🛡️ Alternative Credit Scoring System

Predicting Household Loan Default Risk Using Kenya's 2021 FinAccess Survey

Author: Paul Njoroge | Applied Mathematics & CS (Applied Maths Option) | Kenyatta University, Kenya

---

📌 Project Overview

Kenya has high mobile money penetration yet a significant portion of the population remains "thin-file" — lacking formal credit history. Traditional scoring systems exclude these individuals entirely, creating a massive financial inclusion gap.

This project builds an end-to-end alternative credit scoring engine using the 2021 FinAccess Household Survey (22,000+ households) conducted by KNBS and CBK. Instead of bank statements, it uses behavioral signals — mobile money usage, Chama participation, savings behavior — to predict the probability of loan default.

Metric	Value
Raw Survey Rows	22,024
Cleaned Records	19,476
Feature Variables	26
Models Compared	3

---

🎯 The Problem

• Despite 80%+ mobile money penetration, millions of Kenyans have no formal credit score
• Traditional lenders rely on bank statements and payslips — inaccessible to informal-sector workers
• Poor scoring leads to high Non-Performing Loans (NPLs) or blanket exclusion of creditworthy borrowers
• Core question: Can behavioral and demographic data replace traditional credit history?

---

⚙️ Technical Workflow

Phase 1 — Data Cleaning & Feature Engineering

Step	Detail
Raw Data	22,024 rows × 2,353 columns (FinAccess 2021)
Deduplication	Removed 2,548 duplicate rows
Missing Value Imputation	1,843 income values imputed using median
Target Engineering	is_high_risk: defaulted OR paid late OR delinquent
Encoding	One-hot encoding on gender, education, usage patterns
Final Dataset	19,476 rows × 27 columns → encoded_data.csv

Phase 2 — Exploratory Data Analysis

• Income Skew: Monthly income had extreme positive skew (7.55) — corrected with log(1+x) transformation
• Digital Signal: Daily/weekly mobile money users showed significantly lower default rates
• Social Capital: Chama participants demonstrated higher financial discipline and lower risk
• Age Pattern: Credit risk peaks at 20–35, stabilises with maturity
• Education: Strongest demographic predictor of credit stability across all models

Phase 3 — Modelling & Evaluation

Model	ROC-AUC	Role
Logistic Regression	Baseline	Interpretable baseline; suited for regulated lending
Random Forest	Benchmark	Captures complex feature interactions via Gini splits
XGBoost (Tuned) ✅	0.6681	Champion model — deployed to Streamlit app

---

📊 XGBoost Final Performance

Class	Precision	Recall	F1-Score	Support
Low Risk (0)	0.77	0.58	0.66	2,513
High Risk (1)	0.47	0.68	0.55	1,383
Weighted Avg	0.66	0.61	0.62	3,896

Key insight: The model captures 68% of actual high-risk borrowers (recall = 0.68) — the priority metric in credit risk, where missing a defaulter is more costly than a false positive.

---

🔑 Top Predictive Features (Cross-Model Consensus)

Feature	Logistic Reg.	Random Forest	XGBoost
Mobile Banking Frequency	⭐ Top	⭐ Top	⭐ Top
Savings Behaviour	⭐ Top	High	⭐ Top
Chama Participation	High	High	High
Log Income (B3I_log)	High	⭐ Top	High
Education Level	High	High	High
Age (A19)	Medium	⭐ Top	Medium

---

🛠️ Tech Stack

Layer	Tools
Modelling	Python · XGBoost · Scikit-Learn · Pandas · NumPy
Visualisation	Matplotlib · Seaborn
Deployment	Streamlit · Joblib
Data Source	KNBS FinAccess 2021 Household Survey

---

🚀 Live Demo

👉 Try the Alternative Credit Scoring App

Input demographic and behavioral data, get an instant High/Low risk prediction with feature contribution breakdown.

---

📁 Repository Structure

├── Data_cleaning_phase.ipynb       # Phase 1: Cleaning & feature engineering
├── Exploratory_Data_Analysis_phase.ipynb  # Phase 2: EDA & hypothesis testing
├── Modelling.ipynb                 # Phase 3: Model training & evaluation
├── Alternative_app.py              # Streamlit deployment app
├── xgb_credit_model.joblib         # Trained XGBoost model
├── scaler.joblib                   # StandardScaler for preprocessing
├── requirements.txt                # Dependencies

---

🔭 Future Roadmap

• Interaction Terms: Age × Mobile Banking — captures hidden household economic signals
• Wealth Index: Synthesise income, assets, and education into a single stability score
• SHAP Integration: Full explainability layer for regulated credit decisions
• Target: Push ROC-AUC beyond 0.70 through engineered features

---

📬 Contact

Paul Njoroge | larneymogul@gmail.com | Kenyatta University, Kenya