An Ensemble Classifier-Based Approach for Feature Selection and Prediction of Loan Approval

Abstract

Purpose : The current research addressed the challenges confronted by banks and financial institutions in processing loan requests. With an increase in the number of loans being requested, the risk of non-payment also increases. Thus, an accurate and reliable method was needed that can evaluate the eligibility of the applicant and lessen limit losses. Therefore, this study aimed to implement machine learning (ML) methods to anticipate loan approvals and diagnose potential defaulters.

Methodology : An ensemble-based ML method that uses grid search-based optimized classifiers was proposed. Four different classifiers, namely LightGBM, XGBoost, CatBoost, and AdaBoost, were used in this study. At first, 12 features were utilized to train these models, and the models established their order importance. These features were arranged in descending order of importance, and feature sensitivity analysis was done to find the most optimal set of features for the particular classifier. The performance of the model was compared based on accuracy and F1 score.

Findings : The results depicted that all three ensemble classifiers except Adaboost achieved an accuracy of ~98% trained using only six best features. Furthermore, Catboost outperformed all other models in the prediction of loan approvals with an accuracy and F1 score of 98.68% and 92.632%, respectively. This reflected the importance of selecting optimal features and ensemble learning, which improves the model performance and mitigates the computational complexity.

Practical Implications : The findings recommended that utilizing ML algorithms can improve the loan approval method by reducing the risk of operating inefficiencies and defaults. They also highlighted the necessity of selecting the optimal features to make reliable ML models. The future scope of this research lies in exploring additional data sources, such as transactional and behavioral data, to improve the model’s effectiveness.

Originality : The traditional methods for loan approval rely on subjective assessments and manual judgments. This study provides a data-driven method that can be automated with an improved decision-making process. Furthermore, this study used feature ranking methodology and ensemble classifiers that contribute to different aspects of loan analysis, offering a scalable and efficient alternative to banks and financial institutions.