Method for constructing retail credit risk prediction model and Internet credit business Scoregamma model

Abstract

The invention relates to an internet credit business Scoregamma model and a method for predicting retail credit risk of internet credit business by using the same, and the method comprises the following steps: a data collection step: obtaining retail credit prediction data of a to-be-predicted sample; and a credit default probability calculation step: substituting the retail credit prediction data into the credit default probability model to calculate the credit default probability of the to-be-predicted sample.

Description

Technical Field

[0001] This application relates to a credit risk management system and method. The method and system of this application can assist financial institutions in making more accurate risk decisions and accelerate their digital transformation process. In detail, this application relates to a method for constructing a retail credit risk prediction model for Internet credit business and a retail credit risk prediction model for Internet credit business. Background Art

[0002] In the current environment of booming consumer credit, the manual approval mechanism of some financial institutions can no longer cope with the increasing demand for credit, so there is an urgent need to improve the intelligent risk control capabilities of financial institutions. Financial institutions hope to build a risk mitigation mechanism for the entire credit process, from customer pre-screening, pre-loan review, loan approval, post-loan management to early collection stages.

[0003] If a scoring system can be developed based on the principles of early identification, early warning, early detection and early disposal, and credit business can be monitored and managed quickly and conveniently, the financial institutions' own business volume, competitive advantage and asset quality can be improved while risks are controllable.

[0004] However, the construction of a scoring system is highly dependent on data and technology. The diversity and coverage of data dimensions, modeling techniques and methodology directly affect the final stability and ranking of the scoring system. Some financial institutions have little experience in intelligent risk control of retail business and have weak risk control capabilities. In the actual application process, factors such as lack of data mining and analysis capabilities and weak risk modeling technology have led to financial institutions being unable to fully utilize the value of internal data and effectively improve the accuracy and stability of models, and other control and management technical difficulties. This is also one of the main obstacles faced by small and medium-sized financial institutions in the direction of digital transformation. Summary of the invention

[0005] Internet credit business is a convenient and fast online loan business. Customers can conduct the entire process of online self-service loan business through electronic channels, including real-time application, approval, contract signing, expenditure and repayment, which can all be handled through online channels. In addition to traditional banking institutions, more and more financial institutions are also conducting Internet credit business, such as consumer finance companies, small loan companies, insurance companies and other financial institutions. However, various financial institutions need to continue to improve in order to achieve the rapid transformation of online loan business. Online business is not just about transplanting the registration window to the Internet, but also about achieving pre-examination and pre-submission online. In the case of being unable to verify the real information of customers face to face, it is necessary to use tools that can predict the Internet credit risks of customers to assist in business development. Comprehensive intelligent risk management tools need to be introduced not only before the loan, but also during and after the loan.

[0006] In this context, the credit risk prediction method and system of this application is based on massive data from large banks. It specifically screens customer data with Internet credit business from the massive data, and uses Internet credit business customers in the retail credit customer base as a sample to construct a model for predicting the credit risk of Internet credit business. The model constructed in this application can be applied in a targeted manner in Internet credit business.

[0007] The construction process of other popular scoring models in the market is often limited by unfavorable factors such as small number of modeling samples, relatively single data sources, and high homogeneity of data dimensions. At the same time, since most of the risk assessment models on the market currently use data with weak financial attributes, that is, they are based on non-credit transaction data such as smart terminal device data, social platform data, online shopping mall data, and non-overdue prediction targets for modeling, their prediction results often have a large deviation from the actual credit overdue situation.

[0008] The method and system of the present application are developed based on highly stable and high-coverage data samples, and have made systematic innovations to the relatively mature credit risk control system. At the same time, Internet credit customers in retail credit are used as modeling samples when selecting samples, which can effectively predict the potential credit risks of the retail business of financial institutions, especially the Internet credit business in the retail business, and the credit risk prediction effect is more prominent.

[0009] Compared with existing models on the market, the method and system of the present application have greatly improved the discrimination and stability in predicting Internet credit customers who are overdue for 30 days or more.

[0010] This application involves the following technical solutions:

[0011] 1. A method for calculating retail credit risk, comprising:

[0012] A data collection step, which obtains retail credit prediction data of the sample to be predicted;

[0013] The credit default probability calculation step is to substitute the retail credit prediction data into the credit default probability model to calculate the credit default probability of the sample to be predicted, and preferably the credit default probability is the credit default probability for Internet credit business.

[0014] 2. The method according to claim 1, further comprising:

[0015] After calculating the credit default probability, a step of calculating the credit score of the sample to be predicted is used to calibrate the calculated credit default probability to a standardized score of 0-1000 points.

[0016] 3. The method according to item 1 or 2, wherein:

[0017] The retail credit prediction data includes original retail credit prediction data of the sample to be predicted and derived retail credit prediction data processed based on the original retail credit prediction data;

[0018] Preferably, the original retail credit prediction data includes:

[0019] Credit card basic data, which is based on all available data during the sample user's credit card creation and usage process.

[0020] Basic data on personal loans, which is all available data based on the loan application and usage behavior of sample users.

[0021] Basic customer information data, which is based on the attributes of the sample users themselves but is not directly related to their behavior in financial institutions.

[0022] Basic data on personal financial assets, which includes all other financial assets and financial transaction data of sample users in financial institutions that are not related to credit cards and loans.

[0023] 4. The method according to any one of items 1 to 3, wherein

[0024] The derived retail credit prediction data processed based on the original retail credit prediction data refers to the data obtained by processing the collected original retail credit prediction data based on the time dimension, space dimension, frequency dimension and statistical information dimension;

[0025] Preferably, the derived retail credit forecast data includes but is not limited to:

[0026] Derived retail credit forecast data obtained by processing based on sample relationship length,

[0027] Derivative retail credit forecast data obtained by processing time interval variables,

[0028] Derivative retail credit forecast data obtained by processing based on the frequency of sample behavior,

[0029] Derivative retail credit forecast data obtained by processing the current time point of the sample,

[0030] Derivative retail credit forecast data obtained by processing based on sample continuous behavior,

[0031] Derived retail credit forecast data obtained by processing sample data based on statistical information dimensions.

[0032] 5. The method according to any one of items 1 to 4, wherein

[0033] The retail credit prediction data is selected from one, two, three, four, five or six of the following:

[0034] The customer's region, the remaining available credit limit of the credit card in the current month, whether the customer is a payroll payment customer, the minimum balance of the deposit account at a certain point in time in the past 6 months, the sum of the personal loan credit limit in the current month, and the credit business credit limit utilization rate in the current month.

[0035] 6. The method according to any one of items 1 to 5, wherein

[0036] After performing feature conversion on the retail credit prediction data, the data is substituted into the credit default probability model to calculate the credit default probability of the sample to be predicted. The feature conversion step includes:

[0037] Based on the feature type of the retail credit prediction data that needs to be substituted into the credit default probability model, select the WOE method or the continuous method for feature conversion.

[0038] 7. The method according to claim 6, wherein:

[0039] The continuous method for feature conversion includes the following methods: directly selecting the original value, calculating the square of the original data, calculating the square root of the original data, calculating the cube root of the original data, or calculating the natural logarithm of the original data to perform continuous feature conversion.

[0040] 8. The method according to any one of items 1 to 7, wherein

[0041] The credit default probability model is a model constructed based on sample retail credit prediction data and credit default probability using logistic regression based on the existing user population.

[0042] 9. The method according to any one of items 1 to 8, wherein

[0043] The retail credit prediction data is selected from one, two, three, four, five or six of the following: the customer's region, the remaining available credit limit of the credit card in the current month, whether the customer is a payroll customer, the minimum balance of the deposit account at a certain point in time in the past six months, the sum of the credit limits of personal loans in the current month, and the utilization rate of the credit limit of the credit business in the current month.

[0044] 10. The method according to item 9, wherein the step of calculating the probability of credit default comprises:

[0045] The customer's region, the remaining available credit limit of the credit card in the current month, whether the customer is a payroll customer, the minimum balance of the deposit account at a certain point in time in the past 6 months, the sum of the personal loan credit limit in the current month, and the credit business credit limit utilization rate in the current month are converted into features.

[0046] It is preferred to use the WOE method for conversion based on the customer's region; the continuous conversion method is used for the remaining available credit limit of the credit card in the current month; the WOE method is used for conversion based on whether the customer is receiving payroll; the continuous conversion method is used for the minimum balance of the deposit account at a certain point in time in the past 6 months; the continuous conversion method is used for the sum of the current monthly credit limit of personal loans; and the continuous conversion method is used for the current monthly credit limit utilization rate of credit business;

[0047] Further preferably, the continuous conversion method for the current month's remaining available credit limit of the credit card is a calculation method using the logarithm of the current month's remaining available credit limit of the credit card; the continuous conversion method for the minimum balance of the deposit account at a point in time in the past 6 months is a calculation method using the logarithm of the minimum balance of the deposit account at a point in time in the past 6 months; the continuous conversion method for the sum of the current month's credit limits of personal loans is a calculation method using the cube root of the sum of the current month's credit limits of personal loans; the continuous conversion method for the current month's credit limit utilization rate of the credit business is a calculation method using the square root of the current month's credit limit utilization rate of the credit business.

[0048] 11. The method according to claim 10, wherein:

[0049] The converted values ​​of the six features, namely, the customer's region, the remaining available credit limit of the credit card in the current month, whether the customer is a payroll customer, the minimum balance of the deposit account in the past 6 months, the sum of the personal loan credit limit in the current month, and the utilization rate of the credit limit in the current month, are substituted into the model constructed by logistic regression based on the sample retail credit prediction data and credit default probability to calculate the default probability of the sample to be predicted.

[0050] 12. The method according to claim 11, wherein:

[0051] The model is shown in the following formula 1:

[0052]

[0053] Where k is the number of features entering the model, preferably k is 6,

[0054] α is the intercept, the value range is (-0.1497, 0.2097), and the optimal value is 0.03;

[0055] β1 is the region to which the customer belongs, with a value range of (0.0026, 0.0963) and an optimal value of 0.049;

[0056] β2 is the remaining available credit limit of the credit card in the current month, with a value range of (-0.0483, 0.0566) and an optimal value of 0.004;

[0057] β3 is whether the customer is a payroll agent, the value range is (0.0785, 0.1265), and the optimal value is 0.102;

[0058] β4 is the minimum balance of the deposit account in the past 6 months, with a value range of (-0.1616, 0.1703) and an optimal value of 0.004;

[0059] β5 is the sum of the current monthly credit limits of personal loans, with a value range of (0.0758, 0.1454) and an optimal value of 0.111;

[0060] β6 is the current monthly credit line utilization rate, with a value range of (0.0294, 0.1385) and an optimal value of 0.084;

[0061] x1 is the logarithmic transformation value of the customer's region generated by the feature transformation step;

[0062] x2 is the WOE conversion value of the remaining available credit limit of the credit card in the current month generated in the feature conversion step;

[0063] x3 is the WOE conversion value generated by the feature conversion step, whether the customer is a payroll agent;

[0064] x4 is the logarithmic transformation value of the minimum balance of the deposit account at a certain point in time in the past 6 months generated by the feature conversion step;

[0065] x5 is the cube root conversion value of the sum of the personal loan credit limit for the current month generated in the feature conversion step;

[0066] x6 is the square root conversion value of the current monthly credit limit utilization rate generated by the feature conversion step.

[0067] 13. The method according to claim 12, wherein:

[0068] After calculating the credit default probability, the step of calculating the credit score of the sample to be predicted is to use the following formula to calculate and generate a credit score for characterizing the borrower:

[0069]

[0070]

[0071] Among them, P is the borrower's default probability (P) generated in the credit default probability calculation module, A is 443.9036; B is -72.1348, and the round function rounds the calculated score to the integer value; finally, the score greater than 1000 is set to 1000 points, and the score less than 0 is set to 0 points.

[0072] 14. A device for calculating retail credit risk, comprising:

[0073] A data collection module, which is used to obtain retail credit prediction data of samples to be predicted;

[0074] A credit default probability calculation module is used to substitute the retail credit prediction data into the credit default probability model to calculate the credit default probability of the sample to be predicted, and preferably the credit default probability is the credit default probability for Internet credit business.

[0075] 15. The apparatus according to item 14, wherein the apparatus executes the steps of the method for calculating retail credit risk according to any one of items 1 to 13.

[0076] 16. A system for calculating retail credit risk, characterized in that the system for calculating retail credit risk comprises: a memory, a processor, and a program for the method for calculating retail credit risk stored in the memory and executable on the processor, wherein the program for calculating retail credit risk implements the steps of the method for calculating retail credit risk as described in any one of items 1 to 13 when executed by the processor.

