A scoring modeling method based on the fusion credit data of enterprises in the financial leasing industry

Abstract

The present invention provides a scoring modeling method based on the fusion of corporate credit data in the financial leasing industry. The method includes three steps: corporate public information modeling, corporate credit information modeling, and credit data fusion scoring. By respectively establishing a corporate public information model and a corporate credit information model, the corporate public information score and the corporate credit information score are obtained respectively. Then, the corporate credit fusion score is obtained according to the agreed weight rule, and the business application system displays the final fusion scoring result and key characteristic variables.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a scoring modeling method based on the fusion of corporate credit data in the financial leasing industry. Background Art

[0002] In recent years, credit data, as an important production factor, has played an important role in scenarios such as corporate credit assessment and government supervision. Non-bank financial institutions such as financial leasing, commercial factoring, and microfinance generally face some pain points in the process of credit data collection, processing, and application. It is urgent to adopt a corporate credit data fusion scoring modeling method to solve the data security, compliance, and timeliness requirements. On the one hand, there are multiple data service agencies in the market for public information on business, justice, taxation, intellectual property, and other government affairs, and the data format is not unified. In addition, the text content of this part of information is numerous and requires professional interpretation and analysis; on the other hand, the use of corporate credit information has strict authorization requirements, and the cumbersome approval process will reduce the timeliness of data use, making it difficult to maximize the value of data, and the credit behavior analysis of corporate information subjects faces challenges. Summary of the invention

[0003] The patent of this invention is a method for solving the problem of enterprise credit data fusion scoring modeling based on the financial leasing industry. Through big data software, massive information is quickly interpreted, the underlying correlation of data is mined, and an objective statistical model is constructed. Through model scoring, enterprise risks are scientifically predicted, and the application value and efficiency of data are comprehensively improved. The patent of this invention provides a solution for the modeling data preprocessing, model design and evaluation, multi-model fusion, system integration application and other tasks of enterprise credit data.

[0004] In this method, enterprise public information refers to government information from industry and commerce, judicial, taxation, intellectual property, etc. Enterprise public information has many data items and is scattered and uncentralized; users need to log in to multiple data platforms to obtain the required data, which is inefficient, has a long data service link, and is complicated to operate.

[0005] In this method, corporate credit information refers to information from corporate credit reports, including basic corporate information, repayment performance information, and guarantee information. The use of corporate credit information data requires strict requirements. Under the premise of obtaining authorization from the information subject, the internal application of corporate credit report information in the company also requires strict and standardized management; the approval process for the use of data is complex and the timeliness needs to be improved. Corporate credit report information is generally in PDF format. Credit information is highly professional and requires data structured analysis before professional interpretation and analysis.

[0006] A scoring modeling method based on the fusion of corporate credit data in the financial leasing industry is used for the modeling process of corporate public information and corporate credit information data for corporate credit risk assessment, including technical fields such as data preprocessing, model design and evaluation, multi-model fusion, and system integration application.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] A scoring modeling method based on the fusion of credit data of enterprises in the financial leasing industry includes three steps: enterprise public information modeling, enterprise credit information modeling, and credit data fusion scoring.

[0009] (1) Enterprise public information modeling:

[0010] Based on the identification information of the enterprise information subject, retrieve the enterprise basic information through the enterprise public information query API, and store the retrieved enterprise basic information in the business system memory, wherein the identification information of the enterprise information subject includes but is not limited to the enterprise name and / or unified social credit code;

[0011] Define the modeling target according to the project application requirements, use the logistic regression model as the core modeling technology, use the retrieved enterprise basic information data as sample data, model the enterprise public information, and obtain the enterprise public information model. The model results of the enterprise public information model include variable name, variable meaning, variable value, and percentage score;

[0012] The overall score of the enterprise public information model is finally obtained by taking corresponding values ​​and scoring them on a percentage basis for each variable in the enterprise public information model.

[0013] The corresponding values ​​and percentage scores of each variable in the enterprise public information model are defined in the following table:

[0014]

[0015]

[0016] Specifically, in (1) enterprise public information modeling, a model is built for the enterprise public information to obtain an enterprise public information model; each variable in the enterprise public information model is assigned a value and a score, and finally an overall score of the enterprise public information model is obtained; specifically, the following is included:

[0017] (1.1) Data cleaning and analysis:

[0018] The retrieved enterprise basic information data is cleaned and calculated to obtain characteristic variables in the enterprise public information model. The data cleaning mainly includes: removing duplicate data, removing logical conflict data, completing some single variable calculations, processing noise data, abnormal values ​​and outliers, and processing missing values.

[0019] (1.2) Characteristic variable analysis:

[0020] Conduct statistical characteristics and distribution analysis on the characteristic variables in the obtained enterprise public information model, check extreme values ​​and process them;

[0021] Organize the results of the feature variable analysis into a feature variable table, recording the feature variable name, calculation logic, data coverage, and basic information of data distribution.

[0022] (1.3) Weight of Evidence (WOE) Analysis:

[0023] The logistic regression model is transformed into a standard scorecard format through WOE transformation to obtain the variable values ​​of the feature variables;

[0024] First, all feature variables are automatically binned. Then, the reliability of the automatic binning results is manually checked to see whether they meet business requirements and are interpretable. Then, it is determined whether manual binning is needed.

[0025] The WOE for each category is defined as follows:

[0026]

[0027] The columns Bad Distribution and Good Distribution represent the distribution of "bad customers" and "good customers" in each category, respectively. They are obtained by dividing the frequency count in each category by the total number of "bad customers" or "good customers".

[0028] If the ratio in the brackets is less than 1, the WOE is negative, otherwise the WOE is positive.

[0029] (1.4) Modeling and debugging:

[0030] Initialize a series of model variables, and fit a model based on the current series of variables. The model results of this fitted model include the name of the feature variable, the meaning of the variable, the variable value, and the percentile score. Then determine whether this fitted model is the optimal model.

[0031] If it is judged to be the optimal model, the final model and variables of the enterprise public information model are obtained;

[0032] If it is judged that it is not the optimal model, some variables are added to the model or some variables are deleted, and then a new model is fitted based on the current series of variables to determine whether the refitted model is the optimal model. Until the optimal model is found, the final model and variables of the enterprise public information model are obtained.

[0033] (1.5) Fractional linear transformation:

[0034] The score is linearly converted to 0-100 points, and the distribution characteristics remain unchanged. The conversion formula is:

[0035]

[0036] (1.6) Model Results Presentation

[0037] The final model input variables, variable values ​​and percentage scores of each variable of the enterprise public information model are displayed on the business system.

[0038] The (1.1) data cleaning analysis includes cleaning rules for general data and cleaning rules for specific data.

