Research On Dynamic Credit Risk Evaluation Method In P2P Lending

In recent years,with the rapid development of Internet Finance,lending business has gradually changed from single offline mode to concurrent offline and online mode,peer-to-peer(P2P)lending has become an important component of the multi-layer financial system At the same time,P2P lending has shown some new features and raised some new problems,which bring great challenges to credit risk evaluation.On the one hand,the diversity of credit risk is high in P2P lending.Platform credit risk and borrower credit risk coexist in P2P lending and both have their own characteristics,hence it is necessary to evaluate different types of credit risks.On the other hand,the complexity of credit risk is high in P2P lending.The credit subjects and transaction processes are virtualized in P2P lending and information asymmetry and dynamics are strong,hence it is necessary to have whole process dynamic credit evaluation methods which are timely and can reflect the credit change trend.To this end,this thesis holds the perspective of providing investment decision support service for investors,focuses on two types of credit subjects,i.e.platforms and borrowers,aims at two core points,i.e.credit features and evaluation methods,and carries out our research from the aspects of platform credit features identification,dynamic platform credit risk evaluation method,borrower credit features identification,and dynamic borrower credit risk evaluation method.Specific research contents and contributions are as follows.(1)Credit risk identification and evaluation framework in P2P lending.By analyzing the investment decision-making process in P2P lending with the expected return theory model,we find that comprehensive evaluation for platform credit risk and borrower credit risk can effectively reduce investment risk.On this basis,we analyze the characteristics of platform credit risk and borrower credit risk and their influences on credit feature identification and evaluation method design,and propose a credit risk identification and evaluation framework in P2P lending.(2)Platform credit feature and dynamic platform credit risk evaluation in P2P lending.By synthesizing 5C's of credit,herding effect,and other theories,we systematically analyze the influencing factors of platform credit risk from platform and user perspectives,and analyze and identify multiple dimensions of platform credit features from the aspects of platform characteristic,platform capacity,platform capital,platform conditions,and user word of mouth on this basis.We use the mixture cure method to build dynamic platform credit risk evaluation method and predict the default status and default time of platforms.Besides,we identify multiple effects of platform credit features such as unilateral effects,bilateral effects,and time-varying effects.(3)Dynamic platform credit risk evaluation method based on panel data in P2P lending.Aiming at the dynamics of platform credit features,we propose a dynamic platform credit risk evaluation method based on panel data.The proposed method uses mixture survival modeling to effectively solve the censored problem of observation data and construtes multiple survival observations to correlate credit features with observation status at different times,enabling dynamic information in panel data to be effectively incorporated into the model.Our research shows that the proposed method can accurately predict the default probability of platforms at different times and can effectively depict the credit risk trend of platforms.(4)Borrower credit risk evaluation method combined with soft information in P2P lending.Aiming at unstructured descriptive loan text information,we propose a borrower creditrisk evaluation method combined with semantic soft features.The proposed method uses multiple co-occurrence statistics to identify fixed multi-words expression,uses word embedding model to map the words into vector space,and proposes a semantic cliques extraction algorithm to cluster the words in vector space into semantic cliques and further defines semantic soft features.Then,the extracted semantic soft features are added into the borrower credit risk evaluation models.Our research shows that the proposed method can effectively improve the default prediction performance of borrower credit risk evaluation models and the effectiveness of portfolio selections.(5)Prediction-driven dynamic borrower credit risk evaluation method in P2P lending.In order to improve the accuracy of borrower credit risk evaluation,including the accuracy of default status discrimination and default time estimation,we propose a prediction-driven dynamic borrower credit risk evaluation method.For the accuracy of default status discrimination,we propose an ensemble learning-based default status discrimination model.For the accuracy of default time estimation,in order to consider the interaction effect between credit features and risk time,we propose a time independent hazard-based default time estimation model.Our research shows that the proposed method can effectively the discrimination performance and calibration performance of borrower credit risk evaluation models.Based on above research findings,in theory,this thesis enriches the platform and borrower creditrisk theories,extends existing credit risk evaluation systems,and provides effective credit risk evaluation methods for the dynamic tracking evaluation of the whole life cycle of credit and the whole process of business,and the dynamic prediction of future defaults.In practice,this thesis can provide support for the macro-industrial control of regulatory departments,the risk management of business institutions and the investment decision-making of investors.