Construction method of credit risk assessment model and credit risk assessment model

Abstract

The invention discloses a method for constructing a small and medium-sized enterprise credit risk assessment model based on production and fusion big data. The method comprises the steps of obtaining credit risk assessment multi-source data of small and medium-sized enterprises; performing standardization, normalization and dimension reduction processing on the credit risk assessment multi-source data to obtain a second feature set; respectively training and integrating three credit risk assessment models by using the second feature set to obtain a comprehensive credit scoring model for multi-dimensional accurate assessment of enterprise credit risks; and training a dynamic monitoring early warning model by using the real-time risk assessment data so as to classify the real-time credit risk of the enterprise. According to the method, a dynamic risk monitoring mechanism is designed by introducing multi-source heterogeneous data, so that the problems of single data source, lagging analysis, low precision, lack of an early warning mechanism and the like in a traditional credit risk assessment model for middle and small-sized enterprises are comprehensively solved.

Description

Technical Field

[0001] The present application belongs to the technical field of credit risk assessment for small and medium-sized enterprises, and specifically relates to a method for constructing a credit risk assessment model and a credit risk assessment model. Background Art

[0002] Existing credit risk assessment models for small and medium-sized enterprises mainly rely on the company's historical financial data, such as balance sheets, income statements, and cash flow statements. These data are relatively effective in assessing the company's past performance, but they cannot reflect the company's market dynamics, operating status, and future potential. At the same time, companies often adjust financial data for tax avoidance or other purposes, making it difficult for these data to truly reflect the company's actual operating conditions. The singleness of this data dimension and insufficient information coverage make it difficult for credit risk assessment models to comprehensively and accurately predict the company's credit risk.

[0003] Traditional credit risk assessment methods usually rely on periodically updated data, such as annual financial reports, quarterly financial reports, etc., which makes the model's assessment results lag behind. In the current rapidly changing market environment, the business conditions and credit risks of enterprises may change significantly in the short term, and existing models cannot respond to these changes in a timely manner. Due to the lack of real-time data processing capabilities, the lag in risk monitoring of existing technologies has become a major shortcoming in the credit risk assessment of small and medium-sized enterprises.

[0004] Existing credit risk assessment models mostly use traditional statistical methods such as linear regression and scoring cards. These methods perform well when dealing with simple, linear relationships, but are insufficient when facing complex nonlinear relationships. Credit risk assessment of small and medium-sized enterprises involves data interactions in multiple dimensions, such as market dynamics, social public opinion, and industrial chain relationships. These complex relationships are difficult to effectively capture and quantify through traditional models. In addition, due to data sparsity and the presence of noisy data, traditional methods perform poorly when dealing with high-dimensional data, resulting in low model robustness and prediction accuracy.

[0005] With the development of big data technology, more and more heterogeneous data sources can be used for corporate credit risk assessment, including social media data, e-commerce transaction data, government data (such as tax, water and electricity, social security, etc.). However, when integrating and processing these heterogeneous data, existing technologies face challenges in data standardization, feature selection, noise processing, etc. Due to insufficient data integration capabilities, existing models often find it difficult to effectively utilize these rich data sources, thus failing to improve the accuracy and comprehensiveness of assessment results.

[0006] Existing technologies usually conduct credit risk assessments through fixed periodic updates, lacking real-time monitoring and dynamic early warning functions. This means that when corporate credit risk changes, the model cannot issue early warning signals in a timely manner, causing financial institutions to face delayed information when making decisions. This lack of dynamic monitoring and early warning mechanisms has greatly limited the effectiveness of credit risk management, especially in response to rapid market changes.

[0007] Application Contents

[0008] In order to solve the problems existing in the prior art, the purpose of this application is to provide a credit risk assessment model construction method and a credit risk assessment model, covering multi-source data processing, feature engineering, assessment model construction and optimization, dynamic monitoring and credit scoring system construction, which can conduct in-depth analysis of credit risk changes of small and medium-sized enterprises and implement effective risk prediction and continuous monitoring.

[0009] Specifically, this application involves the following aspects:

[0010] The construction method of the credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data includes:

[0011] Obtain a multi-source data set for credit risk assessment of SMEs containing multiple variables, which includes government data sets, industry data sets, Internet data sets, and data sets owned by SMEs;

[0012] Extracting features of multiple variables in a credit risk assessment multi-source data set to obtain a first feature set; and obtaining a second feature set by screening multiple features in the first feature set;

[0013] Dividing the second feature set into a first feature subset, a second feature subset, and a third feature subset to respectively train a first assessment model, a second assessment model, and a third assessment model; integrating the first assessment model, the second assessment model, and the third assessment model to obtain a comprehensive credit scoring model to perform a multi-dimensional assessment of the credit risk of small and medium-sized enterprises; and

[0014] Establish data connection with third-party data sources to obtain real-time risk assessment data sets for SMEs; use real-time risk assessment data sets to train dynamic monitoring and early warning models to assess the real-time credit risks of SMEs.

[0015] Among them, the first feature subset includes the recent operating status characteristics of SMEs, the second feature subset includes the market competitiveness characteristics of SMEs, and the third feature subset includes the main credit risk characteristics of SMEs.

[0016] 2. The method according to claim 1, wherein

[0017] The government data set includes: business registration information variables, tax payment record variables, water and electricity payment variables, social security payment variables, environmental compliance variables, and / or administrative penalty variables;

[0018] The industry data set includes: industry trend prosperity index variables, market share variables, upstream and downstream supply chain variables, product price index variables and / or industry technology innovation indicator variables;

[0019] The Internet dataset includes: social media public opinion variables, e-commerce platform transaction variables, platform evaluation variables, news media coverage variables, recruitment information analysis variables, corporate official website push variables and / or corporate public account push variables; and

[0020] The SMEs’ ​​own data sets include: financial statement variables, SME internal sales variables, customer relationship management variables, internal management change variables, and / or R&D investment variables.

[0021] 3. The method according to claim 1, wherein

[0022] Acquiring a multi-source data set for credit risk assessment of small and medium-sized enterprises further includes: preprocessing a plurality of variables in the multi-source data set for credit risk assessment; wherein the preprocessing includes:

[0023] Calculating multiple statistical characteristic variables of multiple variables to incorporate into the multiple variables;

[0024] Fill missing values ​​in multiple variables;

[0025] Remove outliers from multiple variables; and

[0026] Merge replicates across multiple variables.

[0027] 4. A method according to claim 3, wherein

[0028] Calculate multiple statistical features of multiple variables including:

[0029] Compute measures of center, dispersion, distribution, time series, and / or location for multiple variables.

[0030] 5. A method according to item 3 or 4, wherein

[0031] Filling missing values ​​in multiple variables includes:

[0032] calculating a first credit assessment quality value or a second credit assessment quality value for each variable of a plurality of variables;

[0033] In response to a variable whose first credit assessment quality value or second credit assessment quality value among the multiple variables is less than a first threshold or a second threshold, a first interpolation variable or a second interpolation variable of the multiple variables is obtained, and the variable whose first credit assessment quality value is less than the first threshold is replaced by the first interpolation variable or the variable whose second credit assessment quality value is less than the second threshold is replaced by the second interpolation variable.

[0034] 6. A method according to item 3 or 4, wherein

[0035] Eliminating outliers in multiple variables includes:

[0036] Calculate the mean and standard deviation of multiple variables, and obtain the abnormal distribution interval based on the mean and standard deviation of multiple variables;

[0037] In response to a variable in the plurality of variables being located in an abnormal interval, a% quantiles of the plurality of variables are obtained, and the variables in the abnormal distribution interval are replaced with the a% quantiles.

[0038] 7. A method according to any one of items 1 to 4, wherein

[0039] Feature extraction of multiple variables in multi-source data sets for credit risk assessment includes:

[0040] Sort multiple variables based on their magnitude;

[0041] Get the first and last variables in a plurality of variables; and

[0042] A standardized value of each of the plurality of variables is calculated based on the first variable and the last variable to replace the value of each of the plurality of variables.

[0043] 8. A method according to any one of items 1 to 4, wherein

[0044] Feature extraction of multiple variables in multi-source data sets for credit risk assessment includes:

[0045] Replace the value of each of multiple variables by taking the logarithm of the value of each of multiple variables.

[0046] 9. A method according to any one of items 1 to 4, wherein

[0047] Feature extraction of multiple variables in multi-source data sets for credit risk assessment includes:

[0048] Calculate means and standard deviations for multiple variables;

[0049] A standard deviation standardized value of each of the plurality of variables is calculated based on a mean value and a standard deviation of the plurality of variables to replace a value of each of the plurality of variables.

[0050] 10. The method according to any one of items 1 to 4, wherein

[0051] The first feature set includes: basic financial features, industrial chain features, transaction behavior features, text data features and key indicator time series features, among which:

[0052] Basic financial characteristics include: profitability indicators, debt-paying ability indicators, operating efficiency indicators and / or growth indicators;

[0053] Industrial chain characteristics include: upstream concentration indicators, downstream customer stability indicators and / or supply chain health indicators;

[0054] Transaction behavior characteristics include: cash flow volatility indicators, abnormal transaction frequency indicators and / or large transaction indicators;

[0055] Text data features include: sentiment scores, TF-IDF keywords and / or topic information; and

[0056] The key indicator time series characteristics include: key indicator moving average, key indicator year-on-year growth rate, key indicator month-on-month growth rate and / or key indicator forecast trend.

