Internet financial enterprise credit risk analysis method based on big data
The dynamic credit risk analysis system, built through big data technology and multi-level models, solves the problems of data complexity and lagging dynamic monitoring in internet finance, enabling accurate prediction and proactive prevention of corporate credit risks and reducing bad debt losses.
Patent Information
- Application Number
- CN202511793847.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-03-17
AI Technical Summary
Traditional corporate credit risk analysis methods suffer from problems in internet finance scenarios, such as complex data sources, inconsistent formats, high missing rates, a large amount of false data, and a lack of dynamic monitoring, resulting in difficulties in risk identification, low assessment accuracy, and lagging monitoring.
Big data technology is used to collect data in layers and build a multi-level dynamic credit risk assessment model, including basic credit scoring, dynamic risk iteration, related risk identification and public opinion risk assessment. Combined with graph neural networks and BERT models, real-time data cleaning and model updates are achieved, and an integrated risk analysis system of 'data-model-scenario' is built.
It enables dynamic perception and accurate prediction of corporate credit risk, significantly reduces bad debt losses, identifies risks through real-time monitoring, prevents chain defaults, and improves data quality and analysis accuracy.
Abstract
Description
Technical Field
[0001] This invention relates to the field of credit risk analysis technology for internet finance companies, specifically a method for credit risk analysis of internet finance companies based on big data. Background Technology
[0002] With the rapid development of the internet finance industry, corporate financing needs are becoming increasingly diversified. As important providers of financing services, internet finance institutions face severe challenges such as difficulty in identifying corporate credit risks, low accuracy in assessment, and lagging monitoring. Traditional corporate credit risk analysis methods have many shortcomings and are no longer suitable for the needs of internet finance scenarios.
[0003] Traditional credit risk analysis relies heavily on static data such as corporate financial statements and central bank credit reports, neglecting dynamic operational data of enterprises in the internet context, such as e-commerce transaction volume, supply chain collaboration data, non-financial data such as corporate public opinion, intellectual property status, and related party data such as shareholder credit and partner performance records.
[0004] In the context of internet finance, enterprise data sources are complex, with problems such as inconsistent data formats, high missing rates, and a large amount of false data. Traditional methods lack effective data cleaning and verification mechanisms, making data integration difficult and leading to distorted analysis results.
[0005] Existing risk analysis models are mostly static statistical models such as logistic regression and decision trees, which rely on historical data for training and cannot adapt to the rapid iteration of internet finance businesses and the dynamic changes in corporate risk patterns. Traditional methods are mostly "one-time pre-loan assessments," lacking real-time monitoring of the company's post-loan operating conditions and making it difficult to detect changes in the company's risk during the lending cycle.
[0006] Therefore, this invention provides a big data-based method for analyzing credit risk in internet finance companies to address the aforementioned problems. Summary of the Invention
[0007] (a) Technical problems to be solved
[0008] To address the shortcomings of existing technologies, this invention provides a big data-based method for analyzing credit risk in internet finance companies. It innovatively integrates cutting-edge technologies such as cross-domain data governance, digital twin modeling, and graph neural network risk projection, constructing a three-in-one enterprise credit risk analysis system encompassing "data-model-scenario." This system is applicable to scenarios such as internet banks, supply chain finance platforms, and cross-border fintech companies, enabling dynamic perception, accurate prediction, and proactive prevention of enterprise credit risk, thus solving the aforementioned problems.
[0009] (II) Technical Solution
[0010] To achieve the above objectives, the present invention provides the following technical solution: a big data-based method for credit risk analysis of internet finance enterprises, comprising the following steps:
[0011] Step 1: Perform stratified data collection:
[0012] This includes core financial data collection, scenario-based operational data collection, and related party data collection;
[0013] Step Two: Constructing Big Data Features and Risk Features:
[0014] Construct the characteristics of a company's financial health, operational capabilities, credit performance, related risks, and public opinion risks;
[0015] Step 3: Construct a multi-level dynamic credit risk assessment model:
[0016] The project involves constructing a basic credit scoring model, a dynamic risk iteration model, a related risk identification model, and a public opinion risk assessment model for enterprises.
[0017] Step 4: Analyze and alert on corporate credit risks in real time using a multi-level dynamic credit risk assessment model.
[0018] Preferably, in step one:
[0019] Core financial data includes corporate credit history, guarantee information, and financial data obtained through compliant interfaces connected to the People's Bank of China Credit Reference Center, Baihang Credit Reference Center, and corporate credit reporting agencies. Specifically, this includes loan balance, repayment history, number of overdue payments, guarantee amount, credit status of the guaranteed company, balance sheet, profit and loss statement, and cash flow statement.
