Credit customer risk label generation and audit sampling method based on audit modeling

Abstract

The application relates to the technical field of credit audit data processing, and particularly discloses a credit customer risk label generation and audit sampling method based on audit modeling, which comprises the following steps: standardizing and correlatively integrating credit audit data from multiple business systems to generate a credit audit basic data set; constructing a credit audit model based on the credit audit basic data set; the credit audit model is a rule parameterization model containing a preset audit rule library, a label weight, an employee portrait parameter and a sampling strategy parameter; and determining a customer risk label set and a customer risk level through the credit audit model. Through the integration of multi-source credit audit data and the construction of a rule parameterization credit audit model that can be updated in feedback, the computer equipment can continuously complete customer risk label generation, employee risk portrait aggregation, audit sample library generation and model version updating, thereby forming a traceable credit audit sampling data processing closed loop.

Description

Technical Field

[0001] This invention belongs to the field of credit audit data processing technology, specifically relating to a method for generating credit customer risk labels and audit sampling based on audit modeling. Background Technology

[0002] As the scale of bank lending expands, commercial banks' internal audits need to verify a large number of customers, loan contracts, repayment records, guarantee information, approval processes, and historical audit issues. Due to the large number of lending customers and the substantial amount of loan business data, auditors typically find it difficult to examine all lending transactions one by one. Therefore, they need to use audit sampling methods to determine audit samples from the lending business data in order to improve the efficiency of audit resource utilization.

[0003] In current technologies, credit audit sampling is usually determined by random sampling, stratified sampling, loan amount, customer type, overdue status, or single risk score. Some audit systems also score customer risk based on preset rules and screen customers to be audited based on customer risk score or risk level, so that the audit sample has a certain risk orientation.

[0004] However, bank credit audit data is usually scattered across credit management systems, core accounting systems, guarantee management systems, audit management systems, and human resources systems. Existing sampling methods are mostly based on a single customer dimension or a single loan dimension, making it difficult to standardize and correlate multi-source credit audit data to form a data foundation that can be used for audit modeling. Customer risk identification results are usually difficult to further aggregate according to the handling staff, resulting in a lack of continuous data processing relationship between customer risk tags, employee risk profiles, and audit sample generation. Furthermore, audit findings are usually stored as audit files or rectification records, making it difficult to use them as model feedback data to feed back into the audit model for updating subsequent risk tag generation, employee profile construction, and audit sample generation processes. Summary of the Invention

[0005] The purpose of this invention is to provide a credit customer risk label generation and audit sampling method based on audit modeling, in order to solve the problems in the existing credit audit sampling process, such as the difficulty in unifying and integrating multi-source credit audit data, the difficulty in associating and aggregating customer risk identification results with the dimension of the handling staff, the reliance on static rules or single risk scores for audit sample generation, and the difficulty in feeding audit findings back to the audit model for subsequent risk identification and sampling strategy updates.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for generating credit customer risk labels and conducting audit sampling based on audit modeling includes: Standardize and integrate credit audit data from multiple business systems to generate a basic credit audit dataset; A credit audit model is constructed based on the basic dataset of credit audit. The credit audit model is a rule-parametric model that includes a preset audit rule base, label weights, employee profile parameters and sampling strategy parameters. Determine the customer risk label set and customer risk level through a credit audit model; Based on the mapping relationship between customers and the staff in charge, the customer risk tags and customer risk levels are aggregated into employee risk profiles. Based on the employee risk profile, key audit employees are identified, and stratified sampling is conducted on the clients under the names of key audit employees to generate an audit sample library; Based on the audit findings, model feedback data is generated, and the credit audit model is updated using the model feedback data. The updated credit audit model is then used for the next round of customer risk label generation, employee risk profile construction, and audit sample generation.

[0007] In one embodiment, a basic dataset for credit auditing is generated, including: Credit audit data is obtained from the credit management system, core accounting system, guarantee management system, audit management system, and human resources system; Perform data cleaning, standardize monetary units, standardize date formats, semantically convert code fields, and convert foreign currencies for credit audit data; Using customer ID, contract ID, and employee ID as the associated indexes, establish the relationships between customer basic information, loan contract information, promissory note information, repayment information, account transaction information, guarantee information, approval process information, audit findings information, and employee information to generate a basic dataset for credit auditing.

[0008] In one embodiment, the credit audit model includes a risk feature identification unit, a customer risk label generation unit, an employee risk profile aggregation unit, an audit sampling decision unit, and a model feedback optimization unit. The risk characteristic identification unit is used to identify customer risk characteristics based on a pre-set audit rule base. The customer risk label generation unit is used to generate a set of customer risk labels based on customer risk characteristics and determine the customer risk level; The employee risk profile aggregation unit is used to aggregate customer risk tag sets and customer risk levels to the employee dimension; The audit sampling decision unit is used to generate an audit sample library based on employee risk profiles; The model feedback optimization unit is used to update the model parameters of the credit audit model based on the model feedback data.

[0009] In one embodiment, a customer risk label set and customer risk level are determined through a credit audit model, including: Identify customer risk characteristics based on a pre-defined audit rule base, and generate customer risk label records based on the identified customer risk characteristics; Customer risk tag records include customer number, tag number, tag name, trigger condition, trigger basis, tag weight, hit flag, and hit time; Calculate the customer's risk score based on the tag weights and hit indicators corresponding to the risk tags matched by the customer; The customer's risk level is determined by comparing the customer's risk score with the risk level threshold.

[0010] In one embodiment, the preset audit rule base includes at least two of the following: fund return rules, related credit rules, insufficient collateral rules, and rules for re-lending after rejection by others; The fund return rules are triggered based on whether the funds return to the borrower's account or related party's account within a preset period after the loan is disbursed, and the proportion of the returned amount to the loan amount. The associated credit granting rules are triggered based on the length of the guarantee loop or the number of associated customer credit grantings within a preset period; The insufficient collateral rule is triggered based on whether the collateral coverage ratio or the loan-to-value ratio exceeds the maximum loan-to-value ratio corresponding to the type of collateral. The rule for re-granting loans after a rejection by another party is triggered based on whether there is a rejection record in the approval process, and the time interval between the re-application and approval after the rejection.

[0011] In one embodiment, based on the mapping relationship between customers and handling employees, the customer risk tag set and customer risk level are aggregated into an employee risk profile, including: Establish a mapping relationship between customers and handling employees based on customer manager number, approver number, responsible customer manager number, handling time and approval node; When the same customer corresponds to multiple staff members, risk contribution weights are assigned to the multiple staff members based on their roles, participation stages, or contributions. When a client manager changes, the risk attribution is determined based on the loan disbursement time, the time when the risk tag is triggered, the time when the audit issue occurs, or the responsible client manager field. Based on the risk attribution results, customer risk labels and customer risk levels are aggregated and assigned to the corresponding employees.

[0012] In one embodiment, the employee risk profile includes at least two of the following employee profile indicators: number of high-risk customers, balance of problem loans, balance of non-performing loans, number of audit findings in history, total number of risk tags hit, proportion of high-risk tags, customer concentration, growth rate of new high-risk customers, rate of failure to rectify, and average risk score. Employee profile metrics were standardized using percentiles. For employee profile indicators where higher indicator values ​​indicate higher risks, the ranking percentile is determined according to the indicator values ​​from largest to smallest. For the customer concentration index, the ranking percentile is determined according to the degree to which the customer concentration deviates from the institution's average. The employee's overall risk score is calculated based on the standardized scores of each employee profile indicator and the corresponding indicator weights, and the employee's risk level is determined based on the overall risk score.

[0013] In one embodiment, generating an audit sample library includes: Employees are ranked according to their comprehensive risk scores, and those who meet the pre-defined ranking criteria are selected as key audit employees. Based on the mapping relationship between customers and the staff in charge, identify the customers under the name of the key audit staff and generate a candidate sample pool containing employee risk information and customer risk information. In the candidate sample pool, stratified sampling is performed according to high-risk customers, medium-risk customers, and low-risk customers; Within the same customer risk level, customer sampling priority is determined from high to low based on the number of risk tags, the total weight of risk tags, the number of historical audit findings, or the loan balance. When multiple clients have the same risk level, the priority of linked sampling is determined based on the comprehensive risk score of the corresponding key audit staff. A coverage verification is performed on the customer sample selected according to the customer sampling priority. The coverage verification includes at least one of the institutional dimension, credit product dimension, and loan amount dimension. When the number of customer samples corresponding to any coverage dimension is lower than the preset minimum sampling quantity, customer samples are drawn from the remaining candidate customers corresponding to that coverage dimension according to the customer sampling priority until the preset minimum sampling quantity is reached. When the same client is associated with multiple key audit staff, the client sample is deduplicated, and the information of the key audit staff associated with that client and the sampling source are retained in the audit sample library.

