关键词: 信用风险评估, 分类, Borderline SMOTE-2, 堆叠模型

Abstract: Credit risk is the core content of today’s risk management, and a lot of research has been done on practical models that support debtor credit assessment, pricing of credit risk instruments, measurement and control of credit risk exposure, and portfolio credit loss analysis. Personal credit risk refers to the possibility of default due to failure to repay debts or loans in time and in full for various reasons, and its degree directly affects the strength of credit. In the context of increasing personal credit default risk, in order to enable enterprises to accurately identify personal credit risks, this paper presents a personal credit risk assessment method based on the improved BS-Stacking model.
This paper obtains bank data related to individual credit risk from the German credit data set published by UCI. The data set has 1000 samples, 700 positive samples, 300 negative samples, and 24 indicator attributes for each original sample. According to the characteristics of the personal credit risk data, we first use the improved Borderline SMOTE 2 algorithm to oversampling the data, and further remove the noise points on the basis of strengthening the identification of the minority sample boundary area, so as to ensure the accurate prediction of the default sample. In addition, for the problems that the classifier in the Stacking algorithm has redundancy and may reduce the prediction performance, grid search is used for parameter adjustment and LR is proposed to analyze the contribution degree of individual learners to obtain the optimal combination of individual learners and achieve the optimal performance of the entire model.
In this study, Accuracy and AUC are used to measure the accuracy of the prediction, and Precision, Recall, F1 score, and Specificity are used to measure the validity of the model. In the initial algorithm, a total of8 models SVC, GBDT, RF, AdaBoost, XGBoost, LightGBM, KNN and LR are selected as individual learners, and LR is selected as the output of individual learners to train. After unbalanced algorithm processing and comparative experiments, the model after preliminary screening is formed. After analyzing the contribution degree of the base model and testing the whole model, the optimal combination model is obtained, and the performance of the integrated model reaches the optimal state. The experiment proves the effectiveness of the unbalanced algorithm and the integrated algorithm from many angles, and also shows that the algorithm can achieve high accuracy and robustness in personal credit risk assessment.

Key words: credit risk assessment, classification, Borderline SMOTE-2, Stacking model