[0057] 11. The method according to any one of items 1 to 4, wherein

[0058] The second feature set obtained by screening multiple features in the first feature set includes:

[0059] Calculate the absolute value of the first credit risk correlation coefficient between each feature of the multiple features and the remaining features respectively;

[0060] Deleting features whose absolute values ​​of the first credit risk correlation coefficients are less than a third threshold value to obtain a plurality of relevant feature sets relative to each feature;

[0061] calculating an absolute value of a second credit risk correlation coefficient between a feature of each of the plurality of relevant feature sets and the target variable; and

[0062] The features with smaller absolute values ​​of the second credit risk correlation coefficients in each relevant feature set are deleted, and multiple relevant feature sets are merged.

[0063] 12. A method according to any one of items 1 to 4, wherein

[0064] The second feature set obtained by screening multiple features in the first feature set includes:

[0065] Calculating a credit risk information value for each of a plurality of features;

[0066] The features whose credit risk information values ​​are less than a fourth threshold value are deleted from the plurality of features.

[0067] 13. A method according to any one of items 1 to 4, wherein

[0068] The second feature set is obtained by screening multiple features in the first feature set, including the following steps:

[0069] Step 1: Input multiple features into a random forest model with multiple decision trees;

[0070] Step 2: Calculate the credit risk importance value of each of the multiple features of each decision tree in the multiple decision trees;

[0071] Step 3: Calculate the average credit risk importance value of the credit risk importance values ​​of each feature in the multiple features of the multiple decision trees, and sort the multiple features based on the size of the average credit risk importance value;

[0072] Step 4: Delete the last a% of features from the multiple features and calculate the performance of the random forest model on the multiple features; and

[0073] Repeat steps 2 to 4 until the performance of the random forest model on the multiple features is no longer improved and / or the number of features of the multiple features is equal to or less than the preset number of features.

[0074] 14. The method according to claim 7, wherein

[0075] The second feature set obtained by screening multiple features in the first feature set includes:

[0076] Obtaining an initial feature matrix of multiple features;

[0077] Calculate the covariance submatrix of each feature in the feature matrix and the remaining features respectively;

[0078] Merging the covariance sub-matrices to obtain a covariance matrix of multiple features, and calculating credit assessment explanatory values ​​of multiple eigenvalues ​​in the covariance matrix;

[0079] sorting a plurality of corresponding features of a plurality of feature values ​​based on the magnitude of the credit assessment explanatory value;

[0080] Calculating a cumulative credit assessment explanatory value of a credit assessment explanatory value of a feature value corresponding to a first corresponding feature among the plurality of corresponding features and a credit assessment explanatory value of a feature value corresponding to a second corresponding feature among the plurality of corresponding features;

[0081] In response to the cumulative credit assessment explanatory value being less than a fifth threshold, the cumulative credit assessment explanatory value is sequentially added to the credit assessment explanatory values ​​of the feature values ​​corresponding to the remaining corresponding features in the plurality of corresponding features until the cumulative credit assessment explanatory value is greater than or equal to the fifth threshold;

[0082] Delete the features in the feature matrix whose credit assessment explanatory values ​​do not add up to the cumulative credit assessment explanatory values.

[0083] 15. A method according to claim 14, wherein

[0084] The second feature set obtained by screening multiple features in the first feature set includes:

[0085] Setting a first weight value for each of the plurality of features in the second feature set;

[0086] Setting second weight values ​​of the plurality of key features; and

[0087] Weighting each of the multiple features in the second feature set and each of the multiple key features based on the first weight value and the second weight value to obtain a weighted feature;

[0088] Among them, the key features include the features with high prediction stability and high prediction explanatory power for credit risk prediction of small and medium-sized enterprises obtained through screening in the first feature set.

[0089] 16. A method according to claim 1, wherein

[0090] The features in the first feature subset include: operating income growth rate, net profit margin, total assets, asset turnover, total liabilities, inventory turnover, accounts receivable turnover, cash flow from operating activities / operating income ratio, industry average growth rate, average daily transaction volume, transaction frequency, large transaction ratio, cross-border transaction record, industry average growth rate, market concentration and / or raw material price index;

[0091] The features in the second feature subset include: enterprise market share, market share growth rate, brand awareness, number of patent applications, number of patent authorizations, number of patent citations, patent family size, patent quality score, customer growth rate, customer retention rate, customer quality score, key customer contribution rate, core technical personnel ratio, employee education composition ratio, talent loss rate, R&D personnel ratio, new product revenue ratio and / or R&D expense / operating revenue ratio; and

[0092] The features in the third feature subset include: personal credit score, debt default record, litigation record, loan history record, number of litigation cases, type of litigation cases, judgment results of litigation cases, information on persons subject to execution, number of affiliated companies, operating conditions of affiliated companies, complexity of equity relations, public opinion index, positive / negative evaluation ratio, topic popularity, frequency of large-scale fund transactions, abnormal consumption pattern index and / or cross-border fund flow records.

[0093] 17. A method according to item 1 or 16, wherein

[0094] The first evaluation model is used to evaluate the business capabilities of small and medium-sized enterprises, which includes: ARIMA model and EWMA model, among which,

[0095] The autoregressive coefficient of the ARIMA model ranges from 0 to 2;

[0096] The difference order of the ARIMA model ranges from 0 to 1;

[0097] The moving average order of the ARIMA model ranges from 2 to 3; and

[0098] The smoothing coefficient of the EWMA model ranges from 0.1 to 0.3.

[0099] 18. A method according to claim 16, wherein

[0100] The features in the first feature subset have weight values ​​in the first evaluation model, where

[0101] The weight value of operating income growth rate is 0.25;

[0102] The weight value of net profit margin is 0.20;

[0103] The weight value of asset turnover is 0.15;

[0104] The weighting value of the ratio of cash flow from operating activities to operating income is 0.25; and

[0105] The total weight value of inventory turnover rate, accounts receivable turnover rate and industry average growth rate is 0.15.

[0106] 19. A method according to claim 16, wherein

[0107] The second evaluation model is used to evaluate the market competitiveness of small and medium-sized enterprises.

[0108] The features in the second feature subset have weight values ​​in the second evaluation model, where

[0109] The total weight value of market share and market share growth rate is 0.30;

[0110] The total weight of patent quantity and patent quality score is 0.25;

[0111] The total weight of customer growth rate and customer quality score is 0.20;

[0112] The weight value of the proportion of core technical personnel is 0.15; and

[0113] The total weight value of the proportion of new product revenue and the ratio of R&D expenses to operating income is 0.10.

[0114] 20. A method according to item 1 or 16, wherein

[0115] The third assessment model is used for the credit risk assessment of small and medium-sized enterprises, which includes: risk knowledge graph, PageRank model and GBDT model, among which,

[0116] The direct association weight of the risk knowledge graph ranges from 0 to 10, the second-degree association weight ranges from 0 to 2, and the third-degree association weight ranges from 0 to 1;

[0117] The damping coefficient of the PageRank model ranges from 0.5 to 1;

[0118] The number of iterations of the PageRank model ranges from 50 to 200;

[0119] The number of decision trees in the GBDT model ranges from 200 to 500;

[0120] The maximum depth of the GBDT model ranges from 0 to 10;

[0121] The learning rate of the GBDT model ranges from 0.05 to 0.1; and

[0122] The sampling ratio of the GBDT model ranges from 0.5 to 2.

[0123] 21. The method according to claim 1, wherein

[0124] Integrating the first assessment model, the second assessment model and the third assessment model to obtain a comprehensive credit scoring model includes:

[0125] Setting a first output weight of the first evaluation model, a second output weight of the second evaluation model, and a third output weight of the third evaluation model;

[0126] The first assessment model, the second assessment model and the third assessment model are integrated into a comprehensive credit scoring model based on the first output weight, the second output weight and the third output weight in an integration method.

[0127] 22. The method according to claim 1, wherein

[0128] Real-time risk assessment datasets include:

[0129] Short-term liquidity indicator dataset, business activity indicator dataset, credit behavior indicator dataset, market and public opinion indicator dataset, industrial chain indicator dataset, macroeconomic indicator dataset and / or abnormal transaction indicator dataset, among which:

[0130] The real-time variables in the short-term liquidity indicator dataset include: the daily average bank deposit balance change rate, the cash inflow and outflow ratio in the past seven days, and / or the overdue proportion of accounts payable;

[0131] The real-time variables in the operating activity indicator data set include: daily operating income change rate, major customer order volume change rate and / or capacity utilization rate;

[0132] The real-time variables in the credit behavior indicator dataset include: bank loan repayment timeliness rate, supplier payment timeliness rate, and / or tax payment timeliness rate;

[0133] Real-time variables in the Market and Sentiment Indicators dataset include: stock price volatility, number of negative news reports, and / or social media sentiment scores;

[0134] The real-time variables in the industrial chain indicator data set include: the on-time delivery rate of major suppliers, the on-time payment rate of core customers, the stock prices of upstream and downstream companies and / or the changes in the credit ratings of upstream and downstream companies;

[0135] The real-time variables in the macroeconomic indicator dataset include: the rate of change of the industry prosperity index, the volatility of relevant commodity prices and / or the volatility of exchange rates; and

[0136] The real-time variables in the abnormal transaction indicator data set include: frequency of large-value fund transactions, abnormal cross-border fund flow index, and / or changes in the proportion of related-party transactions.

[0137] 23. A method according to item 1 or 22, wherein

[0138] Using real-time risk assessment data sets to train dynamic monitoring and early warning models to assess the real-time credit risks of small and medium-sized enterprises includes:

[0139] Setting a conditional dependency relationship between each of the multiple real-time variables in the real-time risk assessment data set and the remaining real-time variables;

[0140] calculating a conditional probability of each of the plurality of real-time variables;

[0141] Calculate the probability of occurrence of the real-time credit risk of the SMEs at multiple risk levels based on the conditional probability of the real-time variables related to the real-time credit risk of the SMEs; and

[0142] Obtain the risk level with the highest probability of occurrence of real-time credit risk for small and medium-sized enterprises.

