mmachelane/South-Africa-risk-model

Explainable credit risk model for South Africa's informal economy — LR + LightGBM + CatBoost stacking ensemble with SHAP explainability, NCA/POPIA/FICA compliance, and cost-optimised decisioning. AUC 0.82 | R71M+ annual value at 100k loans. # 🇿🇦 South African Credit Risk Model ### Explainable AI for Financial Inclusion in Emerging Markets --- ## 🎯 Project Overview **Business Problem:** South Africa faces a **financial inclusion crisis**. With 34% unemployment and 75% of the population having thin or no credit files, traditional credit scoring fails to serve the majority. This creates a R10+ billion opportunity for fintechs who can lend responsibly to the underserved. **Solution:** An **explainable credit risk model** that: - Achieves **82% AUC** using alternative data signals - Treats missing data as **information, not noise** (thin files are the norm) - Provides full **SHAP explainability** for NCA regulatory compliance - Generates **R71M+ annual value** at 100k loans/year scale - Handles SA-specific challenges: informal employment, no car ownership, sparse credit history **Why This Matters:** This isn't just a model—it's a **pathway to financial inclusion** for millions of South Africans locked out of traditional banking. --- ## 📊 Key Results | Metric | Target | Achieved | Business Impact | |--------|--------|----------|-----------------| | **AUC** | >0.80 | 0.82 | Strong risk discrimination | | **Precision** | >0.70 | 0.73 | Minimize false approvals | | **Recall** | >0.65 | 0.68 | Maximize good customers served | | **Annual Value** | — | R71M+ | At 100k loans/year | | **Explainability** | Full | ✅ SHAP | NCA compliant | --- ## 🇿🇦 South African Context ### The Financial Inclusion Gap ``` Population: 60M ├── Banked: 81% (49M) → But most have thin files ├── Unbanked: 19% (11M) └── Unemployment: 34% (20.4M) Credit Access Challenge: ├── 75% have THIN credit files (≤2 products) ├── 40% work informally (no formal employment records) ├── 60% don't own cars (use taxis/public transport) └── Traditional scoring FAILS for this majority ``` ### Regulatory Environment - **NCA (National Credit Act):** Requires affordability assessment and transparency - **POPIA (Protection of Personal Informatio …