Data product · Credit & risk · real-time

Fraud / Transaction-Risk Score

A governed, reusable real-time risk score returned inside the auth window - published as a data product with contract, SLA, entitlements, and SR 26-2 monitoring.

0-99 · ~400 attrs · <1 msvendor-stated

Visa Advanced Authorization / Visa Protect score, trained on billions of VisaNet txns (269B+/yr)

3.2B · 98.83% · ~$33Bvendor-stated

Visa Decision Manager 2023: txns screened, auto-resolved, fraud prevented

>1B txns/dayreported

JPMorgan Chase across 100+ countries (reported, secondary)

−50% FP · +25% detection · ~$250Mreported

JPMC AI fraud models: false positives, detection, annual savings (reported, secondary)

The problem

A real-time Transaction Risk Score must return inside the ~100-200 ms auth window. Class imbalance makes raw accuracy misleading, and chargeback labels land weeks late.

The data-product gap

Without a governed, reusable score product, fraud / auth / disputes / collections each rebuild bespoke models - no contract, SLA, entitlement, audit, or drift monitoring.

The solution, on the platform

1
Inputs: auth message, card/account context, device/channel CNP signals, behavioral store, consortium intel, delayed labels.
2
Features: velocity/aggregation + deviation + graph features in an online feature store (single-digit-ms reads, online/offline parity).
3
Model: tree-ensemble core + sequence/transformer for behavior; two-stage (light filter on every txn, heavy ensemble on flagged).
4
Serving: Kafka → Flink streaming; published as a data product with contract + SLA + row/column entitlements.
5
Governance: every score audited; model monitoring under SR 26-2 (PSI, FPR, recall). Reuse across auth / step-up / disputes / collections / credit-line is the value multiplier.

Online feature storeTree ensemble + transformerTwo-stage scoringKafka → Flink servingSR 26-2 drift monitoring

Caveat

All four JPMC figures are reported / secondary coverage, not a primary Chase disclosure. Vendor figures are vendor-stated. Latency / availability / reuse targets are illustrative.

Success scorecard

Adoption / reuse

›Reuse ratio ≥4 domains (illustrative)
›Time-to-first-score <1 sprint

Time-to-value / SLA

›Latency p99 ≤50-100 ms (illustrative)
›Availability 99.95% (illustrative)
›Feature freshness ≤200 ms

Quality / governance

›PR-AUC + KS tracked
›PSI <0.10 stable, >0.25 retrain
›Auto-resolution ~98% (Visa benchmark)

Business impact

›FPR −50% / detection +25% (JPMC-reported)
›$ net fraud loss prevented

Browse the catalog How the AI analyst is governed

Sources and method

Live · enforced in the database Public · cited Synthetic · illustrative

01
Visa Advanced Authorization / Visa Protect - score 0-99, ~400 attributes, <1 ms, trained on billions of VisaNet txns (269B+/yr) - vendor-stated.
02
Visa Decision Manager 2023 - 3.2B txns screened, 98.83% auto-resolved, ~$33B fraud prevented - vendor-stated (corporate.visa.com).
03
Mastercard Decision Intelligence - real-time, network-agnostic, graph + gen-AI - vendor-stated.
04
JPMorgan Chase fraud AI - >1B txns/day across 100+ countries;

Synthetic cardholder dataFictional partnerIllustrative model metricsCited market context