Real-time card fraud scoring

Context

A large private-sector bank serving tens of millions of customers processes millions of card transactions every day across ATMs, merchant terminals and online channels. Its existing rules engine was tuned for yesterday's attack patterns.

Challenge

Rule-based fraud detection was missing attacks while flooding analysts with false alerts.

What we deployed

Real-time machine-learning fraud scoring across ATM, POS and card-not-present channels.

Typical architecture

1. STEP 01

   Card switch streams authorisation events

   →
2. STEP 02

   QSFS scores each transaction in ~15 ms

   →
3. STEP 03

   Ranked alerts routed to the fraud operations queue

   →
4. STEP 04

   Analyst decisions feed back into monitoring and retraining

Illustrative of how this solution is typically structured and governed; deployment specifics are available under NDA.

Results

• +39% ATM fraud caughtPartner-reported
• −66% false alerts on card-not-presentPartner-reported
• ~15 ms per transaction, ~6M txns/dayClient-validated
• POS fraud caught +18%; card-not-present fraud caught +20%Partner-reported
• False alerts −39% on POS and −28% on ATMPartner-reported
• ~300,000 fewer false alerts per monthClient-validated
• Asian Banker Financial Technology Innovation Award 2024Partner-reported

NoteResults from deployments by GoGenesis and its platform partners. Outcomes vary by portfolio and operating environment; your results are set in the pilot.

Governance approach

• Every score carries readable reason codes for analysts
• Analysts approve, override and record each action
• False-positive rate and drift tracked and reported

Illustrative of how this solution is typically structured and governed; deployment specifics are available under NDA.