[0143] 24. A system for constructing a credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data, which includes:

[0144] A multi-source data collection unit is used to obtain a multi-source data set for credit risk assessment of SMEs containing multiple variables. The multi-source data set for credit risk assessment includes a government data set, an industry data set, an Internet data set, and a data set owned by SMEs.

[0145] The data feature engineering unit extracts features from multiple variables in the credit risk assessment multi-source data set to obtain a first feature set; and obtains a second feature set by screening multiple features in the first feature set;

[0146] a multi-dimensional model building unit, dividing the second feature set into a first feature subset, a second feature subset, and a third feature subset to respectively train a first evaluation model, a second evaluation model, and a third evaluation model; integrating the first evaluation model, the second evaluation model, and the third evaluation model to obtain a comprehensive credit scoring model to perform a multi-dimensional evaluation of the credit risk of small and medium-sized enterprises; and

[0147] The dynamic monitoring and early warning model building unit establishes data connection with third-party data sources to obtain real-time risk assessment data sets for SMEs; the real-time risk assessment data sets are used to train the dynamic monitoring and early warning model to assess the real-time credit risks of SMEs.

[0148] 25. A system according to item 24, wherein the system executes the steps of the method for constructing a credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data in any one of items 1-23.

[0149] 26. An electronic device comprising:

[0150] Processor; and

[0151] A memory, in which computer program instructions are stored, and when the computer program instructions are executed by a processor, the processor executes a method for constructing a credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data, such as any one of items 1-23.

[0152] 27. A credit risk assessment method for small and medium-sized enterprises based on industrial and financial big data, which includes:

[0153] Obtain a multi-source data set for credit risk assessment of SMEs containing multiple variables, which includes government data sets, industry data sets, Internet data sets, and data sets owned by SMEs;

[0154] Perform feature extraction and feature screening on multiple variables in the credit risk assessment multi-source data set to obtain a risk assessment feature set;

[0155] Inputting multiple features in the risk assessment feature set into a comprehensive credit scoring model to assess the credit risk of SMEs; and

[0156] Establish data connection with third-party data sources to obtain real-time risk assessment data sets for SMEs containing multiple real-time variables; input multiple real-time variables in the real-time risk assessment data sets into a dynamic monitoring and early warning model to assess the real-time credit risk of SMEs.

[0157] 28. A method according to claim 27, wherein

[0158] The government data set is obtained by regularly crawling from the data interfaces of the tax bureau, social security bureau and industrial and commercial departments, and includes: business registration information variables, tax payment record variables, water and electricity payment variables, social security payment variables, environmental compliance variables and / or administrative penalty variables;

[0159] The industry data set is obtained regularly from industry associations and third-party research companies, and includes: industry trend prosperity index variables, market share variables, upstream and downstream supply chain variables, product price index variables and / or industry technology innovation indicator variables;

[0160] The Internet dataset uses crawler technology to regularly collect information about small and medium-sized enterprises from social media, news, and e-commerce platforms, including: social media public opinion variables, e-commerce platform transaction variables, platform evaluation variables, news media report variables, recruitment information analysis variables, enterprise official website push variables, and / or enterprise public account push variables; and

[0161] SMEs’ ​​own data sets are regularly exported through the SMEs’ ​​internal ERP system and CRM system, which include: financial statement variables, SMEs’ ​​internal sales variables, customer relationship management variables, internal management change variables and / or R&D investment variables.

[0162] 29. A method according to item 27 or 28, wherein

[0163] Feature extraction and feature screening of multiple variables in the credit risk assessment multi-source data set to obtain the risk assessment feature set includes:

[0164] Extracting an initial set of features from multiple variables;

[0165] The features in the initial feature set are screened for relevance, credit risk information value, credit risk importance value, principal component and feature fusion to obtain a risk assessment feature set.

[0166] 30. A method according to claim 29, wherein

[0167] The initial feature set includes: basic financial features, industry chain features, transaction behavior features, text data features, and key indicator time series features, among which:

[0168] Basic financial characteristics include: profitability indicators, debt-paying ability indicators, operating efficiency indicators and / or growth indicators;

[0169] Industry chain characteristics include: upstream concentration, downstream customer stability and / or supply chain health;

[0170] Transaction behavior characteristics include: cash flow volatility, abnormal transaction frequency index and / or large transaction ratio;

[0171] Text data features include: sentiment scores, TF-IDF keywords and / or topic information; and

[0172] The key indicator time series characteristics include: key indicator moving average, key indicator year-on-year growth rate, key indicator month-on-month growth rate and / or key indicator forecast trend.

[0173] 31. A method according to item 27 or 28, wherein

[0174] Inputting multiple features in the risk assessment feature set into the comprehensive credit scoring model to assess the credit risk of SMEs includes:

[0175] Input the operating income growth rate, net profit margin, asset turnover rate, inventory turnover rate, accounts receivable turnover rate, operating cash flow / operating income ratio and / or industry average growth rate into the comprehensive credit scoring model to obtain the enterprise operation score for the evaluation of the operating status of small and medium-sized enterprises;

[0176] Inputting market share, market share growth rate, number of patents, patent quality score, customer growth rate, customer quality score, proportion of core technical personnel, proportion of new product revenue and / or R&D expense / operating income ratio into the comprehensive credit scoring model to obtain the enterprise competitiveness index for the competitiveness assessment of small and medium-sized enterprises;

[0177] Inputting personal credit scores, debt default records, litigation records, number of affiliated enterprises, operating conditions of affiliated enterprises and / or frequency of large-amount fund transactions into the comprehensive credit scoring model to obtain the risk probability of the business owner for the main risk level assessment of SMEs; and

[0178] The comprehensive credit scoring model assigns weights to the enterprise operation score, enterprise competitiveness index and business owner risk probability and integrates them to obtain a comprehensive credit score for SMEs.

[0179] 32. A method according to claim 27, wherein

[0180] Real-time risk assessment datasets include:

[0181] Short-term liquidity indicator dataset, business activity indicator dataset, credit behavior indicator dataset, market and public opinion indicator dataset, industrial chain indicator dataset, macroeconomic indicator dataset and / or abnormal transaction indicator dataset, among which:

[0182] The real-time variables in the short-term liquidity indicator dataset include: the daily average bank deposit balance change rate, the cash inflow and outflow ratio in the past seven days, and / or the overdue proportion of accounts payable;

[0183] The real-time variables in the operating activity indicator data set include: daily operating income change rate, major customer order volume change rate and / or capacity utilization rate;

[0184] The real-time variables in the credit behavior indicator dataset include: bank loan repayment timeliness rate, supplier payment timeliness rate, and / or tax payment timeliness rate;

[0185] Real-time variables in the Market and Sentiment Indicators dataset include: stock price volatility, number of negative news reports, and / or social media sentiment scores;

[0186] The real-time variables in the industrial chain indicator data set include: the on-time delivery rate of major suppliers, the on-time payment rate of core customers, the stock prices of upstream and downstream companies and / or the changes in the credit ratings of upstream and downstream companies;

[0187] The real-time variables in the macroeconomic indicator dataset include: the rate of change of the industry prosperity index, the volatility of relevant commodity prices and / or the volatility of exchange rates; and

[0188] The real-time variables in the abnormal transaction indicator data set include: frequency of large-value fund transactions, abnormal cross-border fund flow index, and / or changes in the proportion of related-party transactions.

[0189] 33. A method according to item 27 or 32, wherein

[0190] Inputting multiple real-time variables in the real-time risk assessment data set into the dynamic monitoring and early warning model to assess the real-time credit risk of SMEs includes:

[0191] Multiple real-time variables in the real-time risk assessment data set are input into the dynamic monitoring and early warning model to obtain the risk level with the highest probability of occurrence of real-time credit risk of SMEs.

[0192] 34. A method according to claim 27, wherein

[0193] The comprehensive credit scoring model is a comprehensive credit scoring model for the construction method of the credit risk assessment model of small and medium-sized enterprises based on industrial and financial big data in item 1;

[0194] The dynamic monitoring and early warning model is a dynamic monitoring and early warning model for the construction method of the credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data in item 1.

[0195] 35. A credit risk assessment system for small and medium-sized enterprises based on industrial and financial big data, which includes:

[0196] A data collection unit, which obtains a multi-source data set for credit risk assessment of small and medium-sized enterprises containing multiple variables, wherein the multi-source data set for credit risk assessment includes a government data set, an industry data set, an Internet data set, and a data set owned by small and medium-sized enterprises;

[0197] A data processing unit performs feature extraction and feature screening on multiple variables in a multi-source data set for credit risk assessment to obtain a risk assessment feature set;

[0198] A credit risk assessment unit that inputs multiple features in the risk assessment feature set into a comprehensive credit scoring model to assess the credit risk of SMEs; and

[0199] The dynamic monitoring and early warning unit establishes data connection with a third-party data source to obtain a real-time risk assessment data set of SMEs containing multiple real-time variables; multiple real-time variables in the real-time risk assessment data set are input into a dynamic monitoring and early warning model to assess the real-time credit risk of SMEs.

[0200] 36. A system according to item 35, wherein the system executes the steps of the method for credit risk assessment of small and medium-sized enterprises based on industrial and financial big data of any one of items 27-33.

[0201] 37. An electronic device comprising:

[0202] Processor; and

[0203] A memory, in which computer program instructions are stored, and when the computer program instructions are executed by a processor, the processor executes a credit risk assessment method for small and medium-sized enterprises based on industrial and financial big data, such as any one of items 27-34.

