The Research Of Personal Credit Scoring Based On CBR System Optimization

After the openning up and reform China’s banking sector has experienced arapid development and the credit scoring has been gaining increasing importance inthe banking sector. current prevalent credit scoring models globally are built onstatistics or artificial intelligence, which are all double-edged. Statistical modelsoffer tests but fall short of accuracy, while artificial intelligence achieves highaccuracy but fall short of explanatory function. Furthermore, these models arechallenged by reject inference and dynamic changing of credit samples. Rejectinference is the correction of that sample bias where the partial samples of clientswho have received loans are used to train the scoring model applying to the totalcredit population in most scoring organiza itons. That leads to the non-randomsampling bias, which directly affects the valid ity of credit scoring models. Dynamicchanging of credit samples refer to the phenomenon that clients’ credit informationis changing according to their personal economic situations or the overall drift ofthe credit population occurs with the economic and social development. That causesincreasing deviation of scoring models’ outputs from realities. Reject reference andthe dynamic changing of credit samples are urgent problems to be solved in thecredit scoring.Case-based reasoning (CBR) simulates the cognitive process of humanity,enjoying a strong theoretical background and varied applications. CBR has thepotential to be the dynamic credit scoring system that could solve reject reference.First of all, after analyzing the current development of CBR, a traditional CBRcredit scoring system is built and tested with real credit data and results show thatthe application of CBR to credit scoring in China not only shows some advantagesbut also faces constrains of the existing data and of traditional CBR’s hypotheses.Secondly, an improved CBR system is developed from the aspects of both case baseand reasoning cycle concerning those constrains. Case presentation optimization,introduction of rejected cases and a dynamic managing scheme are discussedregarding of the case base. A hybrid retrieving method of neural network and KNNis developed and Bayesian algorithm is employed as the reusing method. Lastly,based on the real data of a commercial bank in Shenzhen, experiments areconducted to test the application of the improved CBR system for personal creditscoring. The results shows that the improved CBR system has a better performanceclassifying those rejected cases and is good at solving problems of reject reference and dynamic changing of credit samples; the improved CBR system is moreaccurate and has obvious progress in stability and explanatory function comparingto traditional CBR system; a supporting function of credit policies is demonstratedby the optimized CBR system which can dynamically adapts to easing or tighteningcredit policies made by commercial banks.