[0077] 17. A computer storage medium, characterized in that a program for calculating retail credit risk is stored on the computer storage medium, and when the program for calculating retail credit risk is executed by a processor, the steps of the method for calculating retail credit risk as described in any one of items 1 to 13 are implemented.

[0078] 18. A method for constructing a retail credit risk prediction model for Internet credit business, comprising:

[0079] A data collection step, which obtains the original retail credit prediction data of the sample used to build the model;

[0080] A data derivation step, which processes derived retail credit forecast data based on the original retail credit forecast data;

[0081] The feature initial screening step is to perform initial screening on all categories, that is, all features, including the original retail credit forecast data and the derived retail credit forecast data, to obtain features after initial screening;

[0082] The initial screening data conversion step is to judge the conversion method of the features after the initial screening to confirm whether to use one of the WOE conversion method, the dummy feature conversion method and the continuous conversion method for feature conversion, and use the best method to convert each feature after the initial screening;

[0083] The feature fine screening step is to perform in-depth screening on the initially screened features after feature conversion to obtain finely screened features;

[0084] The credit default probability modeling step is to select the logistic regression method to build the model based on the relationship between the carefully screened features and the probability of credit default, and confirm the method used to calculate the credit default probability;

[0085] The data samples obtained in the data collection step are samples of customers who have used Internet credit services.

[0086] 19. The method according to claim 18, wherein:

[0087] In the data collection step, the original retail credit prediction data of the samples obtained for building the model include:

[0088] Credit card basic data, which is based on all available data during the sample user's credit card creation and usage process.

[0089] Basic data on personal loans, which is all available data based on the loan application and usage behavior of sample users.

[0090] Basic customer information data, which is based on the attributes of the sample users themselves but is not directly related to their behavior in financial institutions.

[0091] Basic data on personal financial assets, which includes all other financial assets and financial transaction data of sample users in financial institutions that are not related to credit cards and loans.

[0092] 20. The method according to claim 18, wherein:

[0093] In the data derivation step, the derived retail credit prediction data processed based on the original retail credit prediction data refers to data obtained by processing the collected original retail credit prediction data based on the time dimension, space dimension, frequency dimension, and statistical information dimension;

[0094] Preferably, the derived retail credit forecast data includes but is not limited to:

[0095] Derived retail credit forecast data obtained by processing based on sample relationship length,

[0096] Derivative retail credit forecast data obtained by processing time interval variables,

[0097] Derivative retail credit forecast data obtained by processing based on the frequency of sample behavior,

[0098] Derivative retail credit forecast data obtained by processing the current time point of the sample,

[0099] Derivative retail credit forecast data obtained by processing based on sample continuous behavior,

[0100] Derived retail credit forecast data obtained by processing sample data based on statistical information dimensions.

[0101] 21. The method according to any one of items 18 to 20, wherein

[0102] The feature initial screening step includes the following steps:

[0103] The first preliminary screening step is to screen the features based on the missing data of each feature of the sample used to build the model.

[0104] The second initial screening step is to screen the features based on the fact that a single value of a feature sample is too high.

[0105] The third preliminary screening step is to calculate the information IV value of each feature to perform preliminary screening of the features;

[0106] The order of the first preliminary screening step, the second preliminary screening step and the third preliminary screening step can be any order.

[0107] The fourth preliminary screening step uses a stepwise discrimination algorithm to perform preliminary screening of features after the first to third preliminary screenings;

[0108] The fifth preliminary screening step is to conduct preliminary screening of the features after the fourth preliminary screening step based on the consistency between the risk characteristics of each feature itself and the actual real results of the samples used for model construction.

[0109] 22. The method according to any one of items 18 to 21, further comprising:

[0110] The sample selection step is used to screen all users before the data collection step and obtain samples of Internet credit business for model construction.

[0111] 23. A method according to any one of items 18 to 22, wherein, in the initial screening data conversion step, the conversion method of the features after the initial screening is determined based on the concentration and data type of the features after the initial screening.

[0112] 24. The method according to claim 23, wherein:

[0113] The initial screening data conversion step includes the following steps based on the judgment of concentration and data type:

[0114] Classify the data type of each feature into character variables and numeric variables.

[0115] For character variables, dummy feature conversion is used to perform initial screening data conversion.

[0116] The process of further classifying numerical variables includes the following sub-steps:

[0117] If the value of the numeric variable is less than n, the WOE conversion method is used to convert the initial screening data.

[0118] If the value of the numerical variable is more than n, further judge if the value of the continuous variable is large and the concentration of a single value is greater than m%, then the WOE conversion method is used; if the concentration of a single value is less than or equal to m%, then the continuous conversion method is used.

[0119] Preferably, n and m are both positive integers, wherein n=5-10, and m=90-99.

[0120] 25. The method according to claim 24, further comprising:

[0121] For the features that are confirmed to adopt the continuous conversion method, the optimal conversion method is selected based on the correlation between the feature and credit default under different continuous conversion methods to perform the continuous feature conversion of the feature.

[0122] Preferably, the continuous feature conversion is performed by directly selecting the original value, calculating the square of the original data, calculating the square root of the original data, calculating the cube root of the original data, or calculating the natural logarithm of the original data.

[0123] 26. The method according to any one of items 18 to 25, wherein the feature fine screening step comprises:

[0124] The first fine screening step is to screen the features based on the stepwise regression algorithm, the F test and the T test to determine the significance of the features.

[0125] The second fine screening step is to calculate the variance inflation factor based on each feature and eliminate the features with higher variance inflation factors to screen the features.

[0126] The third fine screening step is to analyze the characteristics after the first fine screening step and the second fine screening step based on logistic regression to see whether the characteristic coefficients are consistent with the trend of the prediction results for credit default so as to further perform feature screening.

[0127] 27. A method according to any one of items 18 to 26, wherein the credit default probability modeling step substitutes the features screened in the feature fine screening step into the Sigmoid function to perform logistic regression to calculate the credit default probability model.

[0128] 28. A device for constructing a retail credit risk prediction model for Internet credit business, characterized in that the device comprises:

[0129] A data collection module, which is used to obtain raw retail credit prediction data of samples used to build a model;

[0130] A data derivation module, which is used to process derived retail credit forecast data based on the original retail credit forecast data;

[0131] A feature initial screening module, which is used to perform initial screening on all categories, that is, all features, including the original retail credit forecast data and the derived retail credit forecast data, to obtain features after initial screening;

[0132] The initial screening data conversion module is used to determine the conversion method of the features after the initial screening to confirm whether to use one of the WOE conversion method, the dummy feature conversion method and the continuous conversion method for feature conversion, and to use the best method for feature conversion for each feature after the initial screening;

[0133] A feature fine screening module is used to perform in-depth screening on the initially screened features after feature conversion to obtain finely screened features;

[0134] The credit default probability modeling module is used to select the logistic regression method to build a model based on the probability relationship between the finely screened features and credit default, and confirm the method used to calculate the credit default probability.

[0135] The data samples obtained by the data collection module are samples of customers who have used Internet credit services.

[0136] 29. An apparatus according to item 28, wherein the apparatus executes the steps of the method for constructing a retail credit risk prediction model as described in any one of items 18 to 27.

[0137] 30. A system for constructing a retail credit risk prediction model for Internet credit business, characterized in that the system includes: a memory, a processor, and a program for constructing a retail credit risk prediction model method stored in the memory and executable on the processor, wherein the program for constructing a retail credit risk prediction model method, when executed by the processor, implements the steps of constructing a retail credit risk prediction model method as described in any one of items 18 to 27. BRIEF DESCRIPTION OF THE DRAWINGS

[0138] Figure 1 This is a schematic diagram of the KS curve with a KS value of 58.

[0139] Figure 2 An example of binning to confirm whether the data conforms to business trends is shown. DETAILED DESCRIPTION

[0140] Credit risk is the risk caused by the borrower's reduced willingness to repay or inability to perform the loan contract due to changes in economic ability, rather than the risk of default caused by the borrower's deliberate fraud. Credit default occurs in all kinds of retail credit business scenarios, and is particularly related to the borrower's personal economic situation. The reasons for borrowers' credit defaults can be divided into four categories: 1. Short credit history, such borrowers have little experience in financial management; 2. Borrowers temporarily forget to repay; 3. Over-borrowing, such borrowers have relatively low repayment ability due to a large amount of debt; 4. Affected by major negative factors, such borrowers have long-term impacts on their repayment ability due to major factors such as reduced income, unemployment or divorce. The above different reasons will more or less cause borrowers to default, which may lead to more serious defaults. The purpose of credit risk scoring is to dig out the inherent mathematical relationship between various historical information of customers and the probability of future default, and convert it into a scoring method to quantify the probability of default.

[0141] At present, the credit risk scoring in the existing technology is mainly developed using credit history data and multi-head data (multi-head data refers to the statistical data of borrowers' behavior of applying for loan demands from multiple financial institutions. It is generally believed that the greater the number of multi-heads in a short period of time, the greater the probability of future overdue payments), which reflects their payment behavior and willingness to pay. The data of the samples used to build the model in this application not only includes transaction behavior information of credit business, but also adds asset data that is usually difficult to obtain. While reflecting the borrower's payment behavior and willingness to pay, it also conducts a more comprehensive measurement of their debt repayment ability, personal qualifications, etc., thereby providing more accurate prediction results.

[0142] <Overall description of model building method>

[0143] Specifically, the present application relates to a method for constructing a retail credit risk prediction model for Internet credit business, which includes: a data collection step, which obtains the original retail credit prediction data of the sample used to build the model; a data derivation step, which processes the derived retail credit prediction data based on the original retail credit prediction data; a feature initial screening step, which performs preliminary screening on all categories including the original retail credit prediction data and the derived retail credit prediction data, that is, all features, to obtain the features after the preliminary screening; a preliminary screening data conversion step, which determines the conversion method of the features after the preliminary screening to confirm that one of the WOE conversion method, the dummy feature conversion method and the continuous conversion method is used for feature conversion, and the best method is used for feature conversion for each feature after the preliminary screening; a feature fine screening step, which performs in-depth screening on the features after the preliminary screening after feature conversion to obtain the finely screened features; a credit default probability modeling step, which selects the logistic regression method for model construction based on the finely screened features combined with the probability relationship between the features and the credit default, and confirms the method for calculating the credit default probability.

[0144] In a specific manner, the method for constructing a retail credit risk prediction model for Internet credit business involved in the present application may also include a sample selection step before the data collection step, which is used to screen all users before the data collection step to obtain samples for model construction. Those skilled in the art understand that based on the total number and data of users who can build the model, those skilled in the art can choose whether to eliminate user data that does not meet the requirements for building the model. In addition, those skilled in the art can also first select some users as samples, and then continuously increase the samples for modeling according to the corresponding rules.

[0145] In a specific implementation, the selected samples are customer samples of Internet credit services among customer samples that applied for credit services registration within a certain period of time.

[0146] <Sample selection steps>

[0147] In a specific implementation, the sample selection step is to screen all users before the data collection step to obtain samples for model construction. Specifically, the sample selection step includes classifying all sample users based on a decision tree, and the classification basis includes but is not limited to: whether a certain user is a customer who has applied for credit business registration at a financial institution; whether a certain user is a customer with Internet credit business.

[0148] Furthermore, all sample users are classified based on the decision tree. Since performance variables are required for modeling, the analysis samples are further divided into whether they have applied for credit business registration at the relevant bank, that is, users can be divided into those who have applied for registration and those who have not.

[0149] Credit business is a basic and important asset business of banks. It recovers the principal and interest by issuing bank loans and obtains profits after deducting costs. Generally speaking, bank credit business is an important means for banks to make profits. From the classification of bank credit business, it can be divided into corporate credit business and personal credit business. Among them, corporate credit business includes project loans, working capital loans, small business loans, real estate enterprise loans, etc.; personal credit business includes personal housing loans, personal consumption loans, personal business loans, etc.

[0150] A decision tree is a decision analysis method that uses a decision tree to obtain the probability that the expected value of the net present value is greater than or equal to zero, evaluate project risks, and judge its feasibility based on the known probabilities of various situations. It is a graphical method that intuitively uses probability analysis. When constructing the model, this application uses a decision tree to classify samples used for modeling based on the characteristics of data from large financial institutions. After distinguishing between samples that have applied for credit business and samples that have not applied for credit business, the samples that have applied for credit business can be further classified based on whether they are new customers, and the samples that have applied for credit business can be classified into two categories: customers with Internet credit business and customers without Internet credit business.

