Research On The Credit Risk In Peer-to-peer Lending Based On Text Analysis

As an emerging online lending mode,P2P(Peer-to-Peer)online lending provides a new way to financing for “long tail” customers in the traditional lending market.It can be regarded as a meaningful supplement to the traditional bank lending mode.It has great significance for the realization of inclusive finance.However,as a mode of Internet finance,P2 P online lending is kind of financial instrument,so that financial risks still exist.The most fundamental source of financial risks is the information asymmetry among participants in financial activities.In recent years,accompany with the rapid development of the online P2 P lending market,default events appear,including a large number of default borrowers and a large number of defunct online P2 P lending platforms.These risk events cause great damage to investors and a lot of troubles to the governments and market regulators.Online P2 P lending is a mode of credit lending,and the significant risk it faces is the credit risk from the borrowers.Although there are many studies focus on the credit risk in traditional lending market,researches on credit risk in the online P2 P lending market is still rare.Different from the traditional lending market,online P2 P lending generates a large amount of heterogeneous unstructured data in the lending activities.Recent studies have shown that these unstructured data can also affect the performance of P2 P online lending activities at different stages,including the loan repayment stage.Thus,incorporating unstructured information into the risk model may contribute to more accurate assessment of participants' credit risk.Based on this perspective,comprehensively using the econometric model,machine learning theory,deep learning theory,and text analysis technology,this paper first explore the function of different factors in P2 P online lending activities,especially the role of these factors in the loan repayment stage.Then we construct a credit risk assessment model by incorporating textual information,and finally examine the relationship between comments sentiment from investor community and bankruptcy of the P2 P online lending platform.The research works and related innovations of this paper are briefly described as follows:1.Based on the loan transaction records from Renrendai.com we empirically test the characteristics of loan application and the characteristics of credit certification in different stages--financing,review,and loan repayment--of the lending activities,as well as the relationship between borrowers' characteristics and the outcomes at each stage.In particular,we focus on the role of the borrower's “soft information” in these stages.In addition,we also portray the borrower's group characteristics at different stages--the loan application release stage and the loan repayment stage.We analyze,clearly and completely,the influence and the role of different factors in different stages on the P2 P lending market in China.2.Using dictionary-based text feature representation method,we decompose the loan description into standard information component and specific information component,where the standard information component reflects the commonalities between the current loan description and the recent loan descriptions provided by other borrowers.The specific information component reflects specific information of the borrower.Based on the transaction data from the Renrenda.com,we find that the motivation for borrowers to provide loan descriptions is mainly to hyper their loan applications,rather than to reduce the information asymmetry between them and investors.3.Based on the Transformer-Encoder model,we first extract the credit risk-related text features from the loan description,and then combine it with other traditional quantitative indicators that used to evaluate the borrower's credit risk,to construct a risk assessment model for evaluating borrowers' credit risk in P2 P lending market.The loan records from the “Renrendai.com” and the “Lending Club.com” platform is used to empirically test the predictive performance of the constructed model.The results show that the textual features extracted from the loan description help to assess the borrower's credit risk more accurately,and the model constructed in this paper achieves the best performance in predicting borrowers' default.4.Based on weakly-supervised short text sentiment analysis,we study the relationship between the comment sentiment from investor community and the bankruptcy of online P2 P lending platform.In the sentiment analysis stage,we first use the rating score of each comment from investor community as weakly-supervised training label to train the short text sentiment classification pre-training model.Then,the pre-training model is fine-tuned by using manually labeled investor community comment data.Finally,we use the well-tuned model to assess the sentiment of each comment.Then in the empirical analysis stage,we use the obtained sentiment score of each P2 P lending platform to construct a sentiment score for it.Using the data collected from a third-party platform in China,we analysis the relationship between the comment sentiment of the investor community and the bankruptcy of the P2 P lending platform.It is found that the positive degree of the investor community comment can effectively predict the bankruptcy of the online P2 P lending platform,and the high positive sentiment score corresponds to the lower bankruptcy possibility of the platform in the near future.