Owoko-prog/loan-default-kenya

AI-powered credit risk prediction system for Kenya’s informal economy using alternative financial indicators and machine learning. # loan-default-kenya AI-powered credit risk prediction system for Kenya’s informal economy using alternative financial indicators and machine learning. # Predicting Loan Default in Kenya's Informal Economy ## Overview Binary classification model for a Nairobi-based MFI, predicting loan default risk using alternative data — mobile money behaviour, group lending membership, and repayment track record. ## Files - `loan_default_kenya.ipynb` — Full Jupyter notebook (all analysis + policy note) - `informal_loan_default.csv` — Dataset (place in the same folder as the notebook) ## How to Run ### 1. Install dependencies ```bash pip install numpy pandas matplotlib seaborn scikit-learn jupyter ``` ### 2. Launch ```bash jupyter notebook loan_default_kenya.ipynb ``` ### 3. Run all cells **Kernel → Restart & Run All** Make sure `informal_loan_default.csv` is in the **same folder** as the notebook. --- ## Notebook Structure | Section | Content | |---------|---------| | 0 | Imports & setup | | 1 | Load `informal_loan_default.csv` | | 2 | EDA — distributions, default rates by group | | 3 | Missing value analysis (`mpesa_avg_balance_ksh`) — MNAR pattern + imputation decision | | 4 | Feature engineering (4 new features with justification) | | 5 | Preprocessing pipeline (median imputation + StandardScaler) | | 6 | Model training: Logistic Regression, Random Forest, Gradient Boosting | | 7 | Evaluation — ROC-AUC, Gini, confusion matrix, feature importance | | 8 | Three-tier decision system: Approve / Human Review / Decline | | 9 | Policy note for lending manager | ## Key Design Decisions | Decision | Choice | Reason | |----------|--------|--------| | Missing value imputation | Median (not mean) | `mpesa_avg_balance_ksh` is right-skewed; mean would overestimate | | MNAR handling | Retain `balance_missing` flag | Small but real default-rate difference carries signal | | Primary metric | Recall | Missed defaults = principal loss; higher cost than false positives | | Decision th …