Transaction monitoring: catching fraud inside an alert budget

A fraud model judged the way a monitoring team actually uses it. Built for Standard Chartered's PS-II internship.

47 of 50alerts are real fraud when the team reviews only the top 50 of the day
62%of the day's fraud caught within 200 alerts, out of 140,021 transactions
0frauds the 'flag the largest amounts' rule finds in its first 200 alerts
284,807real transactions; frauds are 0.17% of them

Try it: how big should the alert budget be?

Measured on the second day, which the model never saw. Review cost is an input, not data.

Model vs the rule a bank starts with

Alerts reviewedFrauds caughtAlerts that are real fraudShare of the day's fraud"Largest amounts" rule
504794%22%0
1008383%39%0
20013166%62%0
40015639%74%1
80016320%77%2
160016510%78%5

Validation checks

Charts: fraud caught, precision, calibration, drift

Built by Pavitra Seth (BITS Pilani). Public data used as a stand-in; not affiliated with the company. Code and method: see the repository README.