Model Performance Metrics Calculator

Accuracy weights every correct prediction equally regardless of which class it belongs to, so on an imbalanced dataset -- say 95% negative cases -- a model that predicts "negative" for everything scores 95% accuracy while never correctly catching a single positive case. Precision, recall, and MCC are all more informative in that situation because they expose exactly how the model is failing on the minority class, which accuracy alone hides.

Because they're sensitive to different kinds of mistakes -- precision only gets hurt by false positives, and recall only gets hurt by false negatives. A model tuned to flag more cases as positive (to catch more true positives and raise recall) will typically also pick up more false positives along the way, which lowers precision, and vice versa for a more conservative model. This tradeoff is exactly what F1 score tries to balance into a single figure.

MCC ranges from -1 (total disagreement between predictions and actual labels) through 0 (no better than random guessing) to +1 (perfect prediction), and unlike accuracy or F1, it uses all four confusion-matrix cells including true negatives, which makes it harder to inflate by favoring one class. An MCC near 0 means the classifier's predictions carry essentially no real information about the true labels, even if accuracy looks high due to class imbalance.

Accuracy counts every correct prediction, including true negatives, in its numerator and denominator, so correctly identifying more negative cases directly raises it. Precision and recall are both defined purely in terms of the positive class -- precision only looks at predicted positives, recall only at actual positives -- so neither formula includes true negatives at all, and adding more of them leaves both metrics completely unchanged.