buumba641/July-Study-Jam-Series-African-Credit-Scoring

machine learning model that predicts the probability of loan default. # African Credit Scoring Challenge — Loan Default Prediction **🏆 Rank: 10 / 78  |  Public F1: 0.800  |  Private F1: 0.775** Zindi · July Study Jam Series ·20 July 2026 - 9 August 2026 --- ## Overview This repository contains the full solution for the Zindi African Credit Scoring Challenge, where the goal was to predict the likelihood of a customer defaulting on a loan using financial and macroeconomic data from Kenya and Ghana. The core challenge beyond prediction accuracy was **cross-country generalisation**: models were trained exclusively on Kenyan data yet evaluated on both Kenyan and Ghanaian test records — requiring features and logic that hold across different financial markets. Top-10 finishers were additionally required to submit a **credit scoring function** that translates model outputs into an interpretable 300–850 credit score. --- ## Results | Split | F1 Score | |---|---| | Public leaderboard | 0.8000 | | Private leaderboard | 0.7747 | | Cross-validation (mean) | ~0.79 | --- ## Approach ### 1. Feature Engineering - Date-derived features: loan term in days, disbursement/due weekday, month and year - Financial ratios: repayment ratio, daily repayment amount - Customer-level aggregates: mean and median `Total_Amount_to_Repay` per customer - Log transforms on all monetary columns to reduce skew - Outlier capping at the 90th percentile for `Total_Amount` and `Total_Amount_to_Repay` - FRED macroeconomic indicators (inflation, exchange rate, interest rates, unemployment) as country-level context features ### 2. Handling Class Imbalance - **BorderlineSMOTE** (`sampling_strategy=0.45`) applied inside each training fold only — generates synthetic minority samples near the decision boundary - **`scale_pos_weight`** passed to each booster to further penalise false negatives - **StratifiedKFold** (4 folds) to preserve the class ratio across every fold ### 3. Ensemble Model A majority-vote ensemble of three gradient-boosted classif …