Loan default prediction model for SMES based on FT-CNN

Credit risk is related to the operation and survival of banks. To solve the problems of data distortion, non-obvious features, low feature dimension and weak feature correlation in the loans of small and medium-sized enterprises (SMEs), a convolutional neural network model based on feature transformation (FT-CNN) was proposed for the loan default prediction of SMEs. Firstly, XGBoost algorithm was used to classify the original data by feature subset branch, and the original eigenvalues were hot coded according to the classification results of its leaf nodes. The low-order dense matrix composed of original eigenvalues were converted into high-order sparse matrix to reduce the influence of data distortion and low feature dimension on prediction. Then, the problem of unclear features and weak feature correlation was solved by the local perception virtue of CNN. The prediction model proposed in this paper was compared with two models of XGBoost and CNN, and compared with SVM, XGBoost+LR and XGBoost+SVM models. The model led other models with 95%prediction accuracy, 92.9%F1 score and 91.4%AUC score. For the problem of credit risk of SMEs, this model displays better in prediction results and efficiency, which would be adapted to the task of loan default prediction of SMEs.