[0020] Contextualized operational data includes dynamic operational data collected from enterprises in internet scenarios, such as data from cooperation with e-commerce platforms, enterprise purchase orders, supplier information, logistics data, enterprise business registration information, tax data, social security payment records, and intellectual property data. These data reflect the enterprise's market competitiveness and operational stability, provide a basis for analyzing the enterprise's supply chain coordination capabilities and cash flow, and provide a basis for risk assessment in non-financial dimensions.
[0021] Related party data includes credit data of corporate related parties. It obtains information on shareholders' capital contributions and the operating status of related companies through the corporate credit information disclosure system. It also analyzes social networks to identify whether there are any records of dishonesty in the personal credit records of the actual controller of the company and whether there are any defaults in related companies, thus identifying the transmission path of related risks.
[0022] Preferably, step one further includes cleaning and verifying the collected data, desensitizing and encrypting the data, and standardizing and integrating the data.
[0023] Preferably, the cleaning and verification of the collected data involves: using a combination of rule engine and machine learning for data cleaning; the rule engine is used to process data with inconsistent formats and delete duplicate data; machine learning algorithms such as isolated forest and autoencoder are used to identify abnormal data such as abnormal fluctuations in corporate financial data and false transaction records, and the autoencoder is used to build a feature model of normal transaction data, marking transaction data that deviates from the model as suspicious data and triggering manual review.
[0024] Data anonymization and encryption: Following the principle of minimum necessity, sensitive data such as corporate trade secrets and shareholder ID numbers are anonymized. Corporate bank account numbers are formatted with asterisks added between the first 6 and last 4 digits, and shareholder ID numbers are partially hidden. The AES-256 encryption algorithm is used to encrypt the data transmission and storage process to prevent data leakage.
[0025] Preferably, the data standardization and integration involves: constructing a unified data standard system to convert multi-source heterogeneous data into structured features; adopting standard classification for enterprise industry classification; normalizing operational data; and converting revenue data of enterprises of different sizes into relative indicators such as revenue growth rate and revenue profit margin; and integrating data scattered across different platforms through data lake technology to form a unified credit data view for enterprises, laying the foundation for subsequent analysis.
[0026] Preferably, step two further includes using feature correlation analysis, such as Pearson correlation coefficient, to remove redundant features. When the correlation coefficient between the debt-to-equity ratio and the current ratio exceeds 0.8, the more informative debt-to-equity ratio is retained. Key risk features are screened through feature importance assessment using random forest features to ensure the effectiveness of the model input features and reduce computational costs.
[0027] Financial health characteristics: Based on corporate financial data, traditional financial indicators such as debt-to-equity ratio, current ratio, quick ratio, revenue growth rate, and net profit margin are extracted. At the same time, data on net cash flow from operating activities and short-term debt are introduced as cash flow health indicators, and the consistency between financial statement data and tax declaration data is introduced as a new characteristic of financial data authenticity, so as to comprehensively assess the financial status of enterprises.
[0028] Operational capability characteristics: Extract order volume growth rate, customer retention rate, supplier stability reflected by the length of cooperation with core suppliers, and logistics efficiency reflected by on-time delivery rate from scenario-based operational data to reflect the company's actual operational capabilities;
[0029] Credit performance characteristics: By combining the company's credit history and non-financial performance data, we extract characteristics such as the number of overdue payments, the number of overdue days, the performance rate, and whether there is a record of being a dishonest person subject to enforcement, and assess the company's willingness to perform its credit obligations.
[0030] Associated risk characteristics: By using graph structure data such as enterprise relationship diagrams and supply chain networks, we can extract characteristics such as the number of associated enterprises, the default rate of associated enterprises, and the scale of intra-group fund transactions to identify associated risks;
[0031] Public opinion risk characteristics: Collect public opinion information of enterprises in news media, social platforms, and industry forums, and use natural language processing technology to extract public opinion sentiment, negative public opinion keywords, and public opinion dissemination scope characteristics to assess the impact of public opinion on enterprise credit.
[0032] Preferably, step three also includes establishing a real-time model update mechanism, specifically by collecting new business operation data and risk cases daily, incrementally training the model, and updating feature weights; and by optimizing the model structure and adjusting the associated risk identification algorithm monthly based on newly discovered risk patterns to ensure the model's adaptability to new risks.