[0151] Those skilled in the art can fully understand that other splitting methods can also be selected when performing sample splitting. For example, all samples can be first split into customer groups with credit cards and customer groups without credit cards. The customer sample splitting method when building the model can be completely considered and carried out according to the needs of modeling.

[0152] In the process of constructing this application model, since it is necessary to build a model for Internet credit business, the selected samples are customer samples of Internet credit business among the customer samples that applied for registration of credit business within a certain period of time.

[0153] In a specific implementation, 220,000 customers with Internet credit services are selected from 130 million registered credit services as the modeling sample. Such a sample is more targeted and representative when predicting the credit risk of Internet credit services.

[0154] <Original Retail Credit Forecast Data>

[0155] When constructing the model for this application, firstly, based on the basic data of financial institutions, about 120 types of original retail credit prediction data (i.e. more than 120 basic features) can be obtained, and the retail credit risk points are split or processed as much as possible according to different dimensions under the premise of optimal effect, generating more than 3,000 derivative features in total.

[0156] When constructing the model, first of all, based on all historical data of large financial institutions, the credit card application and behavior data, personal loan application and behavior data, personal financial asset transaction data and personal customer information data of all retail customers over the past four years were initially collected when constructing the model of this application, totaling 730 million people, of which each person has corresponding data to be processed every month. It can be seen that the data system used to construct the model of this application is comprehensive and the data volume is very large. When constructing a model based on such a data system, it is necessary to consider the modeling methodology, otherwise it will be trapped in the huge data, resulting in the special sample groups that need to be paid attention to being covered in the huge data volume and unable to be effectively identified, causing the computer program to run slowly or even unable to run, so that the most suitable prediction model cannot be accurately constructed.

[0157] In the data collection step, the original retail credit prediction data of the sample obtained for building the model include: basic data of credit cards, which are all available data based on the sample's credit card creation process and use process (i.e., retail customers' credit card application and behavior data); basic data of personal loans, which are all available data based on the sample's loan application and usage behavior (i.e., personal loan application and behavior data); basic data of basic customer information, which are data based on the attributes of the sample itself but not directly related to the behavior in the financial institution (i.e., personal customer information data); or basic data of personal financial assets, which are all other financial assets and financial transaction data of the sample in the financial institution that are not related to credit cards and loans (personal financial asset transaction data).

[0158] In a specific implementation, the basic data of credit cards include, but are not limited to: credit card bills, credit card cash withdrawals, credit card installments, interest generated by credit cards, and other information such as accounts, overdue, balances, credit limits, due payments, and actual payments under different dimensions (e.g., time dimension, space dimension, and frequency dimension). The above basic data is not limited to the specific categories listed. With the changes in the credit card business, those skilled in the art can further cover new data types in implementation, that is, all types of data that can be obtained during the user's credit card creation and use process can be used as basic data of credit cards.

[0159] In a specific implementation, basic data of personal loans include but are not limited to: personal loan accounts, overdue personal loans, balances of personal loans, total amount of personal loans, repayments due and actual repayments of personal loans, etc. The above basic data are not limited to the specific categories listed. With the changes in personal loan business, those skilled in the art can further cover new data types in implementation, that is, all types of data that can be obtained based on the user's loan application and behavior can be used as basic data of personal loans.

[0160] In a specific implementation, according to the requirements of the Personal Privacy Protection Act, the basic data of the basic information category of customers only includes: basic information of customers, including gender, age and administrative region information of the business. The above basic data is not limited to the specific categories listed. With the development of customer conditions and social relations, technical personnel in this field can also further cover other or emerging data types in the implementation within the scope of the business application scenario, that is, all types of data based on the attributes of the user sample itself but not directly related to the behavior in the financial institution can be used as basic data of the basic information category of customers.

[0161] In a specific implementation, basic data of personal financial assets include but are not limited to: AUM (i.e., asset management scale), deposits, financial management, and payroll information. The above basic data are not limited to the specific categories listed. With the changes in financial assets, those skilled in the art can also further cover new data types in implementation, that is, all other financial assets and financial transaction data that are not related to credit cards and loans in financial institutions based on samples can be used as basic data of personal financial assets.

[0162] In a specific embodiment, the original retail credit prediction data of the samples obtained for building the model are based on more than 120 data types (i.e., basic variables or basic features) obtained from 730 million people, including but not limited to: credit card bills, credit card withdrawals, credit card installments, interest generated by credit cards, and other accounts, overdue, balances, amounts, due and actual repayments under different dimensions (e.g., time dimension, space dimension, frequency dimension); personal loan accounts, overdue personal loan status, personal loan balances, total amount of personal loans, due and actual repayments of personal loans; basic customer information including gender, age and administrative region to which the business belongs; basic information such as AUM (i.e., asset management scale), deposits, wealth management and payroll payment.

[0163] <Derived Retail Credit Forecast Data>

[0164] In the present application, in the data derivation step, the derived retail credit prediction data processed based on the original retail credit prediction data refers to the data obtained by processing the collected original retail credit prediction data based on the time dimension, space dimension, frequency dimension, and statistical information dimension.

[0165] In a specific embodiment, the derived retail credit prediction data includes, but is not limited to: derived retail credit prediction data obtained by processing based on the sample relationship length, derived retail credit prediction data obtained by processing based on time interval variables, derived retail credit prediction data obtained by processing based on the frequency of sample behavior, derived retail credit prediction data obtained by processing based on the current time point of the sample, derived retail credit prediction data obtained by processing based on the continuous behavior of the sample, or derived retail credit prediction data obtained by processing sample data based on statistical information dimensions. For example, the monthly data of the customer can be obtained and processed based on the monthly data. In this application, processing based on statistical information dimensions includes obtaining the maximum value, minimum value, average value, etc. of the data to describe the data situation.

[0166] In a specific implementation, for example, starting from the time dimension, such as customer relationship length variables: such as the customer account opening time, customer maximum account age, etc. are used as types of derived retail credit prediction data, that is, as derived features or derived variables.

[0167] In a specific implementation, for example, starting from the time dimension, time intervals are considered, such as the number of months from the customer's most recent repayment to the current time point, the number of months from the customer's most recent overdue payment to the current time point, etc. as derived features or derived variables.

[0168] In a specific implementation, for example, starting from the frequency level, the behavior frequency level variable is considered: such as the number of times the customer has repaid > N in the last X months, the number of times the customer's credit limit utilization rate > N in the last X months, etc. as derived features or derived variables. There is no limit on X and N, and they can be any positive integer greater than 0 as long as the business logic is reasonable.

[0169] In a specific implementation, for example, starting from the time dimension, the current time point variables are considered: the customer's current monthly credit limit, the customer's current monthly balance, etc. as derived features or derived variables.

[0170] In a specific approach, starting from the time dimension, continuous behavior variables are considered as derived features or derived variables, such as the maximum number of consecutive overdue payments of a customer > N in the last X months, the number of consecutive repayment rates of a customer > N in the last X months, etc.

[0171] In a specific method, starting from the dimension of statistical information, statistical variables are considered, such as the maximum number of overdue payments of a customer in the last X months, the average credit limit utilization rate of a customer in the last X months, etc. as derived features or derived variables.

[0172] It is clear to those skilled in the art that the above methods for processing derived variables are merely enumerated and can be selected arbitrarily. For example, 120 types of basic data can be processed into more than 3,000 derived data. In this application, data, data types, data types, variables or features are sometimes mixed, and those skilled in the art can understand them based on common sense in statistics.

[0173] In this application, the derivative data can be derived data obtained by simple processing of the original retail credit forecast data, or it can be derived data obtained by complex processing. The simply processed derivative data, such as the current variables such as the salary of the current month and the balance of the current month, can be used directly after the data is aggregated. The complex processed derivative data needs to be based on the variables of the current class based on time slicing and logical processing, and can generate derivative variables such as the maximum salary in the past 3 months and the minimum balance in the past 12 months.

[0174] <Preliminary feature screening>

[0175] In the initial screening data conversion step of the present application, the judgment of the conversion method of the features after the initial screening is based on the concentration and data type of the features after the initial screening. The initial screening data conversion step includes the following steps based on the judgment of the concentration and data type: classify each feature into character variables and numerical variables according to the data type of each feature, use the dummy feature conversion method to perform initial screening data conversion on the character variables, and further classify the numerical variables. The process includes the following sub-steps: if the value of the numerical variable is less than n, use the WOE conversion method to perform initial screening data conversion, if the value of the numerical variable is more than n, further judge if the value of the continuous variable is more and the concentration of the single value is greater than m%, then use the WOE conversion method, if the concentration of the single value is less than or equal to m%, then use the continuous conversion method, preferably, n and m are both positive integers, where n = 5 to 10, m = 90 to 99.

[0176] For example, n is 5, 6, 7, 8, 9 or 10, and m is 90, 91, 92, 93, 94, 95, 96, 97, 98 or 99.

[0177] Specifically, in this application, a variety of different feature initial screening methods are used to screen a large number of features, so that the screening can be effectively performed from the largest dimension of features. Existing credit risk scoring often uses missing rate, concentration and information value IV for feature initial screening.

[0178] In a specific method, features are screened based on the data missingness of each feature of the sample used to build the model. When screening the missing rate, it is generally considered to delete variables with a missing rate greater than 90%, 91%, 92%, 93%, 94%, or 95%. For example, features with a data missing rate exceeding 95% can be eliminated, and features with a data missing rate exceeding 90% can also be eliminated.

[0179] In a specific method, features are screened based on the fact that a single value of a feature sample is too high. When screening by concentration, it is generally considered to delete variables whose single values ​​account for more than 99%, 98%, 97%, 96%, or 95%. For example, features whose single values ​​exceed 99% can be eliminated, and features whose single values ​​exceed 95% can also be eliminated.

[0180] In a specific method, the information IV value of each feature is calculated to perform preliminary screening of the features. During IV screening, the IV value can be used to measure the predictive power of the feature. The larger the IV value, the stronger the predictive power of the feature. The calculation method of the IV value of a single feature is as follows:

[0181]

[0182] Where k is the number of groups after this feature is discretized; y i is the number of non-defaulting customers in group i; s is the total number of customers who have not defaulted; n i is the total number of defaulting customers in group i; n s The quantitative index of IV has the following meanings: when the calculated value of IV is less than 0.02, it means that the predictive power of this feature is very weak; when the calculated value of IV is above 0.02 and below 0.1, it means that the predictive power of this feature is weak; when the calculated value of IV is above 0.1 and below 0.3, it means that the predictive power of this feature is good; when the calculated value of IV is above 0.3, it means that the predictive power of this feature is strong.

[0183] Of course, the IV calculated value deletion threshold may also be set to 0.03, 0.04, 0.05, etc.

[0184] In the model building method of the present application, the step of using missing rate, concentration and information value IV for feature initial screening can be performed in any order, for example, the screening can be performed based on the missing rate first, then based on the concentration, and finally based on the information value IV for feature initial screening. It is also possible to first screen based on the concentration, then screen based on the missing rate, and finally based on the information value IV for feature initial screening. It is also possible to first screen based on the missing rate, then screen based on the information value IV, and finally based on the concentration for feature initial screening. It is also possible to first screen based on the information value IV, then screen based on the missing rate, and finally based on the concentration for feature initial screening. It is also possible to first screen based on the information value IV, then screen based on the missing rate, and finally based on the concentration for feature initial screening. It is also possible to first screen based on the information value IV, then screen based on the concentration, and finally based on the missing rate for feature initial screening. It is also possible to first screen based on the concentration, then screen based on the information value IV, and finally based on the missing rate for feature initial screening. Those skilled in the art can make a choice based on the data situation of the sample, so based on these three methods, features with obvious defects in one aspect can be effectively removed, which can effectively reduce the data dimension and improve the effect of model building.

[0185] In one specific method, after the features have been screened using missing rate, concentration and information value IV respectively, a stepwise discriminant algorithm is used to perform preliminary screening of the features, and then preliminary screening of the features is performed based on the consistency between the risk characteristics of each feature itself and the actual real results of the samples used to build the model.

[0186] On this basis, the technical solution of the present application introduces a stepwise discriminant method to make the initial screening of variables more efficient and accurate in order to improve the overall efficiency of model development. In actual data, there may be a close distribution of good and bad samples on a certain variable, and the ability to distinguish between good and bad is weak. There may also be a class of variables, each of which can distinguish good and bad samples well, but if all are included in the model, they will be redundant due to the duplication of data dimensions covered by the variables. In this regard, a stepwise discriminant analysis method is used in this application, and WILKS'S LAMBDA value is used as the statistical criterion for entry and removal, and variables with weak or redundant discriminant effects in the data are deleted.

