Byabato/financial-inclusion-africa-ml-zindi

End-to-End ML Pipeline & Policy Intervention Simulator for Financial Inclusion in East Africa. Features Stacking Ensembles, Optuna Tuning, and SHAP-driven Explainable AI to map socio-economic barriers to SDG-aligned solutions. # 🌍 Financial Inclusion in Africa — ML Pipeline & Policy Simulator Kelvin Byabato  ·  --- > **86% of East Africans have no bank account.** This pipeline predicts who is excluded, explains why using SHAP, and recommends SDG-aligned interventions — not just a leaderboard score. --- ## 📋 Table of Contents - Problem Statement - Results - Architecture - Feature Engineering - Training Strategy - Explainability Layer - Innovation: Policy Simulator - Setup & Reproduction - File Reference - References --- ## 🎯 Problem Statement **Competition:** Zindi Africa — Financial Inclusion in Africa **Task:** Binary classification — predict bank account ownership (Yes=1, No=0) **Metric:** Mean Absolute Error (MAE) on hard 0/1 predictions **Data:** ~33,600 survey respondents, Kenya · Rwanda · Tanzania · Uganda · 2016–2018 **Class imbalance:** 14% positive (banked) · 86% negative (unbanked) · 6:1 ratio **Why MAE matters here:** MAE on binary labels equals the fraction of misclassified predictions. Minimizing MAE = maximizing accuracy. With class imbalance, the optimal threshold is NOT 0.5 — threshold optimization is critical. --- ## 🏆 Results | Model | OOF MAE | OOF AUC | Notes | |---|---|---|---| | Logistic Regression | ~0.170 | ~0.750 | Baseline | | XGBoost (default) | ~0.130 | ~0.840 | +24% vs baseline | | LightGBM (default) | ~0.130 | ~0.840 | Speed advantage | | CatBoost (default) | ~0.130 | ~0.840 | Native categoricals | | XGBoost (Optuna-tuned) | ~0.120 | ~0.855 | 50-trial Bayesian search | | **Stacking Ensemble** | **0.1117** | **0.8647** | **Final submission** | All scores are **Out-of-Fold** — evaluated only on held-out data, never on training samples. --- ## 🏗️ Architecture ``` financial-inclusion-africa-ml-zindi/ │ ├── src/ ← Library (import, don't modify directly) │ ├── __init__.py │ ├── config.py — Paths, params, encoding maps (change here only) │ ├── features.p …