Felixaustine/african-credit-scoring-zindi

Credit default prediction for the Zindi African Credit Scoring Challenge using group-aware validation, feature engineering, CatBoost and LightGBM, with Kenya–Ghana domain shift analysis. # African Credit Scoring Challenge — Zindi Machine-learning experiments for the **July Study Jam Series: African Credit Scoring Challenge** on Zindi. The task is to predict whether a borrower will default on a loan, evaluated with **F1 score**. **Best public leaderboard F1: `0.800904977`** > This repository documents both the modelling work and the validation/transfer-learning lessons from a difficult cross-country credit-risk problem. The strongest local validation scores did not fully transfer to the leaderboard because the training data is Kenya-only while the test data also includes Ghana. ## Problem Financial institutions need robust default-risk estimates for both existing borrowers and new applicants. The competition asks for a binary prediction: - `1` — loan default - `0` — no default The metric is **F1 score**, making the precision/recall trade-off important under strong class imbalance. Competition: zindi.world ## Approach The project evolved from a CatBoost baseline into a group-aware, feature-rich modelling pipeline. ### 1. Leakage-aware validation Borrowers may occur in multiple rows and across multiple loans. I used **`StratifiedGroupKFold` grouped by `customer_id`** so that the same borrower does not appear in both training and validation folds. ### 2. Loan and customer sequence features Features include: - repayment ratio and implied interest rate - lender profit / lender repayment share - loan duration and calendar features - customer loan number and total number of loans - days since previous loan / until next loan - whether a loan is active at the customer's latest observed loan date - amount relative to the customer's historical amount profile - lender- and loan-type profile statistics ### 3. Robust preprocessing The later LightGBM experiment adds numeric winsorization to reduce the influence of extreme financial values. ### 4. Model families Experiment …