Model Card: Lakeview FCU Transaction Fraud Detection (LFCU-FRAUD-V2)

DEMO ARTIFACT - NOT FOR PRODUCTION USE

This model card describes a synthetic model created solely to demonstrate the Jozu platform (Jozu Hub and Jozu AgentGuard). The underlying company (Lakeview Federal Credit Union), training data, performance metrics, and model artifacts are fabricated. This model is not designed to detect fraud, has not been validated against real-world transaction data, and must not be deployed in any production environment, regulated workflow, or decisioning system. Use is restricted to Jozu demo scenarios, walkthroughs, and integration tests.

---

Model Details

Intended Use

Primary use case

Real-time scoring of card-not-present and card-present debit/credit transactions to flag likely fraudulent activity for downstream review. Scores are consumed by the transaction authorization service and the fraud operations queue.

Intended users

• Authorization decisioning service (automated)
• Fraud operations analysts (manual review queue)
• Member services representatives (context during member-initiated disputes)

Out-of-scope uses

• Credit underwriting or lending decisions
• Account opening / KYC adjudication
• BSA/AML transaction monitoring (handled by separate Verafin pipeline)
• Any use outside Lakeview FCU's transaction processing environment

Inputs and Outputs

Inputs

Feature vector of 87 fields derived from the transaction event and member profile:

• Transaction features: amount, MCC, merchant ID, channel (CP/CNP), entry mode, country code, time-of-day, weekend flag
• Member features: account tenure, average daily balance bucket, 30/90/365-day transaction velocity, typical merchant categories, geographic dispersion
• Device/session features (CNP only): device fingerprint match, IP ASN, IP geolocation distance from member home ZIP, prior device-account pairing count
• Network features: issuer-network risk score, BIN-level fraud rate (rolling 30 days)

PII fields (member name, full PAN, address) are excluded from the feature vector. Member ID is hashed before scoring.

Outputs

• fraud_score: float in [0.0, 1.0]
• decision: one of approve, step_up, decline, review
• top_features: ranked list of the 5 features with highest SHAP contribution to the score
• model_version: string

Training Data

Labeling

Fraud labels derived from confirmed chargebacks (Reg E disputes resolved in member's favor) with a 90-day maturation window. Suspected-but-unconfirmed cases are excluded from training.

Known data limitations

• No labeled examples of authorized push payment (APP) fraud or first-party fraud
• Underrepresentation of high-value wire and ACH fraud (separate models cover these)
• Geographic skew toward Lake County, IL and adjacent Cook County footprint

Evaluation

Holdout performance (synthetic test set)

These numbers are illustrative and generated against a synthetic test set. They do not reflect real-world performance.

Slice performance

Performance was measured across slices (channel, MCC group, member tenure bucket, transaction amount decile). Largest gap: recall on first-30-day-account transactions is ~0.71 vs. ~0.89 for established accounts. Mitigation: lower decline threshold and route to manual review for new-account transactions during the first 30 days.

Limitations and Risks

• Concept drift: Fraud patterns shift continuously. Performance degrades without retraining; current retraining cadence is monthly with weekly threshold recalibration.
• Adversarial pressure: Threat actors actively probe for thresholds. Outputs are not exposed to merchants, members, or external services.
• Population shift: Model was trained on the existing member base. Significant changes to field-of-membership or product mix may degrade performance.
• False positives: At operating threshold, ~1 in 60 declines is a false positive. Member friction is non-trivial and is monitored via the contact-center decline-inquiry rate.
• Disparate impact: Slice analysis includes proxies for protected classes (geography, age bucket via tenure). No statistically significant disparity was identified in synthetic testing, but ECOA-relevant monitoring is required if this model were ever to influence credit decisions. It does not.

Ethical Considerations

• The model influences member access to their own funds. False declines can cause real harm (declined transactions at point of sale, missed bill payments). Human review is required before any account-level action beyond a single transaction decline.
• Members are notified of declined transactions via push notification with a one-tap "this was me" path that immediately re-authorizes and feeds the response back as a label.
• No demographic features (race, gender, religion, national origin, marital status, age) are used as inputs. Geographic features are used and audited quarterly for proxy effects.

Governance and Compliance

• Model risk management aligned to NCUA supervisory guidance and SR 11-7 principles
• Validated by Lakeview FCU Internal Audit prior to release; next validation due 2026-08-14
• Reg E implications reviewed by Compliance (D. Park, BSA Officer)
• Model artifact, training data manifest, evaluation report, and SBOM signed and stored as a ModelKit in Jozu Hub
• Provenance, signatures, and policy attestations enforced at deployment by Jozu AgentGuard

Caveats and Recommendations

• Do not use this model card or artifact as a template for actual production model documentation without substantial additions (vendor model risk review, legal sign-off, member impact assessment).
• All numbers, names, and policies in this document are illustrative and exist to support Jozu platform demonstrations.
• Questions about the demo repo: contact the Jozu team. Questions about Lakeview FCU: there is no Lakeview FCU. It does not exist.

---

Model card maintained alongside the model artifact in the same ModelKit. Updates require a new ModelKit version and re-signing.