Assessing the Performance of Naïve Bayes, Support Vector Machine, and Gradient Boosting Models for Credit Scoring and Bankruptcy Predictions

Abstract

Credit risk assessment is critical to reducing defaults, and selecting the appropriate machine learning technique is crucial for this task. Despite the various research on credit scoring, the choice of machine learning techniques remains inconclusive, the problem to identify the most suitable machine learning techniques and models for credit scoring and bankruptcy prediction in the financial sector, considering the challenges of imbalanced data, interpretability, and feature selection. Also, there is currently limited research on the use of machine learning methods for credit scoring and bankruptcy prediction in the banking and finance sector on the dataset used. To address this gap, this study aims to identify the optimal machine learning model for predicting defaults and evaluate the strengths and weaknesses of three models: Naïve Bayes, Support Vector Machine, and Gradient Boosting. The study utilized credit dataset of selected countries downloaded from the UCI machine learning repository, that is, Australia, Germany, and Japan. The dataset was analyzed using the models, Naïve Bayes, Support Vector Machine, and Gradient Boosting embedded in the RStudio analytic program to make predictions of default. Conclusion on the best learning model was drawn on the accuracy of the models to predict default on the three-credit dataset. The analysis show that, the average accuracy prediction of the models is Gradient Boosting (73.4%), Support Vector Machine (57.3%) and Naïve Bayes (69.8%). As a result, the study concluded that Gradient Boosting model is the best machine learning technique to predict default. Based on this, the study recommends using the Gradient Boosting model as it provides the best balance of accuracy and computational efficiency for this dataset. In addition to the accuracy value, the study recommends considering other evaluation metrics such as precision and recall when selecting a model. Depending on the specific goals of the analysis, these metrics may be more or less important