[0204] This application achieves a comprehensive and accurate assessment of the credit risk of small and medium-sized enterprises by introducing multi-source data, real-time dynamic monitoring and advanced machine learning algorithms. Specifically, the advantages of this application over the existing technology are as follows:

[0205] 1. Improve data diversity and integration capabilities: Introduce government data, industrial data, and Internet data, and deeply integrate these data with machine learning algorithms to ensure that the model can fully cover the company's multi-dimensional business information.

[0206] 2. Real-time monitoring and dynamic early warning: Through real-time data collection and processing, a dynamic risk monitoring and early warning mechanism is established to ensure that financial institutions receive timely warnings before corporate risks occur, thereby improving the effectiveness of credit risk management.

[0207] 3. Processing of complex models and nonlinear relationships: Use advanced algorithms such as GBDT and Logistic regression to effectively process the interactive relationship of multidimensional data and improve the prediction accuracy and stability of the model.

[0208] 4. Heterogeneous data processing and feature engineering: Through standardized processing and feature selection of heterogeneous data, the model's adaptability to multi-source data is improved, the potential information in the data is fully utilized, and the comprehensiveness of credit risk assessment is enhanced.

[0209] This application can effectively overcome the shortcomings of existing prediction methods through the improvement and innovation of the above-mentioned technical solutions, and significantly improve the accuracy and practicality of credit risk assessment for small and medium-sized enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0210] Figure 1 The figure illustrates a functional schematic diagram of a method for assessing the credit risk of small and medium-sized enterprises based on industrial and financial big data according to an embodiment of the present application.

[0211] Figure 2 The figure shows a flow chart of constructing a comprehensive credit assessment model according to the method for constructing a credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data according to an embodiment of the present application;

[0212] Figure 3 The figure shows a flow chart of constructing a dynamic monitoring and early warning model for a credit risk assessment method for small and medium-sized enterprises based on industrial and financial big data according to an embodiment of the present application;

[0213] Figure 4 The figure shows a structural block diagram of a system for constructing a credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data according to an embodiment of the present application;

[0214] Figure 5 The figure illustrates a structural block diagram of a credit risk assessment system for small and medium-sized enterprises based on industrial and financial big data according to an embodiment of the present application. DETAILED DESCRIPTION

[0215] The present application is further described below in conjunction with examples. It should be understood that the examples are only used to further describe and illustrate the present application and are not used to limit the present application.

[0216] Unless otherwise defined, the technical and scientific terms in this specification have the same meaning as those generally understood by those skilled in the art. Although similar or identical methods and materials as described herein may be used in experiments or practical applications, materials and methods are described herein below. In the event of a conflict, the present specification including the definitions therein shall prevail. In addition, materials, methods and examples are provided for illustration only and are not restrictive. The present application is further described below in conjunction with specific embodiments, but is not intended to limit the scope of the present application.

[0217] The existing credit risk assessment models for small and medium-sized enterprises mainly rely on the historical financial data of enterprises, such as balance sheets, income statements and cash flow statements. These data are relatively effective in assessing the credit performance of enterprises in the past period of time, but they cannot reflect the performance of enterprises in more dimensions such as their market dynamics, operating conditions and future development potential; in addition, enterprises often adjust financial data for tax avoidance or other purposes, making it difficult for these data to truly reflect the actual operating conditions of enterprises. The singleness of this data dimension and insufficient information coverage make it difficult for credit risk assessment models to comprehensively and accurately predict the credit risks of enterprises.

[0218] Based on the purpose of data diversity assessment, this application proposes a method for constructing a credit risk assessment model for SMEs based on industrial and financial big data, and constructs a comprehensive credit scoring model for credit risk assessment of SMEs; and a credit risk assessment method for SMEs based on industrial and financial big data, which uses a comprehensive credit scoring model to assess the credit risk of SMEs to obtain a comprehensive credit score for SMEs, measures the business status of the enterprise from multiple angles, and thus more accurately reflects its credit risk.

[0219] The construction process of the comprehensive credit assessment model of the construction method of the small and medium-sized enterprise credit risk assessment model based on industrial and financial big data in this application is as follows Figure 2 As shown, it may include a multi-source data collection step S100, which is used to obtain a multi-source data set for SME credit risk assessment containing multiple variables, so as to mine multiple credit risk features for building a comprehensive credit scoring model in the multi-source data set for SME credit risk assessment.

[0220] Among them, the multi-source data set for SME credit risk assessment may include a variety of SME-related data sets obtained from a variety of different data sources, such as government data sets, industry data sets, Internet data sets and SME-owned data sets.

[0221] In some implementation schemes, the government affairs data set may be derived from corporate information publicly authorized by government departments, such as information obtained through API interfaces connected to the tax bureau, social security bureau, and industrial and commercial departments and captured regularly. It may include obtaining public corporate business registration information, tax payment records, water and electricity payment information, social security payment information, environmental protection compliance information, and / or administrative penalty information; the above information can be used as multiple variables in the government affairs data set and participate in feature extraction to continuously reflect the compliance and operational stability of the enterprise.

[0222] The industry data set can come from the enterprise industry data provided by industry associations and third-party market research companies. For example, a corporate data sharing mechanism can be established with industry associations and third-party research companies and the data can be updated regularly. The data may include industry trend prosperity index, market share information, upstream and downstream supply chain information, product price index and / or industry technology innovation indicators. The above information can be used as multiple variables in the industry data set and participate in feature extraction.

[0223] Internet data sets may include corporate information obtained through the Internet, which may include social media public opinion information, e-commerce platform transaction information, platform evaluation information, news media reporting information, recruitment information analysis, corporate official website push information and / or corporate public account push information; the above information can be used as multiple variables in the Internet data set and participate in feature extraction.

[0224] It should be noted that the method of obtaining corporate information through the Internet is to use Internet information automatic collection technology under the premise of authorization, such as crawler technology such as Srapy to continuously collect corporate information from designated social media, news reports and / or e-commerce platforms within a certain period of time.

[0225] SMEs' own data sets can be exported regularly through the enterprise's internal ERP system and / or CRM system, which may include financial statement information, SME internal sales information, customer relationship management information, internal management change information and / or R&D investment information; the above information can be used as multiple variables in the Internet data set and participate in feature extraction, carrying detailed information on the enterprise's various business operations and corresponding capital flows.

[0226] With the development of big data technology, various types of heterogeneous data sources can be used for credit risk assessment of small and medium-sized enterprises, including the above-mentioned social media data, e-commerce transaction data, government data, etc. However, when integrating and processing these heterogeneous data, existing technologies face challenges in data standardization, feature selection, noise processing, etc.; due to insufficient data integration capabilities, existing models often find it difficult to effectively utilize these rich data sources, thereby failing to improve the accuracy and comprehensiveness of assessment results.

[0227] Therefore, after obtaining the multi-source data set for credit risk assessment of small and medium-sized enterprises, the present application may further include preprocessing multiple variables in the multi-source data set for credit risk assessment, such as step S110, which is used to calculate the deep variables in the multi-source data set for credit risk assessment that have a significant impact on building the model, and reduce redundant information, remove noise information and balance the variable scale to improve the model accuracy and generalization ability.

[0228] Specifically, preprocessing multiple variables in a credit risk assessment multi-source data set may include obtaining statistical feature variables, calculating multiple statistical feature variables of the multiple variables, and incorporating the multiple statistical feature variables into the multiple variables. The multiple statistical feature variables may include central measurement variables, discrete measurement variables, distribution measurement variables, time series variables, and / or position measurement variables of the multiple variables.

[0229] It should be noted that statistical characteristic variables can be directly calculated based on the variable values ​​of numerical variables among multiple variables, such as financial statement information in SMEs’ ​​own data sets, internal sales information of SMEs, etc., which can provide an important basis for the model to analyze and learn the statistical characteristics of variables in multi-source data sets for SME credit risk assessment.

[0230] The central measurement variables of multiple variables may include the number, mean, median and / or mode of multiple variables; the discrete measurement variables of multiple variables may include the variance, standard deviation, interquartile range and / or coefficient of variation of multiple variables; the distribution measurement variables of multiple variables may include the skewness and / or kurtosis of multiple variables; the time series variables of multiple variables may include the trend, seasonality and / or autocorrelation of multiple variables; the location measurement variables of multiple variables may include the percentiles and / or Z-scores of multiple variables.

[0231] In particular, the time series variables of multiple variables may further include: trend variables, such as 3-month, 6-month and 12-month rolling averages, which can calculate the rate of change of the variable by its changing trend slope; seasonal variables, such as seasonal index, which is used to reflect the intensity of cyclical fluctuations to remove seasonal effects; autocorrelation variables, such as the 1-period lagged self-risk correlation coefficient, which can reflect short-term dependence, and the 2-period lagged self-risk correlation coefficient, which can reflect annual cyclicality; volatility variables, such as historical volatility, which can cover the standard deviation of variable values ​​over the past 30 years; conditional volatility, such as the volatility of variables estimated by the GARCH model; stationarity indicators, such as unit root test statistics, preferably the t-statistic of the ADF test, which can reflect the probability of stationarity of the variable; and trend stationarity indicators, such as the variance after detrending, which can reflect the stationarity trend of the variable.

[0232] In particular, after obtaining the Internet data set, it can also include preprocessing multiple variables in the Internet data set. The preprocessing includes using natural language technology to perform sentiment scores, keyword and hot topic extraction on multiple variables in the Internet data set to obtain multiple key Internet information variables and incorporate the multiple key Internet information variables into the multiple variables to improve the sensitivity of the variables in the Internet data set and provide more timely dynamic data support for the model.

