A method for high-dimensional variable screening of credit variable data

Abstract

The present invention relates to the field of credit scoring modeling, and provides a method for high-dimensional variable screening of credit variable data. The purpose is to solve how to fully screen high-dimensional variables, and to try to construct a new variable screening method to assist credit scoring modeling, so as to ensure the optimal accuracy of the obtained modeling model. The main scheme includes obtaining application data, loan data, overdue data, and credit data; intercepting data for a period of time, and preliminarily screening the extracted data according to the cumauc induction method, and selecting data with large AUC values; performing chi-square binning on the preliminarily screened variables, and the chi-square binning makes each data have a separate weight, introduces nonlinearity to the subsequent scoring model, reduces the risk of model overfitting, and selects data with higher weights; performs stepwise regression analysis on the screened data, and finally obtains all data that meet the screening conditions.

Description

Technical Field

[0001] The present invention relates to the field of credit scoring modeling, and in particular to a method for high-dimensional variable screening of credit variable data. Background Art

[0002] Risk control technology is one of the cornerstones of modern finance, and risk control models play a key role in this technology. In the era of internet finance, due to its inclusive nature, targeting individuals and small and micro businesses, automated risk control is the only way to reduce risk costs. Risk control modeling is widely used in automated risk control.

[0003] The effectiveness of risk control modeling depends on the underlying data. This determines the crucial impact of the model's sample set and input variables on its effectiveness. Before risk control modeling, financial institutions obtain a large number of basic and derived variables from users, screening them as input variables for the risk control models. These filtered raw input variables often exhibit some degree of correlation and association, which can affect the accuracy of risk control model training. Therefore, the proper screening of high-dimensional variables is crucial.

[0004] Common methods for credit score variable screening and modeling include IV (Information-Value) screening, KS (Kolmogorov-Smirnov) screening, and PCA dimensionality reduction. In their invention patent, "Automatic Optimization Method and System for Credit Scorecard Systems," patent application number 201910491304.5, Jiang Haibo et al. disclose a method that divides variables into multiple segmentation patterns, calculates the IV value for each segmentation, determines the optimal segmentation pattern and its corresponding maximum IV value, and then selects variables based on the maximum IV value for each variable. This method considers the variation in IV values ​​across different segmentation conditions, but its drawback is that it only considers IV values, and its ability to screen and resolve variables is limited. In the invention patent "A Method for Reducing the Dimensionality of Variables Entering a Risk Control Model," patent application number 202010295782.1, Liu Xingchen et al. disclosed a method of resampling the original samples to obtain modeling samples to ensure the category balance of the modeling sample set; for the modeling sample set, the original variables of the risk control model are selected and preprocessed such as missing value filling and outlier replacement; linear discriminant analysis is performed on the continuous variables in the original variables of the modeling sample set to achieve dimensionality reduction of the continuous variables; the continuous variables after dimensionality reduction are combined with non-continuous variables as the final modeling variables. This method reduces the correlation and association of variables, but the linear discriminant method is not applicable when the data dimension is high and is prone to overfitting. In the invention patent "Credit Assessment Variable Screening Method, Device, Computer Equipment and Storage Medium", patent application number 202011594339.0, Sun Xiangxin disclosed a method that uses variables as sorting factors, calculates the KS index value of each variable based on user data; screens each variable according to the KS index value to obtain a first candidate variable; screens the first candidate variables based on the correlation coefficient between each first candidate variable to obtain a second candidate variable; inputs the second candidate variable into the first credit assessment model to calculate the feature contribution of the second candidate variable; and screens the second candidate variable according to the feature contribution and KS index value of the second candidate variable. This method takes into account the KS value and feature contribution of the variable at the same time, ensuring a certain degree of accuracy, but KS represents the variable's ability to distinguish between good and bad. If there is a situation where the variable's distinguishing ability is consistent, the variable will not be able to be screened, which has certain limitations.

[0005] In view of the problems existing in the above-mentioned common financial scoring modeling variable screening methods, this paper focuses on how to fully screen high-dimensional variables and attempts to construct a new variable screening method to assist credit scoring modeling to ensure the optimal accuracy of the modeling model. Summary of the Invention

[0006] The purpose of the present invention is to solve the collinearity problems existing in traditional single methods and the limitations of single method for screening variables, and to form a more efficient and faster method for screening variables.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A credit scoring method for high-dimensional variable screening includes the following steps:

[0009] Step 1: Obtain all relevant credit variable data of the user, including application data, loan data, overdue data, and credit data. Label the credit data, including normal repayment customers and default customers. Save the credit variable data in matrix format to obtain the credit data set F1 = [x1, x2, ..., x n ], where x represents the credit data variable and the subscript n represents the number of credit data variables.

