The false positive rate (FPR) looks only at the cases where the outcome did not happen. Of all the customers who stayed, all the transactions that were legitimate or all the machines that did not fail, what share did the model flag anyway? In confusion matrix terms, FPR is false positives divided by false positives plus true negatives.
Using the counts published in CUBIG’s UCI Iranian Churn experiment (October 2026), before data preparation, logistic regression flagged 8 of the 531 customers who did not churn, a false positive rate of about 1.5%. After preparation, with the same model settings and evaluation records, it flagged 41 of them, about 7.7%. Over the same change, the model found 78 of 99 churners instead of 42.
Related rates from the same four counts:
- true positive rate, also called recall: the share of actual positives the model flagged
- true negative rate, or specificity: the share of actual negatives the model left alone, equal to 1 minus the false positive rate
- false discovery rate: the share of flags that were wrong, equal to 1 minus precision
How is the false positive rate different from the false discovery rate?
- The false positive rate divides false alarms by all actual negatives. It answers how often the model bothers a case where nothing happened.
- The false discovery rate divides false alarms by all flags. It answers how many of the flags a reviewer receives are wrong, and it equals 1 minus precision.
- In the Iranian Churn counts, the false positive rate after preparation was about 7.7%, while about 34.5% of the 119 flags were wrong. The two rates describe the same false alarms against different totals.
Where does the false positive rate break in AI work?
A low false positive rate can still mean a heavy workload when negatives vastly outnumber positives. If a fraud model screens a million legitimate transactions, a 1% false positive rate still sends ten thousand false alarms to review. Teams that report only the rate can hide the volume a reviewer actually faces.
The false positive rate also depends on the decision threshold and on the evaluation records, so two rates are comparable only at the same threshold on the same evaluation set. A ROC curve plots the true positive rate against the false positive rate across thresholds, which shows the trade-off instead of one point on it. Whether a rise from 1.5% to 7.7% is acceptable depends on what each missed case and each false alarm costs the team.