statistics

Confusion Matrix Calculator

A confusion matrix condenses the outcomes of a binary or multiclass classification model into a table of four key counts: true positives, false positives, false negatives, and true negatives. From these raw counts, you can derive dozens of performance metrics. This calculator computes accuracy, precision, recall, F1 score, sensitivity, specificity, and Matthews correlation coefficient—essential diagnostics for evaluating whether your classifier performs well on your specific problem.

Last updated: May 9, 2026

Creators Wojciech Sas, PhD

Reviewers Anna Szczepanek and Davide Borchia

3,221 people find this calculator helpful

Understanding Confusion Matrix Components

A confusion matrix breaks down classification predictions into four categories based on whether predictions match reality. True positives (TP) are cases correctly predicted as positive. False positives (FP) are cases incorrectly predicted as positive when they were actually negative. False negatives (FN) are cases incorrectly predicted as negative when they were actually positive. True negatives (TN) are cases correctly predicted as negative.

Arranging these four values in a 2×2 matrix reveals patterns in model behaviour. A model may excel at identifying positives but struggle with negatives, or vice versa. The relative sizes of these counts directly influence all downstream metrics, making them the foundation of performance assessment.

Key Performance Metrics Explained

Accuracy measures the proportion of correct predictions overall: how often the model got it right across all cases. It works well when classes are balanced, but can be misleading with imbalanced data.

Precision answers: of all cases predicted positive, how many were actually positive? High precision means few false alarms—critical in applications like medical screening where false positives carry cost.

Recall (also called sensitivity) answers: of all truly positive cases, how many did we catch? High recall matters when missing positives is costly, such as fraud detection.

F1 score balances precision and recall into a single metric, useful when you want both qualities without favouring one over the other.

Specificity measures the true negative rate—how well the model identifies negative cases. False positive rate and false negative rate show the flip side: what fraction of negatives and positives does the model misclassify.

Confusion Matrix Formulas

The following equations derive performance metrics from the four confusion matrix counts:

Accuracy = (TP + TN) / (TP + TN + FP + FN)

Precision = TP / (TP + FP)

Recall = TP / (TP + FN)

F1 Score = (2 × Precision × Recall) / (Precision + Recall)

True Positive Rate = TP / (TP + FN)

True Negative Rate = TN / (TN + FP)

False Positive Rate = FP / (FP + TN)

False Negative Rate = FN / (FN + TP)

Matthews Correlation Coefficient = (TP × TN − FP × FN) / √((TP + FP)(TP + FN)(TN + FP)(TN + FN))

TP — Count of cases correctly predicted as positive
TN — Count of cases correctly predicted as negative
FP — Count of cases incorrectly predicted as positive
FN — Count of cases incorrectly predicted as negative

Interpreting Results with a Practical Example

Suppose a medical diagnostic model tested 200 patients: 80 TP, 30 TN, 20 FP, and 70 FN. Accuracy is (80+30)/200 = 0.55 or 55%—moderate overall performance. Precision is 80/(80+20) = 0.80, meaning 80% of patients flagged as having the disease actually do. Recall is 80/(80+70) ≈ 0.53, revealing that the model only catches about half of truly sick patients. The low recall is alarming in medical contexts, suggesting 70 diseased individuals would be missed. F1 score of approximately 0.64 reflects this imbalance. This example illustrates why aggregate accuracy can mislead: despite moderate accuracy, high false negatives make this model unsuitable for clinical use.

Common Pitfalls When Using Confusion Matrices

Avoid these frequent mistakes when assessing classification model performance.

Ignoring class imbalance — When one class dominates (e.g., 95% negative cases), accuracy alone can be deceptive. A naive model predicting everything as negative achieves 95% accuracy despite being useless. Always examine precision, recall, and F1 score when classes are imbalanced.
Confusing precision and recall — Precision focuses on false alarms among positive predictions, while recall measures missed true positives. A spam filter needs high precision to avoid blocking legitimate mail; a cancer detector needs high recall to avoid missing patients. Know which metric matters for your use case.
Forgetting the cost of errors — Not all mistakes carry equal weight. A false negative in fraud detection costs more than a false positive, while the reverse may be true for email classification. Adjust your threshold and metric priorities based on real business consequences.
Assuming one test set is enough — A confusion matrix from a single hold-out test set can be misleading due to random variation. Cross-validation over multiple folds provides a more robust picture of model reliability and generalization.

Frequently Asked Questions

What distinguishes a confusion matrix from other model evaluation methods?

A confusion matrix provides granular insight into four specific outcomes of a classifier, enabling calculation of dozens of metrics. Unlike a single-number summary like accuracy, it reveals where the model succeeds and fails. This transparency makes it indispensable for diagnosing problems: you immediately see if the model favours false positives over false negatives, or vice versa. Other evaluation methods like ROC curves or precision-recall plots derive from confusion matrix data but aggregate information differently.

When should I prioritise recall over precision?

Recall becomes paramount when the cost of missing positive cases outweighs false alarms. Medical diagnoses, criminal suspect identification, and equipment failure detection all benefit from high recall because a missed positive can cause serious harm. Conversely, choose precision when false positives are expensive—spam filtering, loan approval, or hiring decisions. Many real-world systems require balance; the F1 score helps when you value both equally, but domain knowledge should always guide your choice.

How do I use confusion matrix results to improve my model?

Examine the distribution of errors. If false negatives dominate, your model is too conservative; try lowering the decision threshold or rebalancing training data toward the minority class. If false positives dominate, raise the threshold. Check for systematic errors: does the model fail on specific data subgroups? Collect more examples of underrepresented classes, engineer better features, or try different algorithms. Use the confusion matrix alongside error analysis to guide focused improvements rather than blind hyperparameter tuning.

What does Matthews correlation coefficient measure that accuracy does not?

Matthews correlation coefficient (MCC) is a correlation measure that accounts for all four confusion matrix values and ranges from −1 to +1, where +1 indicates perfect prediction and 0 indicates random guessing. Unlike accuracy, MCC remains informative even with class imbalance. A high accuracy paired with low MCC signals that the model performs well on majority cases but fails on the minority class—a red flag accuracy alone would miss. MCC is particularly useful for imbalanced classification problems.

Can a model have high precision but low recall, and what does this mean?

Yes, and this reveals a conservative classifier. Consider a spam detector with 95% precision but 40% recall: most emails it flags are indeed spam (high precision), but it lets 60% of actual spam through (low recall). This happens when the decision threshold is set high, making the model reluctant to classify anything as positive. For users annoyed by spam, 40% recall is frustrating. Adjusting the threshold lower increases recall but typically decreases precision—a trade-off inherent to most classifiers.

How does class imbalance affect confusion matrix interpretation?

With imbalanced data, accuracy becomes unreliable. If 99% of cases are negative and your model predicts everything negative, it achieves 99% accuracy despite being worthless. The confusion matrix reveals this: you'd see large TN, but TP and FN would both be zero. Precision and recall tell a more honest story. Techniques like stratified cross-validation, class weighting, or resampling help. Always compare metrics on the confusion matrix rather than trusting a single accuracy figure when classes are severely imbalanced.

More statistics calculators (see all)

McNemar's Test Calculator Wilcoxon Rank Sum Test Calculator Decile Calculator Bonferroni Correction Calculator Weibull Distribution Calculator Frequency Polygon Calculator 5 Number Summary Calculator Mode Calculator