[0014] In a preferred embodiment, the sample records in the audit sample library include customer number, customer risk level, customer risk score, risk tag set, main trigger field, employee number, employee risk score, employee risk ranking, linked sampling priority, sampling source, sampling batch, sampling timestamp, and model version number. Sampling sources include at least one of the following: key employee sampling, risk label sampling, regional minimum guarantee sampling, business line minimum guarantee sampling, special attention employee sampling, and random minimum guarantee sampling; Among them, regional minimum sampling and business line minimum sampling are used to ensure that the audit sample covers a preset region or a preset business line.

[0015] In a preferred embodiment, generating model feedback data based on audit findings and updating the credit audit model using the model feedback data includes: Receive audit findings, which include whether the customer has a problem, the type of problem, the severity of the problem, the risk tags involved, the employees involved, the rectification status, and the date the problem was discovered. Model feedback data is generated based on the audit findings. The model feedback data includes the number and severity of problems found for each risk label, the number and severity of problems found for different customer risk levels, the number and severity of problematic customers under the names of each key audit employee, and the sample size and problem distribution for different institutions, loan amount ranges and credit product types. The label weights are updated based on the number and severity of issues found for each risk label. The risk level thresholds are updated based on the number and severity of issues found for different customer risk levels. The number of key audit staff selected (N) or the staff profile parameters are updated based on the number and severity of problematic customers under each key audit staff member. The sampling strategy parameters are updated based on the sample size and issue distribution for different institutions, loan amount ranges, and credit product types. Each update generates a new model version and records the model version number, parameters before the update, parameters after the update, feedback data source, update time, and update basis.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By standardizing and integrating multi-source credit audit data, a rule-parametric credit audit model is constructed, which includes a pre-defined audit rule base, label weights, employee profile parameters, and sampling strategy parameters. This enables computer equipment to generate a set of customer risk labels and customer risk levels with semantic descriptions based on a unified credit audit basic dataset. Furthermore, it aggregates these data according to the mapping relationship between customers and the handling staff to form employee risk profiles. Then, it generates an audit sample library based on the employee risk profiles and customer risk levels. At the same time, the audit findings are fed back to the credit audit model as model feedback data to update the model version. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the credit customer risk label generation and audit sampling method based on audit modeling provided in this embodiment of the invention.

[0018] Figure 2 This is a schematic diagram illustrating the relationship between customer risk information aggregation and employee risk profiling in this invention.

[0019] Figure 3 This is a flowchart illustrating the linked sampling based on employee risk and customer risk in this invention.

[0020] Figure 4 This is a flowchart illustrating how the credit audit model is updated based on audit findings, as per the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0022] Reference Figure 1-4 This embodiment provides a method for generating credit customer risk labels and audit sampling based on audit modeling. This method is executed by computer equipment, which can be a server of the bank's internal audit system, a server cluster of an intelligent audit platform, an audit analysis device deployed in the bank's intranet environment, or a cloud platform node with data processing capabilities.

[0023] In this embodiment, the computer device acquires credit audit data from multiple business systems, standardizes and integrates the credit audit data to generate a basic credit audit dataset, constructs a credit audit model based on the basic credit audit dataset, determines the customer risk label set and customer risk level through the credit audit model, aggregates the customer risk label set and customer risk level into an employee risk profile according to the mapping relationship between customers and handling employees, identifies key audit employees based on the employee risk profile, and performs stratified sampling of customers under the key audit employees to generate an audit sample library, generates model feedback data based on the audit findings, and uses the model feedback data to update the credit audit model so that the updated credit audit model can be used for the next round of customer risk label generation, employee risk profile construction, and audit sample generation.

[0024] In this embodiment, the credit audit model is a rule-parameterized model. The rule-parameterized model is not a simple business judgment rule, nor is it a simple personnel screening rule. Instead, it is composed of model parameter configuration data called by computer equipment and processing units. The model parameter configuration data includes a preset audit rule base, a label weight parameter table, an employee profile parameter table, a sampling strategy parameter table, and a model version record table. The preset audit rule base can be implemented in the form of a risk rule configuration table.

[0025] In this embodiment, the preset audit rule base is used to store risk rules for identifying multi-dimensional risk characteristics of customers. When the preset audit rule base is implemented using a risk rule configuration table, the risk rule configuration table records the rule number, rule name, applicable fields, trigger conditions, trigger thresholds, and corresponding risk labels. The label weight parameter table records the label number, label name, initial weight, current weight, upper weight limit, and lower weight limit. The employee profile parameter table records the profile indicator name, indicator calculation field, indicator standardization method, indicator weight, and employee risk level threshold. The sampling strategy parameter table records the number N of key audit employees, the customer risk level stratification ratio, the minimum sampling size for each coverage dimension, the upper sample size limit, the lower sample size limit, and the minimum sampling rule. The model version record table records the model version number, activation time, parameter source, update basis, and version status.

[0026] In this embodiment, the credit audit model includes a risk feature identification unit, a customer risk tag generation unit, an employee risk profile aggregation unit, an audit sampling decision unit, and a model feedback optimization unit. The risk feature identification unit is used to identify customer risk features based on a preset audit rule base. The customer risk tag generation unit is used to generate a set of customer risk tags and determine the customer risk level based on the customer risk features. The employee risk profile aggregation unit is used to aggregate the set of customer risk tags and the customer risk level to the employee dimension. The audit sampling decision unit is used to generate an audit sample base based on the employee risk profile. The model feedback optimization unit is used to update the model parameters of the credit audit model based on the model feedback data.

[0027] In this embodiment, the computer device sets a model version number for the credit audit model. Each model version corresponds to a set of preset audit rule bases, label weight parameters, risk level thresholds, employee profile parameters, and sampling strategy parameters. The sampling strategy parameters include the number of key audit employees selected (N), the customer risk level stratification ratio, the coverage dimension, and the minimum sampling size corresponding to each coverage dimension. The computer device records the model version number used when generating customer risk labels, constructing employee risk profiles, and generating audit samples, so that the basis for the formation of audit samples, the triggering basis for risk labels, the calculation basis for employee profiles, and the update basis for model feedback can be traced.

[0028] In other implementations, the preset audit rule base, label weight parameter table, employee profile parameter table, sampling strategy parameter table, and model version record table can be stored in a relational database, distributed database, data warehouse, or graph database. The preset audit rule base can be obtained by manual configuration, importing historical audit rules, or updating model feedback data. The label weight parameters can be obtained by initial expert configuration, historical problem hit rate statistics, or model feedback optimization. Employee profile indicators can be enabled or disabled according to business lines, organizational levels, or audit project types. Sampling strategy parameters can be adjusted according to audit resources, regional coverage requirements, or business line coverage requirements.

[0029] In this embodiment, the credit audit model takes the credit audit basic dataset, model parameter configuration data, customer risk label records, employee profile records, audit sample records, and model feedback records as processing objects. The computer device performs field standardization, index association, matching of preset audit rule base, label calculation, profile aggregation, sample generation, and model version update. It can transform audit business rules into an executable, recordable, and traceable computer data processing flow, reducing the risk that the solution will be misunderstood as a simple management method or manual audit rules. Example

[0030] This embodiment describes the method for generating the basic dataset for credit auditing, based on embodiment 1.

[0031] In this embodiment, the computer device obtains credit audit data from the credit management system, the core accounting system, the guarantee management system, the audit management system, and the human resources system. Among them, the credit management system provides loan contract information, basic customer information, approval process records, and approval history records; the core accounting system provides loan receipt information, repayment schedule, actual repayment records, and loan receipt account transaction records; the guarantee management system provides guarantee information and guarantor relationships; the audit management system provides audit findings and audit project information; and the human resources system provides employee information and organizational structure information.