[0039] (A) General data cleaning rules, the specific processing methods are as follows:

[0040] (A1) Date fields: displayed in the format of YYYY-MM-DD;

[0041] (A2) Amount fields: All amounts are converted into numerical format and calculated in RMB 10,000;

[0042] (A3) Ratio fields: All ratios are converted into numerical format, percentage signs are removed, and zero is added if there is no zero before the decimal point;

[0043] (A4) Data duplication: For the same event, there may be multiple duplicate information records in the data table. Data deduplication is mainly identified by the keyword "company name + event unique identifier judgment".

[0044] (B) Cleaning rules for specific data. The specific processing methods are as follows:

[0045] (B1) Enterprise registration: If the enterprise registration date is empty, the date is corrected with the enterprise operation start date; delete the observations with empty enterprise registration date; delete the observations with empty enterprise operation start date; delete the observations with business status of cancellation or revocation, but the cancellation date or revocation date is not empty; delete the observations with business end date not empty, but business end date < business start date; delete the observations with cancellation date not empty, but cancellation date < enterprise registration date; delete the observations with revocation date not empty, but revocation date < enterprise registration date;

[0046] (B2) Main personnel: merge the job titles of the same company and the same person; if a person has multiple job titles and they are written on two lines, merge them into one line;

[0047] (B3) Information on the person subject to enforcement: Delete the observations where the filing date is not empty but the filing date is less than the enterprise registration date;

[0048] (B4) Court Announcement: Delete the observations where the filing date is not empty but the filing date is less than the enterprise registration date;

[0049] (B5) Dishonest debtor: Delete the observations where the publication date is not empty but the publication date is less than the enterprise registration date;

[0050] (B6) Operational anomalies: Delete observations where the inclusion date is not empty but the inclusion date is less than the enterprise registration date; delete observations where the removal date is less than the inclusion date;

[0051] (B7) Administrative penalties: Delete observations where the penalty decision date is not empty, but the penalty decision date is less than the enterprise registration date; if the penalty decision date is empty, replace it with the public announcement date;

[0052] (B8) Administrative license: Delete the observations where the license start date is not empty but the license start date is less than the enterprise registration date; delete the observations where the license end date is not empty but the license end date is less than the license start date;

[0053] (B9) Spot check: delete the observations where the spot check date is not empty but the spot check date is less than the enterprise registration date;

[0054] (B10) Industrial and Commercial Change: Delete the observations where the Industrial and Commercial Change-Change Date is not empty, but the Industrial and Commercial Change-Change Date is <= Enterprise Registration Date;

[0055] (B11) Outward investment: Delete the observations where the opening date is not empty but the opening date is less than the enterprise registration date;

[0056] (B12) Branches: Delete observations where the establishment date is not empty but the establishment date is less than the enterprise registration date;

[0057] (B13) Equity pledge: Delete the observations where the equity pledge registration date is not empty, but the equity pledge registration date is less than the enterprise registration date;

[0058] (B14) Movable Property Mortgage: Delete the observations where the registration date is not empty but the registration date is less than the enterprise registration date.

[0059] Specifically, in the modeling and debugging of the step (1.4), the judgment of whether the fitted model is the optimal model specifically includes: splitting the sample data into a training set and a validation set with a split ratio of 7:3, and during the splitting, it must be ensured that the proportion of bad samples in the training set and the validation set is consistent with the proportion of bad samples in the full data; obtaining the KS curve and the ROC curve of the training set and the validation set, and the one with the best effect is the optimal statistical model for quantitative analysis; calculating the model score of each sample based on the model results of the obtained optimal statistical model and the value of the characteristic variables of each sample data, and obtaining the sample score distribution of the optimal statistical model; reflecting the optimal statistical model's ability to distinguish different samples, stability and possible deviations through the sample score distribution, and judging whether the sample model score can best distinguish the status of good and bad samples (for example: bad samples are concentrated in the low segment, and good samples are concentrated in the high segment) according to the actual application scenario of the model; if the status of good and bad samples can be best distinguished, it is judged to be the optimal model; if the status of good and bad samples cannot be best distinguished, it is judged to be not the optimal model.

[0060] Among them, in the step (1.4) modeling debugging, the initialization of a series of model variables includes: removing variables that are obviously ineffective or have no business significance for modeling; removing variables with too low information values ​​and too high proportion of repeated values; and removing other variables except one retained variable from two or more variables with high Pearson correlation.

[0061] (2) Enterprise credit information modeling

[0062] The enterprise credit information is derived from the enterprise credit report, including: credit reminder information, loan transaction summary information, guarantee transaction summary information, loan account information, etc.;

[0063] The model is layered and key variables are initially screened through the "hierarchical analysis method". The business attributes, relevance, data coverage and other conditions of the variables are comprehensively considered, and scores are assigned through the "expert scoring method". Modeling is carried out for corporate credit information to obtain the corporate credit information model. The model results of the corporate credit information model include variable names, variable meanings, variable values, and percentage scores.

[0064] Based on the identification information of the enterprise information subject, check whether the enterprise information subject has a credit report; if there is a credit report, use the data with the latest report date in the database to determine whether the "Year of the first credit transaction" field in the "Credit reminder information unit" chapter in the credit report is empty. If the field is not empty, calculate the enterprise credit information score according to the enterprise credit information model. If the field is empty, the calculation of the enterprise credit information score is not supported and the enterprise credit information score is treated as empty; if there is no credit report, the calculation of the enterprise credit information score is not supported and the enterprise credit information score is treated as empty.

[0065] The corresponding values ​​and percentage scores of each variable in the enterprise credit information model are defined in the following table:

[0066]

[0067]

[0068] Enterprise credit information comes from enterprise credit reports, with a single data source and standardized structured data, including credit reminder information, loan transaction summary information, guarantee transaction summary information, loan account information, etc. In the aforementioned (2) enterprise credit information modeling, the aforementioned modeling of enterprise credit information to obtain an enterprise credit information model specifically includes:

[0069] (2.1) Obtain characteristic variables in the enterprise credit information model through data cleaning and calculation

[0070] Clean the enterprise credit information data, select important information dimensions, and then process and calculate the variables. If there are multiple records of credit reports being queried for the same enterprise in the credit information data table, sort them in reverse order by the time the credit reports were generated, and take the data of the latest credit report as the modeling sample data;

[0071] Through data cleaning, the retrieved corporate credit information data is re-examined and verified in order to discover and correct errors in the data files and reduce the impact of erroneous data on model performance. The data cleaning mainly includes: removing duplicate data, removing logical conflicting data, completing some single variable calculations, processing noise data, abnormal values ​​and outliers, and processing missing values.

[0072] (2.2) Feature analysis

[0073] Conduct statistical characteristic analysis and distribution analysis on characteristic variables in the enterprise credit information model, check extreme values ​​and process them;

[0074] The results of feature analysis are organized into a feature variable table, which records the feature variable name, calculation logic, data coverage, and basic information on data distribution.

