Improving individual credit classification through hyperparameter tuning by metaheuristics

Using data from 8591 individual customer profiles with demographic characteristics from a Vietnamese commercial bank, this study investigates how metaheuristic algorithms can optimize parameter settings for five machine learning models applied to classification problems. The five machine learning algorithms, Logistic Regression, Decision Tree, Random Forest, Neural Networks, and XGBoost, are investigated in two phases. Phase A classifies customer profiles with default settings, while Phase B applies metaheuristics, Simulated Annealing (SA) and Genetic Algorithm (GA), to find the optimal set of parameters by optimizing the objective function, the Recall score. The empirical results show that Phase B substantially improves Recall performance compared with Phase A, while maintaining competitive classification performance based on additional evaluation metrics. In addition, within Phase B, SA delivers significant improvements over GA for four out of five cases. The optimal algorithm in the experiments is the SA-XGBoost model that achieves a Recall score of 98.23%. The proposed managerial implications recommend leveraging the SA-XGBoost model for enhanced credit scoring and prioritizing key features like loan value and loan term to reduce principal losses.