[0233] In addition, statistical characteristic variables can also be indirectly calculated based on the number of categorical variables in multiple variables or the relationship between categorical variables. For example, the proportion of R&D personnel in an enterprise, the ratio of positive / negative reviews of the enterprise's products on e-commerce platforms, and the topic popularity of the products can be obtained through the ratio of this categorical variable to other categorical variables to measure the distribution characteristics of multiple categorical variables.

[0234] In some embodiments, preprocessing multiple variables in a credit risk assessment multi-source data set may also include missing value processing to fill in missing values ​​in the multiple variables. Specifically, the missing value processing may include the following steps:

[0235] Calculate a first credit assessment quality value or a second credit assessment quality value for each variable among a plurality of variables; in response to a variable among a plurality of variables whose first credit assessment quality value or second credit assessment quality value is less than a first threshold value or a second threshold value, obtain a first interpolation variable or a second interpolation variable of the plurality of variables, and replace the variable whose first credit assessment quality value is less than the first threshold value with the first interpolation variable or replace the variable whose second credit assessment quality value is less than the second threshold value with the second interpolation variable.

[0236] Specifically, for multiple numerical variables among multiple variables, the first credit assessment quality value of each numerical variable is calculated respectively. The first credit assessment quality value can be exemplified as a missing rate, which reflects the proportion of missing values ​​in multiple variables of a multi-source data set for credit risk assessment, and can measure the completeness, accuracy and availability of the multi-source data set for credit risk assessment; similarly, for multiple non-numeric variables among multiple variables, such as categorical variables, the second credit assessment quality value of each categorical variable can be calculated, and the second credit assessment quality value can also be exemplified as a missing rate.

[0237] In some embodiments, the first threshold of the first credit assessment quality value and the second threshold of the second credit assessment quality value can be set to 30% respectively to obtain missing values ​​in multiple variables while keeping the variable dimensions in the available interval. Further, the median of multiple numerical variables in the multiple variables can be obtained as the first interpolation variable and replace the missing values ​​of the first credit assessment quality value greater than the first threshold, and the mode of multiple non-numerical variables in the multiple variables, such as the categorical variable, can be obtained as the second interpolation variable and replace the missing values ​​of the second credit assessment quality value greater than the second threshold.

[0238] In some embodiments, preprocessing multiple variables in the credit risk assessment multi-source data set may also include outlier processing to remove abnormal data from the multiple variables. Specifically, outlier processing may include the following steps:

[0239] Calculate the mean and standard deviation of multiple variables, and obtain the abnormal distribution interval based on the mean and standard deviation of the multiple variables; obtain the centrality value of the multiple variables in response to the variables in the abnormal interval among the multiple variables, and replace the variables in the abnormal distribution interval with the centrality value. It should be noted that the outlier processing can be used for the numerical variables among the multiple variables.

[0240] Specifically, the abnormal distribution interval obtained based on the mean and standard deviation of multiple variables can be a non-intersecting interval of the normal distribution interval of multiple variables, so the abnormal distribution interval can be obtained by calculating the normal distribution interval of multiple variables. In some embodiments, the normal distribution interval of multiple variables can be set to the mean value of multiple variables ± 3 times the standard deviation, so as to replace the variables distributed outside the normal distribution interval with the concentration value of multiple variables.

[0241] The centralization value is used to describe the centralization trend of data distribution in a data set, and it needs to avoid the sensitive influence of extreme variables in multiple variables, such as noise, computational redundancy, etc. Therefore, the 95% quantile of multiple variables can be set as the centralization value, and the extreme values ​​outside the normal distribution interval can be replaced as outliers, so as to ensure that after replacing the outliers in the variables, the replaced variables meet the overall distribution of multiple variables, so as to correctly eliminate the extreme values.

[0242] In addition, in some embodiments, preprocessing multiple variables in a credit risk assessment multi-source data set may also include a consistency check for merging duplicate data in multiple variables. For example, a hash function may be used to identify duplicate variables in multiple variables, and the duplicate variables may be merged to reduce data redundancy. It should be noted that the preprocessing of multiple variables may not be limited to the above-mentioned methods, but the present application preferably uses these methods to optimize the quality of the variables to improve the learning efficiency of the model.

[0243] In order to ensure the comparability of multiple variables in the multi-source data set based on credit risk assessment at different numerical scales and avoid ignoring important features in subsequent feature extraction, this application also performs standardization on multiple variables in the multi-source data set for credit risk assessment before feature extraction. Among them, for variables with relatively uniform distribution, standardization can be performed through the following steps:

[0244] The plurality of variables are sorted based on their sizes; the first variable and the last variable among the plurality of variables are obtained; and the standardized value of each variable among the plurality of variables is calculated based on the first variable and the last variable to replace the value of each variable among the plurality of variables. In this way, the distribution of the corresponding standardized value obtained based on each variable remains unchanged, and the size of the standardized value is unified into the interval [0,1], thereby avoiding the scale influence between the variables.

[0245] In some implementation schemes, for variable types with outliers or a large number of outliers, standardization can also be performed in the following manner:

[0246] Take the logarithm of the value of each variable in multiple variables to replace the value of each variable in multiple variables. For example, for multiple variables related to asset size with a high probability of abnormal values, their standardized values ​​can be obtained by taking the logarithm of their values ​​to stabilize the variance of these variables and improve the linearity of the variables to make them more suitable for subsequent model learning.

[0247] In some implementation schemes, for variable types close to normal distribution, standardization can also be performed by the following steps:

[0248] Calculate the mean and standard deviation of multiple variables; calculate the standard deviation standardized value of each variable in the multiple variables based on the mean and standard deviation of the multiple variables to replace the value of each variable in the multiple variables. For variables with approximate normal distribution, such as variables related to enterprise sales, calculate their standard deviation standardized values ​​to approximate their values ​​to normal distribution, so as to facilitate variable processing and model learning.

[0249] This application extracts features from multiple variables in a multi-source dataset for SME credit risk assessment, and obtains multiple features related to credit risk from government data sets, industry data sets, Internet data sets, and SME-owned data sets to form a first feature set. Therefore, the first feature set contains a wealth of preliminary features for credit risk assessment, and the features in the first feature set can be further extracted to obtain key features.

[0250] Specifically, the first feature set may include the following categories: basic financial features, industrial chain features, transaction behavior features, text data features, and key indicator time series features. Among them, basic financial features may include: profitability indicators, debt repayment ability indicators, operating efficiency indicators and / or growth indicators; industrial chain features may include: upstream concentration indicators, downstream customer stability indicators and / or supply chain health indicators; transaction behavior features may include: cash flow volatility indicators, transaction frequency abnormality indicators and / or large transaction indicators; text data features may include: sentiment scores, TF-IDF keywords and / or topic information; and key indicator time series features may include: key indicator moving average, key indicator year-on-year growth rate, key indicator month-on-month growth rate and / or key indicator forecast trend.

[0251] Those skilled in the art can understand that the indicator of a feature can refer to one or more features under the category of features. For example, the profitability indicator of the basic financial feature can refer to multiple features belonging to profitability indicators such as the enterprise market growth rate, return on assets (ROA) and net profit margin, and the principal component obtained by dimensionality reduction of multiple features in different feature categories in the first feature set or further in different indicator categories may still belong to the feature category, or further belong to the characteristics of the indicator.

[0252] In some implementation schemes, basic financial characteristics are used to reflect the basic financial liquidity of small and medium-sized enterprises at this stage, where profitability indicators may include return on assets (ROA), return on equity (ROE), gross profit margin and / or net profit margin, etc.; debt-paying ability indicators may include current ratio, quick ratio and / or debt-to-asset ratio, etc.; operating efficiency indicators may include accounts receivable turnover rate and / or inventory turnover rate, etc.; growth indicators may include operating income growth rate and / or net profit growth rate, etc.

[0253] In some implementation plans, the characteristics of the industrial chain can reflect the comprehensive market competitiveness of the enterprise and, to a certain extent, reflect the recent financial income and expenditure status of the enterprise. Among them, the upstream concentration can be the Herfindahl-Hirschman Index (HHI); the downstream customer stability can be the proportion of customers who have repeated transactions in the past 12 months; the supply chain health can be calculated by the following formula:

[0254] SH=∑(ω i *f i ) / n

[0255] (Formula 1)

[0256] Among them, SH is the supply chain health, ω i is the weight of the ith supplier, f i is the financial health index of the supplier, and n is the total number of suppliers.

[0257] In some implementation schemes, the transaction behavior characteristics reflect the recent asset flow of small and medium-sized enterprises, where the cash flow volatility can be the coefficient of variation (CV) of cash inflows in the past 6 months; the transaction frequency anomaly index can be calculated by the following formula:

[0258] TAI=(T-μT) / σT

[0259] (Formula 2)

[0260] Among them, TAI is the transaction frequency anomaly index, T is the number of transactions in the current month, μT and σT are the average and standard deviation of the historical transaction numbers respectively; the proportion of large transactions can be the ratio of transactions exceeding twice the average transaction amount to the total transaction amount.

[0261] In some implementation schemes, text data features include relevant information of business platforms and users on the operating status and product quality of small and medium-sized enterprises, wherein the sentiment score can be obtained by utilizing a semantic sentiment analysis model, such as BERT, to extract and calculate sentiment from multiple variables in an Internet data set, and the text sentiment classification result can be further obtained based on the sentiment score, and the classification results include positive sentiment, neutral sentiment, and negative sentiment; TF-IDF keywords are keywords extracted from multiple variables in an Internet data set through a TF-IDF algorithm, and the TF-IDF keywords can be set to the 10 keywords with the highest TF-IDF values; and the topic information is the topic words extracted by performing topic modeling on multiple variables in an Internet data set through a Latent Dirichlet Allocation (LDA) model, and the topic information can be set to the 5 main topic words extracted by the LDA model.