[0032] In this embodiment, the loan contract information includes contract number, customer number, customer type, product code, currency, loan amount, loan balance, loan term, interest rate, disbursement date, maturity date, fund usage code, repayment method, guarantee method, account manager number, handling agency code, business line code, and contract status; the customer basic information includes customer number, customer name, document type, document number, industry classification code, establishment date, registered capital, number of employees, annual operating income, customer credit rating, and customer risk classification; the approval process record includes loan contract number, credit application number, application date, approval node, approver number, approval result, approval opinion, and approval date; the approval history record includes credit application number, historical sequence number, previous approval result, re-application date, and re-approval result.

[0033] In this embodiment, the loan agreement information includes loan agreement number, contract number, customer number, loan amount, loan balance, loan disbursement date, maturity date, interest rate, repayment method, loan status, and account manager number; the repayment schedule includes loan agreement number, repayment period, planned repayment date, planned principal repayment amount, planned interest repayment amount, and planned total repayment amount; the actual repayment record includes loan agreement number, actual repayment date, actual principal repayment amount, actual interest repayment amount, actual total repayment amount, and repayment type; and the loan agreement account transaction record includes loan agreement number, account number, transaction date, transaction time, transaction amount, transaction direction, counterparty account, counterparty name, transaction summary, and transaction channel.

[0034] In this embodiment, the guarantee information includes the guaranteed loan contract number, guarantee type, guarantor number, guarantor name, guarantee amount, collateral type, collateral appraised value, loan-to-value ratio, and appraisal date; the guarantor relationship includes the guarantor number, guaranteed customer number, and relationship type; the audit findings include the issue number, audit project number, involved loan contract number, involved customer number, issue type, issue severity, issue description, responsible account manager number, issue discovery date, and rectification status; the audit project information includes the audit project number, project name, project type, audit scope description, project start date, and project end date; the employee information includes the employee number, employee name, affiliated organization code, business line, position, date of employment, date of most recent job transfer, performance rating, and professional qualifications; and the organizational structure information includes the organization code, organization name, organization level, superior organization code, and regional code.

[0035] In this embodiment, the computer equipment extracts credit audit data from each business system in preset batches. The preset batches can be the T+1 batch at midnight every day, or they can be determined according to the audit project start time, audit plan cycle, or data update cycle. The computer equipment can extract all data or incremental data.

[0036] In this embodiment, the computer equipment performs data cleaning on the extracted credit audit data. Data cleaning includes removing test data, invalid contract data, duplicate loan data, abnormal date data, and data lacking key primary keys. Key primary keys include customer number, contract number, loan number, employee number, approval process number, or audit project number.

[0037] In this embodiment, the computer device standardizes the amount field by unifying the amount unit to RMB yuan and retaining two decimal places. For foreign currency loan data, it converts the amount into RMB according to the preset exchange rate or the median rate of the day. The computer device also unifies the date field to YYYY-MM-DD format and converts the code field into corresponding semantic values ​​through the dimension table. The code field includes the fund use code, guarantee method code, approval result code, institution code, and business line code.

[0038] In this embodiment, the computer device uses customer number, contract number, and employee number as association indexes to establish the association relationship between customer basic information, loan contract information, promissory note information, repayment information, account transaction information, guarantee information, approval process information, audit findings information, and employee information, thereby generating a basic dataset for credit audit.

[0039] In this embodiment, when the computer device generates the basic dataset for credit audit, it prioritizes associating loan contract information, promissory note information, guarantee information, and audit findings information with the contract number; it associates customer basic information, loan contract information, and account transaction information with the customer number; and it associates employee information with the account manager number, approver number, and responsible account manager number. When the same customer corresponds to multiple loan contracts, a one-to-many association record between the customer and the multiple loan contracts is generated. When the same loan contract corresponds to multiple promissory notes or multiple guarantee records, a one-to-many association record between the loan contract and the multiple promissory notes or multiple guarantee records is generated.

[0040] In this embodiment, when there are missing or inconsistent fields in cross-system data, the computer device determines the field values ​​according to preset priorities. The preset priorities include using loan contract information in the credit management system as the source of contract base fields, using the loan balance and repayment records in the core accounting system as the source of balance and repayment fields, and using the responsible account manager number in the audit management system as the source of audit issue attribution fields. When the missing key fields make it impossible to establish a correlation, the computer device writes the corresponding record into the abnormal data record table and does not use the abnormal record when generating the risk label in this round.

[0041] In this embodiment, the basic dataset for credit audit can be stored using relational database tables, wide tables, distributed data tables, or graph data structures. This invention does not limit the specific storage format, as long as it can preserve the relationships between customers, loan contracts, promissory notes, guarantees, approvals, audit issues, and employees, and can be called by the credit audit model.

[0042] In other implementations, the credit audit basic dataset can also generate project-level datasets based on the scope of the audit project, line-level datasets based on business lines, or institution-level datasets based on the institutional hierarchy. When generating the above datasets, the computer equipment still retains the fields of customer number, contract number, employee number, extraction batch, data source, and update time to ensure that subsequent risk labels, employee profiles, audit samples, and model feedback data can be traced back to the original business data. Example

[0043] This embodiment, based on embodiment 1 or embodiment 2, explains the method for determining the customer risk label set and customer risk level.

[0044] In this embodiment, the risk feature identification unit identifies customer risk features according to a preset audit rule base and transmits the identified customer risk features to the customer risk label generation unit, which generates customer risk label records based on the customer risk features.

[0045] In this embodiment, the customer risk label record includes customer ID, label ID, label name, trigger condition, trigger basis, label weight, hit flag, and hit time. Among them, the customer ID is used to identify the customer corresponding to the risk label, the label ID is used to identify the risk label type, the label name is used to express the risk characteristics in a semantic way, the trigger condition is used to represent the rule conditions corresponding to the risk label, the trigger basis is used to record the data field or field value that triggers the risk label, the label weight is used to calculate the customer risk score, the hit flag is used to indicate whether the customer has hit the risk label, and the hit time is used to record the time when the risk label is generated or updated.

[0046] In this embodiment, a preset audit rule base is used to automatically identify multi-dimensional risk characteristics of customers based on internal bank credit business data. These multi-dimensional risk characteristics include one or more of the following: entrepreneurial loans, insufficient repayment ability, re-loan after rejection by others, fund return, related credit lines, insufficient collateral, misappropriation of funds, continuous delinquency, expired assessment, and scale mismatch. After the computer device hits the risk rules in the preset audit rule base, it generates a set of customer risk tags with semantic descriptions, so that the customer risk identification result can not only be represented by a single risk score, but also by multiple interpretable risk sources and risk types.

[0047] In this embodiment, the triggering criteria include the triggering field name, triggering field value, hit rule number, and hit data source. When a customer hits the fund return rule, the triggering criteria include the loan agreement number, account number, transaction date, transaction amount, transaction direction, counterparty account, counterparty name, related party identification result, and corresponding risk rule number. When a customer hits the insufficient collateral rule, the triggering criteria include the guaranteed loan contract number, guarantee amount, collateral valuation, guarantee coverage ratio, collateral type, loan-to-value ratio, and corresponding loan-to-value ratio limit.

[0048] In this embodiment, the preset audit rule base includes at least two of the following: entrepreneurial loan rules, insufficient repayment ability rules, rules for re-issuing loans after rejection by others, fund return rules, related credit rules, insufficient guarantee rules, misappropriation of funds rules, continuous overdue rules, expired assessment rules, and scale mismatch rules.

[0049] In this embodiment, the startup loan rules in the preset audit rule base are triggered based on the fund use field and the operating period field. When the fund use field contains keywords such as startup, operation or opening a store, and the operating period is less than 2 years, the computer device determines that the customer has matched the startup loan tag.

[0050] In this embodiment, the insufficient repayment ability rule in the preset audit rule base is triggered based on the income coverage ratio, asset-liability ratio, or overdue records. When the income coverage ratio is less than 1.2, or the asset-liability ratio is greater than 70%, or the number of consecutive overdue periods is greater than or equal to 3, the computer device determines that the customer has the insufficient repayment ability label.

[0051] In this embodiment, the rule of re-issuing loans after being rejected by others in the preset audit rule base is triggered based on the approval history. When the same credit application was rejected by the superior approval node and was re-applied for and approved within 90 days, the computer device determines that the customer has hit the tag of re-issuing loans after being rejected by others.

[0052] In this embodiment, the fund return rules in the preset audit rule base are triggered based on the account flow after the loan is issued. When the funds return to the borrower's account or related party's account within 30 days after the loan is issued, and the return amount is greater than 80% of the loan amount, the computer device determines that the customer has hit the fund return tag.

