mirror image of recall, also called sensitivity. Recall asks how many real positives the model caught, and specificity asks how many real negatives it correctly cleared. The ROC curve plots the two against each other across every possible confidence threshold.
Why specificity matters
Specificity protects the people and transactions that should pass untouched. A county revenue office that selects returns for audit wants high recall on genuine underpayment, but every legitimate filer pulled into an audit costs staff time and goodwill. Because legitimate returns outnumber problem returns by a wide margin, even a small drop in specificity produces a large pile of unnecessary audits. The same arithmetic applies to fraud screening, medical testing, and quality inspection: when negatives dominate, specificity drives the workload. Report it alongside recall and insist that vendors show both on data with a realistic class balance.