Research On The Construction And Application Of Personal Consumption Credit Score Model

The booming development of the consumer credit industry has forced financial institutions to face huge credit risks while chasing huge profits.Therefore,how to avoid potential credit risks is an important issue the banks and credit institutions facing.The personal credit scoring model is uses statistical analysis methods and data mining techniques to analyze individual basic information data,and converts current personal information into a value that represents a certain credit risk in the future.Early credit scoring techniques were based on basic statistical methods,but with the development of machine learning,more and more scholars use more complex machine learning methods and traditional statistical methods combined with machine learning methods to conduct research.But which method is the best still is no consistent conclusion.Although developed countries have formed a complete system of personal credit evaluation,but they are not suitable for China.Therefore,it is necessary to construct a personal credit information system and find a suitable personal credit evaluation method that conforms to Chinese characteristics.This paper takes the actual data of a financial institution in China as an example to introduce the construction process and evaluation of the personal consumption credit scoring model.Firstly,according to the characteristics of the data,this paper constructed the "RFMC" evaluation index system which fully reflects the data characteristics.After preprocessed the data,three different types of single models are constructed:subjective evaluation methods reflecting expert knowledge and experience-fuzzy analytic hierarchy process,reflect the objective evaluation method of the difference between the different indicators data-gray comprehensive evaluation,and the logistic regression model reflecting the data regular pattern,and use the particle swarm optimization algorithm to combined the above three single models,finally,based on the summary of the paper,the relevant suggestions are put forward,and the future research directions are prospected.By compared the second type of error rate,AUC value and KS value of the four personal consumption credit scoring models,it is found that each model has a certain ability to distinguish.About the single model,the best model is Logistic regression.The second type of error rate is at least 29%.The AUC value and KS value are also the largest of the three single models,AUC value is 0.73 and KS value is 0.36.The second type of error rate of fuzzy analytic hierarchy process is 35%,the AUC value is 0.63,and the KS value is 0.21,it is the worst model among the three single models.After the combined optimization model,the performance has been improved.The combined model reduced the second type of error rate to 25%,which was 4%lower than the optimal single model-Logistic regression model,and the AUC value was optimized to 0.79,although the optimization results were not obvious,but in the credit market,even a slight improvement represents a significant competitive advantage.Compared with the existed research,this paper has made innovations from the construction of the indicator system and the selection of the model:(1)Personal consumption credit evaluation index system.Based on the characteristics of the data-consumption data,this paper introduces the "RFM" model used to measure the importance of customers in the customer management analysis model,and adds the coefficient of variation coefficient "C" that can evaluate consumption volatility.Compared with the index system constructed based on the relationship between variables,the "RFMC" indicator system can fully display the information contained in the consumption data.(2)Personal consumption credit scoring model selection.In this paper,three different types of single evaluation models are selected,and a heterogeneous integration model is constructed by using particle swarm optimization algorithm.After comparing the second type of error rate,AUC value and KS value of the four models,it is found that the combined model is better than the three single models.