统计与数据科学学院知识库

  The difficulty of obtaining the financial support for small/micro enterprises has become a major restrictive factor in their development,in which enterprise credit rating plays an important role in the credit business. Considering the online public data——Polish enterprise financial data as an example, this paper studies the discriminant model of bankruptcy/non-bankruptcy:The multi-scale transformation is used to eliminate abnormal data, the well known 3s criterion is used to shrink and replace individual abnormal data, and the median is used to interpolate missing data.According to the significance test of the mean difference between two independent samples, the average within-class distance two sub-samples and the correlation coefficient between variables, the distinguished variable sets which has important effect on the classification and no collinearity each other are selected. This variable set can be used not only to establish enterprise risk evaluation model, but also to control risk.Six single models, including the classical classification model KNN, SVM, Random Forest and the Boosting Decision Tree algorithm such as XGBoost, GBDT and LightGBM, are studied to study the classification of bankruptcy/non-bankruptcy. It is found that the latter lifting algorithm can improve the classification performance of the decision model to a certain extent. Two ensemble strategy——Voting and Blending,based on the aforementioned foundational model, were studied in the prediction and evaluation of bankruptcy. The results show that the voting method has limited promotion in the classification performance of the model.The Blending method based on XGBoost has a distinguished improvement in the evaluation performance. Compared with other ensemble models based on GBDT and LightGBM, The former also performs well in preventing over fitting of the model. The results show that the ensemble model has a good performance in discriminate and evaluation the bankruptcy/non bankruptcy risk, which is worthy of further research.