[0262] In some embodiments, the key indicator time series characteristics are constructed from key indicators in multiple variables, where the key indicators may include, for example: operating income, net profit, net cash flow from operating activities, debt-to-asset ratio, accounts receivable turnover, inventory turnover, quick ratio, net cash flow from operating activities / current liabilities, interest coverage ratio and / or total asset turnover.

[0263] Furthermore, the moving average of the key indicator in the key indicator time series characteristics can be the moving average of the key indicator in the past 3 months, 6 months, and 12 months; based on these time nodes, the year-on-year growth rate of the key indicator and the month-on-month growth rate of the key indicator can be obtained; the key indicator can also be input into the ARIMA model to predict the key indicator change trend in the next 3 months, and the predicted key indicator change trend can be used as the key indicator prediction trend.

[0264] In response to the problem that existing models have difficulty in processing multi-source heterogeneous data, this application proposes to perform feature selection and dimensionality reduction on multi-source data before training a comprehensive credit assessment model. This can improve the model's adaptability to understanding the various feature information of multi-source data, make full use of the potential information in the data, enhance the comprehensiveness of credit risk assessment for small and medium-sized enterprises, and improve the breadth and accuracy of assessment.

[0265] Specifically, the present application selects multiple features in the first feature set through multiple dimensionality reduction methods to obtain a second feature set containing main features, so that the second feature set is suitable for input into the comprehensive credit assessment model. It can be understood that the first feature set preferably includes the features of the statistical feature variables obtained by analyzing and calculating the numerical variables and categorical variables, such as the asset-liability ratio, the proportion of large transactions, the patent quality score and / or the growth rate / change rate of different variables and other statistical features. These features can contain more complex statistical classification information than the features obtained directly to help improve the model performance. Among them, screening multiple features in the first feature set to obtain the second feature set may include the following steps:

[0266] The absolute value of the first credit risk correlation coefficient between each feature in the multiple features and the remaining features is calculated respectively; the features whose absolute value of the first credit risk correlation coefficient is less than a third threshold are deleted to obtain multiple related feature sets relative to each feature; the absolute value of the second credit risk correlation coefficient between the features of each related feature set in the multiple related feature sets and the target variable is calculated; and the features whose absolute value of the second credit risk correlation coefficient is smaller in each related feature set are deleted, and the multiple related feature sets are merged to obtain a second feature set.

[0267] Specifically, in order to retain features with greater correlation in the first feature set and delete features with less correlation, the first credit risk correlation coefficient of each feature pair in the first feature set may be calculated first, for example, the first credit risk correlation coefficient is set to be the Pearson credit risk correlation coefficient. Feature pairs whose absolute values ​​of the first credit risk correlation coefficients are not less than a third threshold are retained, and feature pairs whose absolute values ​​of the first credit risk correlation coefficients are less than the third threshold are deleted, and features with higher correlation with target variables, such as default probability and other target variables closely related to credit risk assessment, are retained in each retained feature pair.

[0268] In this way, the features screened out after two correlation analyses are relatively independent of each other and have a high positive / negative correlation with the target variable, which helps the model capture different aspects of the input features when training with the second feature set, so as to further reduce redundant information in massive financial data, improve the generalization ability of the model, and avoid changes in individual features causing large fluctuations in the model.

[0269] In some embodiments, the first credit risk correlation coefficient and the second credit risk correlation coefficient are preferably set to the Pearson credit risk correlation coefficient, which is applicable to the correlation between two linear features and can be used for discrete features, such as correlation analysis between categorical features; the third threshold is preferably set to 0.8.

[0270] In some embodiments, screening multiple features in the first feature set to obtain the second feature set may further include the following steps:

[0271] Calculate the credit risk information value of each feature among the multiple features; and delete the features whose credit risk information value is less than a fourth threshold among the multiple features.

[0272] Specifically, in order to measure the amount of information contained in each feature, the credit risk information value of the feature can be set to the variance of the feature. When the variance of the feature is large, its dispersion degree is high among similar features, so the uncertainty of the feature may be higher and retain more information. By calculating the variance of each feature and eliminating features with variances less than the fourth threshold, features with less information can be effectively removed; the fourth threshold is preferably 0.01.

[0273] In some embodiments, screening multiple features in the first feature set to obtain the second feature set may further include the following steps:

[0274] Input multiple features into a random forest model with multiple decision trees; calculate the credit risk importance value of each of the multiple features of each decision tree in the multiple decision trees; calculate the average credit risk importance value of the credit risk importance value of each of the multiple features of the multiple decision trees, and sort the multiple features based on the size of the average credit risk importance value; delete the last a% of the features in the multiple features, and calculate the performance of the random forest model on the multiple features; and repeat the above operations until the performance of the random forest model on the multiple features is no longer improved and / or the number of features of the multiple features is equal to or less than the preset number of features.

[0275] Specifically, the random forest method can be used to rank the importance of multiple high-dimensional features to use the importance to perform feature dimensionality reduction. It is preferred to build a random forest model containing 100 decision trees, and for each decision tree, calculate the credit risk importance value of each feature. Here, feature dimensionality reduction by credit risk importance value is to screen which of the multiple features contributes most to the prediction ability of the model during the training process. Therefore, the credit risk importance value can be set as the Gini credit risk importance value suitable for random forest prediction.

[0276] Next, the credit risk importance values ​​of all features of all decision trees are averaged and the features are sorted from large to small based on the credit risk importance value to remove non-important features.

[0277] Based on the ranking of feature credit risk importance values, feature deletion can be performed by a recursive method: the initial feature set is a plurality of features to be reduced in dimension, and a random forest model is trained on the current feature set and its performance is evaluated. The last a% of features with the lowest credit risk importance values ​​are removed, preferably a = 10. The steps of building a random forest model and removing the features with the lowest credit risk importance values ​​are repeated until the performance of the random forest model is no longer improved or the preset number of features is reached.

[0278] In some implementation schemes, the preferred number of preset features is 10, and the features with the highest credit risk importance values ​​obtained by sorting according to the final feature credit risk importance values ​​are as follows: operating cash flow / total debt ratio, quick ratio, accounts receivable turnover rate, net profit growth rate, debt-to-asset ratio, inventory turnover rate, operating income growth rate, upstream supplier concentration (HHI index), downstream customer stability index, and cash flow volatility (6-month coefficient of variation).

[0279] It should be noted that when screening features by screening multiple features in the first feature set to obtain the second feature set, one or more methods can be selected for screening. When multiple methods are selected, the order of feature screening can also be adjusted according to actual conditions. For example, it is possible to screen based on the amount of feature information through the corresponding method and then perform a secondary screening based on the feature credit risk importance value through the corresponding method, or it is possible to screen based on the feature credit risk importance value and then perform a secondary screening based on the amount of feature information through the corresponding method. This application does not limit the possible combinations of more screening methods.

[0280] In some embodiments, screening multiple features in the first feature set to obtain the second feature set may further include the following steps:

[0281] Obtain an initial feature matrix of multiple features; respectively calculate the covariance submatrix of each feature in the multiple features in the feature matrix with the remaining features; merge the covariance submatrices to obtain the covariance matrix of the multiple features, and calculate the credit assessment explanatory values ​​of the multiple eigenvalues ​​in the covariance matrix; sort the multiple corresponding features of the multiple eigenvalues ​​based on the size of the credit assessment explanatory values; calculate the cumulative credit assessment explanatory value of the credit assessment explanatory value of the eigenvalue corresponding to the first corresponding feature in the multiple corresponding features and the cumulative credit assessment explanatory value of the credit assessment explanatory value of the eigenvalue corresponding to the second corresponding feature in the multiple corresponding features; in response to the cumulative credit assessment explanatory value being less than a fifth threshold, sequentially add the cumulative credit assessment explanatory value to the credit assessment explanatory values ​​of the eigenvalues ​​corresponding to the remaining corresponding features in the multiple corresponding features until the cumulative credit assessment explanatory value is greater than or equal to the fifth threshold; delete the corresponding features of the eigenvalues ​​in the feature matrix whose credit assessment explanatory values ​​are not added to the cumulative credit assessment explanatory value.

[0282] It should be noted that multiple features in the first feature set or multiple features screened by any one or more of the screening methods proposed in this application can be further reduced in dimension to extract main features, such as extracting principal components to effectively reduce the learning cost and complexity of the model, and further filter noise and redundancy.

[0283] Specifically, all features to be reduced in dimension can be standardized and the covariance matrix of the features can be calculated. The eigenvalue of each feature is obtained to construct the corresponding eigenvector and calculate the credit assessment explanatory value. Since the credit assessment explanatory value can reflect the degree to which each feature contains the total amount of information contained in all features to be reduced in dimension, for the credit data features of various small and medium-sized enterprises in this application, the feature variance can be selected to reflect the amount of information it has. For details, please refer to the selection of the feature credit risk information value in the previous text. In this way, this application preferably sets the credit assessment explanatory value corresponding to each feature as the explained variance rate to reflect its contribution to the total variance.

[0284] In some implementation schemes, the cumulative credit assessment explanatory value is finally selected, preferably the principal component feature whose cumulative explained variance rate reaches the fifth threshold, and further preferably the fifth threshold is 95%. In this way, when the cumulative explained variance rate reaches 95%, the first b principal components corresponding to the cumulative credit assessment explanatory value are selected, for example, the first 50 principal components whose cumulative explained variance rate reaches 95%. In this way, the preprocessing, extraction and dimensionality reduction of multiple complex features in complex multi-source heterogeneous data sets are completed, so that the second feature set meets the feature quantity and quality required for model training.