[0010] Step 2: intercept a period of data for variable screening and extraction. First, use the cumauc induction algorithm to screen the credit data characteristic variables, select stable and effective variables, and obtain the new credit data F2 = [x1, x2, ..., x n ];

[0011] Step 3: Perform chi-square binning on the data set F2 to obtain the latest credit data F3 = [x1, x2, ..., x m ], where the subscript m represents the number of credit data variables after the chi-square binning calculation and screening;

[0012] Step 4: For the credit data set F3 = [x1, x2, ..., x n ] Perform stepwise regression to obtain a new credit data set F4 = [x1, x2, ..., x n ], and credit score modeling is performed based on the variables finally screened.

[0013] In the above technical solution, step 2 includes the following steps:

[0014] Step 2.1, calculate the AUC of all individual credit data in the credit data F1 after fitting, and take the feature data with the largest AUC value;

[0015] Step 2.2: Based on step 2.1, add another corresponding credit data, that is, the AUC after the combination of the two variables, and take the feature data with the largest AUC;

[0016] Step 2.3, repeat the steps in step 2.2 until the AUC does not increase and does not meet the standard, that is, the variable screening of cumauc is completed, that is, the credit data variable set after screening is F2 = [x1, x2, ..., x w ].

[0017] By moving from single-variable screening of credit data to combined screening of multiple variables, the most stable and effective credit data variables are screened out.

[0018] In the above technical solution, step 3 includes the following steps:

[0019] Step 3.1: Preset a chi-square threshold value based on each variable in the credit variable data F2;

[0020] Step 3.2: Initialize each individual credit data variable and sort each instance according to its discrete attributes. Each instance belongs to a certain interval.

[0021] Step 3.3: Calculate the chi-square value of each adjacent interval and merge the pair of intervals with the smallest chi-square value.

[0022] Step 3.4: Calculate the post-loan performance of each bin of customers based on the binning results, and select credit variable data F3 whose post-loan performance meets the monotonicity.

[0023] Step 4 includes the following steps:

[0024] Step 4.1: Based on the credit data variable set F3 = [x1, x2, ..., x m ]Establish a univariate regression equation for each credit data independent variable and dependent variable: y = a i x i +b i ,i=1,2,…,n, where a i represents the coefficient of the univariate regression equation established for each credit data, b i represents the intercept of the univariate regression equation proposed for each credit data;

[0025] Step 4.2: Calculate the test statistic F of the regression coefficients in the z univariate regression equations respectively, and find the maximum value. like, Stop filtering, otherwise x k1 Add to the variable set, now you can add x k1 Consider x1 and proceed to step 4.3, where One-sided test results of the univariate regression equation, where the subscript m represents the number of credit data variables involved in the chi-square binning, and x k1 Eligible credit data variables, F a is the F-test function, 1 is the numerator degree of freedom, and n-2 is the denominator degree of freedom.

[0026] Step 4.3: Establish binary regression equations for the independent variables (x1, x2), (x1, x3), ..., (x1, x2) and the dependent variable. In this case, x1 is the x in step 2. k1, calculate x2,x3,...,x in the equation m is the regression coefficient test statistic F, take like Then stop filtering, otherwise x k2 Select the variable set, and then set x k2 Treat it as x2 and iterate in sequence until the maximum F value of the independent variable is less than the critical value.

[0027] In the binary regression equation established in this step, The one-sided test results of the binary regression equation, the subscript m represents the number of credit data variables involved in the chi-square binning binary regression equation, x k2 Credit data variables that meet the conditions of the binary regression equation, F a is the F-test function, 1 is the numerator degree of freedom, and n-2-1 is the denominator degree of freedom.

[0028] The final credit data variable set is the selected variable set F4 = [x1, x2, ..., x n ].

[0029] Because the present invention adopts the above technical solution, it has the following beneficial effects:

[0030] 1. This invention introduces the stepwise regression analysis algorithm into the field of risk control credit, processes high-dimensional variables in risk control modeling, considers the explanatory power of a large amount of credit data, and also considers the explanatory power of existing data after the new data is added. Ultimately, inappropriate data is eliminated from all data, leaving only valid data, which can eliminate the collinearity problem caused by high-dimensional multivariate.

[0031] 2. This invention uses the CUMAUC induction algorithm, which first performs logistic regression to screen credit data variables, then combines iterative backward feature elimination with RFE, and finally calculates the AUC loop. The algorithm first processes individual credit data variables, then combines them with multiple variables for screening, and then screens credit data variables based on the calculated maximum AUC value. The CUMAUC induction algorithm uses a process that proceeds from screening individual credit data to screening multiple credit data combinations, maximizing the accuracy of the subsequent credit data variables entering the model.

