FinPulse-ML: Multi-Rail Payment Fraud & Graph Velocity AI

Calibrated GBDT + Isolation Forest stacking engine cutting multi-rail payment & stablecoin fraud false positives by 92.5% (0.9898 ROC-AUC) with real-time SHAP explainability.

Comment

Inspiration & Problem Statement

Instant payment rails (FedNow_Instant, SEPA_Instant, USDC_Solana_CCTP, Visa_Direct_CrossBorder, and SWIFT_gpi) settle funds irrevocably in seconds. Yet traditional static rule-based screening systems suffer from 12–18% False Positive Rates on legitimate high-value B2B and cross-border settlements, while failing to catch coordinated Threshold Smurfing ($2,800–$2,960 transfers structured just beneath $3,000 Travel Rule limits), Mule Network Fan-In PageRank Bursts, and Account-Takeover Velocity Spikes.

Data Science & Feature Engineering Methodology

We curated and benchmarked a 10,000-transaction multi-rail Fintech settlement dataset ($30.80M volume across 5 rails and 8 global corridors) with 11 engineered domain features:

  1. Rolling 1h & 24h Velocity Z-Scores (velocity_1h_zscore, velocity_24h_zscore): Normalized sender burst intensity.
  2. Counterparty PageRank / In-Degree Burst (counterparty_pagerank_burst): Graph topological fan-in anomaly score capturing mule aggregation accounts.
  3. Benford's Law First-Digit KL Divergence (benford_divergence): Measures leading-digit divergence from $P(d) = \log_{10}(1 + 1/d)$ to detect human-fabricated smurfing numbers.
  4. AML Threshold Structuring Proximity (structuring_proximity): Exponential closeness to $3,000 and $10,000 reporting thresholds.
  5. Merchant Category Code Shannon Entropy (mcc_entropy), Cross-Border Corridor Risk (corridor_risk_score), Device/ASN Geo-Mismatch (device_ip_mismatch), and Cyclic UTC Time Encoding (tod_sin, tod_cos).

Ensemble Modeling & Empirical Results

We implemented and benchmarked 4 models on a chronological 75%/25% train/test split (7,500 train / 2,500 held-out test transactions):

  • Legacy Static Rule Engine (Baseline): ROC-AUC = 0.7805 | PR-AUC = 0.2517 | F1 = 0.3103 | FPR = 15.73% | Brier = 0.1183
  • Unsupervised Isolation Forest (65 Trees): ROC-AUC = 0.9455 | PR-AUC = 0.6926 | F1 = 0.6726 | Brier = 0.0671
  • Gradient Boosted Decision Stumps (Newton-Raphson Boosting): ROC-AUC = 0.9835 | PR-AUC = 0.8292 | F1 = 0.7623 | FPR = 1.52% | Brier = 0.0232
  • FinPulse-ML Calibrated Stacking Ensemble (Ours): ROC-AUC = 0.9898 | PR-AUC = 0.8860 | F1 = 0.8226 | FPR = 1.18% (-92.5% vs Rules) | Brier = 0.0146

Interactive Decision Intelligence Web App

Our live zero-install web dashboard (https://jeffrey-graham.github.io/datanova-finpulse-ml/) includes:

  • Interactive ROC & Precision-Recall curves and Multi-Rail Cohort telemetry tables.
  • Global SHAP / permutation feature importance rankings.
  • A Real-Time Transaction Risk & Local SHAP Waterfall Simulator with 5 one-click preset scenarios allowing judges to test smurfing bursts, mule fan-in rings, and $18,400 corporate treasury wires side-by-side against legacy static rules.

Built With

Share this project:

Updates

Submission history