A Hybrid Credit Risk Evaluation Model Based on Three-Way Decisions and Stacking Ensemble Approach

Credit risk evaluation is a binary classification problem, and machine learning algorithms have achieved remarkable results in this field. However, traditional two-way decisions involve a high risk of making a bad decision with insufficient information. This paper applies the three-way decision method to introduce the delayed decision mechanism and proposes a hybrid credit risk evaluation model based on the stacking ensemble approach. First, the decision loss values in the three-way decision are determined based on cash flow data. A multiobjective optimization model is constructed to determine the three-way decision thresholds by minimizing the decision cost and the size of the boundary region. Second, in the hybrid model, the LightGBM algorithm is used to evaluate the default probabilities of samples in the first step, and partial samples are made delayed decisions. Multiple ensemble learning algorithms are integrated to form a stacking model to achieve further decision-making on delayed decision samples. Experiments on the credit dataset show that the proposed model performed better than a variety of popular machine learning algorithms and could classify samples with high decision costs better.