A stack of mostly ivory classification tokens with a small contrasting green group. Original AI-generated artwork produced for GreenCalc; not a measured diagram or empirical data display.
Start with the simplest competing rule
A classifier can be correct most of the time while missing the cases you built it to find. Suppose a held-out dataset contains 990 ordinary transactions and 10 flagged-category transactions. A rule that always predicts ordinary gets 990 of 1,000 labels right: 99% accuracy. It detects none of the ten positive cases. That baseline should appear beside the model score, not disappear behind a headline percentage.
Work through the four counts
Now consider a model with TP = 8, FN = 2, FP = 12 and TN = 978. It is correct on 986 cases, so accuracy is 98.6%. The score is lower than the always-negative baseline, yet recall is 8/10 = 80%. Precision is 8/20 = 40%. These quantities answer different operational questions: how many labels were correct, how many positive cases were found, and how many positive predictions were correct.
| Rule | Accuracy | Positive recall |
|---|---|---|
| Always negative | 99% | 0% |
| Example classifier | 98.6% | 80% |
Keep the evaluation population visible
Record the positive-class definition, class counts, decision threshold and evaluation period. Use a held-out dataset that represents the intended task. A balanced test sample changes the mixture behind ordinary accuracy; its score cannot simply be treated as the expected deployment accuracy when the real class prevalence differs. Reusing the same observations for model fitting and final reporting can also produce an optimistic comparison.
Choose an operational comparison
The model is not automatically preferable because it catches more positives. Twelve false alerts may be cheap to review or costly enough to outweigh the benefit. Set that cost interpretation before choosing a threshold, then report relevant counts alongside percentages. The Accuracy Calculator gives the compact agreement check; use the Confusion Matrix Calculator for the broader metric report. The guide to balanced accuracy shows how equal class weighting changes the comparison.
Check the model behind your result
Use the Accuracy Calculator for the supported calculation and review its stated assumptions before interpreting the output. Find related tools in the Statistics category, or browse the GreenCalc homepage for other calculation topics.
By GreenCalc Editorial Team · Updated 2026-10-26