ENHANCED SYNTHETIC MINORITY OVERSAMPLING TECHNIQUE (SMOTE) BASED MODEL FOR ENHANCING ACCURACY IN CREDIT MODELLING

Abstract

Credit modelling especially in the financial sector faces significant challenges in deep learning applications. Accurate credit modelling is essential in enabling financial institutions assess credit risk and make viable lending decisions. However, complexities arise due to inaccuracies influenced by datasets, modelling algorithms, and sampling techniques. This research sought to evaluate and validate the effectiveness of the enhanced Synthetic Minority Oversampling Technique (SMOTE) based model in enhancing accuracy in credit Modelling. The enhanced SMOTE-based model integrates traditional and machine learning methods, including decision trees, logistic regression, Neural Networks, Random Forest, and Support Vector Machines. Using a diverse dataset, it incorporates borrower characteristics like age, income, and credit score, and loan details such as amount, interest rate, and term. The model focused on balancing data distribution, creating synthetic samples, addressing the challenge of overfitting and optimizing performance to surpass baseline models across metrics like accuracy, precision, recall. Findings revealed that there was limited adoption of advanced models amongst financial institutions in Meru County, due to their complexity and training demands. Further findings reveal that, applying enhancements to SMOTE based model improved class balance, accuracy and error reduction. Random Forests demonstrated marked improvements with enhanced model. Accuracy increased from 59.19% to 87.70%, and the Kappa statistic improved from 0.0055 to 0.7249, indicating better classification agreement. Error rates showed significant reductions, with the mean absolute error decreasing from 40.81% to 12.30%, and the root mean squared error dropping from 0.6388 to 0.3507. The enhancements in sensitivity (from 78.28% to 91.19%) and specificity (from 22.22% to 80.95%) further underscore the model's effectiveness in handling dataset imbalances with SMOTE. These results suggest that Random Forests, when combined with enhanced SMOTE-based models, can significantly improve the accuracy and precision of credit risk predictions. Adopting enhanced SMOTE-based models with Random Forests offers robust tools for credit risk management, advocating for increased quantitative model adoption and collaboration among financial institutions.