[0053] In this embodiment, the associated credit rules in the preset audit rule base are triggered based on the guarantor's relationship, the length of the guarantee ring, or the number of associated customers granted credit within a preset period. When the length of the guarantee ring is less than or equal to 5, or the number of associated customers granted credit within 90 days is greater than or equal to 3, the computer device determines that the customer has hit the associated credit tag.

[0054] In this embodiment, the length of the guarantee ring used by the associated credit granting rules in the preset audit rule base is determined by the guarantee relationship graph. The computer device generates the guarantee relationship graph based on the guarantor number, the guaranteed customer number, and the relationship type. The guarantee relationship graph uses the customer or guarantor as nodes and the guarantee relationship as edges. The computer device traverses the path starting from the target customer node. When the traversal path returns to the target customer node or to the visited associated customer node, it is determined that a guarantee ring exists. The length of the guarantee ring is the number of edges contained in the closed path. When the length of the guarantee ring is less than or equal to the preset length threshold, it is determined that the associated credit granting rule has been hit.

[0055] In this embodiment, the insufficient collateral rule in the preset audit rule base is triggered based on the collateral coverage ratio or the mortgage ratio. When the collateral coverage ratio is less than 1.0, or the mortgage ratio exceeds the mortgage ratio limit corresponding to the collateral type, the computer device determines that the customer has hit the insufficient collateral label. The mortgage ratio limit corresponding to the collateral type can include 70% for real estate, 60% for land, and 50% for equipment.

[0056] In this embodiment, the misappropriation rules in the preset audit rule base are triggered based on whether the flow of funds is consistent with the agreed purpose. When the abnormal flow amount is greater than 20% of the loan amount or greater than 100,000 yuan, the computer equipment determines that the customer has been flagged as having misappropriated funds.

[0057] In this embodiment, the continuous overdue rule in the preset audit rule base is triggered based on the number of consecutive overdue periods or the cumulative number of overdue days in the past 12 months. When the number of consecutive overdue periods is greater than or equal to 3, or the cumulative number of overdue days is greater than 90 days, the computer device determines that the customer has hit the continuous overdue label.

[0058] In this embodiment, the evaluation expiration rule in the preset audit rule base is triggered based on the evaluation date of the collateral. When the evaluation date of real estate, land or construction in progress collateral is more than 730 days away from the current date, the computer device determines that the customer has hit the evaluation expiration tag.

[0059] In this embodiment, the scale mismatch rule in the preset audit rule base is triggered based on the relationship between the loan balance and the registered capital and the number of employees. When the ratio of the loan balance to the registered capital is greater than 10 times, or the average loan amount per person is greater than 500,000 yuan, the computer equipment determines that the customer has hit the scale mismatch tag.

[0060] In this embodiment, after the customer risk tag set is generated, the computer device calculates the customer risk score based on the tag weights and hit flags corresponding to the risk tags hit by the customer. The customer risk score can be the sum of the tag weights of each hit risk tag, or it can be the cumulative value of the product of the tag weights and the hit flags.

[0061] In this embodiment, each risk label is configured with a basic weight (such as an expert experience value), and the weight can be dynamically adjusted based on the historical audit discovery rate; the customer risk score is calculated using the following formula: Score_customer = Σ(Label weight_i × Label hit flag_i) in: Tag weight_i: The weight value of the i-th risk tag. The initial value is set by experts and optimized based on feedback. Tag hit flag _i: 1 if the customer hits the tag, 0 otherwise.

[0062] Customers are categorized into three risk levels based on their scores: High risk: Score_customer ≥ threshold_H (e.g., 10 points) Medium risk: Threshold_L ≤ Score_customer < Threshold_H (e.g., 5-9 points) Low risk: Score_customer < threshold_L (e.g., below 5 points) Thresholds H and L can be configured according to the bank's risk appetite and business characteristics.

[0063] For example: Client A hit 3 tags with weights of 5, 4, and 6 respectively, and the score = 5 + 4 + 6 = 15, which is classified as high risk; Client B hit two tags with weights of 4 and 3 respectively, and the score = 4 + 3 = 7, which is classified as medium risk.

[0064] Specifically, the computer equipment compares the customer's risk score with the risk level threshold to determine the customer's risk level. When the customer's risk score is greater than or equal to the high-risk threshold, the customer is identified as a high-risk customer. When the customer's risk score is greater than or equal to the low-risk threshold but less than the high-risk threshold, the customer is identified as a medium-risk customer. When the customer's risk score is less than the low-risk threshold, the customer is identified as a low-risk customer.

[0065] In other implementations, the risk rules and thresholds in the preset audit rule base can be configured according to the bank's internal audit system, regulatory requirements, historical audit experience, or actual business characteristics. The risk rules in the preset audit rule base can be enabled or disabled according to customer type, product type, business line, or institutional level. The customer risk score can also be further modified by combining the loan balance, overdue days, severity of the problem, or guarantee coverage. The modified customer risk score is still compared with the risk level threshold to determine the customer risk level. Example

[0066] This embodiment describes the method for generating employee risk profiles based on any of the foregoing embodiments.

[0067] In this embodiment, the computer device establishes a mapping relationship between customers and handling employees based on the customer manager number, approver number, responsible customer manager number, handling time, and approval node.

[0068] In this embodiment, when a customer corresponds to only one account manager, the computer device aggregates the customer's risk label set, customer risk level, customer risk score, and related loan balance to that account manager.

[0069] In this embodiment, when the same customer corresponds to multiple handling employees, the computer device assigns risk contribution weights to the multiple handling employees according to their handling roles, participation links, or contributions. The handling roles include customer managers, approvers, and responsible customer managers. The participation links include credit application, loan approval, loan disbursement, post-loan management, and audit problem rectification.

[0070] In this embodiment, the contribution can be determined by one or more of the following: role weight, participation stage weight, and processing time weight. In a specific example, the risk contribution weight of the account manager is 0.8, the risk contribution weight of the approver is 0.2, and the computer equipment allocates the risk contribution corresponding to the customer risk label set and customer risk level to multiple handling employees according to the above risk contribution weight.

[0071] In this embodiment, when a customer manager changes, the computer device determines the risk attribution based on the loan disbursement time, the risk tag hit time, the audit problem occurrence time, or the responsible customer manager field. For example, if the risk tag hit time is earlier than the customer manager change time, the corresponding risk is attributed to the previous customer manager. If the audit problem record contains a responsible customer manager field, the risk attribution is determined by the responsible customer manager field. If the responsibility attribution cannot be determined, it is allocated proportionally according to the handling time or the participation stage.

[0072] In this embodiment, the computer device aggregates the customer risk tag set and customer risk level to the corresponding employee based on the risk attribution result, and generates an employee risk profile.

[0073] In this embodiment, the employee risk profile includes at least two of the following employee profile indicators: number of high-risk customers, balance of problem loans, balance of non-performing loans, number of audit findings in history, total number of risk tags hit, proportion of high-risk tags, customer concentration, growth rate of new high-risk customers, rate of inadequate rectification, and average risk score.

[0074] In this embodiment, the number of high-risk customers represents the total number of customers under the employee's name whose risk level is high; the outstanding balance of problematic loans represents the total outstanding balance of loans under the employee's name that have been found to have problems in historical audits; the outstanding balance of non-performing loans represents the total outstanding balance of loans under the employee's name classified as substandard, doubtful, or loss; the number of times historical audits have been conducted represents the cumulative number of times problems have been found for customers under the employee's name in audits; the total number of risk tag hits represents the total frequency of risk tag hits for all customers under the employee's name; the proportion of high-risk tags represents the ratio of the number of high-risk tag hits to the total number of tag hits; customer concentration represents the relationship between the number of customers under the employee's name and the average number of customers in the institution; the growth rate of newly added high-risk customers represents the ratio of the number of newly added high-risk customers in the past 90 days to the total number of newly added customers in the past 90 days; the rate of inadequate rectification represents the ratio of the number of unrectified problems to the total number of problems found in historical audits; and the average risk score represents the arithmetic mean of the risk scores of all customers under the employee's name.

[0075] In this embodiment, since the dimensions and numerical ranges of the employee profile indicators are different, the computer device performs percentile standardization processing on the employee profile indicators.

