PAYMENT SIGNALS
One payment contains many signals.
CUSTOMER · MERCHANT · DEVICE · VELOCITY · LOCATION
PAYMENT RISK / CASE STUDY 01
CARDSHIELD
Fraud intelligence without treating every customer like a fraudster.
A bank faces two risks at once: fraud costs money, while blocking genuine payments creates customer friction. CardShield explores how machine learning and decision policy can balance both.
SYNTHETIC DATASETNOT LIVE BANK TRANSACTIONS
EXPLORE THE SYSTEM ↓Built on synthetic transaction data designed to simulate a realistic fraud workflow.
AN OFFLINE FRAUD DECISION STUDY 01 — 04
02 / HOW THE DECISION WORKS
One payment.
Four layers of judgement.
One payment carries signals. A model estimates risk. A separate policy chooses what happens.
FRAUD INTELLIGENCE
The model looks for combinations that resemble fraud.
RISK
The model estimates how suspicious the payment looks.
Calibration makes that score easier to interpret.DECISION POLICY
The risk signal enters a separate policy. MODEL ≠ POLICY.
APPROVE 89.66% · VERIFY 10.00% · REVIEW 0.32% · DECLINE 0.03%MODEL ≠ POLICY
The model estimates risk. A separate, cost-aware policy chooses whether to approve, verify, review or decline.
03 / OFFLINE EXPERIMENT
WHAT DID THE SYSTEM ACHIEVE?
Compared with a consistent baseline using held-out synthetic transactions.
~19.77%SIMULATED FRAUD CAPTURE
~20.9%SIMULATED COST REDUCTION
~£906KSIMULATED NET SAVINGS
CONTROLBASELINE POLICY
TREATMENTLOWER EXPECTED COST
−£5.95 / transactionSIMULATED TREATMENT–CONTROL COST DIFFERENCE
150,886HELD-OUT TRANSACTIONS
PASSGUARDRAILS
PROJECT EVIDENCE
Power BI reporting
04 / POTENTIAL APPLICATION
WHERE COULD THIS FIT IN A BANK?
In a real banking environment, a system like CardShield could sit behind card authorization and help prioritize which transactions need intervention.
CARD PAYMENT→RISK ESTIMATE→POLICY→ACTION→MONITORING
POTENTIAL REAL-WORLD APPLICATIONNOT A LIVE BANK DEPLOYMENT