A Novel Approach to Optimise Credit Risk Prediction with Enhanced SVM and and Hinge Loss

When a borrower doesn't keep their end of the bargain, there is a danger of financial loss known as credit risk. Even though there are numerous factors that contribute to credit risk, careful consideration when approving a loan may lessen the likelihood of the development of non-performing assets (NPA) and fraud. While there are organizations that offer credit rating and credit scoring services to banks for a fee, the researchers continue to investigate different machine learning strategies to increase the accuracy of credit risk assessment. In this study, a novel approach based on a modified version of SVM (Support Vector Machines) is proposed. The proposed methodology incorporates a custom-trained variant of SVM. A function calculates class weights based on the inverse frequency of each class in the training set. The hinge loss is calculated using a function that takes the feature matrix, target labels, weight matrix, and regularization parameter as input. The hinge loss function penalizes the model for misclassifying instances. Further, this work compares three machine learning (ML) algorithms-SVM, XGBoost, and Random Forest with the proposed approach to comprehend the efficacy in spotting credit risk. The proposed custom SVM model is trained with new weights that were found using the current dataset, which provided an accuracy of 96.92% that is higher than the pre-existing models. Also, the findings showed an fl-score of 0.90, a precision value of 0.915, and a recall value of 0.885.