[0075] (2.3) Weight of Evidence (WOE) Analysis

[0076] The variable values ​​of the characteristic variables are obtained through WOE transformation;

[0077] First, all feature variables are automatically binned. Then, the reliability of the automatic binning results is manually checked to see whether they meet business requirements and are interpretable. Then, it is determined whether manual binning is needed.

[0078] The WOE for each category is defined as follows:

[0079]

[0080] The columns Bad Distribution and Good Distribution represent the distribution of "bad customers" and "good customers" in each category, respectively. They are obtained by dividing the frequency count in each category by the total number of "bad customers" or "good customers".

[0081] If the ratio in the brackets is less than 1, the WOE is negative, otherwise the WOE is positive.

[0082] (2.4) Scoring design

[0083] (2.4.1) Overall sample score distribution

[0084] The enterprise credit information model is built according to the expert scoring rules. The score of each sample is obtained according to the scoring range and scoring rules of each variable. The overall sample score distribution analysis helps to identify whether the sample scores are concentrated, dispersed or have outliers. According to the score results, the model variables, variable values, and scores are adjusted in combination with the application scenarios of the financial leasing industry.

[0085] (2.4.2) Distribution of good and bad sample scores

[0086] The modeling goal is to define customers as "good" customers or "bad" customers based on the customer management classification of the actual project user;

[0087] Good and bad samples are distinguished based on the “bad definition” label;

[0088] Because each sample has a model score, we will make a score distribution based on good and bad samples to see whether the score can distinguish good and bad samples, that is, whether good samples are concentrated in high segments and bad samples are concentrated in low segments. According to the score results, we will adjust the model variables, variable values, and scores in combination with the application scenarios of the financial leasing industry.

[0089] Step (2.3) WOE analysis and step (2.4) score design are tuned multiple times to achieve a state where the sample score can best distinguish good and bad samples, and the final model and variables of the enterprise credit information model are obtained;

[0090] (2.5) Model Results Presentation

[0091] The final model input variables, variable values ​​and percentage scores of each variable in the enterprise credit information model are displayed on the business system.

[0092] (3) Credit data integration scoring

[0093] The overall score of the enterprise public information model obtained in step (1) and the enterprise credit information score obtained in step (2) are output to the business application system through the credit data fusion scoring interface according to the agreed weight rules, and the business application system displays the final fusion scoring results and key feature variables;

[0094] The output content of the credit data fusion scoring interface includes but is not limited to: enterprise name, unified social credit code, credit fusion score, enterprise public information score, enterprise credit information score, value and score of each variable of the enterprise public information model, and value and score of each variable of the enterprise credit information model.

[0095] Among them, in (3) credit data fusion scoring, the agreed weight rule can be: when the enterprise credit information score is not empty, the credit data fusion score is calculated according to the weight ratio of the enterprise public information model score to the enterprise credit information score of 4:6; when the enterprise credit information score is empty, the enterprise credit data fusion score is consistent with its enterprise public information model score. Of course, the weight rule can also be modified and adjusted according to business needs.

[0096] The present invention application proposes a scoring modeling method based on the fusion of corporate credit data in the financial leasing industry. First, through in-depth research on corporate public information and corporate credit information, the process and results of data analysis are explained from the perspectives of data sample overview, data preprocessing rules, characteristic variable analysis, and WOE analysis; secondly, according to the sample size, characteristic variable situation, modeling objectives, etc. of corporate public information and corporate credit information, an applicable modeling method is selected to establish a corporate public information model and a corporate credit information model respectively, and then the model parameters and score distribution are optimized; finally, the fusion scoring of the corporate public information model and the corporate credit information model is completed to create a set of total score scoring interface output services.

[0097] Compared with the prior art, the scoring modeling method based on the fusion credit data of enterprises in the financial leasing industry proposed in the present invention has the following beneficial effects:

[0098] 1. Construct a set of enterprise public information modeling methods from the perspective of credit reporting agencies

[0099] This patent application proposes to construct a set of enterprise public information modeling methods based on the enterprise credit data fusion scoring modeling method in the financial leasing industry. Relying on the perspective of a third-party credit reporting agency, based on the macro data analysis of tens of millions of corporate public information, it sorts out the credit risk situation of corporate information subjects at the social public level, analyzes the business-related attributes of government information disclosure, extracts key characteristic variables, and constructs a method for building the entire process of enterprise public information data range, data preprocessing, model design and evaluation through logistic regression model selection.

[0100] 2. Construct a set of enterprise credit information modeling methods from the perspective of financial leasing companies

[0101] The method of this patent application constructs a set of enterprise credit information modeling methods, starting from the actual business needs of the financial leasing company itself, combined with the classification and grading management of customers after the loan, setting the "bad definition" label of the modeling data, sorting out the basic information of the enterprise information subject in the credit report, loan and guarantee transaction information, repayment performance information, extracting key feature variables, and constructing the enterprise credit information model through hierarchical analysis method and expert scoring method.

[0102] 3. Construct a scoring method based on the credit data integration of the financial leasing industry

[0103] Enterprise credit information can only reflect one aspect of enterprise credit risk, while the value of alternative data of enterprise public information is becoming increasingly important. Therefore, modeling based on credit information alone cannot fully meet the needs of enterprise credit risk assessment, and it is urgent to integrate the modeling of enterprise public information. The method of this patent application constructs a method that integrates enterprise public information model and enterprise credit information model to meet the needs of scoring modeling in different scenarios in the future, which has strong scalability.

[0104] Through the above analysis, the patent applied for this time for a scoring modeling method based on the fusion of credit data of financial leasing enterprises has greater novelty, practicality and extensibility compared with other scoring modeling methods on the market.

[0105] The present invention proposes a method for solving the problem of enterprise credit data fusion scoring modeling based on the financial leasing industry. By integrating the data advantages of third-party credit reporting agencies and the application scenarios of financial leasing companies, an objective statistical model is constructed through modeling technology, and enterprise risks are scientifically predicted through model scoring, which comprehensively improves the application value and efficiency of data.

[0106] Social Benefits

[0107] The invention application proposes a method for solving the problem of enterprise credit data fusion scoring modeling based on the financial leasing industry, which comprehensively improves the application value of alternative data. This modeling method combines the behavior records of enterprise entities at the social public level such as industry and commerce, justice, taxation, intellectual property, etc., and combines the performance of enterprise credit information behavior to design a reasonable credit risk assessment model, comprehensively and accurately portray the credit portrait of the information subject, and display the social operation, dishonesty information, credit guarantee and other situations of the information subject through the scoring overview, comprehensively improve the social public credit awareness of the enterprise group, and enhance the application value of alternative data.

