Impact of Normalization Methods on Credit Risk Classification with Imbalanced Data

Abstract

Credit risk classification is an essential task in the financial sector to minimize loan default losses and maintain financial system stability. One of the main challenges in building accurate credit risk models is class imbalance, where the number of non-default cases significantly outnumbers default cases, causing classifiers to be biased toward the majority class. This study investigates the effect of two normalization methods, Z-Score and Min-Max, on the performance of Support Vector Machine (SVM) and Artificial Neural Network (ANN) classifiers for imbalanced credit risk prediction. Six experimental scenarios were designed, incorporating the Synthetic Minority Over-sampling Technique (SMOTE), Chi-Square feature selection, and GridSearchCV hyperparameter tuning. Experiments were conducted on a dataset of 4,801 loan records with a class imbalance ratio of 77:23, divided into 80% training and 20% testing using stratified sampling. Evaluation metrics include Accuracy, Precision, Recall, F1-Score, AUC-ROC, Matthews Correlation Coefficient (MCC), and Balanced Accuracy, with F1-Score as the primary metric given the imbalanced class distribution. The results demonstrate that Z-Score normalization, when applied in isolation, consistently outperforms Min-Max across both classifiers, with ANN achieving an accuracy of 95.42% and F1-Score of 0.9332 on the Z-Score scenario without SMOTE. However, the fully optimized pipeline combining Min-Max normalization, Chi-Square feature selection, SMOTE, and hyperparameter tuning surpassed this result, yielding the best overall performance with an ANN accuracy of 95.73%, F1-Score of 0.9393, and AUC-ROC of 0.9610. This indicates that Min-Max’s sensitivity to outliers can be effectively mitigated through appropriate feature selection and hyperparameter tuning. These findings provide empirical guidance for selecting optimal preprocessing strategies in imbalanced financial classification tasks.