[0187] The stepwise discrimination method uses the WILKS'S LAMBDA criterion to measure the strength of feature capabilities and eliminates features that do not meet the set threshold among the remaining features after three rounds of screening. In the stepwise discrimination process, the variable with the strongest discriminant ability is added first. As the variables in the model gradually increase, the discriminant ability of the variables introduced earlier may also change. If the discriminant ability of a variable in the model is less than the threshold, the variable is removed, and then the process is repeated until the variables contained in the model meet the WILKS'S LAMBDA similarity ratio criterion and other variables do not meet the criteria for entering the model.

[0188] In this application, the step of screening using the stepwise discriminant method is a more critical step. For the purpose of capturing as many credit risk points as possible, the model involved in this application uses a huge amount of data and many data dimensions, so variable screening must be performed to further reduce the time cost of model development. In the existing financial risk model evaluation, the variable importance method is generally used to reduce variables, such as calculating the importance of variables through algorithms such as the Gini index and information entropy, and selecting variables with high importance. The modeling method of screening using the stepwise discriminant method is rarely used. Compared with the variable importance screening scheme commonly used in the industry, the methodology used in this application can retain a large number of variables with relatively weak importance but relatively independent information dimensions.

[0189] In a specific way, more than 3,000 derivative data can be processed based on, for example, 120 types of basic data. After three rounds of screening of missing rate, concentration and information value IV, about 20-30% of the features with poor data can be deleted, but more than 2,000 features will still be retained. If subsequent variable fine screening is performed based on all the above features, it will seriously affect the development efficiency. Therefore, after comparing and testing a variety of different variable deletion schemes, it was finally determined that stepwise discrimination was the best choice.

[0190] For the important features selected by the stepwise discrimination algorithm, further feature selection is performed based on the risk characteristics of each risk point itself (system preset) and the actual bad debt rate of the sample. Determine whether the actual bad debt rate distribution of the remaining features is consistent with the business trend, and eliminate the features whose actual bad debt rate distribution does not conform to the business trend. Specifically, (1) Divide the remaining target features into 10-20 boxes according to the quantile points, and calculate the median value of each box and the corresponding bad debt rate. (2) Use the median value of each box and the previous box and the corresponding bad debt rate to calculate the rate of change (slope). (3) Count the number of boxes greater than 0 and non-0 boxes (non-greater than 0 boxes) in the rate of change between two adjacent boxes, and calculate the percentage of the number of boxes greater than 0 in the rate of change to the number of non-0 boxes in the rate of change. (4) Obtain the risk characteristics preset for this feature, and based on the percentage of the number of boxes greater than 0 in the rate of change to the number of non-0 boxes in the rate of change, check whether the actual bad debt rate change trend is approximately consistent with the business logic bad debt rate change trend. The criteria for whether they are approximately consistent are as follows: First, if the business logic trend of this feature should be: as the value of this feature increases, the bad debt rate increases (such as the credit line utilization rate, etc.). In this module, the features whose number of boxes greater than 0 in the above calculated slope accounts for less than 70% of the number of boxes with a slope other than 0 are eliminated, that is, the features whose feature performance does not conform to the business trend are eliminated. Second, if the business logic trend of this feature should be: as the value of this feature increases, the bad debt rate decreases (such as the deposit amount, etc.). In this module, the features whose number of boxes greater than 0 in the above calculated slope accounts for more than 30% of the number of boxes with a slope other than 0 are eliminated, that is, the features whose feature performance does not conform to the business trend are eliminated.

[0191] For example, the following table gives an example of binning. In this example, the feature values ​​are divided into 10 bins. The value ranges of the bins are also summarized in the following table. At the same time, based on the data in the following table, a graph of the bad debt rate of the feature bins versus the median of the feature bins is drawn, as shown in the following figure: Figure 2 As shown. Figure 2 In the example, there are 7 with slopes greater than 0 and 2 with slopes less than 0. Through the above method, the features can be further screened based on the risk characteristics of each risk point itself (system preset) and the actual bad debt rate of the sample. By further screening features using such a binning method, the features that best match the business development trend can be effectively screened out, thereby further obtaining features suitable for modeling.

[0192]

[0193] <Conversion of eigenvalues>

[0194] The initial feature processing of existing credit risk scoring methods mainly uses WOE conversion (multi-classification) and dummy feature conversion (binary classification) to discretize continuous features (such as age, account age, etc.). The WOE conversion method is an optimal binning scheme based on modeling samples. It discretizes continuous features according to the optimal cutting point and embeds the data of good and bad samples into the WOE value. Therefore, it performs better in model construction, but the binning results may overfit the modeling samples, resulting in a serious decline in model effect when applied to the whole (poor generalization ability). At the same time, due to the normalization operation used in binning, the original features falling into different bins are converted into a single value corresponding to each bin, thus losing the ability to distinguish the risks of people falling into the same interval. The dummy feature conversion method is mainly applied to grouping features. Its advantage is that it can eliminate the difference between good and bad values ​​of different feature values, but it will become extremely complex and redundant when processing continuous variables.

[0195] In the model building method of the present application, the conversion method of the features after preliminary screening is judged to confirm whether to use one of the WOE conversion method, the dummy feature conversion method and the continuous conversion method for feature conversion, and the optimal method is used to perform feature conversion for each feature after preliminary screening.

[0196] The initial screening data conversion step includes the following steps based on the judgment of concentration and data type: classifying each feature into character variables and numeric variables according to the data type of each feature, using dummy feature conversion method to perform initial screening data conversion on character variables, and further classifying the numeric variables includes the following sub-steps: if the value of the numeric variable is less than n, the WOE conversion method is used for initial screening data conversion; if the value of the numeric variable is more than n, further judging that if the value converted to a continuous variable is large and the concentration of a single value is greater than m%, the WOE conversion method is used; if the concentration of a single value is less than or equal to m%, a continuous conversion method is used; preferably, n and m are both positive integers, wherein n=5-10, m=90-99.

[0197] Specifically, taking the character variable as education level, the value of this feature variable can be primary school, middle school, university, graduate school, etc. For numeric variables, if the numeric variable is the number of overdue months in the past three months, the value is 0, 1, 2, 3.

[0198] In the present application, the above n can be 5, 6, 7, 8, 9 or 10, and m can be 90, 91, 92, 93, 94, 95, 96, 97, 98 or 99.

[0199] In a specific implementation, m is selected as 5 and n is selected as 95.

[0200] Specifically, the WOE conversion method is: find the optimal cutting point of the feature, divide the value range of the original feature into multiple bins, calculate the WOE conversion value corresponding to each bin after segmentation according to the good or bad situation in each bin and output it, divide the original feature according to the binning result and output the WOE conversion value. For each bin, the WOE value is calculated as follows:

[0201]

[0202] where y i is the number of non-defaulting customers in group i; s is the total number of customers who have not defaulted; n i is the total number of defaulting customers in group i; n s is the total number of defaulting customers.

[0203] Let's take age as an example to introduce the WOE binning process. The original feature age includes values ​​from 18 to 50 years old. After binning, 5 bins are obtained, namely 18 to 24 years old, 25 to 30 years old, 31 to 35 years old, 36 to 42 years old, and 43 to 50 years old. Then, the WOE conversion value of each bin is calculated according to the number of good and bad customers in each bin. Finally, each user data in each bin is output according to the corresponding conversion value of each bin. For example, 23 years old outputs the WOE conversion value corresponding to the 18 to 24 bins, and 46 years old outputs the WOE conversion value corresponding to the 43 to 50 bins.

[0204] The conversion method of dummy features is: convert a single classification feature into an equal number of dummy features according to the number of values ​​it contains. If a customer belongs to the corresponding value of the generated dummy feature, the corresponding dummy feature value is 1, and the remaining dummy feature values ​​are 0.

[0205] Taking gender as an example, we will introduce the method of dummy feature conversion. The original features include: male, female. After dummy feature conversion, two dummy features, 'gender-male' and 'gender-female', are generated. If the customer's gender is male, 'gender-male' is recorded as 1, and 'gender-female' is recorded as 0. If we take education as an example, the original features include: junior college and below, undergraduate, master and above. After dummy feature conversion, three dummy features, 'education-junior college and below', 'education-bachelor', and 'education-master and above', are generated. If the customer's education is undergraduate, 'education-junior college and below' is recorded as 0, 'education-bachelor' is recorded as 1, and 'education-master and above' is recorded as 0.

[0206] The continuous conversion method is: perform a variety of continuous conversions on the original features (continuous conversion methods include but are not limited to: directly selecting the original value, calculating the square of the original data, calculating the square root of the original data, calculating the cube root of the original data, and calculating the natural logarithm of the original data). Calculate the correlation coefficient r (Correlation Coefficient) between the feature value and the overdue label after continuous conversion, select the conversion method with the largest absolute value of the correlation coefficient, and output the conversion method corresponding to the original feature. The calculation formula of the correlation coefficient is as follows:

[0207]

[0208] Where Σ is the summation symbol in mathematics; n is the total number of observations; x i is the conversion value of the original feature of the i-th observation after continuous conversion; This is the average of the transformed values; where y i is a binary feature indicating whether the i-th observation is in default; It is the average value of this binary feature. The closer the absolute value of the correlation coefficient is to 1, the more relevant the converted value is to the default situation, and the better the effect of this conversion method. The quantitative index of the correlation coefficient r has the following meanings: when the absolute value of the correlation coefficient is above 0 and less than 0.3, it indicates a low correlation; when the absolute value of the correlation coefficient is above 0.3 and less than 0.8, it indicates a moderate correlation; when the absolute value of the correlation coefficient is above 0.8 and less than 1, it indicates a high correlation.

[0209] In the model construction of the present application, for the features that are confirmed to adopt a continuous conversion method, the optimal conversion method is selected to perform continuous feature conversion of the feature based on the correlation of the feature with credit default under different continuous conversion methods. Preferably, the continuous feature conversion is performed by directly selecting the original value, calculating the square of the original data, calculating the square root of the original data, calculating the cube root of the original data, or calculating the natural logarithm of the original data.

[0210] Take age as an example to introduce the continuous conversion method. The original feature age contains values ​​from 18 to 50 years old. The continuous conversion obtains the original value, square, square root, cube root, natural logarithm, etc. of age, and then calculates the WOE value and the absolute value of the correlation coefficient between each conversion value and the good or bad label (prediction result). The conversion method with the largest absolute value of the correlation coefficient is selected for conversion and output. If the cube root of the age feature has the largest absolute value of the correlation coefficient with the good or bad label compared with other conversion methods, the cube root of age is output as the conversion value.

[0211] In the prior art solutions, one of the feature conversion methods, WOE and dummy features, is mainly selected for model construction. For example, in the Chinese patent CN112686749B, the WOE method is used to convert the feature values. The WOE conversion method is an optimal binning scheme based on the modeling sample. It discretizes the continuous features according to the optimal cutting point, so it performs better in model construction, but the binning results may overfit the modeling samples, resulting in a serious decrease in the model effect when applied to the whole (poor generalization ability). At the same time, due to the normalization operation used in the binning, the original features falling into different bins are converted into a single value corresponding to each bin, so the risk differentiation ability of people falling into the same interval is lost.

[0212] The conversion method of dummy features is mainly used for grouping features. Its advantage is that it can eliminate the difference between good and bad values ​​of different feature values. For example, for the industry in which the customer is located, because there is no obvious advantage or disadvantage between the retail industry and the wholesale industry, it is more suitable to use the dummy feature processing method; there is a hierarchical difference between junior colleges and undergraduates, so although dummy features can be used, it is actually more appropriate to use the WOE processing method.

[0213] On the other hand, the continuous conversion method can avoid the overfitting of the modeling samples in the WOE conversion method, has a strong generalization ability for the overall sample, and because no interval mapping is performed, most customer groups rarely fall into a single value. However, it cannot be applied to some features with poor monotonicity or discrete characteristics (such as occupation, position, etc.).

[0214] As described above, the technical solution of the present application takes a different approach and creatively combines three methods: continuous conversion, WOE conversion, and dumb feature conversion. It reprocesses some carefully screened features, combines the advantages and disadvantages of the three conversion methods, and creatively designs a conversion judgment method. It selects the optimal feature conversion method based on parameters such as feature data attributes, missing rate, concentration, and business logic judgment.

