Enter predicted scores and true labels to calculate ROC-AUC and see precision-recall at different thresholds.
ROC-AUC (Area Under the Receiver Operating Characteristic Curve) measures a classifier's ability to distinguish between positive and negative classes across every possible decision threshold, not just one fixed cutoff. An AUC of 1.0 means perfect separation, every positive sample scores higher than every negative sample. An AUC of 0.5 means the model performs no better than random guessing. This makes AUC threshold-independent, unlike accuracy or F1, which depend on where you set the classification cutoff.
ROC-AUC has a genuinely intuitive meaning: it's the probability that a randomly chosen positive sample receives a higher predicted score than a randomly chosen negative sample. This is exactly how this calculator computes it, using the Mann-Whitney U statistic, counting the proportion of all possible positive-negative sample pairs where the model correctly ranks the positive sample higher. An AUC of 0.85 means, given a random positive and random negative example, the model ranks the positive one higher 85% of the time.
Accuracy becomes misleading on imbalanced data, as covered in our Confusion Matrix Calculator article, a model predicting the majority class for everything can score deceptively high accuracy. AUC is more robust here because it evaluates ranking quality across all thresholds rather than one fixed decision point, though for severely imbalanced datasets, precision-recall AUC is often considered even more informative than ROC-AUC, since ROC curves can look deceptively good when negatives vastly outnumber positives.
AUC evaluates ranking quality, not calibration, a model can have excellent AUC while producing poorly calibrated probability estimates (predicting 90% when the true likelihood is closer to 60%). If your application needs trustworthy probability estimates, not just correct ranking, evaluate calibration separately (using tools like reliability diagrams or Brier score) rather than relying on AUC alone.