Improved Credit Scoring Model with Hyperparameter Optimization

This article investigates the application of machine learning techniques for predicting corporate default risk. In the credit scoring domain, the class imbalance problem is prevalent, with defaulted cases typically being much less numerous than non-defaulted ones. To address challenges posed by unbalanced datasets, various calibration methodologies, including data sampling strategies and Synthetic Minority Oversampling Technique (SMOTE), are explored to identify the most effective strategy for the specific case. Additionally, hyperparameter tuning plays a crucial role in enhancing machine learning model performance. Therefore, we employ hyperparameter optimization techniques such as Grid Search, Random Search, Bayesian Optimization, and Adaptive Particle Swarm Optimization. The primary objective is to compare these techniques and determine the optimal strategy for predicting company default events. Our evaluation incorporates standard parameters such as accuracy, balanced accuracy, specificity, sensitivity, and F1 score. Another factor to consider to discriminate the various approaches is the execution time, to be able to develop models that can be implemented in market products. The experimental results reveal which approaches result in significantly improved classification performance.