📱 M-Pesa Transaction Pattern Analysis

End-to-end data science project analysing behavioural patterns, user segmentation, and fraud signals in mobile money transactions.

📊 Analysis Report →

Analysis Pipeline

1. Data Generation — 150K realistic synthetic transactions with diurnal activity models, valid Safaricom phone formats, and Safaricom fee schedules
2. EDA — Temporal, geographic, and distribution analysis (10 figures)
3. Feature Engineering — 30+ behavioural features per user
4. User Segmentation — K-Means clustering with PCA visualisation
5. Anomaly Detection — Isolation Forest + supervised Random Forest

Key Results

• 4 behaviourally distinct user segments identified
• Fraud detection AUC: 0.87+ (5-fold CV)
• Top fraud signals: is_just_below_threshold, pct_night_txns, amount_cv
• Salary week generates +18% transaction volume vs daily average

Skills Demonstrated

Data Simulation EDA Feature Engineering Clustering Anomaly Detection
Scikit-Learn Imbalanced Learning Statistical Visualisation