False Positive Rate Calculator
When a test or model cries wolf, the false positive rate (FPR) tells you how often. It is the share of genuinely negative cases that get flagged as positive: a screening test that alarms 30 healthy people out of 1,000 has an FPR of 3%. This calculator works from a full confusion matrix, from known sensitivity and specificity rates, or from total counts, and adds specificity, a Wilson confidence interval, predictive values and an optional cost analysis for the same run.
Quick Answer
In the pre-filled screening example, 30 of 1,000 actual negatives were flagged by mistake, so the false positive rate is 3.00% and specificity is 97.00%, with a 95% Wilson confidence interval of 2.11% to 4.25%. The formula is FPR = FP ÷ (FP + TN). Enter your own counts, rates or totals above and the same run also returns sensitivity, predictive values and, if you add costs, what the false alarms are costing you.
Key takeaways
- FPR only looks at actual negatives. Prevalence, sensitivity and the number of true positives cannot move it.
- FPR and specificity always add to 100%. A specificity of 97% and an FPR of 3% are the same fact stated twice.
- Do not confuse FPR with the false discovery rate: FPR divides by actual negatives, FDR divides by positive calls.
- A rate from 1,000 negatives is a measurement; a rate from 20 negatives is a guess. The confidence interval tells you which one you have.
- Quick Answer
- How to use the false positive rate calculator
- The false positive rate formula
- Worked example: a screening test on 1,100 people
- False positive rate is not the false discovery rate
- Reading your result: the benchmark bands
- The confidence interval, and what Sample Size really does here
- What the other outputs mean
- Honest limits of this calculator
- What competing calculators miss
How to use the false positive rate calculator
Pick the input method that matches the data you actually have, fill in that section, then click Calculate. The results stay hidden until you do.
- Confusion Matrix is the fullest route. Enter all four outcome counts (true positives, false negatives, false positives, true negatives) from a finished test run or model evaluation, and you get every output the tool offers, including sensitivity, predictive values and accuracy.
- Direct Rates is for published or vendor-supplied figures. Specificity is the only required input, because FPR = 100% minus specificity. Add sensitivity and prevalence if you know them. Add a sample size as well and the tool rebuilds estimated counts and works through the full calculation instead of the direct one.
- Total Counts is the shortest route. Enter the total number of actual negatives and how many of them were false positives. The rate is the second number divided by the first.
- Application Domain and costs are optional context. The domain changes only the benchmark band your result is judged against. The three cost fields turn the run into a cost analysis: what the false positives cost, what the false negatives cost, and the total including testing.
The false positive rate formula
FPR = FP ÷ (FP + TN). Divide the false positives by all the actual negatives, then multiply by 100 for a percentage.
Specificity = True Negatives ÷ (True Negatives + False Positives)
FPR = 1 − Specificity
The denominator is the part people get wrong. It is every case that was truly negative, not every case tested and not every positive call. A model that flags 30 healthy patients out of 1,000 healthy patients has an FPR of 3%, even if it also processed 100 sick patients on the same day. In hypothesis testing, the same quantity is the probability of a Type I error: the chance of declaring a finding when nothing is there. On an ROC curve, FPR is the x-axis, plotted against the true positive rate as the decision threshold moves.
Worked example: a screening test on 1,100 people
A screening programme tests 1,100 people: 100 have the condition and 1,000 do not. The test catches 95 of the 100 cases and wrongly alarms 30 of the healthy group. Those are the pre-filled values above.
| Confusion matrix | Predicted positive | Predicted negative |
|---|---|---|
| Actual positive (100) | TP = 95 | FN = 5 |
| Actual negative (1,000) | FP = 30 | TN = 970 |
FPR = 30 ÷ (30 + 970) = 30 ÷ 1,000 = 3.00%. Specificity = 970 ÷ 1,000 = 97.00%, and the two sum to 100% as they always do. The full run also reports sensitivity of 95.00%, accuracy of 96.82% and an F1-score of 84.44. Under Predictive Values it reports a PPV of 76.00%: of the 125 positive results, 95 were real. That gap between a 97% specific test and a 76% predictive value is prevalence at work, and it is why the next section matters.
| Check (real runs of this calculator) | FPR | Specificity |
|---|---|---|
| Spam filter: FP 100, TN 1,900 (Confusion Matrix) | 5.00% | 95.00% |
| Totals mode: 40 false positives out of 800 negatives | 5.00% | 95.00% |
| Direct Rates: specificity 98.5% entered | 1.50% | 98.50% |
False positive rate is not the false discovery rate
The two rates sound alike and answer different questions. FPR asks: of everything that was really negative, how much did we flag? The false discovery rate asks: of everything we flagged, how much was wrong?
| Rate | Formula | Denominator | In the worked example |
|---|---|---|---|
| False positive rate | FP ÷ (FP + TN) | Actual negatives (1,000) | 30 ÷ 1,000 = 3.00% |
| False discovery rate | FP ÷ (TP + FP) | Positive calls (125) | 30 ÷ 125 = 24.00% |
Same test, same 30 false alarms, two very different headlines: 3% of healthy people were alarmed, and 24% of alarms were false. FDR is the rate that multiple-testing procedures such as Benjamini-Hochberg control, and it has its own page: the false discovery rate calculator works from p-values and study counts when the discovery share is the number you need. Use FPR when you are judging a test against the people who were never positive; use FDR when you are judging a set of positive findings.