[0215] <Logistic regression and deep feature screening based on logistic regression>

[0216] In existing credit risk scoring, due to the requirement of model interpretability, logistic regression model is mainly used for model development, and the software that can be used is generally: SAS, R, Python, etc.

[0217] In a specific embodiment, the present application develops models based on SAS software.

[0218] Specifically, the Sigmoid function is used in logistic regression to fit the probability of predicted default. The Sigmoid function is:

[0219]

[0220] Where Z is a linear combination of the model coefficients and the feature transformation values, and Z is defined as follows:

[0221] Z=α+β1x1+β2x2+...+β k-1 x k-1 +β k x k

[0222] The predicted probability of default is:

[0223] P=P(Y=1|x1,x2,x3,...,x k-1 , x k )

[0224] The fitted prediction for the probability of default is:

[0225]

[0226] From the above formula we can further deduce:

[0227] Substituting the Z value into the above formula can calculate the probability P of predicted default.

[0228] The core of logistic regression model construction is feature screening. The steps of feature screening are as follows: First, batch screen the features according to missing rate, concentration and information value IV. Second, screen all remaining features one by one according to whether the features are in line with business trends, and retain the features with correct business trends. For example: if it is found that the bad debt rate of the customer group decreases with the increase of the characteristic loan balance, it is considered that this feature does not conform to the business trend. In the understanding of credit business, the higher the loan balance, the higher the customer's default risk exposure (EAD) level, and the greater the risk. At this time, this feature will be removed from the feature list. Third, use the stepwise regression function of logistic regression to eliminate the less important features that are highly correlated with other features. Fourth, filter the coefficients according to the positive and negative signs of the training coefficients and the business trend of the feature conversion value, and retain the features whose feature coefficient signs conform to the business logic. In the case where 0 is defined as a good customer and 1 is defined as a bad customer in the Y label, for a feature whose bad debt rate increases monotonically with the increase of the feature value (such as loan balance), the sign of its training coefficient should be positive; conversely, for a feature whose bad debt rate decreases monotonically with the increase of the feature value (such as deposit balance), the sign of its training coefficient should be negative. If it does not meet the above standards, it should be eliminated. Fifth, use variance inflation factor (VIF), correlation coefficient, etc. to further eliminate features with high correlation: for variance inflation factor, eliminate features with the highest VIF and greater than 4 one by one; for features with high correlation, eliminate features with low IV values ​​in the feature group with the highest correlation coefficient and greater than 0.80 one by one. Sixth, use the population stability index (PSI) to eliminate features with large distribution differences at different time points, which lead to instability. For features with PSI>0.25, directly eliminate them, and for features with 0.25>PSI>0.1, carefully eliminate them according to the impact of eliminating this feature on the model's distinguishing ability.

[0229] This application strictly adheres to the above rules to screen features, ensuring the model's interpretability, stability, and ability to distinguish between good and bad customers.

[0230] The feature fine-screening steps of the present application include: a first fine-screening step, based on a stepwise regression algorithm, to screen features based on the F test and the T test for the significance of features; a second fine-screening step, to screen features based on calculating the variance inflation factor for each feature and eliminating features with higher variance inflation factors; a third fine-screening step, based on logistic regression, to analyze whether the feature coefficients of the features after the first fine-screening step and the second fine-screening step are in line with the trend of the prediction results for credit default so as to further perform feature screening.

[0231] Feature fine screening step 1, based on the stepwise regression algorithm, based on the F test and T test method, introduces features in order from high to low significance. Each time a feature is introduced, the selected features are tested one by one. When the originally introduced feature becomes no longer significant due to the introduction of the subsequent features, it is eliminated. This process is repeated until no features above the significance threshold are selected into the equation, and no features below the significance threshold are eliminated from the regression equation.

[0232] Feature fine-screening step 2 is based on the method of eliminating features with high variance inflation factors to further reduce multicollinearity in the model.

[0233] Feature fine screening step 3, based on the risk characteristics of each risk point itself (system preset) and the positive and negative signs of the model training coefficient, determines whether the feature coefficients of the remaining features in the model in the feature fine screening step 3 are consistent with the business trend, removes the features whose model coefficients do not meet the business trend, and iterates again. The specific implementation plan of feature fine screening module 3 is as follows: 1. For features whose feature conversion method is WOE type, the corresponding model training coefficient should be a negative value, and the WOE conversion type features with positive training coefficients should be eliminated. 2. For continuous conversion methods, if the value of this feature increases in business logic, the bad debt rate should increase (such as the credit line utilization rate, etc.), then the corresponding model training coefficient should be a positive value, and such continuous conversion type features with negative training coefficients should be eliminated; if the value of this feature increases in business logic, the bad debt rate should decrease (such as the deposit amount, etc.), then the corresponding model training coefficient should be a negative value, and such continuous conversion type features with positive training coefficients should be eliminated. 3. For the dummy feature conversion method, it is necessary to determine the bad debt rate when the value is 1 and the value is 0. If the bad debt rate when the value is 1 is greater than the bad debt rate when the value is 0, the coefficient should be positive, otherwise it should be negative.

[0234] Feature fine screening step 4, data stability monitoring step, is used to evaluate whether the distribution of individual features and overall scores at different time points has obvious deviations. In this embodiment, features with PSI>0.25 are directly eliminated, and features with 0.25>PSI>0.1 are carefully eliminated based on the impact of eliminating this feature on the model's distinguishing ability.

[0235] Based on the above steps 3 and 4, the feature list is iterated multiple times until no new features are added or removed from the model, and the iteration is stopped to obtain the final feature list and its conversion value. After the above steps, the input variables for constructing the model of this application can be finally obtained.

[0236] The credit default probability modeling step of the present application substitutes the features screened in the feature fine screening step into the Sigmoid function to perform logistic regression to calculate the credit default probability model.

[0237] In the prior art, the core of all scoring models built is whether the data they use is representative. For the existing scoring models in the prior art, due to information security and cost reasons, the sample size and bad labels of most models are small, and the stability of the model cannot be guaranteed, let alone independent modeling of a certain type of customer. At the same time, because the scoring models on the market can obtain information dimensions in the modeling process, mostly multi-loan data and weak financial attribute data (such as smart terminal device data, social platform data, online shopping mall data and other non-credit transaction data), it is impossible to accurately reflect the customer's asset status and repayment ability. The data source used in this application is the full business data of large banks, and the data volume of modeling samples and bad labels is very large. This application designs modeling samples based on different sample groups, which can more finely distinguish the risk differences between such customers, and at the same time ensure the stability of data and models based on cross-time verification and PSI verification. This application uses personal credit history data and asset data for model development, so it has a good reflection of the borrower's repayment willingness and repayment ability.

[0238] For some small and medium-sized financial institutions, the internally built scoring models are relatively small or started late in their retail credit business, and the accumulated historical data is insufficient to develop a credit risk model with stable data and strong differentiation ability. Therefore, they are highly dependent on manual approval when used for credit review business. The efficiency limitations of manual approval have restricted the development of their retail credit business. At the same time, the subjectivity of manual approval has increased the operational risks in the credit review process. The model constructed in this application can assist such financial institutions in making digital decisions, enhance their approval accuracy and speed, and reduce the above-mentioned adverse effects.

[0239] In terms of feature conversion, compared with the traditional WOE conversion method, it is necessary to roughly bin and discretize continuous features according to data performance and the experience of modelers. The results of rough binning are greatly affected by the subjective factors of modelers; and the process of discretizing continuous features may result in too many customers falling into the same interval, which leads to more single values ​​of the calculated scores. This application combines the continuous conversion method, the WOE conversion method and the dummy feature conversion method to encode the original features, which reduces the impact of human factors and single values ​​while enhancing the ability to distinguish scores.

[0240] <Methodology for calculating retail credit risk>

[0241] The present application further relates to a method for calculating the retail credit risk of a sample to be tested using the retail credit risk model constructed in the present application.

[0242] In this application, retail credit risk refers to the probability of credit default in the sample to be tested. In a specific way, the probability of credit default refers to the probability of credit default, specifically, for example, the probability of whether a credit default of 30 days or more will occur, for example, the discrimination and stability of credit default of 30 days or more in a certain customer group.

[0243] The present application relates to a method for calculating retail credit risk, which comprises: a data collection step, which obtains retail credit prediction data of a sample to be predicted; a credit default probability calculation step, which substitutes the retail credit prediction data into a credit default probability model to calculate the credit default probability of the sample to be predicted.

[0244] In this application, the credit default probability covers the credit default probability of various types of credit business. In a specific implementation, the credit default probability is the credit default probability for Internet credit business.

[0245] In the present application, the sample to be predicted may be a sample used to build a model, that is, when the model is built, the sample is already a customer of a certain financial institution holding a credit business (including but not limited to a credit card or personal loan business), and the method for calculating retail credit risk of the present application can be used to calculate its potential retail credit risk in the future. In the present application, the sample to be predicted may be a sample that is not used to build a model, that is, when the model is built, it is not a customer of a certain financial institution holding a credit business, but now it is a customer of a certain financial institution holding a credit business, and the method for calculating retail credit risk of the present application can be used to calculate its potential retail credit risk in the future. In the present application, the sample may also be a customer who was not and is not a customer of the financial institution holding a credit business when the model was built in the past, but is a customer of the financial institution that already holds other financial asset businesses, and the method for calculating retail credit risk of the present application can be used to calculate its potential retail credit risk in the future.

[0246] The present application relates to a method for calculating retail credit risk, which comprises: a data collection step, which obtains retail credit prediction data of a sample to be predicted; a credit default probability calculation step, which substitutes the retail credit prediction data into a credit default probability model to calculate the credit default probability of the sample to be predicted; and after calculating the credit default probability, a step of calculating the credit score of the sample to be predicted.

[0247] The step of calculating the credit score of the sample to be predicted is used to calibrate the calculated credit default probability to a standardized score of 0-1000 points.

[0248] In the method of the present application, the retail credit prediction data includes the original retail credit prediction data of the sample to be predicted and the derived retail credit prediction data processed based on the original retail credit prediction data, wherein the description of the original retail credit prediction data and the derived retail credit prediction data is consistent with the description of the model building part.

[0249] In the present application, the retail credit prediction data is subjected to feature conversion and then substituted into the credit default probability model to calculate the credit default probability of the sample to be predicted. The feature conversion step includes selecting the WOE method or the continuous method for feature conversion based on the feature type of the retail credit prediction data that needs to be substituted into the credit default probability model.

[0250] The method of converting features in WOE mode or continuous mode is well known to those skilled in the art, and the specific conversion method can be referred to the description in the model construction section of this application. However, the existing models in the art only use WOE mode to convert parameters.

[0251] The WOE conversion method is an optimal binning scheme based on the modeling sample. It discretizes the continuous features according to the optimal cutting point. Therefore, it performs better in model construction, but the binning results may overfit the modeling sample, resulting in a serious decline in the model effect when applied to the overall population (poor generalization ability). At the same time, due to the normalization operation used in binning, the original features falling into different bins are converted into a single value corresponding to each bin, thus losing the ability to distinguish the risks of people falling into the same interval. The conversion method of dummy features is mainly applied to grouping features. Its advantage is that it can eliminate the difference between good and bad values ​​of features. For example: for the industry in which the customer is located, because there is no obvious advantage or disadvantage between the retail industry and the wholesale industry, it is more suitable for the dummy feature processing method; there is a hierarchical difference between junior colleges and undergraduates, so although dummy features can be used, it is actually more appropriate to use the WOE processing method. The continuous conversion method can avoid the overfitting of the modeling sample in the WOE conversion method, and has a strong generalization ability for the overall sample. Since no interval mapping is performed, most customer groups rarely fall into a single value. However, it cannot be applied to some features with poor monotonicity or discreteness (such as occupations, positions, etc.). The present application uses a combination of continuous conversion and WOE conversion to process the initial features, combines the advantages and disadvantages of the two conversion methods, designs a conversion judgment module, and selects the optimal feature conversion method based on the parameters such as feature data attributes, missing rate, and concentration, supplemented by business logic judgment. The technical solution of the present application takes a different approach, creatively combining the three methods of continuous conversion, WOE conversion, and dummy feature conversion, reprocesses some carefully screened features, combines the advantages and disadvantages of the three conversion methods, and creatively designs a conversion judgment method, and selects the optimal feature conversion method based on the parameters such as feature data attributes, missing rate, and concentration, supplemented by business logic judgment. The feature conversion method determined in this way is then substituted into the model for calculating retail credit risk, which can more accurately predict the retail credit risk of the sample.

