Card Fraud Detector

Scores every card authorisation for fraud in under 30 ms.

Strictly ConfidentialTier 1 modelFinancial crimeBanking secrecy+1High riskAnomaly Detection · 3.2.0 · card-fraud-detectorOwner Fraud Risk Management

Test in Sandbox

52 1.8M/30d 64

Fraud value detected at 0.5% alert rate

71%

Precision@top 1,000 daily alerts

0.42

False positive ratio

9:1

AUC-PR

0.61

1 more metric

p99 latency

27 ms

About this model

Scores every debit and credit card authorisation in real time for fraud risk, in under 30 ms, with reason codes such as new device plus high-risk merchant category, card-not-present velocity, or geography mismatch. Scores above threshold go to the fraud operations queue or trigger a step-up authentication prompt in the app.

Intended use

Real-time authorisation screening and case prioritisation by Group Financial Crime Compliance fraud operations. Any decline or block is governed by the fraud rules engine and reviewed by analysts; customers can dispute through the standard channel.

Training data lineage

Trained in Vertex AI on 18 months of authorisations from card-transactions-stream in BigQuery, labelled with confirmed-fraud-labels (chargebacks and customer-confirmed fraud). Streaming features computed with Dataflow and served from Vertex AI Feature Store in the Malaysia region. Validated by Model Risk Management as a Tier 1 model; champion-challenger against the incumbent rules runs continuously.

Limitations & bias notes

Label delay of up to 45 days means the most recent fraud patterns are under-represented at each retrain. First-party fraud and friendly fraud are out of scope. Performance on new card products is monitored separately for the first quarter after launch.

#fraud#cards#real-time#anomaly-detection#reason-codes#tier-1

Ownership and sensitivity

Owned by

Group Financial Crime Compliance Fraud Risk Management

Accountable owner: Chloe Tan

Who approves access

Owner, Fraud Risk Management, Chloe Tan
Data Governance Office, Ahmad Faizal
Model Risk Management, Dr. Kavitha Subramaniam

In Fight financial crime

Strictly Confidential

Customer-level data under banking secrecy and PDPA, and models that decide on customers.

Named-user entitlement with the data owner's consent and a stated purpose; reviewed on expiry.

Tier 1 model: Material model: independent validation before use and annual revalidation.
Financial crime: Used in AML, fraud or sanctions controls; tipping-off restrictions apply.
Banking secrecy: Customer account information protected under the Financial Services Act 2013.
Customer PII: Contains or processes personal data about customers (PDPA 2010).

Try it

Live sandbox

auth=AUTH-7731-0482&mcc=5732&cnp=true

fixture

Evaluation only: restricted models run against synthetic or masked sample data until your business unit's access request is approved.

Feedback

Details

Updated: 2026-07-21
Latest version: 3.2.0
Licence: BU Restricted
Access: Restricted
Framework: XGBoost
Language: N/A

Trained on

card-transactions-stream →confirmed-fraud-labels →