Expera-Akakpo/Data-Tour-2025

This repo contains code and methodology for DataTour 2025's credit default prediction challenge. LightGBM binary classification on 17M rows, 62 features; outlier clipping, 5-fold stratified CV, manual hyperparam tuning. OOF ROC AUC: 0.664281. Team AST submission for African financial inclusion. # DataTour 2025 - Credit Risk Prediction ## Overview Repository for DataTour 2025 credit default prediction challenge using LightGBM. Focuses on binary classification for loan default probability. Local OOF ROC AUC: 0.664281. Submitted for National Phase - Final Submission, November 2025. Competition Site | Data Platform ## Team - **Name**: AST - **Members**: - Codjo Ulrich Expéra AKAKPO - Prosper Key SOGBEDJI - Thibaut TCHINHOUN ## Approach Treated as binary classification with LightGBM on large tabular data (17M training rows, 62 features like `enc_paym_*`). Handles missing values natively; aligned with ROC AUC metric. ## Feature Engineering - Outlier clipping to 0.01-0.99 quantiles (numeric columns, excluding target). - Replaced inf/-inf with NaN. ## Model Validation - 5-fold Stratified K-Fold cross-validation (random_state=42, early stopping at 100 iterations). - Preserves class imbalance; OOF AUC: 0.664281. ## Hyperparameter Optimization - Manual selection based on benchmarks. - Final params: `n_estimators=2000`, `learning_rate=0.05`, `num_leaves=63`, `max_depth=-1`, `colsample_bytree=0.7`, `subsample=0.7`, `reg_alpha=0.1`, `reg_lambda=0.1`, `seed=42`. ## Requirements - Python 3.8+ - Libraries: `pandas`, `numpy`, `scikit-learn`, `lightgbm` Install: `pip install pandas numpy scikit-learn lightgbm` ## Usage 1. Place `train.parquet` and `test.parquet` in root (data not included). 2. Run script for feature engineering, training, and predictions. ## Reproducibility Fixed seeds for consistent results; end-to-end executable code. ## License MIT For more, see Methodology PDF.