Mobile Money Fraud & Leakage Assurance System

Project Overview

This project delivers a high-recall machine learning system to detect fraud (SIM Swap/Account Takeover, Agent Collusion) and financial leakage (system errors) within the mobile money platform. The primary goal was to minimize False Negatives (missed fraud), creating a critical safety net for the Assurance department.

• Model Used: Class-Weighted Logistic Regression (Chosen for interpretability).
• Key Result (Recall): $\approx 99%$. The model successfully catches virtually all fraud cases, minimizing financial loss.
• Performance (AUC-ROC): $0.9805$ (Excellent predictive power).

---

Key Features & Actionable Insights

The model's success is driven by custom engineered features and the interpretation of their risk impact (Odds Ratios).

Engineered Risk Factors:

1. Txn_Count_Sender_1H: (Velocity/SIM Swap Risk) Counts transactions by the Sender ID in the last 60 minutes.
2. Agent_Cust_Pair_Count_7D: (Collusion Risk) Tracks repeated transactions between the same Agent/Customer pair.

Top Strategic Recommendations:

The model provides clear rules for where to focus security controls:

1. Critical Risk: P2P Transfers

   • Impact: Odds of fraud are $\mathbf{13,249}$ times higher.
   • Action: Implement mandatory 2FA/step-up authentication for high-value P2P transactions.

2. Engineering Priority: System Leakage

   • Risk Factor: Billing_Status_TIMEOUT
   • Impact: $\mathbf{51.5%}$ increase in risk.
   • Action: Launch urgent investigation to stabilize the billing system, as timeouts are a quantifiable source of leakage.

3. Value-Based Control

   • Risk Factor: Amount_ETB
   • Impact: Odds increase $\mathbf{4.2}$ times for every standard unit increase.
   • Action: Institute dynamic velocity limits based on transaction value.

---

Project Files

• Trained Model: logistic_regression_fraud_model.joblib
• Fitted Scaler: standard_scaler.jobjob (Essential for scoring new data)
• Analysis: final_odds_ratio_analysis.csv (The final list of ranked risk drivers)
• Documentation: [notebooks] containing all EDA, feature engineering, and training steps.

Operational Strategy

The model is optimized for detection over efficiency. All transactions scoring $\mathbf{\geq 0.5}$ must be routed to a Manual Analyst Review Queue. Automatic blocking is not recommended due to the high False Positive rate.