Confusion Matrix vs. ROC Curve

Last Updated July 20, 2026

A Confusion Matrix shows the model's exact correct and incorrect predictions at one chosen threshold. An ROC Curve changes the threshold and shows how catching more positives also changes the false-alarm rate.

  • Use Confusion Matrix when you already have a threshold and need to inspect the exact mistakes made at that setting.
  • Use ROC Curve when you are still comparing thresholds and want to see the trade-off between true positives and false positives.
Confusion Matrix explains one chosen cutoff; ROC Curve compares many possible cutoffs.

Head-to-Head Showdown

Decision Scope

Confusion Matrix: One operating point

ROC Curve: Many operating points

The Implication: A Confusion Matrix evaluates the discrete predictions produced at one selected threshold. An ROC Curve repeats that thresholding process across many cutoffs and compares the resulting operating points.

Information Shown

Confusion Matrix: TP, TN, FP, FN

ROC Curve: TPR against FPR

The Implication: A Confusion Matrix preserves exact class counts, making specific false positives and false negatives visible. The ROC Curve converts those counts into rates so threshold trade-offs can be compared across operating points.

Relationship Between Them

Confusion Matrix: Creates one ROC point

ROC Curve: Combines threshold points

The Implication: At each threshold, the model first produces a Confusion Matrix. Its TP, FP, TN, and FN values generate one TPR/FPR coordinate, and repeating this process creates the ROC Curve.

Selection Criteria

Scenario:A hospital must inspect exactly how many sick patients are missed at the alert threshold currently used in production.

Choose Confusion Matrix:The threshold is already fixed, so the hospital needs exact TP, TN, FP, and FN counts at that operating point. An ROC Curve shows rate trade-offs across thresholds but does not replace the threshold-specific error counts needed for deployment review.

Scenario:A fraud team has not selected a probability cutoff and wants to compare higher detection rates against additional false alarms.

Choose ROC Curve:ROC analysis shows how TPR and FPR change as the cutoff moves, allowing the team to inspect multiple operating choices. A single Confusion Matrix would reveal only one threshold and could hide better trade-offs available elsewhere.

Scenario:Two binary classifiers must be compared before choosing the operating threshold for either model.

Choose ROC Curve:ROC Curves compare model discrimination across many thresholds instead of judging both models at one arbitrary cutoff. After selecting a model and threshold, a Confusion Matrix should still be used to inspect the exact deployed errors.

Side By Side Trace

A medical model assigns Sick probabilities to six patients. The Confusion Matrix applies one selected threshold, 0.500.50, to produce exact TP, TN, FP, and FN counts. The ROC example uses the same scores at three sampled thresholds to produce TPR/FPR operating points; these sampled points demonstrate the process but do not represent every possible threshold.

Data PointActualClassPredictedProbability
P1Sick0.95
P2Healthy0.80
P3Sick0.70
P4Healthy0.40
P5Sick0.30
P6Healthy0.10
TargetThreshold=0.50?

Step 1: Establish Shared Scores

Confusion Matrix

Confusion Matrix receives the six probability scores and selects threshold 0.500.50. Scores at or above the threshold become Positive, while lower scores become Negative.

ROC Curve

ROC Curve receives the same six scores but does not freeze evaluation at one threshold. It repeats the thresholding process at several cutoffs, with each cutoff producing a new Confusion Matrix and one TPR/FPR operating point.

Step 2: Apply Thresholds

Confusion Matrix

At threshold 0.500.50, P1, P2, and P3 are predicted Positive, while P4, P5, and P6 are predicted Negative. Comparing predictions with the actual classes gives TP=2TP=2, FP=1FP=1, TN=2TN=2, and FN=1FN=1.

ROC Curve

The ROC example samples thresholds 0.800.80, 0.500.50, and 0.300.30. Each threshold produces different Positive predictions and therefore different TP, FP, TN, and FN counts.

Step 3: Point of Divergence

Confusion Matrix

Confusion Matrix stops at threshold 0.500.50 and preserves the exact counts TP=2TP=2, FP=1FP=1, TN=2TN=2, and FN=1FN=1. These counts describe one selected operating point.

ROC Curve

ROC analysis converts each threshold-specific Confusion Matrix into rates using TPR=TP/(TP+FN)TPR=TP/(TP+FN) and FPR=FP/(FP+TN)FPR=FP/(FP+TN). Plotting the resulting coordinates shows how sensitivity changes against false-positive rate.

Step 4: Calculate Final Outputs

Confusion Matrix

At threshold 0.500.50, the final Confusion Matrix is TP=2TP=2, TN=2TN=2, FP=1FP=1, and FN=1FN=1.

