Calculator Input
Paste rows as actual,score or id,actual,score.
Labels may be 0/1, yes/no, true/false, or positive/negative.
Scores should be probabilities between 0 and 1.
Example Data Table
This preview shows the first rows currently loaded into the calculator.
| ID | Actual | Score |
|---|---|---|
| 1 | 1 | 0.9800 |
| 2 | 1 | 0.9300 |
| 3 | 0 | 0.8800 |
| 4 | 1 | 0.8400 |
| 5 | 0 | 0.7900 |
| 6 | 1 | 0.7600 |
| 7 | 0 | 0.6700 |
| 8 | 1 | 0.6100 |
| 9 | 0 | 0.5800 |
| 10 | 0 | 0.4600 |
Formula Used
- Precision = TP / (TP + FP)
- Recall / TPR = TP / (TP + FN)
- Specificity = TN / (TN + FP)
- False Positive Rate = FP / (FP + TN)
- Accuracy = (TP + TN) / Total
- F1 = 2 × Precision × Recall / (Precision + Recall)
- F-Beta = (1 + β²)PR / (β²P + R)
- Jaccard = TP / (TP + FP + FN)
- Youden’s J = Recall + Specificity − 1
- MCC = (TP×TN − FP×FN) / √((TP+FP)(TP+FN)(TN+FP)(TN+FN))
- ROC AUC uses trapezoidal integration over FPR and TPR.
- Average Precision sums precision across recall gains.
How to Use This Calculator
- Paste classification results as actual labels and predicted scores.
- Choose a threshold between 0 and 1.
- Set beta if recall or precision deserves more emphasis.
- Optional: add business costs for false positives and false negatives.
- Click Calculate Metrics to generate the confusion matrix, curves, AUC values, and threshold sweep table.
- Use the recommended thresholds to compare different operating goals.
- Export the threshold table as CSV or save the result section as PDF.
FAQs
1. What does this calculator actually measure?
It evaluates binary classifier performance from predicted scores. You get ROC, precision-recall behavior, confusion counts, threshold recommendations, AUC values, and business-cost comparisons in one place.
2. When should I prefer precision over recall?
Prefer precision when false alarms are expensive, such as fraud reviews or manual moderation queues. Prefer recall when missing true positives creates bigger losses, such as medical screening or incident detection.
3. Why can ROC AUC look good while precision is weak?
ROC AUC focuses on ranking quality across thresholds. Precision depends strongly on class imbalance and selected threshold. A model can rank reasonably well while still producing many false positives at deployment.
4. What input format does the dataset accept?
Use rows like 1,0.87 or row15,0,0.21. Labels can be 0/1, yes/no, true/false, or positive/negative. Scores must remain between 0 and 1.
5. What is a good threshold?
There is no universal best threshold. Good thresholds depend on class balance, business costs, review capacity, and your tolerance for false positives versus false negatives.
6. Why is average precision useful?
Average precision summarizes the precision-recall curve into one number. It is especially informative for imbalanced datasets where ROC AUC alone may hide weak positive-class targeting.
7. What does MCC add beyond F1?
MCC uses all four confusion matrix cells and stays informative under imbalance. F1 ignores true negatives, while MCC rewards balanced classification quality more directly.
8. Can I use this with probabilities from any model?
Yes. Any model that outputs binary-class scores or probabilities can be tested here, including logistic regression, gradient boosting, neural networks, and calibrated anomaly detectors.