Fraud operations cockpit

ML scoring converted into a ranked fraud review queue.

A portfolio dashboard for monitoring payment fraud exposure, model quality, threshold trade-offs, and analyst actions.

Critical hold Manual review Auto approve
Selected model random_forest_balanced

Prioritizes fraud recall while keeping the analyst queue operational.

PR-AUC-
Recall-
False negatives-
Critical exposure -

Fraud amount concentrated in transactions recommended for temporary hold or block.

Critical transactions -

Rows assigned to the highest ML risk tier.

Fraud rate -

Dataset-level fraud rarity proves why accuracy is misleading.

Total transactions -

End-to-end data package size after profiling and scoring.

Risk tier exposure

Critical tier carries the business loss concentration.

Action mapped

Operational actions

Risk tiers become fraud team decisions.

Model monitoring

Performance is evaluated with imbalanced-class metrics, not accuracy.

PR-AUC first

Selected model at threshold 0.50

True positive-
False positive-
False negative-
F1 score-

Fraud detection needs recall and queue quality because missed fraud is more expensive than normal classification error.

Threshold simulator

Precision, recall, and queue size trade-off.

SQL investigation

Transaction type risk benchmark.

Rolling fraud trend

Step-level fraud rate and exposure movement from SQL marts.

Window function mart

Analyst review queue

Top scored transactions ready for investigation.

Risk Action Transaction Type Amount Origin Destination

Database analytics proof

SQL marts turn raw transactions into BI layers.

Rule queue sample

SQL rules compared with ML scoring.

Methodology and limitations

Built as a portfolio-grade fraud intelligence system.

  • PaySim is simulated mobile money data, so the dashboard demonstrates fraud analytics workflow rather than live bank production data.
  • The model comparison uses PR-AUC, ROC-AUC, precision, recall, F1, and threshold simulation because fraud is highly imbalanced.
  • The SQL layer exposes CTEs, CASE risk tiers, rolling windows, account-level aggregation, and BI-ready marts.
  • A production version would add drift monitoring, audit trails, explainability review, streaming latency checks, and analyst case status tracking.