AUC
/ AY-yoo-SEE /
AUC — area under the curve — boils a whole ROC curve down to one number: the size of the area underneath it. Since the ROC curve lives in a 1-by-1 box, AUC runs from 0 to 1. A perfect classifier scores 1.0; pure random guessing scores 0.5; and a score below 0.5 means the model is somehow worse than a coin flip (often a sign you've flipped the labels).
There is a lovely plain-English meaning. AUC is the probability that, if you pick one truly positive example and one truly negative example at random, the model gives the positive one a higher score than the negative one. An AUC of 0.85 means it ranks them correctly 85 percent of the time. So AUC measures ranking quality — how well the model separates the two classes — independent of where you set the decision threshold.
That threshold-independence is both its charm and its limit. AUC is handy for comparing models at a glance without committing to a cutoff. But it tells you nothing about whether the model's scores are well-calibrated probabilities, and on rare-event problems it can stay impressively high even when the model is nearly useless in practice. For imbalanced data, the area under the precision-recall curve (PR-AUC) is usually the more honest summary.
Two credit models: A scores AUC 0.78, B scores 0.81. B ranks a random good-payer above a random defaulter slightly more often, suggesting it's the better separator — but you still must pick a threshold to use it, and check precision/recall there before deploying.
AUC ranks models, but you still choose a threshold to actually deploy one.
"AUC" by default means area under the ROC curve, but people also say AUC for the precision-recall curve. They are different numbers with different strengths — on imbalanced data they can disagree sharply. Always confirm which curve someone means.