[0076] In this embodiment, when the computer device performs percentile standardization on the employee profile indicators, it first sorts all employees according to the employee profile indicator values. For employee profile indicators with higher indicator values, the risk is higher. The indicators are sorted from largest to smallest. The standardized score is calculated as (the percentile of the employee's indicator value among all employees) × 100, where the employee ranked 1 corresponds to the highest standardized score.

[0077] In this embodiment, for the customer concentration index, the computer device first calculates the customer concentration deviation value. The customer concentration deviation value is equal to the absolute value of the difference between the result of dividing the number of customers under the employee's name by the average number of customers in the organization and 1. Then, the customers are sorted from largest to smallest according to the customer concentration deviation value and a standardized score is calculated.

[0078] In this embodiment, the computer device calculates the employee's comprehensive risk score based on the standardized scores of each employee profile indicator and the corresponding indicator weights. The employee's comprehensive risk score can be the sum of the products of the standardized scores of each employee profile indicator and the corresponding indicator weights.

[0079] In this embodiment, the formula for calculating the employee's comprehensive risk score is as follows: Score_emp = w1×E01_standardization + w2×E02_standardization + w3×E03_standardization + w4×E04_standardization + w5×E05_standardization + w6×E06_standardization + w7×E07_standardization + w8×E08_standardization + w9×E09_standardization + w10×E10_standardization.

[0080] E01-E10 represent the standardized scores corresponding to the number of high-risk customers, the balance of problem loans, the balance of non-performing loans, the number of times historical audits have been conducted, the total number of risk labels hit, the proportion of high-risk labels, customer concentration, the growth rate of new high-risk customers, the rate of inadequate rectification, and the average risk score, respectively, while w1-w10 represent the weights of the corresponding indicators.

[0081] In this embodiment, w1, w2 and w3 can be 0.15, w4, w5, w6 and w8 can be 0.10, and w7, w9 and w10 can be 0.05. The above weights can be adjusted according to the bank's audit focus or model feedback data.

[0082] In this embodiment, the computer device determines the employee risk level based on the employee's comprehensive risk score. For example, the top 10% of employees with the comprehensive risk score are identified as high-risk employees, the middle 20% are identified as medium-risk employees, and the bottom 70% are identified as low-risk employees. Alternatively, the employee risk level can be determined according to a preset employee risk level threshold.

[0083] In other implementations, employee profile indicators can be adjusted according to the type of audit project, organizational level, or business line. The employee's comprehensive risk score can be calculated using weighted summation, weighted average, or normalized weighted summation. The employee's risk level can be determined using percentile ranking, fixed threshold, or a combination of percentile ranking and fixed threshold. Example

[0084] This embodiment describes the method for generating the audit sample library based on any of the aforementioned embodiments.

[0085] In this embodiment, the computer device selects employees whose risk levels meet the preset level conditions as candidate key audit employees based on their risk levels. The candidate key audit employees are then sorted from high to low according to their comprehensive risk scores, and the top N employees are selected as key audit employees. Here, N is the preset number of key audit employees to be selected. N is an integer greater than or equal to 1 and not greater than the number of candidate key audit employees. The value of N can be configured according to the total amount of audit resources, the scope of audit projects, the audit plan, and risk preferences. In this way, key audit targets with high risk concentration are first determined at the employee level, so that subsequent customer sampling is limited to the customer range corresponding to the key audit employees.

[0086] In this embodiment, when the number of candidate key audit employees whose risk level meets the preset level conditions is less than N, the computer device selects employees from the next employee risk level in descending order of employee risk level, and supplements them to N employees in descending order of employee comprehensive risk score; when the total number of employees is less than N, all employees are identified as key audit employees.

[0087] In one implementation, the number N of key audit employees can be configured directly, or it can be determined by multiplying the total number of employees by a preset selection ratio and then rounding the result. When the credit audit model actually performs sampling, the final integer N is used as the number of key audit employees selected.

[0088] In this embodiment, to avoid insufficient sample coverage due to complete reliance on risk ranking, the computer equipment is equipped with a minimum sampling rule, which includes a regional minimum sampling rule, a business line minimum sampling rule, a special attention employee rule, and a random minimum sampling rule.

[0089] In this embodiment, the regional baseline rule is used to ensure that the audit sample covers a preset region, such as ensuring that at least one employee in each region is included in the scope of key audit employees. The business line baseline rule is used to ensure that the audit sample covers a preset business line, such as ensuring that at least one employee in each business line is included in the scope of key audit employees. The special attention employee rule is used to include recently hired employees, employees who have rotated positions, or employees with a lot of historical problems in the scope of attention. The random baseline rule is used to draw a certain proportion of samples from non-key employees or low-risk customers in addition to risk-oriented sampling.

[0090] In this embodiment, after identifying key audit employees, the computer device establishes a candidate sample pool between key audit employees and their clients based on the mapping relationship between clients and the handling employees. Each candidate record in the candidate sample pool includes employee number, employee comprehensive risk score, employee risk ranking, client number, client risk level, client risk score, risk tag set, and risk contribution weight.

[0091] In this embodiment, the generation of the audit sample library is constrained by both employee risk and customer risk. The computer equipment first determines the key audit employees based on the comprehensive risk score of the employees, and then determines the customer sampling priority among the customers under the key audit employees based on the customer risk level, customer risk score and risk label set, so that the risk aggregation results of the employee dimension and the risk identification results of the customer dimension are linked.

[0092] In this embodiment, after identifying key audit employees, the computer equipment performs stratified sampling of customers under the key audit employees' names. The stratified sampling is carried out according to high-risk customers, medium-risk customers, and low-risk customers. Among them, high-risk customers are sampled first, medium-risk customers are sampled second, and low-risk customers are sampled as a backup sample. The sampling ratio of high-risk customers is higher than that of medium-risk customers, and the sampling ratio of medium-risk customers is higher than that of low-risk customers, so that the customer samples corresponding to key audit employees can form a gradient sampling order according to the customer risk level.

[0093] In this embodiment, the sampling ratio for high-risk customers can be 50%, the sampling ratio for medium-risk customers can be 30%, and the sampling ratio for low-risk customers can be 20%. The above sampling ratios can be configured as sampling strategy parameters.

[0094] In this embodiment, when the computer device generates the audit sample library, it calculates the sample size of each stratum based on the total sample size N configured for the audit project and the stratified sampling ratio. The sample size of high-risk customers is the rounded value of N multiplied by the high-risk customer sampling ratio, the sample size of medium-risk customers is the rounded value of N multiplied by the medium-risk customer sampling ratio, and the sample size of low-risk customers is the remaining value after subtracting the sample sizes of high-risk customers and medium-risk customers from N.

[0095] In this embodiment, within the same customer risk level, the computer device determines the customer sampling priority from high to low based on the number of risk tags, the total weight of risk tags, the number of historical audit findings, or the loan balance. When two customers have the same customer risk level and are both associated with key audit employees, the computer device prioritizes sampling the customer with the higher comprehensive risk score of the associated employee, or determines the linked sampling priority based on the comprehensive risk score of the employee and the risk score of the customer.

[0096] In one specific implementation, the linked sampling priority can be determined jointly based on the employee's comprehensive risk score and the customer's risk score. For example, the linked sampling priority score is equal to the sum of the employee's comprehensive risk score multiplied by a first coefficient and the customer's risk score multiplied by a second coefficient. The first and second coefficients can be configured according to the audit focus. The computer equipment determines the sampling order of candidate customers from high to low according to the linked sampling priority score.

[0097] In this embodiment, when the number of customers at a certain risk level is insufficient to meet the corresponding sample size, the computer equipment supplements the sample size from other risk levels according to the customer sampling priority of the remaining customers. The customer sampling priority is determined based on at least one of the following: the number of risk labels, the sum of risk label weights, the number of historical audit findings, and the loan balance.

[0098] In this embodiment, when the same customer is associated with multiple key audit employees, the computer device deduplicates the customer sample and retains the key audit employee information and sampling source associated with the customer in the audit sample library.

[0099] In this embodiment, the sample records in the audit sample library include customer number, customer risk level, customer risk score, risk tag set, main trigger field, employee number, employee risk score, employee risk ranking, sampling source, sampling batch, sampling timestamp, and model version number. In one embodiment, the sample records also include employee risk source, customer risk source, and linked sampling priority to record the reason why the sample enters the audit sample library based on both employee risk and customer risk.

