Card payments · 80,000 transactions over 60 days · synthetic data

Fraud overview

Loading the latest model results…

Fraud caught by typology

Final period, share of fraud flagged at the production threshold. A new pattern appears from day 45.

Static model, trained before day 39Adaptive model, retrained at day 45

Next to review

View all

Highest-risk payments the adaptive model blocked

    Precision against recall

    Adaptive model, final period. Plain accuracy would read 98%+ by approving everything.

    What drives the score

    Drop in PR-AUC when the feature is shuffled

    Final-period sample · 200 payments

    Alert queue

    Every payment both models scored, with your review verdicts. Replay shows them arriving in order.

    PaymentAmountMerchantChannelCountryStaticAdaptiveDecisionVerdictAction
    Adaptive model · explains every decision

    Score a payment

    Enter a payment or load a case to see the decision and what drove it.

    Payment

    Decision

    Final period · adaptive model · review costs $5 per alert

    Threshold and cost

    Operating point

    Slide, or use the arrow keys, to trade missed fraud against false alarms.

    Total cost by threshold

    Missed fraud plus review cost, in dollars. The blue mark is your threshold.

    Drift-onset period · static model at production threshold

    Fairness

    False-positive rate is the customer-friction metric. Gaps between groups are a flag. Synthetic data: structure only, not a claim about real populations.

    SliceGroupPaymentsFalse-positive rateRecall

    Responsible-use notes

    No personal identifiers or protected attributes are features. Every block has a human-readable reason. Borderline scores go to manual review, not auto-decline. Labels arrive late in reality, so the retraining loop needs monitoring for label bias. Results are on synthetic data and overstate real-world separability.