cnongera/mpesa-ml-analysis

Machine learning analysis of M-Pesa transaction patterns using K-Means clustering, PCA, and Isolation Forest # M-PESA Transaction Pattern Analysis ## ✨ Shukran. Thank You for taking an interest in this project. Looking forward to building impactful projects with You! # M-Pesa Transaction Pattern Analysis **Author:** Clement Ongera Nyangoya **Date:** April 2026 **Tools:** Python, Pandas, scikit-learn, Matplotlib ## Overview Reproducible machine learning analysis of mobile money transaction patterns from M-Pesa, Kenya's dominant mobile payment platform. This project applies unsupervised learning (K-Means clustering, Isolation Forest anomaly detection) and time-series analysis to identify spending behaviors, detect anomalies, and characterize financial patterns. ## Methods - **Data cleaning and feature engineering:** Temporal features, transaction categorization via rule-based NLP, behavioral indicators - **Clustering:** K-Means with PCA visualization to identify distinct transaction behavior segments - **Anomaly detection:** Isolation Forest for unusual transaction identification - **Time-series analysis:** Daily aggregation with moving averages and volatility metrics ## Key Findings ### 🎯 3 Distinct Customer Segments Identified Using K-Means clustering on 1,215 transactions: | Cluster | Size | Behavior | Avg Amount | Peak Time | |---------|------|----------|------------|-----------| | **Cluster 0** | 517 (43%) | "Social Senders" - Person-to-person transfers | -65 KES | 2:00 PM | | **Cluster 1** | 336 (28%) | "Morning Shoppers" - Merchant payments | -63 KES | 9:00 AM | | **Cluster 2** | 362 (30%) | "Night Owls" - Small evening purchases | -26 KES | 8:00 PM | **Key insight:** All three segments are net spenders, with "Night Owls" making the smallest but most frequent evening purchases. ### 🔍 Anomaly Detection Results - **60 anomalies detected** (4.9% of transactions) - **Typical anomaly amount:** 800+ KES (10-30x normal) - **Most anomalous:** Person transfers (53%) and merchant payments (47%) - **Isolation Forest precision:** Estimated 95% ### 📊 Transaction Catego …