konare-dougoukolo/financial-inclusion-prediction

Comparative ML framework predicting financial inclusion in East Africa (23,524 records, 4 countries). Best ROC-AUC 0.8722 (Random Forest). SHAP interpretability. MSc dissertation empirical chapter. # Financial Inclusion Prediction — East Africa > A comparative machine-learning framework that predicts whether an individual > holds a bank account, and explains *why* using SHAP — the empirical core of my > MSc dissertation. ## Overview Using demographic and digital-access survey data from four East African countries, this project benchmarks four classifiers and then opens the "black box" with SHAP to identify the real drivers of financial inclusion. The work is framed around explicit research questions (RQ1–RQ4) covering discrimination performance, explanation consistency, subgroup fairness, and the interpretability-vs-performance trade-off. ## Context Empirical chapter of my **MSc Data Science & Business Analytics dissertation**. Financial inclusion is a key development lever; the goal was not only to predict who is unbanked, but to produce **policy-relevant, explainable** findings a decision-maker could act on. ## Data **Financial Inclusion in Africa** (Zindi) — survey respondents across **Kenya, Rwanda, Tanzania and Uganda**. | Property | Value | |---|---| | Records | **23,524** | | Features | 12 predictors (demographics, location, digital access) | | Target | `bank_account` (Yes / No) | | Class balance | **14.1%** banked vs **85.9%** unbanked (imbalanced) | The dataset is **not** committed — see `data/README.md` for the download link and setup. ## Approach 1. **Preprocessing.** A scikit-learn `ColumnTransformer` (one-hot encoding for categoricals, scaling for numerics) inside a `Pipeline`, with a stratified train/test split and a fixed `RANDOM_STATE = 42` for full reproducibility. 2. **Modelling.** Four classifiers — **Logistic Regression, Decision Tree, Random Forest, XGBoost** — evaluated on accuracy, precision, recall, F1 and ROC-AUC, with **5-fold stratified cross-validation** for stability. 3. **Robustness.** Threshold tuning, learning curves, and subgroup analysis (by gender and country) to check performance stability. 4. **Explainability.** …