[0033] Basic Credit Scoring Model: Based on the XGBoost algorithm, a basic credit scoring model is constructed using core financial data and static features of enterprises. It outputs a basic credit score of 0-1000 for enterprises. By learning the financial characteristics of enterprises with historical non-performing loans, the model can identify high-risk feature combinations such as a debt-to-equity ratio of over 80% and a current ratio of less than 1, and give low credit scores to enterprises that meet the combination.
[0034] Dynamic Risk Iteration Model: Introducing Long Short-Term Memory Network and Temporal Convolutional Network to perform time-series analysis on dynamic business data of enterprises, capturing real-time changes in the business situation of enterprises, and learning the time series trend of enterprise revenue through LSTM model. When the model predicts that the revenue will decline by more than 20% month-on-month in the next 3 months, the enterprise credit score will be automatically downgraded and the enterprise will be marked as a potential business risk enterprise.
[0035] Related Risk Identification Model: A graph neural network is used to construct a graph of enterprise relationships, where nodes represent enterprises and related parties, and edges represent relationship types, including equity relationships and supply chain relationships. The GraphSAGE algorithm is used to learn the characteristics of related nodes and to explore the transmission path of related risks. When an enterprise defaults on its credit, the model can automatically identify downstream enterprises that are related to it in the supply chain, assess the default risk of these enterprises, and trigger a related risk warning.
[0036] Public opinion risk assessment model: Based on the BERT pre-trained model, the model performs sentiment analysis and risk classification on corporate public opinion texts and outputs the public opinion risk level, including three levels: low, medium and high. When a company has negative public opinion about unpaid wages being complained about by employees and the spread covers multiple industry media, the model judges it as a high public opinion risk and incorporates this feature into the overall risk assessment.
[0037] (III) Beneficial Effects
[0038] Compared with existing technologies, this invention provides a big data-based method for credit risk analysis of internet finance companies, which has the following beneficial effects:
[0039] 1. This big data-based credit risk analysis method for internet finance companies comprehensively characterizes corporate credit risk through multi-source data fusion and multi-level models, effectively identifies risk points missed by traditional methods, significantly reduces bad debt losses, and effectively solves the problems of inconsistent data formats and a large amount of false data through big data cleaning and verification mechanisms. The data missing rate is reduced and the false data identification rate is improved, providing reliable data support for risk analysis.
[0040] 2. This big data-based credit risk analysis method for internet finance companies upgrades risk monitoring from static one-time assessment to dynamic continuous monitoring through real-time data collection and model update mechanisms. It identifies risks in advance and freezes unused credit in a timely manner, preventing overdue payments. The correlation risk identification model constructed by graph neural networks can effectively capture the risk transmission between enterprises and prevent chain defaults. Detailed Implementation
[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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.
[0042] Example:
[0043] A big data-based method for credit risk analysis of internet finance companies is characterized by the following steps:
[0044] Step 1: Perform stratified data collection:
[0045] This includes core financial data collection, scenario-based operational data collection, and related party data collection;
[0046] Core financial data includes corporate credit history, guarantee information, and financial data obtained through compliant interfaces connected to the People's Bank of China Credit Reference Center, Baihang Credit Reference Center, and corporate credit reporting agencies. Specifically, this includes loan balance, repayment history, number of overdue payments, guarantee amount, credit status of the guaranteed company, balance sheet, profit and loss statement, and cash flow statement.
[0047] Contextualized operational data includes dynamic operational data collected from enterprises in internet scenarios, such as data from cooperation with e-commerce platforms, enterprise purchase orders, supplier information, logistics data, enterprise business registration information, tax data, social security payment records, and intellectual property data. These data reflect the enterprise's market competitiveness and operational stability, provide a basis for analyzing the enterprise's supply chain coordination capabilities and cash flow, and provide a basis for risk assessment in non-financial dimensions.
[0048] Related party data includes credit data of corporate related parties. It obtains information on shareholders' capital contributions and the operating status of related companies through the corporate credit information disclosure system. It also analyzes social networks to identify whether there are any records of dishonesty in the personal credit records of the actual controller of the enterprise and whether there are any defaults in related companies, thus identifying the transmission path of related risks.
[0049] This also includes cleaning and verifying the collected data, data anonymization and encryption, and data standardization and integration:
[0050] Data cleaning and verification after collection: Data cleaning is performed using a combination of rule engine and machine learning. The rule engine is used to handle data with inconsistent formats and delete duplicate data. Machine learning algorithms such as isolated forest and autoencoder are used to identify abnormal data such as abnormal fluctuations in corporate financial data and false transaction records. The autoencoder is used to build a feature model of normal transaction data. Transaction data that deviates from the model is marked as suspicious data and triggers manual review.
