Financial anti-fraud model and device based on enterprise finance and tax data

By screening features from multiple subdivided dimensions and building a score card model, the problem of incomplete assessment of tax fraud risk in existing technology small and medium-sized enterprises is solved, and a comprehensive assessment and accurate identification of financial fraud risks in corporate fiscal and tax data is achieved.

CN118967168APending Publication Date: 2024-11-15NANJING BINGJIAN INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411142499.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing financial anti-fraud model has limitations in identifying and evaluating the tax fraud risks of small and micro enterprises. It fails to derive features from multiple subdivided dimensions, resulting in incomplete risk assessment, degraded model performance, and inability to accurately identify potential fraud behaviors. .

Method used

Receive corporate financial and tax data through the protocol interface, perform data analysis and feature derivation, and perform feature screening from six subdivided dimensions, including basic information, amount abnormality, declaration concentration, timeliness of payment, tax violation records, and fiscal and tax deviation. Use the card The square binning method draws an IV diagram, and combines business logic to build a score card model. All feature scores are set to subtraction form, and a logistic regression model is constructed and converted into a score card model to perform model evaluation and adjustment.

Benefits of technology

A comprehensive assessment of the risk of financial fraud in corporate fiscal and taxation data is achieved, and enterprises with serious fraud risks can be accurately identified, improving the identification ability and accuracy of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967168A_ABST
    Figure CN118967168A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of finance and taxation systems, and discloses a financial anti-fraud model and device based on enterprise finance and taxation data, and the financial anti-fraud model based on enterprise finance and taxation data comprises the following steps: a, receiving structural data related to tax and finance from an enterprise through a protocol interface; b, carrying out data analysis on the structured data; and c, performing feature derivation on the analyzed data from six subdivision dimensions including basic information, amount abnormity, declaration concentration degree, payment withstanding timeliness, tax law violation records and finance and tax deviation degree. According to the scheme, the structured data related to tax and finance from an enterprise is received through a protocol interface, then data analysis is performed on the structured data, and financial fraud risks in the finance and taxation data are comprehensively mined from six subdivision dimensions including basic information, amount abnormity, declaration concentration degree, payment withstanding timeliness, tax illegal records and finance and taxation deviation degree.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of financial and taxation systems, and in particular to a financial anti-fraud model and device based on corporate financial and taxation data. Background Art

[0002] In the current financial anti-fraud field, traditional risk assessment models mainly rely on corporate invoice data to identify tax fraud risks. However, this single data source approach has obvious limitations because it fails to cover potential risk signals in tax data, especially possible fraud in the tax declaration process of small and micro enterprises. Due to their size and complexity, small and micro enterprises often have a higher risk of fraud in tax declarations, but existing risk control models have not been able to effectively identify and assess this risk.

[0003] In addition, in the process of building existing risk control models, features are often not derived from multiple subdivision dimensions for machine learning model training. This practice may result in the omission of features of certain key dimensions when screening features for model input, making it impossible for the model to assess risks from a more comprehensive perspective. This one-sided risk assessment may lead to a decline in model performance and an inability to accurately identify and predict potential fraud.

[0004] Therefore, there is a need for a financial anti-fraud model and device based on corporate financial and tax data to perform feature screening from multiple subdivision dimensions to more comprehensively assess the financial fraud risks in corporate financial and tax data. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a financial anti-fraud model and device based on corporate financial and tax data.

[0006] To solve the above problems, the present invention adopts the following technical solutions:

[0007] A financial anti-fraud model based on corporate financial and tax data includes the following steps:

[0008] a. Receive structured data related to taxation and finance from enterprises through protocol interfaces;

[0009] b. performing data parsing on the structured data;

[0010] c. Derive features from the parsed data from six subdivision dimensions: basic information, abnormal amount, declaration concentration, timeliness of declaration and payment, tax violation record, and fiscal and tax deviation;

[0011] d. Concatenate the processed feature data with the sample label y, and use the chi-square binning method to draw an IV graph for analysis of the derived features;

[0012] e. Combine the IV diagram and business logic to gradually filter the input features, build a logistic regression model, and convert it into a scorecard model. The scores of all features in the scorecard are set to deduct points;

[0013] f. When calculating the enterprise score, add 1,000 basic points to get the final score;

[0014] g. Use the constructed scorecard model to score the samples, and evaluate and adjust the model based on the model scores;

[0015] h. Divide the final model evaluation score into five levels: A, B, C, D, and E according to a certain sample proportion.