[0252] The model used in the present application to calculate the probability of credit default is a model constructed based on sample retail credit prediction data and the probability of credit default using logistic regression based on an existing user population, that is, a model constructed using the method described in the present application.

[0253] <Device, system and computer storage medium for calculating retail credit risk>

[0254] The present application relates to a device for calculating retail credit risk, which comprises: a data acquisition module for acquiring retail credit prediction data of a sample to be predicted; and a credit default probability calculation module for substituting the retail credit prediction data into a credit default probability model to calculate the credit default probability of the sample to be predicted.

[0255] The device for calculating retail credit risk of the present application can execute the steps of the method for calculating retail credit risk of the present application.

[0256] The present application also relates to a system for calculating retail credit risk, characterized in that the system for calculating retail credit risk comprises: a memory, a processor, and a program for calculating retail credit risk stored in the memory and executable on the processor, wherein the program for calculating retail credit risk implements the steps of the method for calculating retail credit risk as described in the present application when executed by the processor.

[0257] The present application relates to a computer storage medium, characterized in that a program for a method for calculating retail credit risk is stored on the computer storage medium, and when the program for the method for calculating retail credit risk is executed by a processor, the steps of the method for calculating retail credit risk as described in the present application are implemented.

[0258] All the contents described above for the method for calculating retail credit risk can be fully applied to the device, system and computer storage medium for calculating retail credit risk.

[0259] The method for calculating retail credit risk in the present application can avoid the technical shortcomings of overfitting when building a model with complete WOE variables and the inability to adapt to categorical variables well when building a model with complete continuous variables. Therefore, when the model is used for retail credit risk calculation, it can better cover the retail credit risk prediction needs of various types of customer groups and meet the demand for obtaining more accurate prediction results when predicting credit risk as a general score.

[0260] Example

[0261] Example 1 Collection of modeling samples

[0262] In the process of constructing this embodiment, credit card application and behavior data, personal loan application and behavior data, personal financial asset transaction data, and personal customer information data of all retail customers of China Construction Bank between 2017 and 2021 were collected, totaling 730 million people. The modeling sample was confirmed through a professional model design scheme, and 620 million data in 2018 and 650 million data in 2019 were selected as analysis samples. Since performance variables are required for modeling and the performance period of Internet credit business is relatively short, 2019 was selected as the observation period, and the 650 million analysis samples were divided into two parts: registered credit business and unregistered credit business, with sample sizes of 130 million and 520 million people respectively.

[0263] The specific Gamma model is a scenario scoring for Internet credit business. When designing the model, 220,000 customers who have applied for Internet credit business are analyzed and designed among the 130 million customers who have applied. After the model is developed and put online, the results will be applied to the total number of 730 million people who have applied and those who have not applied, or all other potential customer groups.

[0264] The model design includes 1) exclusion rules: such as excluding customer data that has been settled, closed, has no performance, or has special circumstances. After exclusion, the modeling population is 180,000 customers; 2) time window setting: determine the sample to use 2019 data as the modeling sample, and use the next 12 months as the performance period of Y; 3) sample sampling: use a good / bad sample number of 10:1 for sampling modeling. Different segmentation schemes are designed through the decision tree method, and the parent-child model comparison method is used to confirm the final model segmentation plan, and then each segmentation model is modeled separately. Through actual data analysis combined with professional business experience, this application has found that the differences in Internet credit business volume and customer base are relatively small, and no further segmentation is performed. All customer groups use the same main model for development and application.

[0265] In this embodiment, the segmentation model is no longer divided, and all samples are developed uniformly.

[0266] Specifically, the customer sample size used to construct the Gamma model is about 180,000, and its customer base mainly consists of customers with Internet credit business. The model is used to predict the probability of their credit being overdue for more than 30 days.

[0267] Based on the customer samples confirmed above, historical data information of borrowers is obtained: 1) Credit card category, basic fields include account, overdue, balance, quota, repayment, actual repayment and other information under different dimensions such as bills, cash withdrawals, installments, interest, etc.; 2) Personal loan category, including account, overdue, balance, quota, repayment, actual repayment and other information; 3) Basic customer information, including gender, age and administrative region information of the business; 4) Personal financial assets, including AUM, deposits, wealth management and payroll information.

[0268] According to various risk points in the credit business, macro credit risk is split according to information dimensions, time slices, etc. Variables are derived through professional characteristic variable construction methods. The basic methods are: 1) Customer relationship length variables: such as customer account opening time, customer maximum account age, etc. 2) Time interval variables: such as the number of months from the customer's last repayment to the observation point, the number of months from the customer's last overdue payment, etc.; 3) Behavior frequency degree variables: such as the number of times the customer's repayments are > N in the last X months, the number of times the customer's credit utilization rate is > N in the last X months, etc.; 4) Current time point variables: customer's current monthly credit limit, customer's current monthly balance, etc.; 5) Statistical value variables: such as the customer's maximum number of overdue periods in the last X months, the customer's average credit utilization rate in the last X months; 6) Continuous behavior variables: such as the maximum number of consecutive overdue periods > N in the last X months, the number of consecutive repayment rates > N in the last X months, and other characteristic variables. Finally, 3067 features with potential predictive power for customer overdue payment were derived, such as 'current repayment amount', 'average number of overdue periods in the past 6 months', and 'number of credit card installments in the past 12 months'. In this article, observation point refers to the time point when samples are collected until modeling, and current also refers to the sampling cutoff time. Observation point and current have the same meaning.

[0269] Table 1 Summary of basic variables and derived variables used in the examples

[0270]

[0271]

[0272] Example 2 Feature Screening

[0273] A preliminary screening was performed on the 3067 features (variables) collected in Example 1 that have potential predictive power for customer overdue payments.

[0274] The preliminary screening of the Gamma model features in Example 1 was performed as follows:

[0275] In the first round of preliminary screening, features with a missing rate of more than 95% in the data collected in Example 1 were first eliminated, and a total of 352 variables were deleted, leaving 2715 variables.

[0276] In the second round of preliminary screening, for the 2715 features that passed the first round of preliminary screening, the features with single values ​​exceeding 99% among the remaining features were eliminated, and a total of 514 variables were eliminated, leaving 2201 variables.

[0277] In the third round of preliminary screening, the 2201 features after the second round of preliminary screening are sorted according to the feature values ​​(the sorting method is specifically that if the feature is a character variable, each value is a separate box, if the feature is a numeric variable, it is sorted from small to large according to the value size), and then divided into 10-20 boxes according to the quantile point, and the feature IV value is calculated. If the feature IV value is lower than 0.02, it is eliminated. In this embodiment, since the variable attributes are strong financial and asset variables, the prediction effect is strong. Here, variables with IV values ​​lower than 0.05 are eliminated. A total of 1080 features are eliminated, and 1121 feature variables remain.

[0278] In the fourth round of preliminary screening, for the 1121 features after the third round of screening, the features are further screened based on the stepwise discrimination algorithm. After this round of screening, 141 important feature variables can be quickly screened out. The following is an introduction to the stepwise discrimination method. In actual data, there may be little difference in the mean values ​​of different categories on a certain variable. In this case, the effect of using this variable for classification will not be very good; there is also a type of variable that can indeed distinguish different categories in the data well when considered independently, but after including these variables in the model variables, they may appear redundant. Therefore, the embodiment pioneered the use of the stepwise discrimination method to better select more important feature variables under the same dimension, greatly reducing the workload of judging and screening one by one according to the trend of the variables in the fifth round of screening, and fully improving the efficiency of model development without affecting the overall model effect.

[0279] The stepwise discrimination method uses the WILKS'S LAMBDA criterion to measure the strength of feature capabilities and eliminates features that do not meet the set threshold among the remaining features after three rounds of screening. In the stepwise discrimination process, the variable with the strongest discriminant ability is added first. As the variables in the model gradually increase, the discriminant ability of the variables introduced earlier may also change. If the discriminant ability of a variable in the model is less than the threshold, the variable is removed, and then the process is repeated until the variables contained in the model meet the WILKS'S LAMBDA similarity ratio criterion and other variables do not meet the criteria for entering the model.

[0280] The fifth round of preliminary screening is to further screen the 141 important features after the fourth round of screening based on the risk characteristics of each risk point itself (system preset) and the actual bad debt rate of the sample.

[0281] Determine whether the actual bad debt rate distribution of the remaining features is consistent with the business trend, and remove the features whose actual bad debt rate distribution of the samples does not conform to the business trend. Specifically,

[0282] (1) Divide the remaining target features into 10-20 boxes according to the quantiles, and calculate the median of each box and the corresponding bad debt rate.

[0283] (2) Calculate the rate of change (slope) using the median of the values ​​of each box compared to the previous box and the corresponding bad debt rate.

[0284] (3) Count the number of boxes with a change rate greater than 0 and the number of boxes with a change rate greater than 0 between two adjacent boxes, and calculate the percentage of boxes with a change rate greater than 0 to the number of boxes with a change rate greater than 0.

[0285] (4) Obtain the risk characteristics preset for this feature, and based on the percentage of the number of boxes greater than 0 in the above-calculated change rate to the number of non-zero boxes in the change rate, determine whether the actual bad debt rate change trend is approximately consistent with the business logic bad debt rate change trend.

[0286] The criteria for approximate consistency are as follows:

[0287] First, if the business logic trend of this feature is: as the value of this feature increases, the bad debt rate increases (such as the credit line utilization rate, etc.), in this module, the features whose number of boxes with slope greater than 0 in the above calculation accounts for less than 70% of the number of boxes with slope non-0 are eliminated, that is, the features that do not conform to the business trend are eliminated.

[0288] Second, if the business logic trend of this feature is: as the value of this feature increases, the bad debt rate decreases (for example, the amount of deposits, etc.), in this module, the features whose number of boxes with slope greater than 0 in the above calculation accounts for more than 30% of the number of boxes with slope not 0 are eliminated, that is, the features that do not conform to the business trend are eliminated.

[0289] After the fifth round of screening, 92 features were eliminated and 49 features remained.

[0290] Example 3 Conversion of features after initial screening

[0291] In the feature judgment step, the optimal conversion method is selected according to the concentration of features after the initial feature screening step, data type, etc. When judging the conversion method, data types are generally divided into two categories, one is character variables, and the other is numeric variables. For character variables, the dummy variable conversion method is generally used for variable conversion; for numeric variables, if the variable value is less than 5 or so, the WOE conversion method will be used. If the variable value is more, the WOE or continuous conversion method will be used (the optimal conversion form is selected by the correlation with the target variable). In this process, it is necessary to comprehensively consider the situation of variable concentration. For example, if the continuous variable has many values, but the concentration of a single value exceeds 95%, the continuous processing will not be performed, and the WOE conversion method can be used directly. Features using different conversion methods are grouped and divided into different data sets. Specifically,

[0292] First, the remaining 49 features in Example 2 are used to determine the conversion methods of these features, and the following three methods are selected according to the concentration of the features, data types, etc.

[0293] Feature conversion method 1 is used to perform WOE conversion on the features obtained in the data acquisition module and judged in the feature judgment module that the optimal conversion method is WOE conversion.

[0294] Feature conversion method 2 is used to perform dummy feature conversion on the features obtained in the data acquisition module and judged in the feature judgment module that the optimal conversion method is dummy feature conversion.

[0295] Feature conversion method 3 is used to obtain the features in the data acquisition module that are determined to be optimal for continuous conversion in the feature judgment module, select the optimal continuous conversion method, and perform continuous conversion.

[0296] The feature merging module is used to horizontally splice the data of feature conversion mode 1, feature conversion mode 2 and feature conversion mode 3.

[0297] Through this embodiment, 49 features are converted, of which 15 features are converted into WOE, 7 features are converted into dummy features (expanded into 18 variables), and 27 features are converted into continuous types, for a total of 60 features.

[0298] Example 4 Feature Depth Screening (Feature Fine Screening Step)

[0299] The feature fine screening step mainly includes the following four steps in this embodiment. Steps 1 and 2 are based on stepwise regression and calculation of variance inflation factor to eliminate features with large multicollinearity in the feature merging module to enhance the robustness of the model. Step 3 eliminates features whose positive and negative signs of the training coefficients in the model do not conform to the business trend. Step 4 uses the population stability coefficient (PSI) to eliminate unstable features. This embodiment uses the LOGISTIC process in SAS for deep screening.

[0300] Feature fine screening step 1, based on the stepwise regression algorithm, based on the F test and T test method, introduces features in order from high to low significance. Each time a feature is introduced, the selected features are tested one by one. When the originally introduced feature becomes no longer significant due to the introduction of the subsequent features, it is eliminated. This process is repeated until no features above the significance threshold are selected into the equation, and no features below the significance threshold are eliminated from the regression equation. After this step, 60 features are eliminated to 22 features.

