Abstract:

The purpose of credit evaluation is to reduce information asymmetry and reduce the credit risk of banks. However， the existing credit risk evaluation models are prone to result in the “Score-Risk” mismatch phenomenon of “high credit score sometimes matches high level of default risk”. To solve this mismatch， inversion numbers are introduced to measure the level of mismatch. Then， a nonlinear credit risk evaluation model is developed by minimizing the inversion number of related loss given default sequence. There are some innovations and features in this study. （1） The functional relationship among the indicators of the credit evaluation of small and medium-sized enterprises （SMEs）， the inversion number IN{V}_Y(\mathit{P}) of its corresponding default state sequence \mathit{P}， and the inversion number IN{V}_{LGD}({\mathit{P}}_{LGD}) of its corresponding loss given default sequence {\mathit{P}}_{LGD} are constructed. （2） Inspired by the concept of AUC and GINI coefficients and to measure the mismatch between credit score and credit risk， this paper establishes a kind of indicator based on the distribution of default lenders， which is used to describe the mismatch level of “Score-Risk”. （3） Particle Swarm Optimization（PSO） algorithm is used to solve the nonlinear programming model and find that the model would result in a low level of mismatch with the initial number of particles， the number of iterations， and the diversity of initial positions of particles increasing.To compare with other weighting methods， the loan data of 1231 small enterprises in China are analyzed， the empirical results show that there is indeed a “Score-Risk” mismatch in the existing credit evaluation methods and the model established in this paper can improve the “Score-Risk” mismatch well. Further， the proposed model is compared with six prediction measurements （AUC， Accuracy， etc） and seven other models （Support Vector Machine， Decision Tree， etc） in the credit data of Germany， Australia and Chinese Taiwan published by UCI. By post-hoc analysis， it is found that although there is a small difference in predictive performance among these models， the proposed model still has certain advantages over the artificial intelligence credit evaluation model. Different from existing models which focus on the accuracy of prediction， it aims to solve the problem of “Score-Risk” mismatch in this paper， which may provide a new perspective for credit evaluation.

Key words: credit rating, default state, loss given default, inversion number, mismatch of credit score & credit risk