Default Identification of P2P Lending Based on Stacking Ensemble Learning

With the development of Internet finance, P2P lending (Peer-to-Peer Lending) have been popularized rapidly. However, due to the lack of a perfect credit risk identification system, the extremely high rate of bad debts has caused huge loss to operators and investors of P2P platforms. Aiming at the problem of high latitude and large scale of network loan data and low accuracy of single classifier’s default identification. The Lending data of the fourth quarter of 2018 on Lending Club platform in the United States was taken as the research object. Firstly, IV information value and Spearman-Boruta algorithm were combined to extract features, and a credit evaluation index system was established based on 5C credit evaluation principle widely recognized in the financial sector. Then, super parameter tuning was performed by the method combining Random search and 5 fold crossover validation. Finally, Stacking algorithm is adopted to integrate four base classifiers, Namely ANN (Artificial neural networks), RF (Random forest), AdaBoost (Adaptive Boosting) and XGBoost (Gradient boosting), to realize the default discrimination of network loans. The results show that the Stacking had better performance overall than the single classifier. Compared with the best performance XGBoost in the base model, the Accuracy, Precision, Recall, F1-score and AUC of Stacking model are improved by 0.41%, 0.45%, 0.28%, 0.37% and 0.42%, respectively. The Stacking model based on the fusion of multiple models can effectively reduce the error rate of misjudging defaulting customers as non-defaulting customers, and it is a default identification model that can meet the demand of lending business better, providing a scientific basis for the credit default prediction in financial sector.