[0051] Data anonymization and encryption: Following the principle of minimum necessity, sensitive data such as corporate trade secrets and shareholder ID numbers are anonymized. Corporate bank account numbers are formatted with asterisks added between the first 6 and last 4 digits, and shareholder ID numbers are partially hidden. The AES-256 encryption algorithm is used to encrypt the data transmission and storage process to prevent data leakage.
[0052] Data Standardization and Integration: Construct a unified data standard system to convert multi-source heterogeneous data into structured features, adopt standard classification for enterprise industry classification, normalize operating data, and convert revenue data of enterprises of different sizes into relative indicators such as revenue growth rate and revenue profit margin; integrate data scattered across different platforms through data lake technology to form a unified credit data view for enterprises, laying the foundation for subsequent analysis.
[0053] By integrating multi-source data and multi-level models, we comprehensively characterize corporate credit risk, effectively identify risk points missed by traditional methods, significantly reduce bad debt losses, and effectively solve the problems of inconsistent data formats and a lot of false data through big data cleaning and verification mechanisms. The data missing rate is reduced and the false data identification rate is improved, providing reliable data support for risk analysis.
[0054] Step Two: Constructing Big Data Features and Risk Features:
[0055] The collected data is used to extract the enterprise's financial health characteristics, operational capability characteristics, credit performance characteristics, related risk characteristics, and public opinion risk characteristics.
[0056] Financial health characteristics: Based on corporate financial data, traditional financial indicators such as debt-to-equity ratio, current ratio, quick ratio, revenue growth rate, and net profit margin are extracted. At the same time, data on net cash flow from operating activities and short-term debt are introduced as cash flow health indicators, and the consistency between financial statement data and tax declaration data is introduced as a new characteristic of financial data authenticity, so as to comprehensively assess the financial status of enterprises.
[0057] Operational capability characteristics: Extract order volume growth rate, customer retention rate, supplier stability reflected by the length of cooperation with core suppliers, and logistics efficiency reflected by on-time delivery rate from scenario-based operational data to reflect the company's actual operational capabilities;
[0058] Credit performance characteristics: By combining the company's credit history and non-financial performance data, we extract characteristics such as the number of overdue payments, the number of overdue days, the performance rate, and whether there is a record of being a dishonest person subject to enforcement, and assess the company's willingness to perform its credit obligations.
[0059] Associated risk characteristics: By using graph structure data such as enterprise relationship diagrams and supply chain networks, we can extract characteristics such as the number of associated enterprises, the default rate of associated enterprises, and the scale of intra-group fund transactions to identify associated risks;
[0060] Public opinion risk characteristics: Collect public opinion information of enterprises in news media, social platforms, and industry forums, and use natural language processing technology to extract public opinion sentiment, negative public opinion keywords, and public opinion dissemination scope characteristics to assess the impact of public opinion on enterprise credit.
[0061] It also includes using feature correlation analysis, such as Pearson correlation coefficient, to remove redundant features. When the correlation coefficient between the debt-to-equity ratio and the current ratio exceeds 0.8, the more informative debt-to-equity ratio is retained. Key risk features are screened through feature importance assessment using random forest to ensure the effectiveness of the input features to the model and reduce computational costs.
[0062] Step 3: Construct a multi-level dynamic credit risk assessment model:
[0063] The project involves constructing a basic credit scoring model, a dynamic risk iteration model, a related risk identification model, and a public opinion risk assessment model for enterprises.
[0064] Basic Credit Scoring Model: Based on the XGBoost algorithm, a basic credit scoring model is constructed using core financial data and static features of enterprises. It outputs a basic credit score of 0-1000 for enterprises. By learning the financial characteristics of enterprises with historical non-performing loans, the model can identify high-risk feature combinations such as a debt-to-equity ratio of over 80% and a current ratio of less than 1, and give low credit scores to enterprises that meet the combination.
[0065] Dynamic Risk Iteration Model: Introducing Long Short-Term Memory Network and Temporal Convolutional Network to perform time-series analysis on dynamic business data of enterprises, capturing real-time changes in the business situation of enterprises, and learning the time series trend of enterprise revenue through LSTM model. When the model predicts that the revenue will decline by more than 20% month-on-month in the next 3 months, the enterprise credit score will be automatically downgraded and the enterprise will be marked as a potential business risk enterprise.