[0108] The present invention application proposes a method for solving the problem of enterprise credit data fusion scoring modeling based on the financial leasing industry, which effectively improves the standardization and efficiency of credit data modeling. The present invention has conducted in-depth thinking and research from the aspects of data preprocessing, model design and evaluation, multi-model fusion, system integration application, etc., and provides reference value for other non-bank financial institutions (financial leasing, commercial factoring, micro loans, consumer finance, etc.), and also provides a set of effective modeling methods for data modeling in the corporate credit industry to a certain extent. The present invention provides a reliable credit data fusion scoring modeling method in terms of preventing credit risks and improving the level of performance, helping other non-bank financial institutions to quickly build scoring models that adapt to their own business needs.

[0109] The invention application proposes a method for solving the enterprise credit data fusion scoring modeling based on the financial leasing industry, and strengthens the construction of a credit system for a trustworthy society. The invention creates a response to the country's requirements for standardizing credit information management, explores the path of standardizing commercial (enterprise) credit information, and jointly improves the "cross-domain" joint incentive for trustworthiness and joint punishment for dishonesty, constrains the enterprise information subject to operate in good faith in accordance with the law, maintains the normal order of the market, creates a trustworthy social environment, and promotes the development of inclusive finance. It will gradually improve the risk control level of inclusive financial institutions and make greater contributions to the construction of an industry risk control system with industry self-discipline and social supervision. This will inevitably produce huge social benefits and play an important role in creating a good credit environment and building a credit system.

[0110] Economic Benefits

[0111] The present invention application proposes a method for solving the problem of enterprise credit data fusion scoring modeling based on the financial leasing industry, which improves the risk control management efficiency of the financial leasing industry. With the continuous development of various types of credit business and the fierce competition in the industry, more and more financial leasing institutions have realized that gradually improving the level of enterprise risk control is a necessary basis for the steady advancement of business. Use data analysis tools to conduct in-depth mining and analysis of risk data to identify potential risk patterns and trends, integrate massive data into simple scoring data through modeling methods, and identify and predict risks through objective statistical values. By establishing a scoring monitoring system, ongoing risks can be tracked and evaluated in real time. The reference application of scoring in risk control can effectively simplify and optimize the risk control management process, reduce unnecessary steps and links, and improve work efficiency.

[0112] The present invention application proposes a method for solving the enterprise credit data fusion scoring modeling based on the financial leasing industry, which reduces transaction costs and improves social operation efficiency. The application of credit data fusion scoring can effectively track and monitor the defaulters through credit data fusion scoring, so that they are restricted in all aspects, such as restricting their application for financial products such as loans, or restricting their participation in commercial activities such as bidding. This will make it difficult for defaulters to move forward and greatly increase the cost of default. Trustworthy people can get more trust and opportunities, such as lower interest rates, faster approval speed, etc., to expand the value of trustworthiness, which can encourage more companies to become trustworthy people. Reduce information asymmetry, reduce transaction costs, improve social operation efficiency, and thus improve social economic benefits.

[0113] The present invention application proposes a method for solving the problem of enterprise credit data fusion scoring modeling based on the financial leasing industry, wherein enterprise public information refers to government public information from industry and commerce, judicial, taxation, intellectual property rights, etc. Enterprise public information has many data items and is scattered and unconcentrated; users need to log in to multiple data platforms to obtain the required data, which is inefficient, has a long data service link, and is complicated to operate.

[0114] The invention application proposes a method for solving the enterprise credit data fusion scoring modeling based on the financial leasing industry, in which the use of enterprise credit information is strictly required. On the premise of obtaining the authorization of the information subject, the credit report within the company also needs strict and standardized management; the approval process for the use of data is complex and the timeliness needs to be improved. Credit reports are generally in PDF format, and credit information is highly professional, so it is necessary to first analyze the data in a structured manner and then conduct professional interpretation and analysis. DETAILED DESCRIPTION

[0115] Other advantages and features of the present invention are shown by the following description of an embodiment of the present invention in conjunction with the accompanying drawings, which embodiment is given in the form of an example but is not limited thereto.

[0116] A scoring modeling method based on the fusion of credit data of enterprises in the financial leasing industry includes three steps: enterprise public information modeling, enterprise credit information modeling, and credit data fusion scoring.

[0117] (1) Enterprise public information modeling:

[0118] Based on the identification information of the enterprise information subject, retrieve the enterprise basic information through the enterprise public information query API, and store the retrieved enterprise basic information in the business system memory, wherein the identification information of the enterprise information subject includes but is not limited to the enterprise name and / or unified social credit code;

[0119] Define the modeling target according to the project application requirements, use the logistic regression model as the core modeling technology, use the retrieved enterprise basic information data as sample data, model the enterprise public information, and obtain the enterprise public information model;

[0120] By taking corresponding values ​​and scores for each variable in the enterprise public information model, the overall score of the enterprise public information model is finally derived.

[0121] The enterprise public information data referred to in this method refers to: government public information such as industry and commerce, judicial, taxation, intellectual property, etc., including but not limited to: enterprise registration information, shareholder information, key personnel, industrial and commercial changes, annual reports, administrative licenses, administrative penalties, equity pledges, and corporate overseas investments.

[0122] The specific information types and main fields of the enterprise public information data are shown in Table 1 below:

[0123] Table 1: Enterprise public information types and fields

[0124]

[0125]

[0126]

[0127]

[0128] In data modeling, the general modeling goal is to build a model to achieve the goal or purpose. Modeling goals can be diverse, generally used for prediction and decision support, a common modeling goal is to estimate the probability of a certain bad event occurring within a period of time.

[0129] For corporate public information modeling, the definition of bad samples may vary depending on the specific situation. The general definition of bad samples and their advantages and disadvantages are shown in Table 2:

[0130] Table 2: Comparison of the enterprise public information-modeling target general scheme

[0131]

[0132] The selection of modeling objectives generally requires a comprehensive consideration of the project application scenario and sample data distribution characteristics. In order to enhance the specificity of modeling objectives, modeling objectives can be defined based on project application requirements.

[0133] For example:

[0134] The definition of the modeling target variable ("good" or "bad") can be based on the actual application needs of the project and the post-loan management classification of the enterprise customer base, such as whether it is overdue, the overdue situation, and the five-level risk classification.

[0135] The modeling target in this method is based on the actual application of the financial leasing industry. Customers classified as "N1 / N2 / N3" are defined as "good" customers, and customers classified as "A / B / C" are defined as "bad" customers.

[0136] N1 is a normal project;

[0137] N2 is a project under construction, and it is uncertain whether it has risks;

[0138] N3 has some minor operational defects, such as failure to timely report electricity revenue in the last month;

[0139] Category C means risks have occurred and require the intervention of the account manager;

[0140] Category B means risks arise and the company needs to intervene, such as debt collection;

[0141] Category A is when risks arise and legal proceedings and other means are required to resolve the problem;

[0142] The degree of project deterioration is N1→N2→N3→C→B→A.

