Matthews Correlation Coefficient

Matthews Correlation Coefficient (MCC) is a classification metric that uses all four confusion matrix values — true positives, true negatives, false positives, and false negatives — to produce a single balanced score from -1 to +1.

What It Is

When you evaluate a classifier using precision, recall, and F1, you get a useful picture — but sometimes an incomplete one. If your dataset has far more negative samples than positive ones (say, 95% legitimate transactions and 5% fraud), a model that rarely flags anything can still look decent on accuracy alone. MCC exists to close that blind spot. It gives you a single number that only scores high when a classifier handles both the positive and the negative class well.

MCC pulls together all four cells of the confusion matrix into one formula. Think of it like a report card that grades every type of answer — correct acceptances, correct rejections, false alarms, and missed detections. According to scikit-learn Docs, the formula is:

MCC = (TP x TN - FP x FN) / sqrt((TP+FP)(TP+FN)(TN+FP)(TN+FN))

The result lands between -1 and +1. A score of +1 means every prediction matched the true label. Zero means the model did no better than flipping a coin. And -1 means every prediction was inverted — the model consistently got it backwards.

What separates MCC from F1? F1 combines precision and recall but ignores true negatives entirely. That works fine when both classes are roughly equal in size. But when one class dominates — which is common in real-world tasks like spam detection, medical screening, or fraud flagging — F1 can overrate a model that handles the majority class well but botches the minority class. According to Chicco & Jurman (2020), MCC produces a high score only when all four quadrants of the confusion matrix show good results, making it a more reliable single-number summary for imbalanced classification.

Brian W. Matthews originally proposed this metric in 1975 for predicting protein secondary structures — a biology problem, not a machine learning one. The math is identical to the phi coefficient in statistics, which measures association between two binary variables. That dual heritage explains why MCC is trusted across fields: it rests on well-understood correlation mathematics rather than a domain-specific heuristic.

How It’s Used in Practice

The most common place you encounter MCC is in model evaluation reports, especially when teams benchmark classifiers on imbalanced datasets. If you are a product manager reviewing a vendor’s classification model — for churn prediction, content moderation, or anomaly detection — and the dataset is skewed, asking for MCC alongside precision, recall, and F1 gives you a more honest performance picture.

Computing MCC takes one line of code. According to scikit-learn Docs, you call sklearn.metrics.matthews_corrcoef(y_true, y_pred) and get a single float back. Most teams add it to their evaluation dashboard right next to accuracy, precision, recall, and F1 so they can spot disagreements between metrics at a glance.

Pro Tip: If a model’s F1 looks strong but MCC is mediocre, that’s a warning sign. It usually means the model catches the positive class well but generates too many false positives or misses too many true negatives. Ask your data team to break out the full confusion matrix before making deployment decisions.

When to Use / When Not

Common Misconception

Myth: MCC is only useful for imbalanced datasets and adds nothing when classes are roughly equal. Reality: MCC is informative regardless of class balance. On balanced datasets it correlates closely with accuracy and F1, which confirms those metrics are telling the truth. On imbalanced datasets, divergence between MCC and F1 reveals problems that F1 alone would hide. It always works as a sanity check — the question is whether the other metrics are already reliable enough on their own.

One Sentence to Remember

When precision, recall, and F1 paint a rosy picture but your dataset is heavily skewed, MCC is the metric that keeps everyone honest — it only rewards classifiers that get all four types of predictions right.

FAQ

Q: How is MCC different from F1 score? A: F1 combines precision and recall but ignores true negatives entirely. MCC factors in all four confusion matrix values, which makes it a more reliable single metric when class sizes differ significantly.

Q: Can MCC handle multi-class classification problems? A: Not directly — MCC is defined for binary classification. For multi-class problems, you can compute MCC per class using one-vs-rest binarization and then average the per-class results.

Q: What MCC score counts as good? A: No universal cutoff exists, but scores above 0.7 generally indicate strong agreement between predictions and labels. Scores near 0 mean the model performs no better than random chance.

Sources

• scikit-learn Docs: matthews_corrcoef — scikit-learn documentation - Official API reference with formula and usage examples
• Chicco & Jurman (2020): The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation - Peer-reviewed analysis comparing MCC with F1 and accuracy