[0100] In this embodiment, the main triggering fields correspond to the triggering basis in the customer risk label record. The main triggering fields include the risk label number that leads the customer to enter the audit sample library, the triggering field name, the triggering field value, and the sampling source. When the same customer enters the audit sample library based on multiple risk labels or multiple key audit employees, the computer device retains all triggering basis and uses the risk label with the highest weight or the associated employee with the highest risk ranking as the main triggering field.

[0101] In this embodiment, the sampling sources include at least one of key employee sampling, risk label sampling, regional guaranteed sampling, business line guaranteed sampling, special attention employee sampling, and random guaranteed sampling. The sampling batch is used to distinguish different audit projects or different sampling periods. The sampling timestamp is used to record the sample generation time. The model version number is used to record the credit audit model version used when generating the sample. The main trigger field is used to record the main risk basis for the customer being selected.

[0102] In other implementations, the audit sample library can be generated separately based on the audit project, audit period, institutional level, or business line. The customer sampling priority can be determined by prioritizing the number of risk labels, the total weight of risk labels, the number of historical audit findings, the loan balance, or a multi-indicator weighted ranking method. The stratified sampling ratio can be adjusted based on audit resources, model feedback data, or regulatory focus.

[0103] Through the above sample records, auditors can trace the reason why each sample entered the audit sample library and determine the risk label, employee risk profile, sampling source and model version corresponding to the sample. Example

[0104] This embodiment describes the feedback update method of the credit audit model based on any of the foregoing embodiments.

[0105] In this embodiment, after completing on-site inspections or online verifications, auditors will enter the audit findings into the computer. The audit findings include whether the customer has a problem, the type of problem, the severity of the problem, the risk tags involved, the employees involved, the rectification status, and the date the problem was discovered.

[0106] In this embodiment, the computer device generates model feedback data based on the audit findings. The model feedback data includes risk label hit rate, problem detection rate, high-risk hit rate, employee hit rate, and coverage rate.

[0107] In this embodiment, the risk label hit rate represents the ratio between the number of samples that hit a certain risk label and are confirmed to have problems by audit, and the total number of samples that hit that risk label; the problem detection rate represents the ratio between the number of samples with problems and the total number of audited samples; the high-risk hit rate represents the ratio between the number of samples with problems in the high-risk samples and the total number of high-risk samples; the employee hit rate represents the ratio between the number of samples with problems in the key audited employee samples and the total number of key audited employee samples; and the coverage rate represents the ratio between the number of business lines covered by the audited samples and the total number of business lines in the entire industry, or the ratio between the number of regions covered by the audited samples and the total number of regions.

[0108] In this embodiment, the computer device updates the label weight based on the risk label hit rate. When the actual hit rate of a certain risk label is higher than the expected hit rate, the label weight of the risk label is increased. When the actual hit rate of a certain risk label is lower than the expected hit rate, the label weight of the risk label is decreased.

[0109] In this embodiment, the weight adjustment formula is as follows: New weight = old weight × (1 + α × (actual hit rate - expected hit rate)), where α is the learning rate parameter used to control the magnitude of weight adjustment, and the expected hit rate is the preset target hit rate. For example, if α is 0.5, the expected hit rate is 30%.

[0110] In this embodiment, to avoid significant fluctuations in label weights due to insufficient audit samples or abnormal samples in a single audit, the computer equipment sets weight adjustment boundaries. When the actual hit rate is higher than the expected hit rate, the single weight increase shall not exceed 20%. When the actual hit rate is lower than the expected hit rate, the single weight decrease shall not exceed 20%. When the actual hit rate is 0%, the risk label weight is frozen and manual review is triggered. When the actual hit rate is 100%, the risk label weight is limited to a preset multiple of the initial weight, for example, not exceeding twice the initial weight.

[0111] In this embodiment, the computer device updates the risk level threshold based on the problem detection rate and the high-risk hit rate. When the problem detection rate is lower than the preset problem detection rate and the high-risk hit rate is lower than the preset high-risk hit rate, the high-risk threshold is lowered and the number of key audit employees selected, N, is increased. When the problem detection rate is lower than the preset problem detection rate and the high-risk hit rate reaches the preset high-risk hit rate, the low-risk threshold is lowered and the sampling ratio of medium-risk customers is increased. When the problem detection rate reaches the preset problem detection rate and the high-risk hit rate is lower than the preset high-risk hit rate, the number of key audit employees selected, N, is increased and the label weight of the high-risk label is adjusted. When both the problem detection rate and the high-risk hit rate reach the corresponding preset values, the current sampling strategy is maintained and only the label weight is fine-tuned.

[0112] In this embodiment, the high-risk threshold is adjusted by 1 point each time, and the high-risk threshold is maintained within the range of 8 to 15 points. The low-risk threshold is adjusted by 1 point each time, and the low-risk threshold is maintained within the range of 3 to 8 points. The number of key audit employees N is adjusted by 2% each time, and the number of key audit employees N is maintained within the range of 3% to 15%. The sampling ratio of high-risk customers is adjusted by no more than 5% each time, the sampling ratio of medium-risk customers is adjusted by no more than 3% each time, and the sampling ratio of low-risk customers is adjusted accordingly based on the adjustment results of the sampling ratios of high-risk customers and medium-risk customers.

[0113] In this embodiment, the computer device updates the number N of key audit employees or the employee profile parameters based on the number of problematic customers under each key audit employee and the severity of the problems. When the number of problematic customers corresponding to a key audit employee is lower than the preset value, the number N of key audit employees can be increased. When the number of problematic customers corresponding to a key audit employee is large and the severity of the problems is high, the number N of key audit employees can be maintained or reduced, and the weight of the corresponding employee profile indicators can be adjusted to increase the influence of the number of historical audit findings, the balance of problematic loans, or the proportion of high-risk labels in the employee's comprehensive risk score.

[0114] In this embodiment, the computer device updates the minimum sampling rule based on the coverage rate. When the coverage rate is lower than the preset coverage rate threshold, the minimum number of people in the region or the minimum number of people in the business line is increased, or the regional granularity and business line granularity are refined.

[0115] In this embodiment, after each model update, the computer device writes the preset audit rule base, label weight parameters, risk level threshold, employee profile parameters, and sampling strategy parameters from before the update into the historical model version record, and writes the updated preset audit rule base, label weight parameters, risk level threshold, employee profile parameters, and sampling strategy parameters into the current model version record. The audit project number, sample batch, feedback data statistical period, and update personnel or update task identifier are recorded as feedback data sources, so that a traceable relationship is formed between audit samples, audit findings, and model parameter updates.

[0116] In this embodiment, during the next round of audit sampling, the computer device calls the updated credit audit model to re-execute customer risk label generation, employee risk profile construction, and audit sample generation, so that the audit findings are fed back to the credit audit model and used to adjust subsequent risk identification and sampling decisions.

[0117] In other implementations, model feedback data can be statistically analyzed according to audit project, audit period, business line, institutional level, or risk label type. The credit audit model can update model parameters by means of automatic update, manual confirmation after update, or automatic generation of update suggestions followed by confirmation by auditors. The model version can be set to enabled, disabled, and rollback states so that it can be restored to the historical model version when the parameter update effect does not meet the preset conditions. Example

[0118] In this embodiment, the computer equipment extracts the credit audit data that was added or changed the previous day from the credit management system, core accounting system, guarantee management system, audit management system and human resources system every morning at midnight. The extracted data is then cleaned, standardized and field-transformed. The basic credit audit dataset is then generated using customer number, contract number and employee number as associated indexes.

[0119] In this embodiment, the computer device calls the current version of the credit audit model based on the credit audit basic dataset. The current model version includes a preset audit rule base, label weight parameters, risk level thresholds, employee profile parameters, and sampling strategy parameters.

[0120] In this embodiment, for customer A, the computer device reads its loan purpose field, establishment date, debt-to-asset ratio, and account transaction information. If customer A's loan purpose field includes opening a store and operating for one year, then customer A is determined to have met the entrepreneurial loan label. If customer A's debt-to-asset ratio is 75%, then customer A is determined to have met the insufficient repayment ability label. If more than 80% of the loan amount flows back to the related party's account within 30 days after the loan is disbursed, then customer A is determined to have met the fund return label.

[0121] In this embodiment, the computer device writes the entrepreneurial loan tag, insufficient repayment ability tag, and capital return tag corresponding to customer A into the customer risk tag set, and calculates the customer risk score according to the tag weight of each risk tag. If customer A's customer risk score is greater than or equal to the high risk threshold, then customer A is determined to be a high-risk customer.