[0016] As a further description of the above technical solution: the data parsing step includes cleaning, formatting and standardizing the structured data.

[0017] As a further description of the above technical solution: the feature derivation step includes statistical analysis and anomaly detection on the data of each subdivision dimension.

[0018] As a further description of the above technical solution: the chi-square binning method includes discretizing continuous variables and performing a chi-square test on the discretized variables.

[0019] As a further description of the above technical solution: the feature scores in the scorecard model are set according to their weights and business logic in the model.

[0020] As a further description of the above technical solution: the model evaluation and adjustment step includes using cross-validation and ROC curve analysis to analyze model performance.

[0021] The present invention also provides a financial anti-fraud device based on corporate financial and tax data, comprising:

[0022] A data receiving module, used for receiving structured data related to taxation and finance from enterprises through a protocol interface;

[0023] A data parsing module, used for performing data parsing on the structured data;

[0024] The feature derivation module is used to derive features from six subdivisions of the parsed data, including basic information, abnormal amount, declaration concentration, timeliness of declaration and payment, tax violation records, and fiscal and tax deviation;

[0025] The feature screening module is used to splice the processed feature data with the sample label y, and use the chi-square binning method to draw an IV graph for analysis of the derived features. Combine the IV graph with the business logic to gradually screen the features to be entered into the model;

[0026] The model building module is used to build a logistic regression model and convert it into a scorecard model. The scores of all features in the scorecard are set to deduct points.

[0027] The scoring calculation module is used to add 1000 points to the basic score when calculating the enterprise score to get the final score;

[0028] The model evaluation module is used to score samples using the constructed scorecard model and to evaluate and adjust the model based on the model score;

[0029] The grading module is used to divide the final model evaluation scores into five grades: A, B, C, D, and E according to a certain sample proportion.

[0030] As a further description of the above technical solution: the data analysis module also includes a processing unit for data cleaning, formatting and standardization, and the feature derivation module also includes a statistical analysis and anomaly detection unit.

[0031] As a further description of the above technical solution: the feature screening module also includes a chi-square binning and chi-square test unit; the model building module also includes a feature score setting unit.

[0032] As a further description of the above technical solution: the model evaluation module also includes a cross-validation and ROC curve analysis unit.

[0033] Compared with the prior art, the advantages of the present invention are:

[0034] 1. This solution receives structured data related to taxation and finance from enterprises through a protocol interface, and then performs data analysis on the structured data, and comprehensively mines the financial fraud risks in financial and taxation data from six subdivision dimensions, including basic information, abnormal amounts, declaration concentration, timeliness of payment, tax violation records, and financial and tax deviation.

[0035] 2. The present invention combines the IV graph trend with business experience to screen out the input features, build a machine learning model, and finally convert the model into a score card. All features are set to reduce points to identify companies with serious fraud risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the financial anti-fraud model based on corporate financial and tax data of the present invention;

[0037] Figure 2 This is a system diagram of the financial anti-fraud device based on corporate financial and tax data of the present invention. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments, and all other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work are within the scope of protection of the present invention.

[0039] See also Figure 1 , a financial anti-fraud model based on corporate financial and tax data, aims to build a model that combines machine learning and business experience from the entire process, so as to fully explore the financial fraud risks existing in corporate financial and tax data, including the following steps:

[0040] a. Receive structured data related to taxation and finance from enterprises through protocol interfaces.

[0041] b. Perform data analysis on structured data, which includes cleaning, formatting and standardizing structured data.

[0042] c. Based on the parsed data, feature derivation is performed from six sub-dimensions, namely basic information, abnormal amount, declaration concentration, timeliness of declaration and payment, tax violation record, and fiscal and tax deviation. Feature derivation includes statistical analysis and anomaly detection of the data in each sub-dimension.

