A Stacked Ensemble Framework for Credit Approval Predictions

Credit approval prediction is a critical task for financial institutions, balancing default risk mitigation with financial inclusion. Traditional credit scoring relies on rigid rule-based systems and linear models that fail to capture complex, non-linear relationships, often resulting in sub-optimal or biased decisions. To address this, we propose a stacked ensemble machine learning framework combining three distinct base classifiers: Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM), with a Gradient Boosting Machine (GBM) meta-classifier. Evaluated on a Kaggle dataset of 45,000 loan applications with 14 demographic and financial features, the framework incorporates missing value imputation, categorical encoding, Z-score normalization, and class imbalance mitigation via SMOTE. Models were tuned using 5-fold cross-validation with GridSearchCV on an 80/10/10 train-validation-test split. The Stacking Ensemble achieved the highest overall accuracy at 92.33%, while Random Forest delivered the top ROC-AUC (0.9729) and F1-score (92.24%), both significantly outperforming the baseline Logistic Regression (86.60%). These findings confirm that multi-tiered ensemble learning effectively enhances credit risk assessment and supports automated financial decision-making.