[0122] In this embodiment, the computer device maps the risk label and risk level of customer A to the corresponding employees based on the customer manager number and approver number in customer A's loan contract. If customer A corresponds to customer manager A and approver B, then risk contribution weights are assigned to A and B according to their respective roles, and the corresponding risk contributions are included in their employee risk profiles.

[0123] In this embodiment, the computer device counts the number of high-risk customers under each employee's name, the balance of problem loans, the balance of non-performing loans, the number of times historical audits have been conducted, the total number of risk tags hit, the proportion of high-risk tags, customer concentration, the growth rate of new high-risk customers, the rate of non-rectification, and the average risk score. The above indicators are then standardized for percentiles to calculate the employee's comprehensive risk score.

[0124] In this embodiment, the computer equipment first filters employees whose risk levels meet preset criteria as candidate key audit employees based on their risk levels. Then, it sorts the candidate key audit employees according to their comprehensive risk scores from highest to lowest, and selects the top N employees as key audit employees (N is 5 in this embodiment). The computer equipment determines the clients under the key audit employees based on the mapping relationship between clients and the handling employees, and generates a candidate sample pool containing employee risk information and client risk information. Stratified sampling is performed in the candidate sample pool according to high-risk clients, medium-risk clients, and low-risk clients. Within the same client risk level, sampling is performed based on the number of risk tags, the sum of risk tag weights, and historical audit data. Customer sampling priorities are determined by ranking customers by the number of discovery events or loan balance from highest to lowest. When multiple customers have the same risk level, the linked sampling priority is determined based on the comprehensive risk score of the corresponding key audit staff. The computer equipment also performs coverage verification on the initially selected customer samples according to the customer's institution, credit product type, and loan amount range, and sets the minimum sampling size for each institution, product type, or amount range. When the number of customer samples corresponding to any coverage dimension is lower than its minimum sampling size, additional samples are drawn from the remaining candidate customers corresponding to that coverage dimension according to the customer sampling priority until the corresponding minimum sampling size is reached, so that the audit sample takes into account both risk orientation and business coverage.

[0125] In this embodiment, the computer device performs coverage verification on the initial customer samples selected according to the customer sampling priority. The coverage verification includes the institutional dimension, credit product dimension, and loan amount dimension. When the number of customer samples corresponding to any coverage dimension is lower than the preset minimum sampling quantity, customer samples are supplemented from the remaining candidate customers corresponding to that coverage dimension according to the customer sampling priority until the preset minimum sampling quantity is reached. After the supplemented sampling, duplicate customer samples are deduplicated to generate an audit sample library.

[0126] In this embodiment, the institutional dimension can be divided according to the head office, branch, sub-branch, or region; the product dimension can be divided according to working capital loans, fixed asset loans, personal business loans, consumer loans, or other credit products; and the amount dimension can be divided according to a preset loan amount range. Different coverage dimensions can be configured with minimum sampling sizes. When the same customer meets the supplementary conditions of multiple coverage dimensions at the same time, the customer is only included in the audit sample library once, and the institutional coverage, product coverage, and amount coverage conditions met by the customer are recorded separately.

[0127] In this process, whether Client A is included in the audit sample library depends not only on Client A's own client risk level and risk label set, but also on whether its corresponding staff member is identified as a key audit staff member. Thus, the staff risk profile results and client risk label results work together to generate the audit sample.

[0128] In this embodiment, the computer device generates an audit sample library. The sample records in the audit sample library include customer number, customer risk level, customer risk score, risk tag set, main trigger field, employee number, employee risk score, employee risk ranking, sampling source, sampling batch, sampling timestamp, and model version number.

[0129] In this embodiment, after completing the audit, the auditors enter the information such as whether the customer has a problem, the type of problem, the severity of the problem, the risk tags involved, the employees involved, the rectification status, and the date the problem was discovered into the computer device.

[0130] In this embodiment, the computer equipment, based on the audit findings, counts the number and severity of problems found for each risk label and adjusts the label weight accordingly; counts the number and severity of problems found among high-risk, medium-risk, and low-risk customers and adjusts the risk level threshold accordingly; counts the number and severity of problematic customers under each key audit employee and adjusts the number N of key audit employees selected or the employee profile parameters accordingly; and counts the sample size and problem distribution for different institutions, loan amounts, and credit products and adjusts the minimum sampling rule accordingly.

[0131] In this embodiment, after the computer device completes the adjustment, it generates a new model version. In the next round of audit sampling, the computer device uses the new model version to regenerate customer risk labels, construct employee risk profiles, and generate audit samples. Example

[0132] This embodiment provides a credit customer risk label generation and audit sampling device based on audit modeling. The device may include a data processing module, an audit modeling module, a customer risk label generation module, an employee risk profile module, a sampling generation module, and a model feedback optimization module.

[0133] In this embodiment, the data processing module is used to standardize and integrate credit audit data from multiple business systems to generate a basic credit audit dataset.

[0134] In this embodiment, the audit modeling module is used to build a credit audit model based on the credit audit basic dataset. The credit audit model is a rule parameterized model that includes a preset audit rule base, label weights, employee profile parameters and sampling strategy parameters.

[0135] In this embodiment, the customer risk label generation module is used to determine the set of customer risk labels and the customer risk level through a credit audit model.

[0136] In this embodiment, the employee risk profiling module is used to aggregate customer risk tags and customer risk levels into an employee risk profile according to the mapping relationship between customers and the handling employees.

[0137] In this embodiment, the sampling generation module is used to identify key audit employees based on employee risk profiles and to perform stratified sampling of the clients under the key audit employees to generate an audit sample library.

[0138] In this embodiment, the model feedback optimization module is used to generate model feedback data based on the audit findings, and to update the credit audit model using the model feedback data, so that the updated credit audit model can be used for the next round of customer risk label generation, employee risk profile construction and audit sample generation.

[0139] For specific limitations of each module in the above-mentioned device, please refer to the limitations of the corresponding steps in the foregoing method embodiments, which will not be repeated here. Example

[0140] This embodiment provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the credit customer risk label generation and audit sampling method based on audit modeling of any of the foregoing embodiments.

[0141] This embodiment also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the credit customer risk label generation and audit sampling method based on audit modeling of any of the foregoing embodiments.

[0142] This embodiment also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the credit customer risk label generation and audit sampling method based on audit modeling of any of the foregoing embodiments.

[0143] This invention standardizes and integrates credit audit data from multiple business systems to generate a basic credit audit dataset. This dataset enables customer, contract, IOU, repayment, transaction history, guarantee, approval, audit findings, and employee data scattered across credit management systems, core accounting systems, guarantee management systems, audit management systems, and human resources systems to form a unified data processing foundation using associated indexes such as customer ID, contract ID, and employee ID. This avoids subsequent risk identification and audit sampling relying on data from a single system or manually pieced-together data, and improves the consistency and traceability of credit audit data during model processing.

[0144] This invention constructs a rule-parametric credit audit model based on a basic credit audit dataset. This model includes a pre-defined audit rule base, label weights, employee profile parameters, and sampling strategy parameters. Customer risk identification no longer simply outputs a single risk score; instead, it identifies multi-dimensional risk characteristics based on the pre-defined audit rule base, such as entrepreneurial loans, insufficient repayment ability, cash flow problems, related-party lending, insufficient collateral, and re-lending after rejection by others. It generates a set of customer risk labels with semantic descriptions and calculates customer risk scores and determines customer risk levels through label weights and hit indicators. This allows customer risk identification results to participate in subsequent calculations and also enables tracing the source of risk through label names, triggering conditions, and triggering bases.

[0145] This invention aggregates customer risk tags and customer risk levels into employee risk profiles based on the mapping relationship between customers and their responsible employees. Then, it identifies key audit employees based on these profiles and performs stratified sampling among the customers under these key audit employees according to their risk levels. This ensures that the generation of audit samples is constrained by both employee and customer risks. In this way, it links the risk concentration at the employee level with the risk tag results at the customer level, avoiding the generation of samples based solely on single-point customer risks or employee experience. This improves the structuring and risk-oriented nature of the sample sources in the audit sample library.

[0146] This invention generates model feedback data based on audit findings and uses this data to update the label weights, risk level thresholds, employee profile parameters, and sampling strategy parameters in the credit audit model. This ensures that audit findings are not merely stored as audit files but are fed back into the next round of customer risk label generation, employee risk profile construction, and audit sample generation. Simultaneously, the model version number records the parameters before and after the update, the source of the feedback data, the update time, and the basis for the update, creating a traceable data processing closed loop between audit samples, audit findings, and model parameter updates.