[0301] Feature fine-screening step 2, which is based on the method of removing features with high variance inflation factors, further reduces multicollinearity in the model. After this step, 22 features are eliminated to 18 features.

[0302] Feature fine screening step 3 compares the risk characteristics of each risk point itself (system preset) with the positive and negative signs of the model training coefficients to determine whether the feature coefficients of the remaining features in the model in step 3 are consistent with the business trend, eliminates the features whose model coefficients do not conform to the business trend, and iterates again.

[0303] The specific implementation scheme of the feature fine screening module 3 is as follows:

[0304] 1. For features whose feature conversion method is WOE type, the corresponding model training coefficient should be a negative value, and the WOE conversion type features with positive training coefficients should be eliminated.

[0305] 2. For continuous conversion methods, if the value of this feature increases in business logic, the bad debt rate should increase (such as the credit limit utilization rate, etc.), then the corresponding model training coefficient should be positive, and such continuous conversion features with negative training coefficients should be eliminated; if the value of this feature increases in business logic, the bad debt rate should decrease (such as the deposit amount, etc.), then the corresponding model training coefficient should be negative, and such continuous conversion features with positive training coefficients should be eliminated.

[0306] 3. For the dummy feature conversion method, it is necessary to determine the bad debt rate when the value is 1 and the value is 0. If the bad debt rate when the value is 1 is greater than the bad debt rate when the value is 0, the coefficient should be positive, otherwise it should be negative.

[0307] Feature fine screening step 4, data stability monitoring step, is used to evaluate whether the distribution of individual features and overall scores at different time points has obvious deviations. In this embodiment, features with PSI>0.25 are directly eliminated, and features with 0.25>PSI>0.1 are carefully eliminated based on the impact of eliminating this feature on the model's distinguishing ability.

[0308] Based on the above steps 3 and 4, the feature list is iterated multiple times until no new features are added or removed from the model, and the iteration is stopped to obtain the final feature list and its conversion value. After the above steps, 18 features are eliminated to the final 6 features that are entered into the model.

[0309] Example 5 Construction of Gamma Model

[0310] Generally speaking, the stronger the correlation between the characteristic variable and the target variable, the more accurate the final model will be. In this embodiment, the sample size of the Gamma model is about 180,000, and its customer base is mainly customers with Internet credit business. The model is constructed to predict the probability of credit overdue for more than 30 days (target variable).

[0311] In this embodiment 5, it is finally confirmed that the final modeling result is described by taking the six features of the customer's region, the remaining available credit limit of the credit card in the current month, whether it is a payroll customer, the minimum balance of the deposit account at a time point in the past 6 months, the sum of the current monthly credit limit of personal loans, and the current monthly credit limit utilization rate of the credit business as an example. In the feature conversion step, the 6 modeling variables have been converted in a corresponding form according to the correlation between the feature and the target variable. In the credit default probability modeling step, the 6 features screened by the feature fine screening step are substituted into the Sigmoid function (using the LOGISTIC process of SAS software) for logistic regression to calculate the credit default probability model.

[0312] The region where the customer belongs reflects the risk level of the region where the borrower is located. For customers with Internet credit business, the region where they choose to handle the business or the region where the customer is located is very important. Generally speaking, the customer risk performance levels vary from region to region. If the customer's region is historically a high-risk region, then the possibility of overdue payment in the future will be very high. The relationship between this variable and the performance of customer risk is that the lower the historical performance risk of the customer's region, the lower the possibility of serious default in the future; conversely, if the historical performance risk of the customer's region is high, the possibility of serious default in the future will also increase. The conversion method is WOE conversion.

[0313] The remaining available credit limit of the credit card in the current month reflects the borrower's credit demand. For customers with Internet credit business, a large amount of credit demand in the near future may lead to excessive debt, which is directly reflected in the continuous reduction of the remaining available credit limit on the credit card, and eventually leads to overdue payments. The relationship between this variable and customer risk is that the lower the remaining available credit limit of the credit card in the current month, the stronger the borrower's credit demand, the greater the repayment pressure, and the higher the possibility of serious default; conversely, the higher the remaining available credit limit of the credit card in the current month, the weaker the borrower's credit demand, the lower the repayment pressure, and the lower the possibility of serious default. The conversion method is logarithmic conversion.

[0314] Whether or not the client is a payroll agent reflects the stability of the borrower's income. For clients with Internet credit business, if they are payroll agents, then they have a stable source of income and are less likely to default in the future. The relationship between this variable and the performance of customer risk is that if the value of whether or not the client is a payroll agent is No, the client does not have a stable salary income, and the possibility of a serious default is higher; conversely, if the value of whether or not the client is a payroll agent is No, the client has a stable salary income, and the possibility of a serious default is also reduced. The conversion method is WOE conversion.

[0315] The minimum balance of the deposit account at a certain point in the past six months reflects the borrower's deposit level and repayment ability. For customers with Internet credit business, the more pressure the borrower has to repay the debt, the higher the possibility of overdue payment. The relationship between this variable and customer risk is that the larger the minimum balance of the deposit account at a certain point in the past six months, the better the borrower's deposit level and the stronger the repayment ability, and the lower the possibility of serious default; conversely, the smaller the minimum balance of the deposit account at a certain point in the past six months, the worse the borrower's deposit level and the weaker the repayment ability, and the higher the possibility of serious default. The conversion method is logarithmic conversion.

[0316] The sum of the current monthly credit lines for personal loans reflects the financial institution's recognition of the borrower's risk level. For customers with Internet credit business, if the sum of personal loan lines is very high, on the one hand, it means that the customer has collateral such as collateral for credit, and on the other hand, it means that the financial institution fully recognizes the customer's repayment ability, and the possibility of overdue payment is lower. The relationship between this variable and customer risk is that the larger the sum of the current monthly credit lines for personal loans, the stronger the financial institution's recognition of the borrower, and the lower the possibility of serious default; conversely, the smaller the sum of the current monthly credit lines for personal loans, the weaker the financial institution's recognition of the borrower, and the higher the possibility of serious default. The conversion method is cube root conversion.

[0317] The current monthly credit line utilization rate reflects the borrower's credit demand. For customers with Internet credit business, a large amount of short-term credit demand may lead to excessive debt. The manifestation of revolving loans is that the credit line utilization rate is often too high in the short term, and the possibility of overdue payments in the future will be very high. The relationship between this variable and customer risk is that the smaller the average credit line utilization rate of the revolving loan in the past three months, the lower the borrower's credit demand, and the lower the possibility of serious default in the future; conversely, the larger the average credit line utilization rate of the revolving loan in the past three months, the higher the borrower's credit demand, and the possibility of serious default will also increase. The conversion method is square root conversion.

[0318] In the step of calculating the probability of credit default, based on the data and coefficients obtained in the feature conversion step, the following model formula is used to predict the probability (P) that the borrower is in default:

[0319]

[0320] Where k is the number of features entering the model, and in Formula 1, k is 6.

[0321] α is the intercept, with a range of (-0.1497, 0.2097) and an optimal value of 0.03; β1 is the region to which the customer belongs, with a range of (0.0026, 0.0963) and an optimal value of 0.049; β2 is the remaining available credit limit of the credit card in the current month, with a range of (-0.0483, 0.0566) and an optimal value of 0.004; β3 is whether the customer is a payroll customer, with a range of (0.0785, 0.1265 ), the optimal value is 0.102; β4 is the minimum balance of the deposit account in the past 6 months, the value range is (-0.1616, 0.1703), the optimal value is 0.004; β5 is the sum of the current monthly credit limit of personal loans, the value range is (0.0758, 0.1454), the optimal value is 0.111; β6 is the current monthly credit limit utilization rate of credit business, the value range is (0.0294, 0.1385), the optimal value is 0.084. (Note: The value range comes from the 95% confidence interval, that is, the 95% CI in the table below)

[0322] x1 is the logarithmic conversion value of the customer's region generated by the feature conversion step; x2 is the WOE conversion value of the remaining available credit limit of the credit card in the current month generated by the feature conversion step; x3 is the WOE conversion value of whether the customer is a payroll customer generated by the feature conversion step; x4 is the logarithmic conversion value of the minimum balance of the deposit account at a certain point in time in the past 6 months generated by the feature conversion step; x5 is the cube root conversion value of the sum of the credit limits of personal loans in the current month generated by the feature conversion step; x6 is the square root conversion value of the credit limit utilization rate of the current month generated by the feature conversion step. The model performance of some features is shown in Table 2 below.

[0323] Table 2

[0324]

[0325] The P values ​​of all model characteristics are less than 0.05, indicating that the above characteristics are significantly correlated with default performance.

[0326] Example 6

[0327] The P values ​​calculated by the above formulas can be further used to calculate the score of any customer.

[0328] The scoring calculation step is used to convert the calculated credit default probability into a score of 0-1000 using a pre-stored default score conversion code.

[0329] In the scoring calculation module, the following formula is used to calculate and generate the credit score used to characterize the borrower:

[0330]

[0331]

[0332] Among them, P is the borrower's default probability (P) generated in the credit default probability calculation module, A is 443.9036; B is -72.1348, and the round function rounds the calculated score to the integer value; finally, the score greater than 1000 is set to 1000 points, and the score less than 0 is set to 0 points.

[0333] In this embodiment, the P value calculated in the above embodiments can be substituted to calculate the score.

[0334] The KS (Kolmogorov-Smirnov) statistic was proposed by two Soviet mathematicians, ANKolmogorov and NVSmirnov. In risk control, KS is often used to evaluate the discrimination of models. The greater the discrimination, the stronger the risk ranking ability of the model.

[0335] The KS statistic is based on the empirical cumulative distribution function (ECDF), which is generally defined as:

[0336] KS=max(|cum(bad_rate)-cum(good_rate)|)

[0337] According to the prediction model, a KS curve is drawn. The KS of this embodiment is 35.11. It can be seen that the model constructed in this embodiment performs well in assessing customer credit risk. Figure 1 An example of a KS graph is shown with a KS value of 58.

[0338] The following shows the results of the KS values ​​of the model applied to the development samples and the reserved validation samples. From Table 3, it can be seen that the KS values ​​of the model in the training set and the validation set (where the sample size ratio of the training set and the validation set is 6:4) and the overall sample are all above 30, and the discrimination effect is good.

[0339] Table 3

[0340]

Claims

1. A method for calculating retail credit risk, comprising: A data collection step, which obtains retail credit prediction data of the sample to be predicted; The credit default probability calculation step is to substitute the retail credit prediction data into the credit default probability model to calculate the credit default probability of the sample to be predicted, and preferably the credit default probability is the credit default probability for Internet credit business.

2. The method according to claim 1, further comprising: After calculating the credit default probability, a step of calculating the credit score of the sample to be predicted is used to calibrate the calculated credit default probability to a standardized score of 0-1000 points.

3. The method according to claim 1 or 2, wherein: The retail credit prediction data includes original retail credit prediction data of the sample to be predicted and derived retail credit prediction data processed based on the original retail credit prediction data; Preferably, the original retail credit prediction data includes: Credit card basic data, which is based on all available data during the sample user's credit card creation and usage process. Basic data on personal loans, which is all available data based on the loan application and usage behavior of sample users. Basic customer information data, which is based on the attributes of the sample users themselves but is not directly related to their behavior in financial institutions. Basic data on personal financial assets, which includes all other financial assets and financial transaction data of sample users in financial institutions that are not related to credit cards and loans.

4. The method according to any one of claims 1 to 3, wherein: The derived retail credit prediction data processed based on the original retail credit prediction data refers to the data obtained by processing the collected original retail credit prediction data based on the time dimension, space dimension, frequency dimension and statistical information dimension; Preferably, the derived retail credit forecast data includes but is not limited to: Derived retail credit forecast data obtained by processing based on sample relationship length, Derivative retail credit forecast data obtained by processing time interval variables, Derivative retail credit forecast data obtained by processing based on the frequency of sample behavior, Derivative retail credit forecast data obtained by processing the current time point of the sample, Derivative retail credit forecast data obtained by processing based on sample continuous behavior, Derived retail credit forecast data obtained by processing sample data based on statistical information dimensions.

5. The method according to any one of claims 1 to 4, wherein: The retail credit prediction data is selected from one, two, three, four, five or six of the following: The customer's region, the remaining available credit limit of the credit card in the current month, whether the customer is a payroll payment customer, the minimum balance of the deposit account at a certain point in time in the past 6 months, the sum of the personal loan credit limit in the current month, and the credit business credit limit utilization rate in the current month.

