a novel multi-algorithm stacking framework for enhanced credit risk management

Credit risk prediction remains a central challenge for financial institutions because inaccurate assessments can cause substantial financial losses and systemic instability. this study introduces a multi‑level stacking ensemble that combines gradient boosting, extreme gradient boosting (xgboost), and random forest as base learners with logistic regression as the meta‑learner. to address class imbalance, we do not use synthetic resampling; instead, we apply a class‑management protocol based on fold‑wise class‑weighting, probability calibration, and operating‑point tuning to ensure fair treatment of the minority (default) class without introducing synthetic examples. the approach was evaluated on two uci benchmark datasets (german and australian credit) using a fixed train/test split and stratified 10‑fold cross‑validation on the training set for model selection; final models were retrained on the full training set and assessed on a held‑out test set. results show the stacked ensemble consistently outperforms individual base learners on balanced metrics including f1 and matthews correlation coefficient (mcc) while preserving interpretability via calibrated base‑learner probabilities and inspectable logistic meta‑coefficients. an empirical analysis of principal component analysis (pca) reveals dataset‑dependent effects: pca can benefit simpler classifiers but may reduce performance for interaction‑sensitive ensembles. the paper provides a practical deployment blueprint covering class‑management placement, probability calibration before meta‑learning, and cost‑aware evaluation tailored to credit‑risk operations.