An independent fairness and performance audit of an XGBoost fraud detection model trained on 6.3 million mobile money transactions, evaluated against BoG CISD 2026, NIST AI RMF 1.0, and the EU AI Act.
The model achieves 99.27% overall fraud recall but provides significantly weaker protection to low-balance users. Low-Balance TPR = 0.6667 versus High-Balance TPR = 0.9972. Miss rate disparity: 33.3% vs 0.3%, a 110-fold gap. Equalized Odds Difference = 0.3305, exceeding the 0.10 regulatory threshold by three times. BoG CISD 2026 Annexure E §l(i) mandatory suspension triggered.
| Section | Description | Standard | Outcome | Artifact |
|---|---|---|---|---|
| Scope and Regulatory Alignment | Audit objectives, proxy variable design, regulatory framework mapping | BoG CISD 2026 Annexure E §e(i)(3) | COMPLETE | View |
| Environment Setup and Thresholds | Library setup, pass/fail thresholds defined before any results seen | BoG CISD 2026 §122(1) | COMPLETE | View |
| Data Audit and Integrity Checks | Missing values, duplicates, leakage detection, column exclusions documented | BoG CISD 2026 Annexure E §f(i) | CLEAN | View |
| Pre-Training Proxy Detection | Chi-square and Cramer's V confirming structural fraud imbalance before training | BoG CISD 2026 Annexure E §g(iii)(3) | IMBALANCE CONFIRMED | View |
| Exploratory Data Analysis | Transaction type analysis, temporal patterns, amount distributions, zero-balance logic, correlation heatmap, Invisible 41 discovery | NIST AI RMF MAP 1.1 | 41 LOW-BALANCE FRAUD CASES | View |
| Feature Engineering Audit | 13 features engineered and documented. Implicit wealth correlation flagged on balance-derived features. | BoG CISD 2026 Annexure E §j(ii) | RISK DOCUMENTED | View |
| Section | Description | Standard | Outcome | Artifact |
|---|---|---|---|---|
| Model Training Audit | XGBoost 300 trees, SMOTE applied. 164:1 High-Balance to Low-Balance fraud ratio in training confirmed. SMOTE limitation documented. | BoG CISD 2026 Annexure E §g(iii)(1) | STRUCTURAL GAP CONFIRMED | View |
| Section | Description | Standard | Outcome | Artifact |
|---|---|---|---|---|
| Performance Evaluation | MCC 0.44 FAIL. PR-AUC 0.9357 PASS. KS 0.9916 PASS. Log Loss 0.0135 PASS. Per-group MCC confirms Low-Balance underperformance. | NIST AI RMF MEASURE 2.3 | MCC FAIL · 3 of 4 PASS | View |
| Fairness and Bias Audit | Transaction type PASS. Balance tier EOD = 0.3305, three times over threshold. Low-Balance TPR = 0.6667 vs High-Balance 0.9972. | BoG CISD 2026 Annexure E §l(i) | HIGH RISK · SUSPENSION TRIGGERED | View |
| Error Analysis | Low-Balance miss rate 33.3% vs High-Balance 0.3%. 110-fold disparity confirmed. | NIST AI RMF MEASURE 2.11 | STRUCTURAL PROTECTION GAP | View |
| Explainability | SHAP analysis. Top features: newbalanceOrig, oldbalanceOrg, amount. Wealth correlation in balance features confirmed. | BoG CISD 2026 Annexure E §j(i) | RISK DOCUMENTED | View |
| Deployment Readiness Assessment | Threshold sensitivity analysis. 8 of 8 deployment prerequisites unmet under BoG CISD 2026. | BoG CISD 2026 §100 | NOT DEPLOYMENT READY | View |
| Monitoring and Drift Plan | Per-group monitoring framework, retraining triggers, and drift detection requirements documented. | BoG CISD 2026 §100(6) | FRAMEWORK DOCUMENTED | View |
| Final Scorecard and Risk Register | 5 of 8 metrics pass. EOD balance tier FAIL. Model not deployment-ready. BoG notification required. | BoG CISD 2026 §115(2)(b) | DO NOT DEPLOY | View |