[0043] The financial and tax data of enterprises in the parsing protocol interface will obtain basic information such as taxpayer status, enterprise registration time, legal representative information, etc., as well as specific financial and tax data such as enterprise tax return, tax arrears information form, enterprise balance sheet, income statement, cash flow statement, etc. From these different report information, based on business experience, the derivative logic of each feature is determined from the perspectives of the number of changes in financial and tax data over the years and abnormal situations, such as: the number of records with a total sales revenue growth rate of less than 0 for each period of cumulative income tax in the past year, the number of times tax returns were not made on time every month or quarter in the past year, etc. Then use the python programming language to write the derivative logic code, and then calculate all the features for all sample data. According to the characteristics of these features, they are divided into 6 dimensions, which can fully cover the information dimensions existing in financial and tax data.

[0044] d. Concatenate the processed feature data with the sample label y, and use the chi-square binning method to draw an iv diagram for analysis of the derived features.

[0045] The sample data are concatenated with the corresponding labels (y) to construct an industry-specific training set. For each derived feature, the continuous variable is discretized using the chi-square binning method, the IV value is calculated, the IV graph is drawn, and the bad debt rate corresponding to each bin is calculated.

[0046] Among them, chi-square binning is a commonly used feature discretization method, which is used to convert continuous variables into ordered discrete variables and perform chi-square test on the discretized variables. The optimal binning boundary is determined by comparing the difference between the observed frequency and the expected frequency. This method can effectively reduce the impact of outliers on the model.

[0047] Chi-square binning includes the following key steps:

[0048] 1) Divide the continuous variable into initial binning intervals according to its value range;

[0049] 2) Count the number of samples in each bin interval to form an initial observation frequency matrix;

[0050] 3) According to the distribution of the overall sample, calculate the expected frequency of each bin interval for subsequent chi-square test;

[0051] 4) Use the chi-square test to measure the difference between the observed frequency and the expected frequency, calculate the chi-square statistic, and measure the goodness of fit of the overall model;

[0052] 5) Consider adjacent bins one by one from left to right, and calculate the chi-square statistic after merging. If the merger can significantly reduce the chi-square statistic, merge the two bins into a new bin;

[0053] 6) Repeat the merging operation of adjacent bins until the chi-square statistic cannot be further reduced;

[0054] 7) Determine the conditions for stopping merging based on business needs and model effects, such as the number of bins, the minimum number of samples, etc.;

[0055] 8) The final binning boundary is obtained based on the merging results to form discretized features.

[0056] The IV value is used to encode the features and evaluate their predictive power. The ratio of the current positive example yi to the global positive example ys and the ratio of the current group negative example ni to the global negative example ns are calculated in each bin, and then calculated using Formula 1. The bad rate refers to the ratio of bad samples (such as defaults, arrears, etc.) observed in each bin to the total number of samples, that is, the ratio of the current group negative example ni to the global negative example ns.

[0057]

[0058] With the group number as the horizontal axis and the number of each group and the bad rate as the double vertical axes, the corresponding IV graph can be drawn. The IV graph can intuitively show the trend of the proportion of counterexamples corresponding to the change of the feature value. The feature trend here refers to the degree of correlation between the change trend and direction between the feature and the prediction target. In order to measure this degree of correlation, the correlation coefficient can be used as a statistical indicator. If the change trend of the feature is consistent with the change trend of the prediction target, and the correlation coefficient is high (close to 1), it indicates that the trend of the feature is good. If the change trend of the feature is opposite to the change trend of the prediction target, and the correlation coefficient is low (close to -1), it also indicates that the trend is good. However, if the change trend of the prediction target is consistent with the change trend of the feature in some feature value intervals, and opposite in other intervals, the correlation is low and the trend is poor.

[0059] e. Combine the IV diagram and business logic to gradually screen the features to be included in the model, build a logistic regression model, and convert it into a scoring card model. The scores of all features in the scoring card are set to a deduction form. The feature scores in the scoring card model are set according to their weights in the model and business logic.