[0032] 3. The cumauc classification algorithm of this invention incorporates RFE. The entire iterative process eliminates features, removing irrelevant credit data while retaining an optimized subset of credit data features that are relatively important for discrimination. This optimizes the credit data feature subset and improves discrimination accuracy. This effectively improves the validity and stability of credit data input into the model.

[0033] 4. This invention uses the chi-square binning algorithm, which discretizes data from the bottom up, ensuring that relative frequencies are completely consistent within a bin. If two adjacent bins have very similar class distributions, they can be merged. Otherwise, they should be separated. Low chi-square values ​​indicate similar class distributions, ensuring good interpretability of the data into the mold.

[0034] 5. Compared to traditional feature selection methods, this method uses a combination of the CuMauc induction algorithm, binning, and stepwise regression to screen features from credit data. The screened data is then used for subsequent credit scoring modeling. For practical purposes, this method overcomes the collinearity and single-method limitations of traditional single-method credit data variable selection, resulting in a more discriminatory credit scoring model. DETAILED DESCRIPTION

[0035] Extract customer application information data, loan data, loan overdue information data, and People's Bank of China credit information data to form high-dimensional credit variable data as basic credit data variables. The user transaction flow data is as follows:

[0036] time XXXX Name XXXX ID number XXXX Household registration address XXXX Application amount XXXX Loan amount XXXX Overdue days XXXX ... XXXX Principal repaid XXXX Total loan amount XXXX

[0037] The basic format of a single data is shown in the table above. A single user behavior sequence data consists of a series of the above-mentioned flow data.

[0038] The process is as follows:

[0039] Step 1: Obtain all relevant user application data, loan data, overdue data, and credit data. Label the credit data with labels for regular repayments and defaults. Save the data file as a 100,000 × 430 matrix, and obtain the credit variable dataset F1 = [name, ID number, ..., total loan amount].

[0040] Step 2: intercept a period of data for variable screening and extraction. First, use the cumauc induction algorithm to screen the credit data characteristic variables, select stable and effective variables, and obtain new credit data F2 = [name, ID number, ..., total loan amount];

[0041] Step 3: Perform chi-square binning on the credit variable data F2 to obtain the latest variable set F3 = [name, ID number, ..., total loan amount];

[0042] Step 4: Perform a stepwise regression analysis on the credit variable data set F3 = [name, ID number, ..., total loan amount] to obtain new credit variable data F4. Based on the final screened credit variable data, perform credit scoring modeling.

[0043] In the above technical solution, step 2 includes the following steps:

[0044] Step 2.1: Calculate the AUC of all individual variables in the credit data in F1 after fitting. There are 430 variables involved in this calculation. Select the variables with AUC values ​​greater than 0.64.

[0045] Step 2.2: Add another variable based on step 1, that is, the AUC of the combination of two variables. Select the variable with the AUC value of the combination variable greater than 0.64 again, and the number of remaining variables is 80;

[0046] Step 2.3, repeat step 2 until the increase in AUC does not meet the standard, completing the credit data variable screening of the cumauc induction algorithm, that is, the filtered variable set is F2 = [total loan principal, total number of loan periods, ..., current number of overdue credit reports].

[0047] By moving from single-variable screening to multi-variable combination screening, the most stable and effective variables are screened out.

[0048] In the above technical solution, step 3 includes the following steps:

[0049] Step 3.1: Select the credit data variable "Total Loan Principal" in F2 and set the chi-square threshold to 3.8;

[0050] Step 3.2: Initialize, sort the credit data variables according to the discrete attributes, and each instance belongs to an interval.

[0051] Step 3.3: Calculate the chi-square value of each adjacent interval and merge the pair of intervals with the smallest chi-square value.

[0052] Step 3.4: Merge the intervals of the credit data variable "Total Loan Principal" and divide it into 7 boxes. The post-loan performance of the customers in each box does not conform to monotonicity.

[0053] Calculate the credit data variables in F2 in sequence, and select the credit data variables of each box of customers whose post-loan performance meets the monotonicity and enter them into the credit data variables F3.

[0054] In the above technical solution, step 4 includes the following steps: Vespa serrata

[0055] Step 4.1: Based on the credit variable dataset F3 = [total loan principal, housing provident fund contribution base, ..., current overdue credit number], establish a univariate regression equation for the total loan principal: y = 0.76x1 + 11, and calculate the regression equation for each variable in turn.