ROC Curve

At threshold 0.800.80, TP=1TP=1, FP=1FP=1, TN=2TN=2, and FN=2FN=2, giving TPR=1/3TPR=1/3 and FPR=1/3FPR=1/3. At 0.500.50, TPR=2/3TPR=2/3 and FPR=1/3FPR=1/3. At 0.300.30, P1, P2, P3, P4, and P5 are Positive, giving TP=3TP=3, FP=2FP=2, TN=1TN=1, FN=0FN=0, TPR=1TPR=1, and FPR=2/3FPR=2/3.

Step 5: Compare the Evaluation Scope

Confusion Matrix

Confusion Matrix produces one exact error snapshot for the chosen deployment threshold.

ROC Curve

ROC analysis produces multiple TPR/FPR operating points and reveals how threshold changes exchange additional true positives for additional false positives.

Final Result

Confusion Matrix:Confusion Matrix evaluates threshold 0.500.50 and records the exact outcomes TP=2TP=2, TN=2TN=2, FP=1FP=1, and FN=1FN=1. It explains the model's errors at one selected operating point.

ROC Curve:ROC analysis uses the same scores to calculate sampled operating points at thresholds 0.800.80, 0.500.50, and 0.300.30. The points are (FPR=1/3,TPR=1/3)(FPR=1/3,TPR=1/3), (FPR=1/3,TPR=2/3)(FPR=1/3,TPR=2/3), and (FPR=2/3,TPR=1)(FPR=2/3,TPR=1). Each point comes from the Confusion Matrix produced at that threshold.

Common Pitfalls & Exam Mistakes

  • Thinking ROC Curve replaces Confusion Matrix.

    The Mistake: Students assume threshold-wide ROC analysis makes threshold-specific confusion counts unnecessary.

    Why It's Wrong: ROC Curve compares TPR/FPR trade-offs across thresholds, while Confusion Matrix shows exact TP, TN, FP, and FN outcomes at one selected threshold. In practice, ROC analysis helps choose an operating point, and a Confusion Matrix then explains the errors at that point.

  • Building an ROC Curve from one threshold.

    The Mistake: Students calculate one TPR/FPR pair and treat that single coordinate as the complete ROC Curve.

    Why It's Wrong: One threshold produces one Confusion Matrix and therefore one ROC operating point. A curve requires operating points from multiple thresholds, usually including boundary points across the score range.

  • Assuming one Confusion Matrix predicts other thresholds.

    The Mistake: Students inspect one matrix and infer how TP, FP, TN, and FN would change at different cutoffs.

    Why It's Wrong: One Confusion Matrix contains only the predictions produced at its selected threshold. New thresholds require new predictions, new confusion counts, and therefore new ROC operating points.

Comparative Analysis

AttributeConfusion MatrixROC Curve
Decision ScopeOne operating pointMany operating points
Main OutputTP, TN, FP, FNTPR/FPR curve
Threshold UseThreshold already selectedThreshold varies
Primary PurposeInspect exact errorsCompare threshold trade-offs
Multiclass UseDirect multiclass matrixRequires multiclass extension
Summary MetricsPrecision, recall, F1ROC-AUC

Common Questions & Edge Cases

  • Can ROC Curve replace Confusion Matrix after a deployment threshold is selected?

    No. ROC Curve shows TPR/FPR trade-offs across thresholds but does not display the final deployed TP, TN, FP, and FN counts. Confusion Matrix remains necessary for inspecting exact prediction outcomes at the selected operating point.

  • Can multiple Confusion Matrices be used to construct an ROC Curve?

    Yes. Each threshold produces discrete predictions, a Confusion Matrix, and one TPR/FPR coordinate. Repeating this calculation across thresholds creates the operating points used to form the ROC Curve.

  • Can Confusion Matrix and ROC Curve evaluate the same prediction scores?

    Yes. Confusion Matrix thresholds the scores once to produce one set of discrete predictions. ROC Curve applies multiple thresholds to those same scores and compares the resulting TPR/FPR operating points.

  • Is ROC Curve always the best threshold-wide evaluation for imbalanced data?

    No. ROC Curve can remain useful, but a large negative class may make its false-positive rate appear reassuring even when many false positives occur. Precision-Recall analysis is often more informative when the rare positive class is the primary concern.

Explore the Algorithms in Action

Open the theory pages or try the interactive solvers for the algorithms compared above.

Try the Confusion Matrix Calculator

Apply one threshold and inspect TP, TN, FP, and FN counts behind a single ROC operating point.

Confusion Matrix Theory

Review how thresholded predictions create confusion counts, then connect those counts to ROC Curve TPR/FPR coordinates directly.