Default Prediction of SMEs Based on STUNT Method

The default prediction of small and medium-sized enterprises (SMEs) provides decision-making basis for bank loans. The current research needs a variety of and a large amount of data, but the financial information of SMEs is incomplete, and the amount of data is small. We solve the two problems of predicting corporate defaults based only on invoice information and improving the accuracy of predictions while minimizing reliance on corporate data and loan records. To solve these problems, we first established three categories of 17 original feature for the invoice information and performed feature selection, and then proposed using the Self-generated Tasks from UNlabeled Tables (STUNT) method: first, meta-learning a large number of companies without credit information, and then use the companies with credit information to fine-tune the model, and use the accuracy rate to stop in advance to ensure the optimal model. Compared with the other machine learning methods: support vector machine (SVM), decision trees (DT), random forests (RF) and gradient boosting trees (GBT), the STUNT model has an advantage that in most cases the F-score and Average Accuracy scores are 20% higher than the baseline method.