[0143] Specifically, in (1) enterprise public information modeling, a model is built for the enterprise public information to obtain an enterprise public information model; each variable in the enterprise public information model is assigned a value and a score, and finally an overall score of the enterprise public information model is obtained; specifically, the following is included:

[0144] (1.1) Data cleaning and analysis:

[0145] The retrieved enterprise basic information data is cleaned and calculated to obtain characteristic variables in the enterprise public information model. The data cleaning mainly includes: removing duplicate data, removing logical conflict data, completing some single variable calculations, processing noise data, abnormal values ​​and outliers, and processing missing values.

[0146] Data cleaning analysis, including cleaning rules for general data and cleaning rules for specific data.

[0147] (A) General data cleaning rules, the specific processing methods are as follows:

[0148] (A1) Date fields: displayed in the format of YYYY-MM-DD;

[0149] (A2) Amount fields: All amounts are converted into numerical format and calculated in RMB 10,000;

[0150] (A3) Ratio fields: All ratios are converted into numerical format, percentage signs are removed, and zero is added if there is no zero before the decimal point;

[0151] (A4) Data duplication: For the same event, there may be multiple duplicate information records in the data table. Data deduplication is mainly identified by the keyword "company name + event unique identifier judgment". The judgment keywords for deduplication in each table are detailed in Table 3 below.

[0152] Table 3: Uniqueness identification of enterprise public information records

[0153]

[0154] (B) Cleaning rules for specific data. The specific processing methods are as follows:

[0155] (B1) Enterprise registration: If the enterprise registration date is empty, the date is corrected with the enterprise operation start date; delete the observations with empty enterprise registration date; delete the observations with empty enterprise operation start date; delete the observations with business status of cancellation or revocation, but the cancellation date or revocation date is not empty; delete the observations with business end date not empty, but business end date < business start date; delete the observations with cancellation date not empty, but cancellation date < enterprise registration date; delete the observations with revocation date not empty, but revocation date < enterprise registration date;

[0156] (B2) Main personnel: merge the job titles of the same company and the same person; if a person has multiple job titles and they are written on two lines, merge them into one line;

[0157] (B3) Information on the person subject to enforcement: Delete the observations where the filing date is not empty but the filing date is less than the enterprise registration date;

[0158] (B4) Court Announcement: Delete the observations where the filing date is not empty but the filing date is less than the enterprise registration date;

[0159] (B5) Dishonest debtor: Delete the observations where the publication date is not empty but the publication date is less than the enterprise registration date;

[0160] (B6) Operational anomalies: Delete observations where the inclusion date is not empty but the inclusion date is less than the enterprise registration date; delete observations where the removal date is less than the inclusion date;

[0161] (B7) Administrative penalties: Delete observations where the penalty decision date is not empty, but the penalty decision date is less than the enterprise registration date; if the penalty decision date is empty, replace it with the public announcement date;

[0162] (B8) Administrative license: Delete the observations where the license start date is not empty but the license start date is less than the enterprise registration date; delete the observations where the license end date is not empty but the license end date is less than the license start date;

[0163] (B9) Spot check: delete the observations where the spot check date is not empty but the spot check date is less than the enterprise registration date;

[0164] (B10) Industrial and Commercial Change: Delete the observations where the Industrial and Commercial Change-Change Date is not empty, but the Industrial and Commercial Change-Change Date is <= Enterprise Registration Date;

[0165] (B11) Outward investment: Delete the observations where the opening date is not empty but the opening date is less than the enterprise registration date;

[0166] (B12) Branches: Delete observations where the establishment date is not empty but the establishment date is less than the enterprise registration date;

[0167] (B13) Equity pledge: Delete the observations where the equity pledge registration date is not empty, but the equity pledge registration date is less than the enterprise registration date;

[0168] (B14) Movable Property Mortgage: Delete the observations where the registration date is not empty but the registration date is less than the enterprise registration date.

[0169] (1.2) Characteristic variable analysis:

[0170] Conduct statistical characteristics and distribution analysis on the characteristic variables in the obtained enterprise public information model, check extreme values ​​and process them;

[0171] Organize the results of the feature variable analysis into a feature variable table, recording the feature variable name, calculation logic, data coverage, and basic information of data distribution.

[0172] (1.3) Weight of Evidence (WOE) Analysis:

[0173] The logistic regression model is transformed into a standard scorecard format through WOE transformation to obtain the variable values ​​of the feature variables;

[0174] First, all feature variables are automatically binned. Then, the reliability of the automatic binning results is manually checked to see whether they meet business requirements and are interpretable. Then, it is determined whether manual binning is needed.

[0175] The WOE for each category is defined as follows:

[0176]

[0177] The columns Bad Distribution and Good Distribution represent the distribution of "bad customers" and "good customers" in each category, respectively. They are obtained by dividing the frequency count in each category by the total number of "bad customers" or "good customers".

[0178] If the ratio in the brackets is less than 1, the WOE is negative, otherwise the WOE is positive.

[0179] (1.4) Modeling and debugging:

[0180] Initialize a series of model variables, fit a model based on the current series of variables, and the model results of this fitted model include the name of the feature variable, the meaning of the variable, the variable value, and the percentage score, and then determine whether this fitted model is the optimal model. If it is determined to be the optimal model, the final model and variables of the enterprise public information model are obtained; if it is determined not to be the optimal model, some variables are added to the model or some variables are deleted, and then a model is refitted based on the current series of variables, and it is determined whether this refitted model is the optimal model, until the optimal model is found, and the final model and variables of the enterprise public information model are obtained.

[0181] Initialize model variables, including:

[0182] ① Remove variables that have no obvious effect on modeling

[0183] Manually remove variables in the original data of corporate public information that are obviously useless for modeling or have no business significance, such as: unified social credit code, registration authority, legal representative, licensed business items, business scope, etc.

[0184] ②Remove variables with too low information value and too high proportion of repeated values

[0185] For example, in the modeling and debugging of a certain model, as shown in Table 4 below, the information value (info_value) of the characteristic variable "Company Name" is less than 0.02, which is too low and is removed. The identical_rate of the "Number of Copyrights of Works", "Number of Judicial Auction Information", "Number of Dishonest Persons Debtors", "Number of Tax Arrears Information", "Number of Intellectual Property Rights Pledges", "Number of Administrative Penalties", and "Number of Judgment Documents in the Past Two Years" is greater than 0.95, which is too large and is removed.

[0186] Table 4: Information value IV and repeated value ratio IR of characteristic variables in data modeling

[0187]

[0188] ③Remove variables with high Pearson correlation

[0189] For example, the Pearson correlation coefficient between "Annual Report" and "Years of Establishment" is 0.813. In the modeling process, "Annual Report" is removed and only the "Years of Establishment" variable is retained.

