Fraud operations · alert budget

Fraud Alert Budget Explorer

A fraud team can only review so many transactions a day. Pick a review budget and see how much fraud each detector would catch in a test week of 324,037 transactions containing 637 frauds worth $416,079.

Source code on GitHub Loan Scorecard Studio Merchant Churn Study Treasury Cash Forecast AML Investigation Console

Share of transactions reviewed 0.50%

Detector	Fraud $ caught	Frauds caught	Precision	Net savings

What the comparison shows

Rules miss most card testing. A fixed threshold can't tell fraudsters' micro-charges from commuters' transit taps.
Isolation Forest needs no labels and beats the rules, a sensible first line when confirmed fraud is scarce.
Supervised gradient boosting ranks best even though only 1,011 of 1,545 training frauds were confirmed by the training cutoff (chargebacks arrive late).
The hybrid catches the most dollars. It adds the rule hits and anomaly score, and its training weights each fraud by amount, so the queue puts money at risk first.

How it was tested

Train on days 1–20, skip two days, test on days 23–30. Features look only at an account's past: transaction counts in the last 10 minutes, hour and 24 hours, spend against its own history, new devices, foreign IPs and bursts of small charges. Net savings subtract $4 of analyst time per reviewed alert.

All data is synthetic: three fraud patterns (account takeover, card testing, mule cash-out) plus legitimate behaviour that looks suspicious. Featurising and scoring the full 1.2M-transaction month takes about 20 seconds on a laptop. Built with AI-assisted coding (Claude); modeling choices, validation and review by Martins Ogiekpolor.