Reading your result: the benchmark bands
The badge on your result compares your FPR with the band table built into this tool for the domain you selected. The cutoffs below are the tool’s own reference values, stated plainly so you can see exactly what your verdict rests on. They are a starting point for judgement, not a clinical or regulatory standard.
| Domain | Excellent ≤ | Good ≤ | Acceptable ≤ | Poor ≤ |
|---|---|---|---|---|
| Medical Diagnosis | 1% | 3% | 5% | 10% |
| Drug Testing | 0.5% | 2% | 5% | 10% |
| Security Screening | 2% | 5% | 10% | 20% |
| Fraud Detection | 1% | 5% | 10% | 20% |
| Spam Filtering | 0.1% | 1% | 5% | 10% |
| Malware Detection | 0.5% | 2% | 5% | 10% |
| Quality Control | 1% | 3% | 5% | 10% |
| Machine Learning Model | 1% | 5% | 10% | 20% |
| Intrusion Detection | 1% | 3% | 5% | 10% |
| Biometric Authentication | 0.1% | 1% | 3% | 5% |
| A/B Testing | 1% | 5% | 10% | 20% |
| Statistical Test | 1% | 5% | 10% | 20% |
| Custom Application | 1% | 5% | 10% | 20% |
Above the Poor cutoff the verdict is Very Poor. Separately, the Performance Assessment line grades specificity alone, the same way for every domain: under 1% is Outstanding, under 5% Good, under 10% Moderate, under 20% Poor, and 20% or more Very poor. The two readings can disagree, and that is informative: a 4% FPR is a Good performance in the abstract yet only Acceptable for spam filtering, where a false alarm means a lost customer email.
The confidence interval, and what Sample Size really does here
The interval is a Wilson score interval, the standard fix for proportions near 0% or 100% where the simple normal interval misbehaves. It answers: given this many actual negatives, how precisely does the observed FPR pin down the true one?
One behaviour of this tool deserves to be stated plainly, because it is easy to misread. Entering a Sample Size switches the interval on, but the interval’s width is computed from the number of actual negatives in your data (FP + TN), not from the number you typed. Run the worked example with a sample size of 50 or of 100,000 and you get the same 2.11% to 4.25% interval, because the data behind it has 1,000 actual negatives either way. Statistically that is the defensible choice: the precision of an FPR estimate is set by how many negatives you observed, and the estimate cannot get sharper just because a bigger number was typed into a box. If you want the general method for any proportion, the confidence interval calculator works through it directly.
What the other outputs mean
Three result lines repay a closer look: predictive values, the F1-score, and the cost analysis.
Predictive values: where prevalence bites
PPV and NPV answer the question the person holding a result actually asks: given this result, how likely is it right? They are Bayes’ theorem applied to the same four counts, which is why a test with 97% specificity can still produce a PPV of only 76% when positives are uncommon, as the worked example shows. FPR itself never moves with prevalence; predictive values do almost nothing else. The mechanics are the same ones the conditional probability calculator walks through: update a prior with evidence, then read the revised probability.
Sensitivity, accuracy and F1
Sensitivity is the mirror metric on the positive side: TP ÷ (TP + FN), the share of real cases caught. Accuracy counts every correct call over all cases and flatters models on imbalanced data, which is why it sits in a supporting line rather than the headline. The F1-score is the harmonic mean of precision (PPV) and sensitivity, a single number for the positive-class trade-off. None of the three changes the FPR; together they stop a low FPR from hiding a test that catches nothing.
Cost analysis: putting a price on false alarms
Fill in the three cost fields and the run prices itself. In a fraud-style run of this calculator (TP 80, FN 20, FP 10, TN 890) with a false positive costing $25, a false negative $400 and testing $10 per case, the tool reports a total FP cost of $250.00, a total FN cost of $8,000.00 and a total cost of $18,250.00:
The shape of that chart is the usual story: the false alarms themselves are cheap, the misses are not, and the testing bill dwarfs both. It is also why the recommendations line watches the cost ratio: when one false positive costs more than twice a false negative, the tool suggests raising the decision threshold.
Honest limits of this calculator
Four behaviours are worth knowing before you rely on a number from it.
- Total Counts mode needs only two numbers. The rate runs on total negatives and the FP count alone, which is all the formula needs, so there is no total positive cases field: it was recorded but never used, and it has been removed.
- Totals mode does not report NPV or accuracy. With no TP or FN counts, NPV would always read 100% and accuracy would simply equal specificity, so the totals run no longer prints either figure. Use Confusion Matrix mode when you need predictive values or accuracy.