[0190] Determine whether the fitted model is the optimal model, specifically including: splitting the sample data into training set and validation set with a split ratio of 7:3. When splitting, the proportion of bad samples in the training set and validation set must be kept consistent with the proportion of bad samples in the full data; obtain the KS curve and ROC curve of the training set and validation set, and the one with the best effect is the optimal statistical model for quantitative analysis; calculate the model score of each sample based on the model results of the optimal statistical model obtained and the value of the characteristic variables of each sample data, and obtain the sample score distribution of the optimal statistical model; reflect the optimal statistical model's ability to distinguish different samples, stability and possible deviations through the sample score distribution, and judge whether the sample model score can best distinguish the status of good and bad samples according to the actual application scenario of the model, for example: bad samples are concentrated in the low segment, and good samples are concentrated in the high segment; if the status of good and bad samples can be best distinguished, it is judged to be the optimal model, that is, the final model and variables of the enterprise public information model are obtained; if the status of good and bad samples cannot be best distinguished, it is judged not to be the optimal model.

[0191] When it is judged that it is not the optimal model, some variables are added to the model or some variables are deleted, and then a new model is fitted based on the current series of variables. The above method is used to continue to determine whether the refitted model is the optimal model until the optimal model is found, and the final model and variables of the enterprise public information model are obtained.

[0192] (1.5) Fractional linear transformation

[0193] The score is linearly converted to 0-100 points, and the distribution characteristics remain unchanged. The conversion formula is:

[0194]

[0195] (1.6) Model Results Presentation

[0196] The final model input variables, variable values ​​and percentage scores of each variable of the enterprise public information model are displayed on the business system.

[0197] Table 5: Enterprise public information model

[0198]

[0199] (2) Enterprise credit information modeling

[0200] The enterprise credit information is derived from the enterprise credit report, including: credit reminder information, loan transaction summary information, guarantee transaction summary information, loan account information, etc.;

[0201] The model is layered and key variables are initially screened through the "hierarchical analysis method". The business attributes, relevance, data coverage and other conditions of the variables are comprehensively considered, and scores are assigned through the "expert scoring method". Modeling is carried out for corporate credit information to obtain the corporate credit information model.

[0202] Based on the identification information of the enterprise information subject, check whether the enterprise information subject has a credit report; if there is a credit report, use the data with the latest report date in the database to determine whether the "Year of the first credit transaction" field in the "Credit reminder information unit" chapter in the credit report is empty. If the field is not empty, calculate the enterprise credit information score according to the enterprise credit information model. If the field is empty, the calculation of the enterprise credit information score is not supported and the enterprise credit information score is treated as empty; if there is no credit report, the calculation of the enterprise credit information score is not supported and the enterprise credit information score is treated as empty.

[0203] Enterprise credit information comes from enterprise credit reports, with a single data source and standardized structured data, including credit reminder information, loan transaction summary information, guarantee transaction summary information, loan account information, etc. In the aforementioned (2) enterprise credit information modeling, the aforementioned modeling of enterprise credit information to obtain an enterprise credit information model specifically includes:

[0204] (2.1) Obtain characteristic variables in the enterprise credit information model through data cleaning and calculation

[0205] Clean the enterprise credit information data, select important information dimensions, and then process and calculate the variables. If there are multiple records of credit reports being queried for the same enterprise in the credit information data table, sort them in reverse order by the time the credit reports were generated, and take the data of the latest credit report as the modeling sample data;

[0206] Through data cleaning, the retrieved corporate credit information data is re-examined and verified in order to discover and correct errors in the data files and reduce the impact of erroneous data on model performance. The data cleaning mainly includes: removing duplicate data, removing logical conflicting data, completing some single variable calculations, processing noise data, abnormal values ​​and outliers, and processing missing values.

[0207] (2.2) Feature analysis

[0208] Conduct statistical characteristic analysis and distribution analysis on characteristic variables in the enterprise credit information model, check extreme values ​​and process them;

[0209] The results of feature analysis are organized into a feature variable table, which records the feature variable name, calculation logic, data coverage, and basic information on data distribution.

[0210] (2.3) Weight of Evidence (WOE) Analysis

[0211] The variable values ​​of the characteristic variables are obtained through WOE transformation;

[0212] First, all feature variables are automatically binned. Then, the reliability of the automatic binning results is manually checked to see whether they meet business requirements and are interpretable. Then, it is determined whether manual binning is needed.

[0213] The WOE for each category is defined as follows:

[0214]

[0215] The columns Bad Distribution and Good Distribution represent the distribution of "bad customers" and "good customers" in each category, respectively. They are obtained by dividing the frequency count in each category by the total number of "bad customers" or "good customers".

[0216] If the ratio in the brackets is less than 1, the WOE is negative, otherwise the WOE is positive.

[0217] (2.4) Scoring design

[0218] (2.4.1) Overall sample score distribution

[0219] The enterprise credit information model is built according to the expert scoring rules. The score of each sample is obtained according to the scoring range and scoring rules of each variable. The overall sample score distribution analysis helps to identify whether the sample scores are concentrated, dispersed or have outliers. According to the score results, the model variables, variable values, and scores are adjusted in combination with the application scenarios of the financial leasing industry.

[0220] (2.4.2) Distribution of good and bad sample scores

[0221] The modeling goal is to define customers as "good" customers or "bad" customers based on the customer management classification of the actual project user;

[0222] Good and bad samples are distinguished based on the “bad definition” label;

[0223] Because each sample has a model score, we will make a score distribution based on good and bad samples to see whether the score can distinguish good and bad samples, that is, whether good samples are concentrated in high segments and bad samples are concentrated in low segments. According to the score results, we will adjust the model variables, variable values, and scores in combination with the application scenarios of the financial leasing industry.

[0224] Step (2.3) WOE analysis and step (2.4) score design are tuned multiple times to achieve a state where the sample score can best distinguish good and bad samples, and the final model and variables of the enterprise credit information model are obtained;

[0225] (2.5) Model Results Presentation

[0226] The final model input variables, variable values ​​and percentage scores of each variable in the enterprise credit information model are displayed on the business system.

[0227] Table 6: Enterprise credit information model

[0228]

[0229]

[0230] (3) Credit data integration scoring

[0231] The overall score of the enterprise public information model and the enterprise credit information score obtained are output to the business application system through the interface according to the agreed weight rules; the business system displays the final integrated scoring results and key characteristic variables.

[0232] In actual business scenarios, each enterprise information subject can calculate the enterprise public information score, but may not be able to calculate the enterprise credit information score due to the lack of credit information. The enterprise public information score and enterprise credit information score are processed and calculated in the system, and then the fusion score is calculated according to the weight ratio of 4:6.

