Abstract: P2P lending is a new type of loan mode formed by the intersection of Internet and traditional finance. It provides a more convenient loan platform and has been developing rapidly in China.However, the phenomenon of collapse in P2P is getting worse as P2P loans is facing default risk and bad debt losses seriously. Credit evaluation is an important basis for managing loan default risk and supporting lending decision. Compared with traditional loans, the financial data of borrowers collected by P2P platform is limited, which is also called the hard the information.However,there is lots of soft information generated during the loan application, such as loan description text,also involving some information about loans and borrowers. Therefore, a default prediction method combined with soft informationfor P2P lending is proposed. Firstly, the soft information is categorized according to the characteristics of P2P, and the LDA topic model is used to quantify valuable factors in the text of soft information. Secondly, some regression analysis and contrast experiments are performed to test the effect of soft information on P2P default probability. Moreover, a two-stage method is designed to selecteffective variablesets for default modeling, and the default prediction model is constructed through the random forest (RF) method.Finally, based on the data from a Chinese P2P platform—eloan.com, an experimental research is conducted to verify the effectiveness of methods we proposed.The results show that the soft information can improve the recognition rate of loan default, which can be used as the basis of P2P credit evaluation. The feature combination selection method proposed in this paper and the credit evaluation model based on Random Forest have achieved good classification accuracy.And the proposed method can improve predictionperformancesobviously compared withthe platform's own rating method, which has certain reference significance for the credit evaluation of P2P network lending.

Key words: P2P lending, default prediction, soft information, topic model, variable selection, RF model