- Negative counts are rejected. A count below zero in any count field now shows an error and produces no result, instead of returning a negative FPR or a specificity above 100%. Real counts cannot be negative.
- Direct Rates mode skips costs. The direct path from specificity always reports “No cost data provided”, even if the cost fields are filled. Run Confusion Matrix mode for a priced analysis.
What competing calculators miss
Most FPR tools stop at the single division. The gaps in the usual offerings are consistent:
- Two-box calculators (CalculatorsHub and similar): FP and TN in, a percentage out. Correct arithmetic, but no specificity alongside, no interval, no sense of whether the rate is good or bad for your field, and no route in when all you know is a published specificity.
- Suite calculators (OmniCalculator’s confusion matrix and sensitivity/specificity pages, OnlyCalculators’ accuracy suite): broad metric sets done well, but FPR is one line among many, the outputs assume a medical framing, and none of them prices the errors or benchmarks the rate by domain.
- Clinical software (MedCalc’s diagnostic test evaluation): the professional standard, with likelihood ratios and intervals, but it is paid desktop software documentation territory, medical-only, and it will not take a specificity you already know and hand back the FPR.
- A naming error worth flagging: at least one established statistics package’s help pages use “false positive rate” for the probability that a positive result is wrong. That quantity is 1 minus PPV, a discovery-style rate with positive calls in the denominator. It is not FP ÷ (FP + TN), and mixing the two is how a 3% FPR gets quoted as a 24% scandal.
- The gap this page fills: three ways in (counts, rates, totals), the rate judged against your own domain’s bands, a Wilson interval with its sample-size behaviour disclosed rather than hidden, and a cost line that turns the percentage into a budget decision.
Frequently asked questions
It is the share of actual negative cases that a test or model wrongly calls positive: FPR = FP / (FP + TN). With 30 false positives and 970 true negatives, FPR = 30 / 1,000 = 3.00%. It answers one question: of everything that was really negative, how much did we flag by mistake?
Subtract specificity from 100%. FPR = 1 – specificity, so a test with 97% specificity has an FPR of 3%. That is exactly what Direct Rates mode does: enter the specificity and the FPR comes back, with sensitivity and prevalence reported too if you add them.
The denominators differ. FPR divides false positives by all actual negatives (FP + TN); the false discovery rate divides them by all positive calls (TP + FP). In the worked example the same 30 false alarms give an FPR of 3.00% and a discovery share of 24.00%. FPR judges a test against people who were never positive; FDR judges a set of positive findings.
It depends on the cost of a false alarm in your field, and this tool judges your result against the band table built into it: medical diagnosis counts 1% or under as excellent and up to 5% as acceptable, spam filtering demands 0.1% for excellent because a false alarm loses a real email, and security screening tolerates up to 10% as acceptable. Treat the bands as a starting point, then weigh what a false alarm actually costs you.
The interval shows how precisely your counts pin down the true rate: an FPR of 3% from 1,000 negatives is a measurement, while 3% from 30 negatives is a guess. This tool uses a Wilson score interval. Entering any sample size switches the interval on, but its width is computed from the actual negatives in your data (FP + TN), not from the number typed, so the same data always gives the same interval.
Yes. If a test produces no false positives, FP is 0 and the FPR is 0.00% whatever the TN count. The confidence interval still earns its place in that case: zero observed false alarms in 200 negatives is consistent with a true rate of a percent or so, and the interval says how much room is left.
No. FPR is computed only from actual negatives, so how common the condition is cannot move it. Prevalence moves the predictive values instead: the rarer the condition, the more of your positive results are false alarms, which is why the worked example pairs a 3.00% FPR with a PPV of only 76.00%.
In hypothesis testing, yes. The false positive rate is the probability of a Type I error: declaring an effect or a positive when the null situation is true. A test run at a 5% significance level is designed to produce false positives 5% of the time when nothing is going on, which is exactly an FPR of 5%.
Add This Calculator to Your Website
Embed this calculator on your own site with the code below. Pick a size to preview it live, then copy the code.
Powered by Calculator Zilo
References
- Altman DG, Bland JM. Diagnostic tests 1: sensitivity and specificity. BMJ, 1994. Accessed October 21, 2026.
- Bossuyt PM, Reitsma JB, Bruns DE, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ, 2015. Accessed October 21, 2026.
- National Institute of Standards and Technology. NIST/SEMATECH e-Handbook of Statistical Methods: Critical values and p values. Accessed October 21, 2026.
- U.S. Food and Drug Administration. Statistical Guidance on Reporting Results from Studies Evaluating Diagnostic Tests. Guidance for Industry and FDA Staff, 2007. Accessed October 21, 2026.
- Brown LD, Cai TT, DasGupta A. Interval Estimation for a Binomial Proportion. Statistical Science, 2001. Accessed October 21, 2026.
- scikit-learn developers. Metrics and scoring: quantifying the quality of predictions. Accessed October 21, 2026.