머신러닝 기반 기업부도 예측모델의 이상치 분류기준 재정의에 따른 성능 비교연구

Recently, banks have been facing difficulties in developing segmentation models due to the extreme imbalance between the number of observations in the minority category and the number of observations in the majority category in the development of machine learning models. In this study, we propose a method to resolve the imbalance by changing the criteria for classifying minority categories without increasing or decreasing the number of observations as in previously studied sampling techniques. For the study, we used corporate financial information and credit rating information from Nice Information Service co., Ltd. data. We extracted 22 financial indicators from 5,705 companies and used them as independent variables, and redefined the definition of default by dividing it into Korean corporate credit ratings, and used them as dependent variables to conduct experimental research using XG BOOST algorithms in Wise Profit's automated machine learning platform. The training and validation ratio of the machine learning algorithm was 7:3, and the experimental results were measured by the error matrix. The experiments were divided into two parts. The first experiment evaluated the performance of the machine learning model trained on the data of each default override. The results of the first experiment showed that the AUROC value increased (0.5 → 0.831) as the data imbalance ratio was reduced by redefining the outlier classification criteria. The second experiment compared the performance of each model created in the first experiment using the same balanced data generated from actual defaults. The results of the second experiment showed that the model trained with the data that resolved the imbalance through outlier redefinition performed better than the model trained with the original data without redefinition.