ColbertN/African-Credit-Scoring-Challenge

# African Credit Scoring Challenge Python workflow for training, scoring, credit score generation, feature importance, and monitoring. ## Setup ```bash pip install -r requirements.txt ``` ## Train ```bash python train.py ``` This trains an ensemble of LightGBM, XGBoost, and CatBoost with engineered financial, date, customer-history, lender, loan, and macroeconomic features. It also applies BorderlineSMOTE inside each CV fold, tunes the F1 threshold from out-of-fold probabilities, learns loan-level post-processing, and saves the trained bundle. Main artifacts: - `model/model_bundle.pkl` - full scoring bundle - `model/lightgbm_model.pkl` - legacy LightGBM model export - `model/preprocessor_pipeline.pkl` - legacy preprocessor export - `reports/cv_fold_metrics.csv` - fold metrics - `reports/postprocessing_results.csv` - loan-level post-processing comparison - `reports/segment_thresholds.csv` - learned country/loan-type thresholds - `reports/feature_importance.csv` and `reports/feature_importance.png` - `reports/training_summary.json` Current training summary: ```text OOF F1: 0.8963 Best threshold: 0.7158 Best loan post-processing: max probability per tbl_loan_id ``` Full training can take a while because it fits 4 CV folds plus final versions of all 3 boosted-tree models. ## Score Test Data ```bash python score.py ``` This reads `data/Test.csv`, loads `model/model_bundle.pkl`, and writes: - `submission.csv` - Zindi submission with `ID,Target` - `reports/scored_loans.csv` - probability of default, prediction, credit score, score band, and risk category - `reports/monitoring/data_drift.csv` - train-vs-score feature drift using PSI - `reports/monitoring/prediction_summary.csv` - default-rate and score distribution summary - `reports/monitoring/score_band_distribution.csv` Current scored test output has 18,594 rows and predicts 680 defaults. ## Credit Score Bands Scores are capped between 300 and 850: - `300-579`: Poor - `580-669`: Fair - `670-739`: Good …