[0285] It should be noted that the extraction and dimensionality reduction of multiple features are not limited to the above-mentioned methods, but this application preferably adopts these methods to optimize the quality of variables to further improve the learning efficiency of the model.

[0286] Furthermore, in order to improve the interpretability of the principal component obtained by dimensionality reduction in the second feature set, the features in the second feature set can also be fused with specific features to obtain the principal component containing key information. Specifically, a first weight value can be set for each of the multiple features in the second feature set, preferably each feature in the principal component obtained by dimensionality reduction, and a second weight value can be set for multiple selected key features; based on the first weight value and the second weight value, each feature in the second feature set and each key feature in the multiple key features are weighted and fused to finally obtain a weighted feature.

[0287] Among them, the key features are obtained by multi-dimensional screening of multiple features in the first feature set, and have stronger predictive ability, stability and explainability for the assessment of corporate credit risk than other features. The specific steps of obtaining the key features by multi-dimensional screening of multiple features in the first feature set may include:

[0288] The multiple features in the first feature set are screened using one or more of the screening methods for screening multiple features in the first feature set to obtain the second feature set in the present application, for example, the methods may include performing correlation analysis screening using the credit risk correlation coefficient of the features, performing sorting screening using the credit risk importance value of the features, performing feature screening using LASSO regression, etc. as proposed in the present application; performing stability test screening on multiple features, for example, obtaining and screening features with high stability in different time periods through time series cross-validation; and reviewing multiple features by business experts to ensure high business relevance and explainability of key features.

[0289] Preferably, the key features may include: operating cash flow / total debt ratio, quick ratio, debt-to-asset ratio, which can reflect the financial health of the enterprise; inventory turnover rate, accounts receivable turnover rate, and asset turnover rate, which can reflect the operating efficiency of the enterprise; net profit margin, gross profit margin, and EBITDA profit margin, which can reflect the profitability of the enterprise; the market share of the enterprise, the market share growth rate, and the ratio of R&D expenses / operating income, which can reflect the market competitiveness of the enterprise; the bank loan repayment timeliness rate and the supplier payment timeliness rate can be used as the performance of the enterprise's credit behavior; the HHI index and the downstream customer stability index can reflect the development of the enterprise's industrial chain; the GDP growth rate sensitivity and the industry prosperity index sensitivity can reflect the enterprise's macroeconomic sensitivity; the management stability index and the complexity of the equity structure can reflect the level of corporate governance; the number of valid patents and the proportion of new product revenue can reflect the innovation ability of the enterprise; and the frequency of negative news reports and the frequency of abnormal large-scale fund transactions can reflect the current risks and risk warnings of the enterprise.

[0290] In some embodiments, weighted fusion of the principal components and the key features not involved in feature dimensionality reduction may be performed by the following method:

[0291]

[0292] Where F is the weighted feature obtained by weighted fusion, PC i is the i-th principal component obtained by feature dimensionality reduction in the second feature set, K j is the jth selected key feature, ω i and The first weight value and the second weight value corresponding to the i-th principal component and the j-th selected key feature are respectively. It should be noted that the weighted features obtained by weighted fusion can belong to the same category as one or more features in the key features, or the same category as the principal component. The key features can be used to set the category of the model prediction performance.

[0293] Existing credit risk assessment models mostly use traditional statistical methods such as linear regression and scoring cards. These methods perform well when dealing with simple, linear relationships, but are insufficient when facing complex nonlinear relationships. The credit risk assessment of small and medium-sized enterprises involves data interactions in multiple dimensions, such as market dynamics, social public opinion, and industrial chain relationships. These complex relationships are difficult to effectively capture and quantify through traditional statistical models. In addition, due to data sparsity and the presence of noisy data, traditional methods perform poorly when dealing with high-dimensional data, resulting in low model robustness and prediction accuracy.

[0294] Therefore, the comprehensive credit scoring model proposed in this application consists of three core models: the first evaluation model is used to evaluate the business ability of enterprises; the second evaluation model is used to evaluate the competitiveness of enterprises; and the third evaluation model is used to evaluate the risk of business owners. The three evaluation models are independently modeled, and finally the outputs of the three evaluation models are combined through an integrated learning method to generate a comprehensive credit score for the enterprise. In this way, the comprehensive credit scoring model can selectively measure the importance of multi-dimensional heterogeneous data to the credit ability assessment of small and medium-sized enterprises, so as to improve the accuracy of credit risk prediction.

[0295] In order to train three evaluation models, the second feature set can be divided into three feature subsets to train three evaluation models respectively, such as step S120. In some embodiments, the features in the first feature subset include relevant indicators of the recent operating status of small and medium-sized enterprises, so the first feature subset may include: operating income growth rate, net profit margin, total assets, asset turnover, total liabilities, inventory turnover, accounts receivable turnover, operating cash flow / operating income ratio, industry average growth rate, average daily transaction amount, transaction frequency, large transaction ratio, cross-border transaction record, industry average growth rate, market concentration, raw material price index and / or first weighted feature. Among them, the first weighted feature represents the business operation related features in the weighted features, for example, the main component or key feature constituting the weighted feature is the business operation related feature.

[0296] The first evaluation model trained using the first feature subset is the enterprise operation evaluation model, which should be based on the longitudinal operation time series analysis, and can then be constructed by combining the AutoRegressive Integrated Moving Average (ARIMA) model and the Exponentially Weighted Moving Average (EWMA) model to predict the current key operating indicators of the enterprise.

[0297] The general form of the ARIMA model is:

[0298]

[0299] Among them, φi and θ j are the autoregressive coefficient and the moving average order, p and q are the difference order and the moving average order of the autoregressive model and the moving average model, respectively. t-i is the time series y t The historical value of t is a random error term. By using the ARIMA model to approximate the time series data formed by the predicted operating conditions over time to identify the sequence characteristics, the future value within a certain period of time can be predicted from the past value and present value of the first characteristic subset of the enterprise operation within a certain period of time to achieve business forecasting.

[0300] In particular, the autoregressive coefficient of the ARIMA model ranges from 0 to 2, preferably the autoregressive coefficient is 2; the difference order of the ARIMA model ranges from 0 to 1, preferably the difference order is 1; the moving average order of the ARIMA model ranges from 2 to 3, preferably the moving average order is 2.

[0301] In addition, the first evaluation model also includes an EWMA model, which is used to smooth sudden changes in time series. The update formula is:

[0302] S t =α·x t +(1-α)·S t-1

[0303] (Formula 5)

[0304] Among them, S t-1 To predict the time series S t The historical value of α is the smoothing coefficient, which usually ranges from 0.1 to 0.3 and is used to adjust the weight of historical data; the preferred smoothing coefficient in this application is 0.3. By combining the ARIMA model with the EWMA model, the business forecast level and the volatility level of the forecast value can be obtained, which helps small and medium-sized enterprises to manage risks and allocate assets, thereby reducing credit risks.

[0305] In some implementations, the first evaluation model can output a business operation score between 0 and 100 to reflect the business forecast. The higher the business operation score, the better the business situation of the enterprise at this stage. In this way, the first evaluation model can independently help enterprises and creditors to judge their credit risk tolerance from an operational perspective, and use the forecast results as one of the important indicators of the final comprehensive credit evaluation.

[0306] Specifically, the business condition level can be set based on the business operation score: business operation score of 80-100 points: excellent business condition; business operation score of 60-79 points: good business condition; business operation score of 40-59 points: average business condition; business operation score of 20-39 points: poor business condition; and business operation score of 0-19 points: dangerous business condition.

[0307] In particular, the features in the first feature subset can also adjust the learning weight value of the features by setting a certain learning rate when training the first evaluation model to help the model learn important features and improve training efficiency. According to the first evaluation model proposed in this application, the weight value of the operating income growth rate is preferably 0.25; the weight value of the net profit margin is preferably 0.20; the weight value of the asset turnover rate is preferably 0.15; the weight value of the operating cash flow / operating income ratio is preferably 0.25; and the total weight value of the inventory turnover rate, accounts receivable turnover rate and industry average growth rate is preferably 0.15.

[0308] In some embodiments, the features in the second feature subset include indicators related to the market competitiveness of small and medium-sized enterprises, and thus may include: enterprise market share, market share growth rate, brand awareness, number of patent applications, number of patent authorizations, number of patent citations, patent family size, patent quality score, customer growth rate, customer retention rate, customer quality score, key customer contribution rate, core technical personnel ratio, employee education composition ratio, talent loss rate, R&D personnel ratio, new product revenue ratio, R&D expense / operating income ratio and / or second weighted features. Similarly, the second weighted feature represents the enterprise competitiveness-related features in the weighted features, for example, the main component or key feature constituting the weighted feature is the enterprise competitiveness-related feature.

[0309] The second evaluation model can be an enterprise competitiveness evaluation model, which uses the analytic hierarchy process (AHP) to perform weighted calculations on the characteristics of different dimensions to measure the market competitiveness of the enterprise in the industry. Therefore, the second evaluation model selects and considers characteristic indicators of multiple dimensions such as market share, customer growth rate, product innovation capability, etc. and calculates the competitiveness index, which can be calculated by the following formula:

[0310]

[0311] Among them, ω i is the weight of each feature, I i is the characteristic value of each feature. It should be noted that the features analyzed by AHP are standardized features. To ensure the scientific nature of the weight setting, the second evaluation model can further include the introduction of consistency checks, such as the CR value method, to ensure that the AHP hierarchy meets the consistency requirements and the quality and reliability of the selected features.

