Skip to main content

Learn · Fraud Threshold

Fraud detection: the threshold trade-off

A fraud model gives each transaction a risk score. You still have to pick the line at which you block it. Slide that line and watch the trade-off every fincrime team lives with: catch more fraud and you flood analysts with false alarms; cut the false alarms and you let real fraud through. The maths below is real — only the distributions are illustrative.

Legitimate Fraud (scaled by count) Fraud shape, magnified Decision threshold

Illustrative distributions. Because fraud is rare, its true curve (magenta) is tiny next to the legitimate one — that is class imbalance. The dashed gold outline shows the fraud curve's shape magnified so you can see where it sits.

0.50

Block any transaction scoring at or above this line.

2.0%

Share of transactions that are actually fraud (0.5%–10%).

At this threshold

Precision
Recall
F1 score
False alarms

Confusion matrix

Out of an illustrative population of 10,000 transactions (200 actually fraud).

Predicted fraud
Predicted legit
Actually fraud
True positive
False negative (missed fraud)
Actually legit
False positive (false alarm)
True negative

The honest caveat: real transaction scores are never two neat bell curves, fraud patterns drift, and the "right" threshold depends on the cost of a missed case versus a blocked customer — a business and ethics decision, not just a maths one. That is exactly the kind of responsible-AI trade-off we work through with fincrime teams at bigspark. New to precision and recall? See the Learn shelf.