[0066] Related Risk Identification Model: A graph neural network is used to construct a graph of enterprise relationships, where nodes represent enterprises and related parties, and edges represent relationship types, including equity relationships and supply chain relationships. The GraphSAGE algorithm is used to learn the characteristics of related nodes and to explore the transmission path of related risks. When an enterprise defaults on its credit, the model can automatically identify downstream enterprises that are related to it in the supply chain, assess the default risk of these enterprises, and trigger a related risk warning.
[0067] Public opinion risk assessment model: Based on the BERT pre-trained model, the model performs sentiment analysis and risk classification on corporate public opinion texts and outputs the public opinion risk level, including three levels: low, medium and high. When a company has negative public opinion about unpaid wages being complained about by employees and the spread covers multiple industry media, the model judges it as a high public opinion risk and incorporates this feature into the overall risk assessment.
[0068] This also includes establishing a real-time model update mechanism, specifically by collecting new business operation data and risk cases daily, incrementally training the model, and updating feature weights; and by optimizing the model structure and adjusting the associated risk identification algorithm monthly based on newly discovered risk patterns to ensure the model's adaptability to new risks.
[0069] Step 4: Analyze and alert on corporate credit risks in real time using a multi-level dynamic credit risk assessment model.
[0070] The beneficial effects of this invention are as follows: This big data-based method for analyzing credit risk in internet finance enterprises comprehensively characterizes enterprise credit risk through multi-source data fusion and multi-level models, effectively identifies risk points missed by traditional methods, significantly reduces bad debt losses, and effectively solves the problems of inconsistent data formats and numerous false data through big data cleaning and verification mechanisms. The data missing rate is reduced, and the false data identification rate is improved, providing reliable data support for risk analysis. The real-time data collection and model update mechanism upgrades risk monitoring from static one-time assessment to dynamic continuous monitoring, identifies risks in advance, and freezes unused credit in a timely manner to avoid overdue payments.
[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A big data-based credit risk analysis method for an Internet financial enterprise, characterized in that, Comprise the following steps: Step one: stratified data collection: Including core financial data collection, scenario-based business data collection, associated party data collection; Step two: build big data features and risk features: Corresponding to the collected data of the enterprise's financial health characteristics, operating ability characteristics, credit performance characteristics, associated risk characteristics and public opinion risk characteristics; Step three: build a multi-level dynamic credit risk assessment model: The construction of the enterprise's basic credit scoring model, dynamic risk iteration model, associated risk identification model and public opinion risk assessment model; Step four: real-time analysis and warning of enterprise credit risk through multi-level dynamic credit risk assessment model. 2.The big data-based Internet financial enterprise credit investigation risk analysis method according to claim 1, characterized in that, In step one: The core financial data includes the enterprise credit history, guarantee information and financial data obtained by connecting the central bank credit center, Baihang credit and enterprise credit institutions through compliance interface, specifically loan balance, repayment record, overdue times, guarantee amount, credit status of guaranteed enterprises, balance sheet, profit table and cash flow table; Scenario-based business data includes collecting dynamic business data of enterprises in Internet scenarios, such as data cooperation with e-commerce platforms, enterprise procurement orders, supplier information and logistics data, enterprise business registration information, tax data, social security payment records and intellectual property data, which reflect the market competitiveness and operating stability of enterprises, analyze the supply chain coordination ability and fund flow status of enterprises, and provide basis for risk assessment of non-financial dimensions; The associated party data includes the credit data of the enterprise's associated parties, the investment situation of the shareholders and the operating status of the associated enterprises are obtained through the enterprise credit information public system, and whether the personal credit record of the actual controller of the enterprise exists in the social network analysis and mining of the enterprise and the associated enterprise default situation, and the associated risk transmission path is identified. 3.The big data-based Internet financial enterprise credit investigation risk analysis method according to claim 1, characterized in that, The step one also includes data cleaning and verification, data desensitization and encryption, and data standardization and integration after data collection.
