blaise-fonguh/MOMOKASH-Loan-Scoring-model

A production-ready credit risk–scoring system built for MOMOKASH, a digital micro-lending platform in Cameroon. The project automates credit-limit assignment using an end-to-end data pipeline, behavioural feature engineering, and an unsupervised K-Means model (Silhouette ≈ 0.376). MOMOKASH Behavioural Loan Scoring Engine A production-grade, unsupervised behavioural credit scoring system developed for MOMOKASH, a micro-lending platform serving 20,000+ users across Cameroon. The system automates credit-limit assignment using historical behavioural data and machine learning, enabling consistent, fair, and scalable lending decisions. 🚀 Project Overview Traditional manual credit review slowed MOMOKASH’s loan approvals and exposed the platform to inconsistent decisions. This project solves that by: Building an end-to-end data integration pipeline (loans, refunds, penalties, debts – 3 years of history). Engineering a behavioural feature layer capturing real repayment behaviour. Training an unsupervised K-Means model with automated K-selection. Packaging the entire system into a deployable .pkl scoring engine. The model clusters borrowers into risk tiers and maps them into credit-limit bands (500–10,000 FCFA). 📊 Key Features 1. Data Engineering & Cleaning Consolidated multi-source data into a unified integrated_data.csv. Enforced a strict modelling window to avoid leakage (Sept 2022 — Sept 2025). Cleaned missing values, inconsistencies, and abnormal borrower histories. 2. Feature Engineering Constructed a behavioural feature layer including: Repayment rate Debt-to-limit ratio Borrowing frequency Penalty patterns Refund consistency Tenure and usage patterns 3. Unsupervised Risk Modelling Trained a K-Means clustering model. Evaluated clusters using Silhouette, Davies–Bouldin, and Calinski–Harabasz. 📌 Final Silhouette Score: 0.376 (Indicates strong behavioural separation for risk segmentation.) 4. Credit-Limit Mapping Risk clusters are mapped to limit bands: Cluster Risk Level Assigned Limit 0 High 500–2,000 FCFA 1 Medium 3,000–5,000 FCFA 2 Low 6,000–10,000 FCFA 5. Production Scoring Engine Packaged as: scoring_engine.pkl Includes: StandardScaler preprocessing K-Means model Mapping dictionary Predict → Assign Limit → E …