[0060] The input features are gradually screened from the derived features. The features in the binary classification model are mainly screened by combining the IV value and the feature trend. The top 20 features with IV values ​​greater than 0.2 and good feature trends are screened as input features. The logistic regression model is constructed using these input features, and the coefficients in the logistic regression model are converted into scores in the score card. WOE (Weight of Evidence) and IV (Information Value) are usually used for conversion. WOE conversion: WOE is used to measure the degree of influence of feature values ​​on the target variable. First, the continuous features are discretized into a set of different bins, each bin representing a specific value interval. Then, the good-bad ratio of the target variable in each bin is calculated (the ratio of the number of good customers to the number of bad customers). Then, the WOE of each bin is calculated, that is, the difference of the good-bad ratio of the natural logarithm. WOE indicates the degree of influence of the feature value on the target variable. A positive value indicates a positive impact on good customers, and a negative value indicates a positive impact on bad customers. IV conversion: IV is a measure of the amount of information of the feature in the prediction model. By weighting the WOE values ​​of all bins, the information value (IV) of the feature can be calculated. The higher the IV, the stronger the predictive power of the feature for the target variable.

[0061] According to the converted coefficients and feature values, the score of each feature is calculated. Generally speaking, the scorecard model uses a linear weighted method to calculate the score. For each feature, the WOE corresponding to its value is multiplied by the corresponding coefficient, and then the intercept term of the model is added to finally obtain the score of each feature. This is the first version of the scorecard model. Here, combined with the business rules we set, all feature scores in the scorecard are designed in the form of deduction. In the process of setting the feature deduction items, the Pearson correlation coefficient is first used to measure the correlation between features. The Pearson correlation coefficient is a statistical indicator used to measure the strength and direction of the linear relationship between two continuous variables. Its value range is between -1 and 1. When the coefficient is 1, it indicates a complete positive correlation; when the coefficient is -1, it indicates a complete negative correlation; when the coefficient is close to 0, it indicates that there is no linear relationship between the two variables. The features with the absolute value of the correlation coefficient between the two features greater than 0.5 are screened out and classified into a rule feature group. The features with the absolute value of the correlation coefficient less than 0.5 are each a rule feature group. The deduction weight of each rule feature group is the same, and the deduction weight within the same feature group is evenly distributed. In this way, in the process of rule deduction, the enterprises that hit the same type of rules will not be deducted too much, thereby reducing the accuracy of the model.

[0062] f. When calculating the enterprise score, add 1,000 basic points to get the final score;

[0063] g. Use the constructed scorecard model to score the samples and evaluate and adjust the model based on the model score, including using cross-validation and ROC curve to analyze model performance.

[0064] h. The final model evaluation score is divided into five levels: A, B, C, D, and E according to a certain sample ratio. The number of samples in each level is divided according to a certain ratio to the total number of samples in the test set. For example, the A level samples account for 10%, the B level samples account for 25%, the C level samples account for 30%, the D level samples account for 25%, and the E level samples account for 10%. In this way, the score range corresponding to each level can be obtained.

[0065] See also Figure 2 , a financial anti-fraud device based on corporate financial and tax data, including a data receiving module, a data parsing module, a feature derivation module, a feature derivation module, a model building module, a scoring calculation module, a model evaluation module and a grade classification module.

[0066] The data receiving module is used to receive structured data related to taxation and finance from enterprises through a protocol interface.

[0067] The data analysis module is used to analyze structured data; the data analysis module includes processing units for data cleaning, formatting and standardization.

[0068] The feature derivation module is used to derive features from six sub-dimensions of the analyzed data, namely basic information, abnormal amount, declaration concentration, timeliness of payment, tax violation record, and fiscal and tax deviation; the feature derivation module includes statistical analysis and anomaly detection units.

[0069] The feature derivation module is used to concatenate the processed feature data with the sample label y, and use the chi-square binning method to draw an IV diagram for analysis of the derived features. Combining the IV diagram with business logic, the features to be input into the model are gradually screened. The feature screening module also includes chi-square binning and chi-square test units.

[0070] The model building module is used to build a logistic regression model and convert it into a scorecard model. The scores of all features in the scorecard are set in a deduction form; the model building module also includes a feature score setting unit.

[0071] The score calculation module is used to calculate the enterprise score by adding 1000 basic points to get the final score;

[0072] The model evaluation module is used to score samples using the constructed scorecard model, and to evaluate and adjust the model based on the model score; the model evaluation module also includes cross-validation and ROC curve analysis units.

