nathanaeldereje/mobile-money-credit-score

Behavior-based proxy credit scoring for unbanked users from mobile money transaction patterns, with a hybrid PaySim + heuristic-label approach calibrated to Ethiopian telebirr statistics. # Behavior-Based Proxy Credit Scoring for Mobile Money Users 💳 **Project:** mobile-money-credit-score **Role:** Data Scientist / ML Engineer **Status:** ✅ Complete ## Project Overview This project builds a **behavior-based proxy credit scoring model** for individuals who use mobile money platforms but have no formal credit history — a common gap in markets like Ethiopia, where mobile money adoption has outpaced traditional credit infrastructure. Using simulated mobile money transaction data (PaySim) combined with a self-defined, statistically-calibrated risk label, the project investigates whether early behavioral patterns — cash-flow volatility, cash-in cadence, counterparty diversity — can serve as a usable proxy signal for creditworthiness when bureau data doesn't exist. - The pipeline aggregates raw transaction logs into a per-user behavioral profile, applies a strict temporal split to prevent target leakage, and trains a classifier to predict risk from early-window behavior alone. - **Goal:** Demonstrate a defensible, explainable approach to alternative credit scoring — from raw transaction logs to a risk score with SHAP-based reasoning — under an honestly-scoped data limitation (real transaction patterns, simulated outcome labels). --- ## 📖 Problem Statement Ethiopia has made significant progress in financial inclusion, but many active mobile money users still have little or no formal credit history. Traditional lenders rely on bureau records that first-time borrowers and informal workers simply don't have — even though platforms like telebirr already capture rich behavioral signal (transaction frequency, cash-flow patterns, account usage) that could stand in for it. **The Problem:** Real, transaction-level Ethiopian mobile money data isn't publicly available. Furthermore, open-source sets like PaySim simulate individual anomalous transactions, lacking the continuous longitudinal user histories required for behavioral scoring. **The Approach:** Use P …