Recall is, of all the fraud that actually happened, the share your system caught. It is the coverage counterpart to precision and a direct proxy for how much fraud you stop, so it answers the question that keeps managers up at night: how much are we missing?
What is recall, in plain English?
Recall measures coverage: of all the fraud that truly happened, how much did your system actually catch? If a hundred fraudulent transactions occurred and you flagged eighty-five of them, your recall is eighty-five percent. It is computed over all real fraud, caught and missed, which makes it a direct read on how much you are stopping versus letting through.
Recall is the metric managers feel in their gut, because its shortfall is loss. The fifteen percent you did not catch is not an abstraction; it is chargebacks, fraud reports, and money out the door. Recall is the flip side of the false negative rate: what you catch plus what you miss is the whole.
In the detection stack recall is one half of the classic pair with precision. Recall is about coverage, precision is about clean alerts, and the two pull against each other, so neither is read alone.
Recall versus precision
Raising recall usually means flagging more aggressively, which costs you on precision. The two are tuned together.
What changes | Recall | Precision |
Question it answers | Of all fraud, how much did I catch? | Of what I flagged, how much is fraud? |
What it protects | Coverage, loss prevention | Analyst time, alert trust |
Push it too hard | Flag too much, precision falls | Get conservative, miss fraud |
Hard part | Misses are invisible to measure | Sensitive to the base rate |
Flagging more aggressively to lift recall lowers precision and raises review cost and friction, so the two are tuned together against your loss appetite rather than maximized alone.
What it looks like in practice
In practice
A team reports ninety-two percent recall and celebrates. The number is computed against the fraud they know about: confirmed chargebacks and reported cases. On paper, they are catching almost everything.
Then the loss rate creeps up, which should be impossible if recall is really that high. The gap is the denominator. Their recall only counted the fraud they eventually identified; the fraud they never caught and no one reported was never in the calculation at all. Real recall was lower than the reported figure, because the misses were invisible. The reported ninety-two percent looked great precisely because it excluded the very fraud it was supposed to be measuring.
Why the denominator hides your misses
Recall has a built-in measurement problem: to compute it honestly you need to know all the fraud that happened, but the fraud you never caught is invisible by definition. It leaves no flag and often no report, so it never enters the denominator. That means reported recall is systematically optimistic, counting your catches against only the fraud you eventually learned about, not the fraud that truly occurred.
The practical consequence is that real recall often looks better on paper than it is in the losses. A team can post a high recall figure while the loss rate quietly rises, because the misses that drive those losses were never counted. That is why recall is best read alongside loss rate and matured cohorts, which capture damage the label-based recall number cannot see. Treat a high recall figure with healthy suspicion until the losses agree with it.
What to watch in the data
- The invisible denominator. Recall computed only against known fraud overstates coverage; the fraud you never caught is missing from the math.
- Recall high, loss rate rising. A contradiction that usually means real recall is lower than reported and misses are going uncounted.
- Immature cohorts. Recall on fresh data looks strong because the misses have not surfaced yet; read matured cohorts.
- Precision trade-off. A sudden recall jump often came at the cost of precision and higher review load; check both moved sensibly.
- Segment coverage. A strong overall recall can hide a segment or channel where coverage is poor and misses concentrate.
Quick questions
What is the difference between recall and precision?
Recall is the share of all fraud you caught, a measure of coverage. Precision is the share of your flags that are truly fraud, a measure of clean alerts. They trade off and are read together.
Is recall the same as the true positive rate?
Yes, they are the same quantity: the share of actual fraud that your system correctly flagged. Recall is the more common term in fraud work, and the false negative rate is its complement.
Why is recall hard to measure?
Because the fraud you never caught is invisible, so it does not enter the denominator. Reported recall counts catches against only the fraud you eventually learned about, which makes the figure optimistic.
How do I raise recall?
Usually by flagging more aggressively, which catches more fraud but lowers precision and raises review cost and friction. So you tune recall and precision together against your loss appetite, not in isolation.
Why does high recall not match my losses?
Because the reported figure excludes uncounted misses. If recall looks high while loss rate rises, the missing fraud that drives those losses was never in the recall calculation.
Go deeper
- FFIEC BSA/AML Examination Manual ↗ — The manual US examiners use to assess BSA and AML programs.
- FTC Consumer Advice: Scams ↗ — US consumer guidance on current scams and fraud, and how to report them.