Credit card default prediction remains a central challenge for financial institutions, particularly due to data irregularities, class imbalance, and outlier-driven distortions. While Deep Neural Networks (DNNs) have been widely adopted, their performance is often sensitive to noisy data and computationally demanding. This study investigates whether integrating z-score–based outlier removal with a Multilayer Perceptron (MLP) can enhance predictive performance on the UCI Default of Credit Card Clients dataset. The methodology involved removing outliers using the z-score method per class, standardizing the dataset, partitioning the data into training and test sets, and developing a two-hidden-layer MLP model. After detecting and removing outliers separately within each class, a total of 3,785 anomalous observations were eliminated, resulting in 26,215 usable samples. The cleaned dataset was standardized, split using a stratified 70:30 ratio, and used to train an MLP with two hidden layers. Model performance was evaluated using accuracy, precision, recall, F1-score, specificity, G-mean, and AUC. The MLP achieved an accuracy of 83%, precision of 0.78, recall of 0.67, and F1-score of 0.70, strong specificity (0.96), confirming its ability to correctly classify non-defaulters. AUC score of 0.79 which indicates good overall discriminative power. The confusion matrix showed that while most non-defaulters were correctly identified, a notable number of defaulters were misclassified as non-defaulters. Compared to the DNN, the MLP outperformed across all metrics, with the most substantial improvements observed in recall (0.67 vs. 0.39) and F1-score (0.70 vs. 0.48). These findings demonstrate that a relatively simple MLP architecture, when combined with robust outlier handling, can outperform more complex models such as DNNs.
Yakubu et al. (Tue,) studied this question.