Predicting Default Situations in the P2P Lending

Abstract

As a flexible and efficient new financial format, P2P lending suffers from breach of contract and lack of trust due to the uneven credit, income, and region of borrowers. Therefore, we plan to use machine learning algorithms to predict the default situation in the P2P market in the future, and compare the prediction accuracy of various models to find the optimal default prediction model. The research data in this article includes P2P lending data from 33,105 users in 50 states in the United States. It includes variables such as investment income loss percentage, borrower income, and loan term. To simplify subsequent analysis, missing values were cleaned and data on borrower state and loan date were classified and simplified. T-test and chi-square test were used to preliminarily analyze data-type data and categorical-type data, and the results showed that all relevant variables are statistically significant and need to be considered in subsequent research. To further determine the significance of each variable in the default situation, a logistic regression model was introduced, which has practical significance for lending platforms in user selection. Finally, four types of models were used for constructing default prediction models, which are logistic regression, decision trees, random forests, and GBDT. The ACC and AUC values of different models on the training and testing sets were compared. The conclusion is that the GBDT model has the highest prediction accuracy and a high AUC value, which can serve as a prediction model for future lending platforms to predict user default situations.