[0233] For example: the enterprise public information model score is 80, the credit information model score is 60, and the credit data fusion score is 80*0.4+60*0.6=68.

[0234] For example: if the enterprise public information model score is 70 and there is no credit information data, the credit data fusion score is 70.

[0235] The output content of the credit data fusion scoring interface includes: enterprise name, unified social credit code, credit fusion score, enterprise public information score, enterprise credit information score, the value and score of each variable in the enterprise public information model, and the value and score of each variable in the enterprise credit information model.

[0236] Although the present invention has been described above based on the preferred embodiments, this does not mean that the scope of the present invention is limited to the above-mentioned structure. As long as technicians in this technical field can easily develop equivalent alternative structures after reading the above description, all equivalent changes and modifications made without departing from the spirit and scope of the present invention should be included in the patent scope of the present invention.

Claims

1. A scoring modeling method based on the fusion of credit data of enterprises in the financial leasing industry, characterized in that: The following steps are involved: (1) Enterprise public information modeling: Based on the identification information of the enterprise information subject, retrieve the enterprise basic information through the enterprise public information query API, and store the retrieved enterprise basic information in the business system memory, wherein the identification information of the enterprise information subject includes but is not limited to the enterprise name and / or unified social credit code; Define the modeling target according to the project application requirements, use the logistic regression model as the core modeling technology, use the retrieved enterprise basic information data as sample data, model the enterprise public information, and obtain the enterprise public information model. The model results of the enterprise public information model include variable name, variable meaning, variable value, and percentage score; Through the corresponding values ​​and percentage scores of each variable in the enterprise public information model, the overall score of the enterprise public information model is finally obtained; The corresponding values ​​and percentage scores of each variable in the enterprise public information model are defined in the following table: (2) Enterprise credit information modeling The enterprise credit information is derived from the enterprise credit report, including: credit reminder information, loan transaction summary information, guarantee transaction summary information, loan account information, etc.; The model is layered and key variables are initially screened through the "hierarchy analysis method". The business attributes, relevance, data coverage and other conditions of the variables are comprehensively considered, and scores are assigned through the "expert scoring method". Modeling is carried out for corporate credit information to obtain the corporate credit information model. The model results of the corporate credit information model include variable names, variable meanings, variable values, and percentage scores. Based on the identification information of the enterprise information subject, check whether the enterprise information subject has a credit report; if there is a credit report, use the latest data in the database with the report date to determine whether the "Year of First Credit Transaction" field in the "Credit Prompt Information Unit" section of the credit report is empty. If the field is not empty, calculate the enterprise credit information score according to the enterprise credit information model; if the field is empty, the enterprise credit information score is not supported and the enterprise credit information score is treated as empty; if there is no credit report, the enterprise credit information score is not supported and the enterprise credit information score is treated as empty; The corresponding values ​​and percentage scores of each variable in the enterprise credit information model are defined in the following table: (3) Credit data integration scoring The overall score of the enterprise public information model obtained in step (1) and the enterprise credit information score obtained in step (2) are output to the business application system through the credit data fusion scoring interface according to the agreed weight rules, and the business application system displays the final fusion scoring results and key feature variables; The output content of the credit data fusion scoring interface includes but is not limited to: enterprise name, unified social credit code, credit fusion score, enterprise public information score, enterprise credit information score, value and score of each variable of the enterprise public information model, and value and score of each variable of the enterprise credit information model.

2. The scoring modeling method based on the fusion credit data of financial leasing industry enterprises according to claim 1 is characterized by: The said modeling is performed on the enterprise public information to obtain the enterprise public information model; By taking corresponding values ​​and scores for each variable in the enterprise public information model, the overall score of the enterprise public information model is finally obtained; Specifically include: (1.1) Data cleaning and analysis: Clean and calculate the retrieved enterprise basic information data to obtain characteristic variables in the enterprise public information model. The data cleaning mainly includes: removing duplicate data, removing logical conflict data, completing some single variable calculations, processing noise data, abnormal values ​​and outliers, and processing missing values; (1.2) Characteristic variable analysis: Conduct statistical characteristics and distribution analysis on the characteristic variables in the obtained enterprise public information model, check extreme values ​​and process them; Organize the results of feature variable analysis into a feature variable table, recording the feature variable name, calculation logic, data coverage, and basic data distribution information; (1.3) Weight of Evidence (WOE) Analysis: The logistic regression model is transformed into a standard scorecard format through WOE transformation to obtain the variable values ​​of the feature variables; First, all feature variables are automatically binned. Then, the reliability of the automatic binning results is manually checked to see whether they meet business requirements and are interpretable. Then, it is determined whether manual binning is needed. The WOE for each category is defined as follows: The columns Bad Distribution and Good Distribution represent the distribution of "bad customers" and "good customers" in each category, respectively. They are obtained by dividing the frequency count in each category by the total number of "bad customers" or "good customers". If the ratio in the brackets is less than 1, the WOE is negative, otherwise the WOE is positive; (1.4) Modeling and debugging: Initialize a series of model variables, fit a model based on the current series of variables, and the model results of the fitted model include the feature variable name, variable meaning, variable value, and percentile score, and then determine whether the fitted model is the optimal model; If it is judged to be the optimal model, the final model and variables of the enterprise public information model are obtained; If it is judged that it is not the optimal model, some variables are added to the model or some variables are deleted, and then a new model is fitted based on the current series of variables to determine whether the refitted model is the optimal model. Until the optimal model is found, the final model and variables of the enterprise public information model are obtained; (1.5) Fractional linear transformation: The score is linearly converted to 0-100 points, and the distribution characteristics remain unchanged. The conversion formula is: (1.6) Model Results Presentation The final model input variables, variable values ​​and percentage scores of each variable of the enterprise public information model are displayed on the business system.

