Evans-Ataaya/ghana-credit-risk-scorer
Domain:
socioeconomicRecord type:
software
Creator:
Eva
Host:
ML-powered credit risk assessment tool for Ghana's microfinance sector | XGBoost + SHAP Explainability | Streamlit App
# Ghana Credit Risk Scorer
## Overview
A machine learning-powered credit risk assessment tool built
specifically for Ghana's microfinance and retail lending sector.
This project predicts the probability of loan default using
XGBoost with full SHAP explainability, deployed as an interactive
Streamlit web application.
## Live Demo
Run locally with:
```bash
streamlit run app.py
```
## Project Structure
```
ghana-credit-risk-scorer/
├── README.md
├── requirements.txt
├── ghana_credit_risk_scorer.ipynb
├── app.py
├── data/
│ ├── raw/
│ │ └── loan_data.csv
│ └── processed/
├── models/
│ ├── logistic_regression.pkl
│ ├── random_forest.pkl
│ ├── xgboost.pkl
│ ├── xgboost_tuned.pkl
│ └── scaler.pkl
└── outputs/
└── figures/
```
## Model Performance
| Model | AUC-ROC | Precision | Recall | F1-Score |
|---|---|---|---|---|
| Logistic Regression | 0.6977 | 0.2149 | **0.6578** | **0.3240** |
| Random Forest | 0.6786 | 0.2104 | 0.3688 | 0.2680 |
| XGBoost (original) | 0.6891 | 0.4808 | 0.0951 | 0.1587 |
| XGBoost (tuned) | **0.7037** | 0.3478 | 0.2738 | 0.3064 |
## Top Default Risk Factors (SHAP)
| Rank | Feature | Mean SHAP |
|---|---|---|
| 1 | num_open_accounts | 0.465 |
| 2 | loan_tenure_months | 0.403 |
| 3 | loan_type | 0.358 |
| 4 | income | 0.355 |
| 5 | age | 0.276 |
## Quick Reference Guide
For non-technical users and loan officers, see the
Borrower Risk Assessment Guide
for plain-English explanation of risk tiers, borrower
profiles, and how to interpret app decisions.
## Key Findings
- Number of concurrent open accounts is the strongest
default predictor — reflecting over-indebtedness in
Ghana's fragmented microfinance sector
- Loan tenure and loan type are the second and third
most important features
- Traditional delinquency metrics rank surprisingly low
- Credit utilisation ratio — dominant in Western scoring
systems — is the weakest predictor in the Ghana context
## Tech Stack
- Python 3.10
- XGBoost 1.7.6
- SHAP 0.43.0
- S …