4. The big data-based Internet financial enterprise credit investigation risk analysis method according to claim 3, characterized in that, The cleaning and verification of the collected data: rule engine and machine learning are combined to clean the data; Rule engine is used to process data with inconsistent format and delete duplicate data; Machine learning algorithms such as isolation forest and autoencoder are used to identify abnormal data such as abnormal fluctuations in enterprise financial data and false transaction records, and the feature model of normal transaction data is constructed through autoencoder, and the transaction data deviating from the model is marked as suspicious data and triggers manual review; Data desensitization and encryption: follow the principle of minimum necessity, desensitize sensitive data such as enterprise business secrets and shareholder ID number, use the format of adding stars in the middle of the first 6 and last 4 of the enterprise bank account number, and use part of the field to hide the shareholder ID number; AES-256 encryption algorithm is used to encrypt the data transmission and storage process to prevent data leakage. 5.The big data-based Internet financial enterprise credit investigation risk analysis method according to claim 4, characterized in that, The data standardization and integration: construct a unified data standard system, convert multi-source heterogeneous data into structured features, use standard classification for enterprise industry classification, normalize operating data, and convert operating data of enterprises of different sizes into relative indicators such as operating income growth rate and operating profit rate; integrate the data scattered in different platforms through data lake technology to form a unified credit investigation data view of the enterprise, laying a foundation for subsequent analysis. 6.The big data-based Internet financial enterprise credit investigation risk analysis method according to claim 1, characterized in that, The step two also includes removing redundant features by using feature correlation analysis such as Pearson correlation coefficient, when the correlation coefficient of asset-liability ratio and liquidity ratio exceeds 0.8, the asset-liability ratio with more information is retained; key risk features are screened through feature importance evaluation of random forest feature importance, ensuring the effectiveness of model input features and reducing calculation cost; Financial health features: based on enterprise financial data, traditional financial indicators such as asset-liability ratio, liquidity ratio, quick ratio, operating income growth rate, and net profit rate are extracted, and data of net cash flow from operating activities and short-term debt are introduced as new features of cash flow health and consistency of financial report data and tax return data as financial data authenticity, to comprehensively evaluate the financial condition of the enterprise; Operating ability features: order quantity growth rate, customer retention rate, supplier stability reflected by core supplier cooperation time, and logistics efficiency features reflected by on-time delivery rate are extracted from scenario-based operating data to reflect the actual operating ability of the enterprise; Credit performance features: combining enterprise credit history records and non-financial performance data, extracting overdue times, overdue days, performance rate, and whether there is a record of dishonesty enforcement person features to evaluate the credit performance willingness of the enterprise; Correlation risk features: through enterprise correlation graph and graph structure data of supply chain network, extracting correlation enterprise quantity, correlation enterprise default rate, and group internal fund transfer scale features to identify correlation risks; Public opinion risk features: collecting public opinion information of enterprises in news media, social platforms, and industry forums, extracting public opinion sentiment tendency, negative public opinion keywords, and public opinion spread range features through natural language processing technology to evaluate the impact of public opinion on enterprise credit. 7.The big data-based Internet financial enterprise credit investigation risk analysis method according to claim 1, characterized in that, The step three also includes establishing a real-time model updating mechanism, specifically collecting new enterprise operating data and risk cases daily, incrementally training the model, and updating feature weights; Based on newly discovered risk patterns, optimize the model structure and adjust the correlation risk identification algorithm every month to ensure the adaptability of the model to new risks; Basic credit scoring model: based on XGBoost algorithm, using enterprise core financial data and static features, a basic credit scoring model is constructed to output enterprise basic credit score 0-1000, through learning the financial features of historical bad loan enterprises, the model can identify high-risk feature combinations such as asset-liability ratio exceeding 80% + liquidity ratio less than 1, and give low credit score to enterprises meeting the combination; Dynamic risk iteration model: Introduce long short-term memory network and time series convolution network to conduct time series analysis on enterprise dynamic operation data, capture real-time changes in enterprise operating conditions, learn time series trends of enterprise revenue through LSTM model, and automatically reduce enterprise credit score when model predicts that revenue in next three months will decrease by more than 20%, and mark it as potential operating risk enterprise; Correlation risk identification model: Use graph neural network to construct enterprise correlation graph, whose nodes are enterprises and associated parties, and edges are correlation types, including equity correlation and supply chain correlation. Learn the features of associated nodes through GraphSAGE algorithm to mine the transmission path of associated risks. When an enterprise defaults on credit, the model can automatically identify downstream enterprises with supply chain correlation and assess the default risk of these enterprises, and trigger a correlation risk warning; Public opinion risk assessment model: Based on BERT pre-training model, conduct sentiment analysis and risk classification on enterprise public opinion text, output public opinion risk level, including low, medium and high levels. When an enterprise has negative public opinion of employee complaints about unpaid wages, and the spread covers multiple industry media, the model determines it as high public opinion risk, and includes this feature in the overall risk assessment.