Comparative Analysis of Traditional Statistical and Machine Learning Approaches in Credit Scoring Applications

Abstract

This paper compares the performance of traditional statistical approaches, such as logistic regression (LR), and machine learning approaches, like support vector machines (SVMs), in credit scoring. In this paper, a dataset is simulated containing borrower characteristics like income, wealth, repayment history, length of requested loans, and total debt, which can be altered to represent different macroeconomic scenarios. Mathematica is used to train or fit and ultimately test both LR and SVM models on the dataset, focusing on evaluation metrics such as accuracy, precision, and recall to assess their performance. Results show that SVM consistently outperforms LR, with recall being 31.8% higher, precision being 15.6% higher, and accuracy being only 4.6% higher. This suggests that banks should consider implementing machine learning methods for credit scoring, as long as they have access to large datasets and sufficient computational power. Traditional approaches like LR should not be dismissed, as they offer transparency and interpretability, which are essential for financial institutions due to the fact that they are regulated entities.