3. The scoring modeling method based on the fusion credit data of financial leasing industry enterprises according to claim 2 is characterized by: The (1.1) data cleaning analysis specifically includes: (A) General data cleaning rules, the specific processing methods are as follows: (A1) Date fields: displayed in the format of YYYY-MM-DD; (A2) Amount fields: All amounts are converted into numerical format and calculated in RMB 10,000; (A3) Ratio fields: All ratios are converted into numerical format, percentage signs are removed, and zero is added if there is no zero before the decimal point; (A4) Data duplication: For the same event, there may be multiple duplicate information records in the data table. Data deduplication is mainly identified by the keyword "company name + event unique identifier judgment"; (B) Cleaning rules for specific data. The specific processing methods are as follows: (B1) Enterprise registration: If the enterprise registration date is empty, the date is corrected with the enterprise operation start date; delete the observations with empty enterprise registration date; delete the observations with empty enterprise operation start date; delete the observations with business status of cancellation or revocation, but the cancellation date or revocation date is not empty; delete the observations with business end date not empty, but business end date < business start date; delete the observations with cancellation date not empty, but cancellation date < enterprise registration date; delete the observations with revocation date not empty, but revocation date < enterprise registration date; (B2) Main personnel: merge the job titles of the same company and the same person; if a person has multiple job titles and they are written on two lines, merge them into one line; (B3) Information on the person subject to enforcement: Delete the observations where the filing date is not empty but the filing date is less than the enterprise registration date; (B4) Court Announcement: Delete the observations where the filing date is not empty but the filing date is less than the enterprise registration date; (B5) Dishonest debtor: Delete the observations where the publication date is not empty but the publication date is less than the enterprise registration date; (B6) Operational anomalies: Delete observations where the inclusion date is not empty but the inclusion date is less than the enterprise registration date; delete observations where the removal date is less than the inclusion date; (B7) Administrative penalties: Delete observations where the penalty decision date is not empty, but the penalty decision date is less than the enterprise registration date; if the penalty decision date is empty, replace it with the public announcement date; (B8) Administrative license: Delete the observations where the license start date is not empty but the license start date is less than the enterprise registration date; delete the observations where the license end date is not empty but the license end date is less than the license start date; (B9) Spot check: delete the observations where the spot check date is not empty but the spot check date is less than the enterprise registration date; (B10) Industrial and Commercial Change: Delete the observations where the Industrial and Commercial Change-Change Date is not empty, but the Industrial and Commercial Change-Change Date is <= Enterprise Registration Date; (B11) Outward investment: Delete the observations where the opening date is not empty but the opening date is less than the enterprise registration date; (B12) Branches: Delete observations where the establishment date is not empty but the establishment date is less than the enterprise registration date; (B13) Equity pledge: Delete the observations where the equity pledge registration date is not empty, but the equity pledge registration date is less than the enterprise registration date; (B14) Movable Property Mortgage: Delete the observations where the registration date is not empty but the registration date is less than the enterprise registration date.

4. The scoring modeling method based on the fusion credit data of financial leasing industry enterprises according to claim 2 is characterized by: In the step (1.4) of modeling and debugging, the step of judging whether the fitted model is the optimal model specifically includes: The sample data is split into a training set and a validation set with a split ratio of 7:

3. When splitting, the proportion of bad samples in the training set and validation set must be consistent with the proportion of bad samples in the full data. Obtain the KS curve and ROC curve of the training set and validation set, and the best one is the optimal statistical model for quantitative analysis; According to the model results of the optimal statistical model obtained and the characteristic variable values ​​of each sample data, the model score of each sample is calculated to obtain the sample score distribution of the optimal statistical model; The sample score distribution reflects the optimal statistical model's ability to distinguish different samples, stability, and possible deviations. Based on the actual application scenario of the model, it is determined whether the sample model score can best distinguish between good and bad samples. If the state of good and bad samples can be best distinguished, it is judged as the optimal model. If the state of good and bad samples cannot be best distinguished, it is judged as not the optimal model.

5. The scoring modeling method based on the fusion credit data of financial leasing industry enterprises according to claim 2 is characterized by: In the step (1.4) of modeling and debugging, the initialization of a series of model variables includes: Remove variables that are obviously not useful for modeling or have no business significance; Remove variables with low information values ​​and high proportion of duplicate values; and Remove all but one of the two or more variables with high Pearson correlations.

6. The scoring modeling method based on the fusion credit data of financial leasing industry enterprises according to claim 1 is characterized by: In the aforementioned (2) enterprise credit information modeling, the aforementioned modeling of enterprise credit information to obtain an enterprise credit information model specifically includes: (2.1) Obtain characteristic variables in the enterprise credit information model through data cleaning and calculation Clean the enterprise credit information data, select important information dimensions, and then process and calculate the variables. If there are multiple records of credit reports being queried for the same enterprise in the credit information data table, sort them in reverse order by the time the credit reports were generated, and take the data of the latest credit report as the modeling sample data; Re-examine and verify the retrieved enterprise credit information data through data cleaning to find and correct errors in the data files and reduce the impact of erroneous data on model performance. The data cleaning mainly includes: removing duplicate data, removing logical conflict data, completing some single variable calculations, processing noise data, abnormal values ​​and outliers, and processing missing values; (2.2) Feature analysis Conduct statistical characteristic analysis and distribution analysis on characteristic variables in the enterprise credit information model, check extreme values ​​and handle them; The results of feature analysis are compiled into a feature variable table, which records the feature variable name, calculation logic, data coverage, and basic data distribution information; (2.3) Weight of Evidence (WOE) Analysis The variable values ​​of the characteristic variables are obtained through WOE transformation; First, all feature variables are automatically binned. Then, the reliability of the automatic binning results is manually checked to see whether they meet business requirements and are interpretable. Then, it is determined whether manual binning is needed. The WOE for each category is defined as follows: The columns Bad Distribution and Good Distribution represent the distribution of "bad customers" and "good customers" in each category, respectively. They are obtained by dividing the frequency count in each category by the total number of "bad customers" or "good customers". If the ratio in the brackets is less than 1, the WOE is negative, otherwise the WOE is positive; (2.4) Scoring design (2.4.1) Overall sample score distribution The enterprise credit information model is built according to the expert scoring rules. The score of each sample is obtained according to the scoring range and scoring rules of each variable. The overall sample score distribution analysis helps to identify whether the sample scores are concentrated, dispersed or have outliers. According to the score results, the model variables, variable values, and scores are adjusted in combination with the application scenarios of the financial leasing industry. (2.4.2) Distribution of good and bad sample scores The modeling goal is to define customers as "good" customers or "bad" customers based on the customer management classification of the actual project user; Good and bad samples are distinguished based on the "bad definition" label; Because each sample has a model score, we will make a score distribution based on good and bad samples to see whether the score can distinguish good and bad samples, that is, whether good samples are concentrated in high segments and bad samples are concentrated in low segments. According to the score results, we will adjust the model variables, variable values, and scores in combination with the application scenarios of the financial leasing industry. Step (2.3) WOE analysis and step (2.4) score design are tuned multiple times to achieve a state where the sample score can best distinguish good and bad samples, and the final model and variables of the enterprise credit information model are obtained; (2.5) Model Results Presentation The final model input variables, variable values ​​and percentage scores of each variable in the enterprise credit information model are displayed on the business system.

7. The scoring modeling method based on the fusion credit data of financial leasing industry enterprises according to claim 1 is characterized by: In the aforementioned (3) credit data fusion scoring, the agreed weighting rules are: When the enterprise credit information score is not empty, the credit data fusion score is calculated based on the weight ratio of the enterprise public information model score to the enterprise credit information score of 4:6; When the enterprise credit information score is empty, the enterprise credit data fusion score is consistent with its enterprise public information model score.

Images