Credit Card Default Prediction: A Comparative Study Of Machine Learning Models Based On Accuracy, Sensitivity, And Specificity

Abstract

Objective: This analysis aimed to compare three machine learning models—Logistic Regression, Naive Bayes, and Linear Discriminant Analysis (LDA)—for their ability to predict credit card default.

Methods: Each model's performance was evaluated based on accuracy, sensitivity, and specificity metrics using a dataset of credit card holders.

Results: All three models demonstrated similar accuracy levels, between 0.969 and 0.9715, indicating a good ability to correctly classify cases overall. In terms of sensitivity, or the ability to correctly identify non-default accounts, all models achieved high scores (0.9953 to 0.9979). However, there were differences in the specificity, or the ability to correctly identify default accounts. The Logistic Regression model showed a higher specificity (0.2754) compared to the Naive Bayes and LDA models (both 0.2319), suggesting a better performance in identifying default accounts.

Implications: While all three models showed high accuracy and sensitivity, the Logistic Regression model outperformed in terms of specificity, making it the preferred model for this task. However, all models exhibited relatively weak performance in identifying default accounts, indicating a potential need for further optimization, consideration of other metrics, or different modeling approaches, especially given the high cost associated with misclassifying defaulting accounts. However, all models exhibited relatively weak performance in identifying default accounts. This is likely due to the imbalanced nature of the data. Therefore, different modeling approaches or techniques to handle imbalanced data might be necessary to improve the identification of defaults.