This repository contains an empirical credit risk modelling study focused on predicting corporate defaults using multiple classification algorithms. The objective is to evaluate whether advanced machine learning methods outperform a baseline logistic regression model.
We model one-year corporate default probability using financial ratios:
- WKTA (Working Capital / Total Assets)
- RETA (Retained Earnings / Total Assets)
- EBITTA (EBIT / Total Assets)
- MV (Market Value proxy)
Models evaluated:
- Logistic Regression (baseline)
- Decision Tree
- Neural Network
- Gradient Boosting Classifier
- Missing value treatment
- Outlier analysis using IQR method
- Correlation analysis
- 60/20/20 stratified train/validation/test split
- Hyperparameter tuning on validation set
- Evaluation using:
- Accuracy
- Precision
- Recall
- F1 Score
- Fβ (β = 2)
F1 and Fβ metrics were prioritized due to class imbalance (~10% defaults), where recall is economically more important than raw accuracy.
- Decision Tree: tuned
max_depth - Neural Network:
- ReLU activation
- Dropout = 0.2
- 50 epochs
- Batch size = 10
- Hyperparameter search over hidden layers and nodes
- Gradient Boosting:
- Learning rate = 0.8
- Max depth = 2
- Tuned
n_estimators
- Logistic Regression achieved the strongest overall performance.
- More complex models did not outperform the linear baseline, likely due to:
- Limited dataset size
- Relatively linear feature relationships
- Class imbalance constraints
- Reduced-feature decision trees required deeper structures and underperformed relative to logit.