Precision is, of the items you flagged, the share that are truly fraud. It is the core read on how clean your alerts are, so high precision means analysts waste little time chasing false alarms.
What is precision, in plain English?
Precision answers a narrow, practical question: of everything my system flagged, how much was really fraud? If you raised a hundred alerts and eighty turned out to be genuine fraud, your precision is eighty percent. It is a measure of how clean your flagged pile is, computed only over the cases you flagged, not over all traffic.
Operationally, precision is about analyst time and trust. High precision means most alerts are worth working, so reviewers spend their day on real fraud rather than clearing false alarms. Low precision buries the team in noise, and worse, it teaches analysts to stop trusting the alerts.
In the detection stack precision is one half of a pair. It tells you how clean your catches are, but it says nothing about how much fraud you missed. For that you need its counterpart, recall, which is why the two are always read together.
Precision versus recall
Precision and recall pull in opposite directions. Chasing one alone distorts the other.
What changes | Precision | Recall |
Question it answers | Of what I flagged, how much is fraud? | Of all fraud, how much did I catch? |
What it protects | Analyst time, alert trust | Coverage, loss prevention |
Push it too hard | Get conservative, miss fraud | Flag too much, drown in false alarms |
Read it with | Recall | Precision |
Chasing precision alone makes you conservative and lets fraud through, which is exactly why it is always read alongside recall, its coverage counterpart. The two are tuned together against your loss appetite.
What it looks like in practice
In practice
A new fraud model launches and precision comes in at forty percent, well below the seventy percent a vendor benchmark quoted. Leadership is disappointed and starts questioning the model.
An analyst digs in and finds the reason is the base rate. On this product, real fraud is a tiny fraction of traffic, so even a strong model produces several false alarms for every genuine catch, and precision looks modest as a result. On a higher-fraud population the same model would post much higher precision. The benchmark was measured on a very different prevalence, so comparing to it punished a good model for a hard population. Judged against actual fraud prevalence, the model was performing well.
Why base rates change the picture
Precision is unusually sensitive to how common fraud actually is. When fraud is rare, even a well-built model will show modest precision, because a handful of false alarms weighs heavily against the small number of real catches available. The same model dropped onto a higher-fraud population would look far more precise, with no change to the model at all.
That is why a precision number means little without knowing the base rate behind it. A generic benchmark, or a competitor's figure measured on a different population, can make a strong model look weak or a weak one look strong. Judge precision against your actual fraud prevalence, and read it next to recall, or you will draw the wrong conclusion about a model that is doing exactly what it should on a genuinely difficult population.
What to watch in the data
- Base rate context. A precision figure is meaningless without the fraud prevalence behind it; low base rates naturally depress precision.
- Precision without recall. High precision alone can mean you are flagging only the obvious cases and missing the rest; always pair the two.
- Benchmark mismatch. A vendor or peer precision number measured on a different population tells you little about your own.
- Analyst trust erosion. Falling precision floods the queue with false alarms and trains reviewers to ignore alerts, which is its own risk.
- Segment variation. Precision can swing hard by product or channel because base rates differ; a blended number hides that.
Quick questions
What is the difference between precision and recall?
Precision is the share of your flags that are truly fraud. Recall is the share of all fraud that you caught. Precision is about clean alerts; recall is about coverage. They trade off and are read together.
Why is my precision low even with a good model?
Often because fraud is rare on your population. When true fraud is a tiny fraction of traffic, a few false alarms weigh heavily and precision looks modest, even for a strong model. Judge it against your actual prevalence.
Can I just maximize precision?
Not safely. Pushing precision up usually means flagging only the most obvious fraud and letting the rest through, which craters recall. You tune the two together against your loss appetite.
Why does precision matter to analysts?
Because it sets how much of their time is spent on real fraud versus false alarms. Low precision buries the team in noise and erodes trust in the alerts, which is a real operational cost.
Is a benchmark precision number useful?
Only if it was measured on a similar fraud prevalence. Precision is so sensitive to base rate that a benchmark from a different population can badly mislead you about your own performance.
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.