[0056] Step 4.2: Calculate the test statistic F for each regression coefficient in the univariate regression equation. 47 variables satisfy the test statistic for the univariate regression equation. The remaining 47 variables are used as new data and proceed to step 4.3.

[0057] Step 4.3: Calculate the test statistic F of the regression coefficients in the binary regression equation for the remaining credit data variables. 15 variables satisfy the test statistic for the binary regression equation. The final credit data set is the screened credit data set F4 = [total loan principal, housing provident fund contribution base, ..., current number of overdue credit reports]. The remaining 15 variables will be used for scorecard modeling.

Claims

1. A method for high-dimensional variable screening of credit variable data, characterized in that: The following steps are involved: Step 1: Obtain all relevant credit variable data of the user, including application data, loan data, overdue data, and credit data. Label the credit data, including normal repayment customers and default customers. Save the credit variable data in matrix format to obtain the credit data set F1 = [x1, x2, ..., x n ], where x represents the credit data variable and the subscript n represents the number of credit data variables; Step 2: intercept a period of credit variable data for variable screening and extraction. First, use the cumauc induction algorithm to screen the credit variable data for credit data characteristic variables, select stable and effective variables, and obtain new credit data F2 = [x1, x2, ..., x n ]; Step 3: Perform chi-square binning on the dataset F2 to obtain the latest credit data F3 = [x1,x2,……,x m ] The subscript m represents the number of credit data variables after the chi-square binning calculation and screening; Step 4: For the credit data set F3 = [x1, x2, ..., x n ] Perform stepwise regression to obtain new data set variables, and perform credit score modeling based on the final filtered data set variables; Step 2 includes the following steps: Step 2.

1. Calculate the AUC of all credit data in the credit dataset F1 after fitting, and take the feature data with the largest AUC value. Step 2.2: Based on step 2.1, add another corresponding credit data, that is, the AUC after the combination of the two variables, and take the feature data with the largest AUC; Step 2.3, repeat the steps in step 2.2 until the AUC does not increase and does not meet the standard, that is, the variable screening of cumauc is completed, that is, the credit data variable set after screening is F2 = [x1, x2, ..., x w ]; By moving from single-variable screening of credit data to combined screening of multiple variables, the most stable and effective credit data variables are screened out.

2. The method for high-dimensional variable screening of credit variable data according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1: Preset a chi-square threshold value based on each variable in the credit variable data F2; Step 3.2: Initialize each individual credit data variable and sort each instance according to its discrete attributes. Each instance belongs to a certain interval. Step 3.3: Calculate the chi-square value of each adjacent interval and merge the pair of intervals with the smallest chi-square value; Step 3.4: Calculate the post-loan performance of each bin of customers based on the binning results, and select credit variable data F3 whose post-loan performance meets the monotonicity.

3. The method for high-dimensional variable screening of credit variable data according to claim 1, characterized in that Step 4 includes the following steps: Step 4.1: Based on the credit data variable set F3 = [x1, x2, ..., x m ]Establish a univariate regression equation for each credit data independent variable and dependent variable: y = a i x i +b i ,i=1,2,…,n, where a i represents the coefficient of the univariate regression equation established for each credit data, b i represents the intercept of the univariate regression equation proposed for each credit data; Step 4.2: Calculate the regression coefficient a in each of the m univariate regression equations i The test statistic F of , and find the maximum value like Stop filtering, otherwise it will correspond Credit data variable x k1 Add to the variable set, now x k1 Consider x1 and proceed to step 4.3, where One-sided test results of the univariate regression equation, where the subscript m represents the number of credit data variables involved in the chi-square binning, and x k1 is a qualified credit data variable, F a is the F test function, 1 is the numerator degree of freedom, and n-2 is the denominator degree of freedom; Step 4.3, respectively, the independent variables (x1, x2), (x1, x3), ..., (x1, x m ) and the dependent variable to establish a binary regression equation, where x1 is the x in step 4.

2. k1 , calculate x2,x3,...,x in the equation m is the regression coefficient test statistic F, take like Then stop filtering, otherwise x k2 Select the variable set, and then set x k2 Treat it as x2 and iterate until the maximum F value of the independent variable is less than the critical value; In the binary regression equation established in this step, The one-sided test results of the binary regression equation, the subscript m represents the number of credit data variables involved in the chi-square binning binary regression equation, x k2 is a credit data variable that meets the conditions of the binary regression equation, F a is the F test function, 1 is the numerator degree of freedom, n-2-1 is the denominator degree of freedom, and the final credit data variable set is the selected variable set F4=[x1,x2,……,x n ]。

Images