[0073] The grading module is used to divide the final model evaluation scores into five levels: A, B, C, D, and E according to a certain sample proportion.

[0074] In summary: the present invention receives the actual financial and tax data from the enterprise through the access protocol interface, and the construction of the entire model is based on more accurate and reliable data, so the entire model is more convincing.

[0075] The present invention will perform feature screening from six subdivision dimensions, and each dimension is developed around the fraud risk in the tax data, and the feature score items are set to negative scores in order to better identify the fraud risk of the enterprise.

[0076] After training the scoring card model, the present invention divides the evaluation scores of the samples into five levels: A, B, C, D, and E, and the score threshold of the level is calculated based on a certain sample proportion, so that the final scoring level conforms to the business logic.

[0077] The above is only a preferred specific implementation of the present invention; however, the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solution and its improved conception within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A financial anti-fraud model based on corporate financial and tax data, characterized by: The following steps are involved: a. Receive structured data related to taxation and finance from enterprises through protocol interfaces; b. performing data parsing on the structured data; c. Derive features from the parsed data from six subdivision dimensions: basic information, abnormal amount, declaration concentration, timeliness of declaration and payment, tax violation record, and fiscal and tax deviation; d. Concatenate the processed feature data with the sample label y, and use the chi-square binning method to draw an IV graph for analysis of the derived features; e. Combine the IV diagram and business logic to gradually filter the input features, build a logistic regression model, and convert it into a scorecard model. The scores of all features in the scorecard are set to deduct points; f. When calculating the enterprise score, add 1,000 basic points to get the final score; g. Use the constructed scorecard model to score the samples, and evaluate and adjust the model based on the model scores; h. Divide the final model evaluation score into five levels: A, B, C, D, and E according to a certain sample proportion.

2. A financial anti-fraud model based on corporate tax data according to claim 1, characterized in that: The data parsing step includes cleaning, formatting and standardizing the structured data.

3. The financial anti-fraud model based on corporate tax data according to claim 1 is characterized by: The feature derivation step includes statistical analysis and anomaly detection on the data of each segmentation dimension.

4. The financial anti-fraud model based on corporate tax data according to claim 1 is characterized by: The chi-square binning method includes discretizing continuous variables and performing a chi-square test on the discretized variables.

5. The financial anti-fraud model based on corporate tax data according to claim 1 is characterized by: The feature scores in the scorecard model are set according to their weights and business logic in the model.

6. The financial anti-fraud model based on corporate tax data according to claim 1 is characterized by: The model evaluation and tuning steps include analyzing model performance using cross-validation and ROC curves.

7. A financial anti-fraud device based on corporate tax data according to any one of claims 1 to 6, characterized in that: include: A data receiving module, used for receiving structured data related to taxation and finance from enterprises through a protocol interface; A data parsing module, used for performing data parsing on the structured data; The feature derivation module is used to derive features from six subdivisions of the parsed data, including basic information, abnormal amount, declaration concentration, timeliness of declaration and payment, tax violation records, and fiscal and tax deviation; The feature screening module is used to splice the processed feature data with the sample label y, and use the chi-square binning method to draw an IV graph for analysis of the derived features. Combine the IV graph with the business logic to gradually screen the features to be entered into the model; The model building module is used to build a logistic regression model and convert it into a scorecard model. The scores of all features in the scorecard are set to deduct points. The scoring calculation module is used to add 1000 points to the basic score when calculating the enterprise score to get the final score; The model evaluation module is used to score samples using the constructed scorecard model and to evaluate and adjust the model based on the model score; The grading module is used to divide the final model evaluation scores into five grades: A, B, C, D, and E according to a certain sample proportion.

8. A financial anti-fraud device based on corporate tax data according to claim 7, characterized in that: The data analysis module also includes a processing unit for data cleaning, formatting and standardization, and the feature derivation module also includes a statistical analysis and anomaly detection unit.

9. A financial anti-fraud device based on corporate tax data according to claim 7, characterized in that: The feature screening module also includes a chi-square binning and chi-square test unit; the model building module also includes a feature score setting unit.

10. A financial anti-fraud device based on corporate tax data according to claim 7, characterized in that: The model evaluation module also includes a cross-validation and ROC curve analysis unit.