[0147] Those skilled in the art will understand that all or part of the processes in the foregoing embodiments can be implemented by computer program instructions and related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the foregoing method embodiments.

[0148] The above embodiments only illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, without departing from the concept of the present invention, modifications or substitutions can be made to the data source, risk rules, tag weights, employee profile indicators, sampling strategy parameters, feedback update methods, model version management methods, and data storage methods. All such modifications or substitutions should fall within the scope of protection of the present invention.

Claims

1. A method for generating credit customer risk labels and conducting audit sampling based on audit modeling, characterized in that, include: Standardize and integrate credit audit data from multiple business systems to generate a basic credit audit dataset; A credit audit model is constructed based on the basic dataset of credit audit. The credit audit model is a rule-parametric model that includes a preset audit rule base, label weights, employee profile parameters and sampling strategy parameters. Determine the customer risk label set and customer risk level through a credit audit model; Based on the mapping relationship between customers and the staff in charge, the customer risk tags and customer risk levels are aggregated into employee risk profiles. Based on the employee risk profile, key audit employees are identified, and stratified sampling is conducted on the clients under the names of key audit employees to generate an audit sample library; Based on the audit findings, model feedback data is generated, and the credit audit model is updated using the model feedback data. The updated credit audit model is then used for the next round of customer risk label generation, employee risk profile construction, and audit sample generation.

2. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 1, characterized in that, The generated credit audit basic dataset includes: Credit audit data is obtained from the credit management system, core accounting system, guarantee management system, audit management system, and human resources system; Perform data cleaning, standardize monetary units, standardize date formats, semantically convert code fields, and convert foreign currencies for credit audit data; Using customer ID, contract ID, and employee ID as the associated indexes, establish the relationships between customer basic information, loan contract information, promissory note information, repayment information, account transaction information, guarantee information, approval process information, audit findings information, and employee information to generate a basic dataset for credit auditing.

3. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 1, characterized in that, The credit audit model includes a risk feature identification unit, a customer risk label generation unit, an employee risk profile aggregation unit, an audit sampling decision unit, and a model feedback optimization unit. The risk feature identification unit is used to identify customer risk features based on a preset audit rule base. The customer risk label generation unit is used to generate a set of customer risk labels based on customer risk characteristics and determine the customer risk level. The employee risk profile aggregation unit is used to aggregate customer risk tag sets and customer risk levels to the employee dimension; The audit sampling decision unit is used to generate an audit sample library based on employee risk profiles; The model feedback optimization unit is used to update the model parameters of the credit audit model based on the model feedback data.

4. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 1, characterized in that, The credit audit model is used to determine the set of customer risk labels and customer risk levels, including: Identify customer risk characteristics based on a pre-defined audit rule base, and generate customer risk label records based on the identified customer risk characteristics; The customer risk tag record includes customer number, tag number, tag name, trigger condition, trigger basis, tag weight, hit flag, and hit time; Calculate the customer's risk score based on the tag weights and hit indicators corresponding to the risk tags matched by the customer; The customer's risk level is determined by comparing the customer's risk score with the risk level threshold.

5. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 1, characterized in that, The preset audit rule base includes at least two of the following: fund return rules, related credit rules, insufficient collateral rules, and rules for re-lending after rejection by others; The fund return rules are triggered based on whether the funds return to the borrower's account or related party's account within a preset period after the loan is disbursed, and the proportion of the return amount to the loan amount. The associated credit granting rules are triggered based on the length of the guarantee loop or the number of associated customer credit grantings within a preset period; The insufficient collateral rule is triggered based on whether the collateral coverage ratio or the loan-to-value ratio exceeds the maximum loan-to-value ratio corresponding to the type of collateral. The rule for re-granting loans after a rejection by another party is triggered based on whether there is a rejection record in the approval process and the time interval between the re-application and approval after a rejection.

6. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 1, characterized in that, The process of aggregating customer risk tags and customer risk levels into employee risk profiles based on the mapping relationship between customers and their assigned employees includes: Establish a mapping relationship between customers and handling employees based on customer manager number, approver number, responsible customer manager number, handling time and approval node; When the same customer corresponds to multiple staff members, risk contribution weights are assigned to the multiple staff members based on their roles, participation stages, or contributions. When a client manager changes, the risk attribution is determined based on the loan disbursement time, the time when the risk tag is triggered, the time when the audit issue occurs, or the responsible client manager field. Based on the risk attribution results, customer risk labels and customer risk levels are aggregated and assigned to the corresponding employees.

7. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 6, characterized in that, The employee risk profile includes at least two of the following employee profile indicators: number of high-risk customers, balance of problem loans, balance of non-performing loans, number of audit findings in history, total number of risk tags hit, percentage of high-risk tags, customer concentration, growth rate of new high-risk customers, rate of failure to rectify issues, and average risk score. Employee profile metrics were standardized using percentiles. For employee profile indicators where higher indicator values ​​indicate higher risks, the ranking percentile is determined according to the indicator values ​​from largest to smallest. For the customer concentration index, the ranking percentile is determined according to the degree to which the customer concentration deviates from the institution's average. The employee's overall risk score is calculated based on the standardized scores of each employee profile indicator and the corresponding indicator weights, and the employee's risk level is determined based on the overall risk score.

8. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 7, characterized in that, The generation of the audit sample library includes: Employees whose risk levels meet the preset criteria are selected as candidates for key audits. Candidate key audit employees are ranked according to their comprehensive risk scores, and those who meet the preset criteria are selected as key audit employees. Based on the mapping relationship between customers and the staff in charge, identify the customers under the name of the key audit staff and generate a candidate sample pool containing employee risk information and customer risk information. In the candidate sample pool, stratified sampling is performed according to high-risk customers, medium-risk customers, and low-risk customers; Within the same customer risk level, customer sampling priority is determined from high to low based on the number of risk tags, the total weight of risk tags, the number of historical audit findings, or the loan balance. When multiple clients have the same risk level, the priority of linked sampling is determined based on the comprehensive risk score of the corresponding key audit staff. A coverage verification is performed on the customer sample selected according to the customer sampling priority. The coverage verification includes at least one of the institutional dimension, credit product dimension, and loan amount dimension. When the number of customer samples corresponding to any coverage dimension is lower than the preset minimum sampling quantity, customer samples are drawn from the remaining candidate customers corresponding to that coverage dimension according to the customer sampling priority until the preset minimum sampling quantity is reached. When the same client is associated with multiple key audit staff, the client sample is deduplicated, and the information of the key audit staff associated with that client and the sampling source are retained in the audit sample library.

9. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 8, characterized in that, The sample records in the audit sample library include customer number, customer risk level, customer risk score, risk tag set, main trigger field, employee number, employee risk score, employee risk ranking, linked sampling priority, sampling source, sampling batch, sampling timestamp, and model version number. The sampling sources include at least one of the following: key employee sampling, risk label sampling, regional minimum guarantee sampling, business line minimum guarantee sampling, special attention employee sampling, and random minimum guarantee sampling; Among them, regional minimum sampling and business line minimum sampling are used to ensure that the audit sample covers a preset region or a preset business line.

10. The method for generating credit customer risk labels and conducting audit sampling based on audit modeling as described in claim 1, characterized in that, The step of generating model feedback data based on audit findings and updating the credit audit model using the model feedback data includes: Receive audit findings, which include whether the customer has a problem, the type of problem, the severity of the problem, the risk tags involved, the employees involved, the rectification status, and the date the problem was discovered. Model feedback data is generated based on the audit findings. The model feedback data includes the number and severity of problems found for each risk label, the number and severity of problems found for different customer risk levels, the number and severity of problematic customers under the names of each key audit employee, and the sample size and problem distribution for different institutions, loan amount ranges and credit product types. The label weights are updated based on the number and severity of issues found for each risk label. The risk level thresholds are updated based on the number and severity of issues found for different customer risk levels. The number of key audit staff selected (N) or the staff profile parameters are updated based on the number and severity of problematic customers under each key audit staff member. The sampling strategy parameters are updated based on the sample size and issue distribution for different institutions, loan amount ranges, and credit product types. Each update generates a new model version and records the model version number, parameters before the update, parameters after the update, feedback data source, update time, and update basis.

Images