6. The method according to any one of claims 1 to 5, wherein: After performing feature conversion on the retail credit prediction data, the data is substituted into the credit default probability model to calculate the credit default probability of the sample to be predicted. The feature conversion step includes: Based on the feature type of the retail credit prediction data that needs to be substituted into the credit default probability model, select the WOE method or the continuous method for feature conversion.

7. The method according to claim 6, wherein: The continuous method for feature conversion includes the following methods: directly selecting the original value, calculating the square of the original data, calculating the square root of the original data, calculating the cube root of the original data, or calculating the natural logarithm of the original data to perform continuous feature conversion.

8. The method according to any one of claims 1 to 7, wherein: The credit default probability model is a model constructed based on sample retail credit prediction data and credit default probability using logistic regression based on the existing user population.

9. The method according to any one of claims 1 to 8, wherein: The retail credit prediction data is selected from one, two, three, four, five or six of the following: the customer's region, the remaining available credit limit of the credit card in the current month, whether the customer is a payroll customer, the minimum balance of the deposit account at a certain point in time in the past six months, the sum of the credit limits of personal loans in the current month, and the utilization rate of the credit limit of the credit business in the current month.

10. The method according to claim 9, wherein: The steps to calculate the probability of credit default include: The customer's region, the remaining available credit limit of the credit card in the current month, whether the customer is a payroll customer, the minimum balance of the deposit account at a certain point in time in the past 6 months, the sum of the personal loan credit limit in the current month, and the credit business credit limit utilization rate in the current month are converted into features. It is preferred to use the WOE method for conversion based on the customer's region; the continuous conversion method is used for the remaining available credit limit of the credit card in the current month; the WOE method is used for conversion based on whether the customer is receiving payroll; the continuous conversion method is used for the minimum balance of the deposit account at a certain point in time in the past 6 months; the continuous conversion method is used for the sum of the current monthly credit limit of personal loans; and the continuous conversion method is used for the current monthly credit limit utilization rate of credit business; Further preferably, the continuous conversion method for the current month's remaining available credit limit of the credit card is a calculation method using the logarithm of the current month's remaining available credit limit of the credit card; the continuous conversion method for the minimum balance of the deposit account at a point in time in the past 6 months is a calculation method using the logarithm of the minimum balance of the deposit account at a point in time in the past 6 months; the continuous conversion method for the sum of the current month's credit limits of personal loans is a calculation method using the cube root of the sum of the current month's credit limits of personal loans; the continuous conversion method for the current month's credit limit utilization rate of the credit business is a calculation method using the square root of the current month's credit limit utilization rate of the credit business.

11. The method according to claim 10, wherein: The converted values ​​of the six features, namely, the customer's region, the remaining available credit limit of the credit card in the current month, whether the customer is a payroll customer, the minimum balance of the deposit account in the past 6 months, the sum of the personal loan credit limit in the current month, and the utilization rate of the credit limit in the current month, are substituted into the model constructed by logistic regression based on the sample retail credit prediction data and credit default probability to calculate the default probability of the sample to be predicted.

12. The method according to claim 11, wherein: The model is shown in the following formula 1: Where k is the number of features entering the model, preferably k is 6, α is the intercept, the value range is (-0.1497, 0.2097), and the optimal value is 0.03; β1 is the region to which the customer belongs, with a value range of (0.0026, 0.0963) and an optimal value of 0.049; β2 is the remaining available credit limit of the credit card in the current month, with a value range of (-0.0483, 0.0566) and an optimal value of 0.004; β3 is whether the customer is a payroll client, the value range is (0.0785, 0.1265), and the optimal value is 0.102; β4 is the minimum balance of the deposit account in the past 6 months, with a value range of (-0.1616, 0.1703) and an optimal value of 0.004; β5 is the sum of the current monthly credit limits of personal loans, with a value range of (0.0758, 0.1454) and an optimal value of 0.111; β6 is the current monthly credit line utilization rate, with a value range of (0.0294, 0.1385) and an optimal value of 0.084; x1 is the logarithmic transformation value of the customer's region generated by the feature transformation step; x2 is the WOE conversion value of the remaining available credit limit of the credit card in the current month generated in the feature conversion step; x3 is the WOE conversion value generated by the feature conversion step, whether the customer is a payroll customer; x4 is the logarithmic transformation value of the minimum balance of the deposit account at a certain point in time in the past 6 months generated by the feature conversion step; x5 is the cube root conversion value of the sum of the personal loan credit limit for the current month generated in the feature conversion step; x6 is the square root conversion value of the current monthly credit limit utilization rate generated by the feature conversion step.

13. The method according to claim 12, wherein: After calculating the credit default probability, the step of calculating the credit score of the sample to be predicted is to use the following formula to calculate and generate a credit score for characterizing the borrower: Among them, P is the borrower's default probability (P) generated in the credit default probability calculation module, A is 443.9036; B is -72.1348, and the round function rounds the calculated score to the integer value; finally, the score greater than 1000 is set to 1000 points, and the score less than 0 is set to 0 points.

14. A device for calculating retail credit risk, comprising: A data collection module, which is used to obtain retail credit prediction data of samples to be predicted; A credit default probability calculation module is used to substitute the retail credit prediction data into the credit default probability model to calculate the credit default probability of the sample to be predicted, and preferably the credit default probability is the credit default probability for Internet credit business.

15. The device according to claim 14, wherein: The device executes the steps of the method for calculating retail credit risk according to any one of claims 1 to 13.

16. A system for calculating retail credit risk, characterized in that: The system for calculating retail credit risk comprises: a memory, a processor, and a program for the method for calculating retail credit risk stored in the memory and executable on the processor. When the program for calculating retail credit risk is executed by the processor, the steps of the method for calculating retail credit risk as described in any one of claims 1 to 13 are implemented.

17. A computer storage medium, characterized in that: The computer storage medium stores a program for calculating the retail credit risk method. When the program for calculating the retail credit risk method is executed by the processor, the steps of the method for calculating the retail credit risk according to any one of claims 1 to 13 are implemented.

18. A method for constructing a retail credit risk prediction model for Internet credit business, comprising: A data collection step, which obtains the original retail credit prediction data of the sample used to build the model; A data derivation step, which processes derived retail credit forecast data based on the original retail credit forecast data; The feature initial screening step is to perform initial screening on all categories, that is, all features, including the original retail credit forecast data and the derived retail credit forecast data, to obtain features after initial screening; The initial screening data conversion step is to judge the conversion method of the features after the initial screening to confirm whether to use one of the WOE conversion method, the dummy feature conversion method and the continuous conversion method for feature conversion, and use the best method to convert each feature after the initial screening; The feature fine screening step is to perform in-depth screening on the initially screened features after feature conversion to obtain finely screened features; The credit default probability modeling step is to select the logistic regression method to build the model based on the relationship between the carefully screened features and the probability of credit default, and confirm the method used to calculate the credit default probability; The data samples obtained in the data collection step are samples of customers who have used Internet credit services.

19. The method according to claim 18, wherein: In the data collection step, the original retail credit prediction data of the samples obtained for building the model include: Credit card basic data, which is based on all available data during the sample user's credit card creation and usage process. Basic data on personal loans, which is all available data based on the loan application and usage behavior of sample users. Basic customer information data, which is based on the attributes of the sample users themselves but is not directly related to their behavior in financial institutions. Basic data on personal financial assets, which includes all other financial assets and financial transaction data of sample users in financial institutions that are not related to credit cards and loans.

20. The method according to claim 18, wherein: In the data derivation step, the derived retail credit prediction data processed based on the original retail credit prediction data refers to data obtained by processing the collected original retail credit prediction data based on the time dimension, space dimension, frequency dimension, and statistical information dimension; Preferably, the derived retail credit forecast data includes but is not limited to: Derived retail credit forecast data obtained by processing based on sample relationship length, Derivative retail credit forecast data obtained by processing time interval variables, Derivative retail credit forecast data obtained by processing based on the frequency of sample behavior, Derivative retail credit forecast data obtained by processing the current time point of the sample, Derivative retail credit forecast data obtained by processing based on sample continuous behavior, Derived retail credit forecast data obtained by processing sample data based on statistical information dimensions.

21. The method according to any one of claims 18 to 20, wherein: The feature initial screening step includes the following steps: The first preliminary screening step is to screen the features based on the missing data of each feature of the sample used to build the model. The second initial screening step is to screen the features based on the fact that a single value of a feature sample is too high. The third preliminary screening step is to calculate the information IV value of each feature to perform preliminary screening of the features; The order of the first preliminary screening step, the second preliminary screening step and the third preliminary screening step can be any order. The fourth preliminary screening step uses a stepwise discrimination algorithm to perform preliminary screening of features after the first to third preliminary screenings; The fifth preliminary screening step is to conduct preliminary screening of the features after the fourth preliminary screening step based on the consistency between the risk characteristics of each feature itself and the actual real results of the samples used for model construction.

22. The method according to any one of claims 18 to 21, wherein: Also includes: The sample selection step is used to screen all users before the data collection step and obtain samples of Internet credit business for model construction.

23. The method according to any one of claims 18 to 22, wherein: In the initial screening data conversion step, the conversion method of the features after the initial screening is determined based on the concentration and data type of the features after the initial screening.

24. The method according to claim 23, wherein: The initial screening data conversion step includes the following steps based on the judgment of concentration and data type: Classify the data type of each feature into character variables and numeric variables. For character variables, dummy feature conversion is used to perform initial screening data conversion. The process of further classifying numerical variables includes the following sub-steps: If the value of the numeric variable is less than n, the WOE conversion method is used to convert the initial screening data. If the value of the numerical variable is more than n, further judge if the value of the continuous variable is large and the concentration of a single value is greater than m%, then the WOE conversion method is used; if the concentration of a single value is less than or equal to m%, then the continuous conversion method is used. Preferably, n and m are both positive integers, wherein n=5-10, and m=90-99.

25. The method according to claim 24, wherein: Also includes: For the features that are confirmed to adopt the continuous conversion method, the optimal conversion method is selected based on the correlation between the feature and credit default under different continuous conversion methods to perform the continuous feature conversion of the feature. Preferably, the continuous feature conversion is performed by directly selecting the original value, calculating the square of the original data, calculating the square root of the original data, calculating the cube root of the original data, or calculating the natural logarithm of the original data.

26. The method according to any one of claims 18 to 25, wherein: The feature fine screening steps include: The first fine screening step is to screen the features based on the stepwise regression algorithm, the F test and the T test to determine the significance of the features. The second fine screening step is to calculate the variance inflation factor based on each feature and eliminate the features with higher variance inflation factors to screen the features. The third fine screening step is to analyze the characteristics after the first fine screening step and the second fine screening step based on logistic regression to see whether the characteristic coefficients are consistent with the trend of the prediction results for credit default so as to further perform feature screening.

27. The method according to any one of claims 18 to 26, wherein: The credit default probability modeling step substitutes the features selected in the feature fine screening step into the Sigmoid function to perform logistic regression to calculate the credit default probability model.

28. A device for constructing a retail credit risk prediction model for Internet credit business, characterized in that: The device comprises: A data collection module, which is used to obtain raw retail credit prediction data of samples used to build a model; A data derivation module, which is used to process derived retail credit forecast data based on the original retail credit forecast data; A feature initial screening module, which is used to perform initial screening on all categories, that is, all features, including the original retail credit forecast data and the derived retail credit forecast data, to obtain features after initial screening; The initial screening data conversion module is used to determine the conversion method of the features after the initial screening to confirm whether to use one of the WOE conversion method, the dummy feature conversion method and the continuous conversion method for feature conversion, and to use the best method for feature conversion for each feature after the initial screening; A feature fine screening module is used to perform in-depth screening on the initially screened features after feature conversion to obtain finely screened features; The credit default probability modeling module is used to select the logistic regression method to build a model based on the probability relationship between the finely screened features and credit default, and confirm the method used to calculate the credit default probability. The data samples obtained by the data collection module are samples of customers who have used Internet credit services.

29. The device according to claim 28, wherein The device executes the steps of the method for constructing a retail credit risk prediction model according to any one of claims 18 to 27.

30. A system for constructing a retail credit risk prediction model for Internet credit business, characterized in that: The system includes: a memory, a processor, and a program for constructing a retail credit risk prediction model stored in the memory and executable on the processor. When the program for constructing a retail credit risk prediction model is executed by the processor, the steps of constructing a retail credit risk prediction model as described in any one of claims 18 to 27 are implemented.

Images