[0312] In some embodiments, the second evaluation model may output an enterprise competitiveness index with a value between 0 and 10 to reflect the predicted competitiveness of the enterprise. The higher the enterprise competitiveness index, the stronger the enterprise competitiveness. Specifically, the enterprise competitiveness level can be set based on the enterprise competitiveness index: the enterprise competitiveness index is 8.0-10.0 points: extremely strong competitiveness; the enterprise competitiveness index is 6.0-7.9 points: very strong competitiveness; the enterprise competitiveness index is 4.0-5.9 points: medium competitiveness; the enterprise competitiveness index is 2.0-3.9 points: relatively weak competitiveness; and the enterprise competitiveness index is 0.0-1.9 points: extremely weak competitiveness.

[0313] In particular, the features in the second feature subset can also adjust the learning weight value of the features by setting the learning rate when training the second evaluation model to help the model learn important features and improve training efficiency. According to the second evaluation model proposed in this application, the total weight value of market share and market share growth rate is preferably 0.30; the total weight value of the number of patents and patent quality score is preferably 0.25; the total weight value of customer growth rate and customer quality score is preferably 0.20; the weight value of the proportion of core technical personnel is preferably 0.15; and the total weight value of the proportion of new product revenue and the ratio of R&D expenses to operating income is preferably 0.10.

[0314] It should be noted that the patent quality score in the second evaluation model can be obtained by setting the patent quality score standard. In some implementation schemes, the patent quality score standard can be set as: 8-10 points for patent application, 5-7 points for utility model and 3-4 points for design. In addition, due to the sensitivity of AHP to features, adjusting the weights of each dimension in AHP will significantly affect the model performance and the final market competitiveness prediction results, so it is even more necessary to ensure the accuracy of the weight values ​​of each feature.

[0315] In some embodiments, the features in the third feature subset include highly relevant indicators of the main credit risk of small and medium-sized enterprises, and thus may include: personal credit score, debt default record, litigation record, loan history record, number of litigation cases, type of litigation cases, judgment results of litigation cases, information on the person subject to execution, number of related enterprises, operating conditions of related enterprises, complexity of equity relations, public opinion index, positive / negative evaluation ratio, topic popularity, frequency of large-amount fund transactions, consumption pattern abnormality index, cross-border fund flow records and / or third weighted features. Similarly, the third weighted feature represents the enterprise credit risk-related features in the weighted features.

[0316] The third assessment model can be an enterprise owner risk assessment model, which focuses on direct assessment of SME risks in terms of multiple loans, financial risks, legal risks, etc. The model can analyze the correlation between enterprise owner risks and other risk nodes by constructing a knowledge graph, and perform risk prediction in combination with a supervised learning model. The specific structure of the third assessment model may include:

[0317] Using the risk association network analysis method, such as the PageRank algorithm, each risk node in the risk network composed of the third feature subset can be ranked by importance. The implementation method is to calculate the importance factor of the risk node:

[0318]

[0319] Among them, d is the damping coefficient, which is usually between 0.5 and 1, PR(j) is the set pointing to node j, L(j) is the out-degree of node j, and i is the number of iterations. In this way, the risk importance ranking corresponding to each feature can be obtained to obtain the training weight of the feature.

[0320] In addition, the Gradient Boosting Decision Tree (GBDT) model is used to perform multiple rounds of iterative training for supervised learning to continuously optimize the prediction results. The objective function of GBDT is:

[0321]

[0322] in, is the loss function of GBDT, which can be exemplified by mean square error (MSE) or logarithmic loss (LogLoss), etc. k ) is a regularization term used to control the complexity of GBDT and avoid overfitting. Through the risk association network analysis method and GBDT supervised learning, training weights can be effectively assigned to input features to improve the risk prediction results of GBDT for business owners; in addition, the weight relationship between features can be further extracted by combining the risk knowledge graph constructed according to the third feature subset to help GBDT achieve the optimal fit with the input features.

[0323] In some implementation schemes, the direct association weight of the risk knowledge graph ranges from 0 to 10, the second-degree association weight ranges from 0 to 2, and the third-degree association weight ranges from 0 to 1, preferably 1, 0.5, and 0.2, respectively; the damping coefficient of the PageRank model ranges from 0.5 to 1, preferably 0.85; the number of iterations of the PageRank model ranges from 50 to 200, preferably 100; the number of decision trees of the GBDT model ranges from 200 to 500, preferably 500; the maximum depth of the GBDT model ranges from 0 to 10, preferably 5; the learning rate of the GBDT model ranges from 0.05 to 0.1, preferably 0.05; and the sampling ratio of the feature of the GBDT model ranges from 0.5 to 2, preferably 0.8.

[0324] It is particularly important to note that the third assessment model may also include a logistic regression model to further improve the model's ability to model and assign weights for binary classification problems based on business owner risk prediction, such as whether the enterprise has a default risk, so as to further improve the classification performance of the third assessment model.

[0325] In some implementation schemes, the third assessment model may output a business owner risk probability with a value between 0 and 11 to reflect the current risk prediction of the enterprise. The higher the business owner risk probability, the greater the probability that the enterprise will have credit risk for the main related features of the third feature subset. Specifically, the business owner risk level can be set based on the business owner risk probability: the business owner risk probability is 0.0-0.2: low risk state; the business owner risk probability is 0.2-0.4: medium-low risk state; the business owner risk probability is 0.4-0.6: medium risk state; the business owner risk probability is 0.6-0.8: medium-high risk state; the business owner risk probability is 0.8-1.0: high risk state.

[0326] In particular, the highest threshold of low risk status can be set to 0.3, that is, when the risk probability of the business owner is not higher than 0.3, it can be considered that it is currently in a low credit risk state; the lowest threshold of high risk status can be set to 0.3, that is, when the risk probability of the business owner is not lower than 0.7, it can be considered that it is currently in a high credit risk state. In this way, the third assessment model can use the logistic regression model to independently conduct a two-class assessment of the credit risk of SMEs, helping SMEs to quickly estimate their approximate credit risk situation.

[0327] After obtaining the first evaluation model, the second evaluation model and the third evaluation model, the present application may further include integrating the first evaluation model, the second evaluation model and the third evaluation model to obtain a comprehensive credit scoring model, and the integration steps are as follows:

[0328] Set and update the first output weight of the first evaluation model, the second output weight of the second evaluation model, and the third output weight of the third evaluation model; integrate the first evaluation model, the second evaluation model, and the third evaluation model into a comprehensive credit scoring model based on the first output weight, the second output weight, and the third output weight using an integration method.

[0329] Specifically, the comprehensive credit scoring model includes three independent models: a first evaluation model, a second evaluation model, and a third evaluation model. At its output end, the outputs of the three independent models are fused into a comprehensive credit score. The fusion method can be exemplified by calculating by the following weighted average method:

[0330]

[0331] Among them, ω i is the weight of each evaluation model, f i (x) is the characteristic function output of each evaluation model, which can be exemplified by enterprise operation score, enterprise competitiveness index, enterprise owner risk probability or any one or more of the characteristic representations in the corresponding evaluation model. In addition, the parameters in each evaluation model can be further adjusted through hyperparameter tuning methods such as cross-validation and grid search to ensure that the comprehensive credit scoring model can achieve better performance in different data scenarios and improve the generalization ability of the model in complex credit data, especially sparse data.

[0332] In some implementations, the comprehensive credit scoring model can output a comprehensive credit score, and classify the predicted current credit risk of the enterprise into multiple risk levels such as low risk, medium-low risk, medium risk, medium-high risk and high risk according to the value of the comprehensive credit score of the SME. The specific risk level classification may include: low-risk enterprises, with a comprehensive credit score of 80-100 points; medium-low risk enterprises, with a comprehensive credit score of 60-79 points; medium-risk enterprises, with a comprehensive credit score of 40-59 points; medium-high risk enterprises, with a comprehensive credit score of 20-39 points; and high-risk enterprises, with a comprehensive credit score of 0-19 points.

[0333] Correspondingly, for the current credit risk level of the enterprise, the comprehensive credit scoring model can be set to output the corresponding enterprise recommendation strategy, which can be: for low-risk enterprises, it is recommended to increase the credit limit and reduce the complexity of the approval process; for medium-risk enterprises, it is recommended to conduct regular reviews and moderately increase the collateral requirements; for medium- and high-risk enterprises, it is recommended to reduce the credit limit and require additional guarantees; for high-risk enterprises, it is recommended to suspend credit or strengthen supervision, and take mitigation measures depending on the specific situation. In this way, the comprehensive credit scoring model can provide reasonable credit planning suggestions for the assessed SMEs based on the evaluation results, so as to help enterprises specify multiple risk control plans such as industrial model adjustment, market public opinion control, and financial revenue and expenditure adjustment under different risk conditions.

[0334] Traditional credit risk assessment methods usually rely on periodically updated data, such as annual financial reports, quarterly financial reports, etc., which makes the model's assessment results lag behind. In the current rapidly changing market environment, the business conditions and credit risks of enterprises may change significantly in the short term, and existing models cannot respond to these changes in a timely manner. Due to the lack of real-time data processing capabilities, the lag in risk monitoring of existing technologies has become a major shortcoming in the credit risk assessment of small and medium-sized enterprises.

[0335] Therefore, when corporate credit risk changes, the existing model cannot issue early warning signals in time, causing financial institutions to face information lag when making decisions. This lack of dynamic monitoring and early warning mechanisms has greatly limited the timeliness of credit risk management, especially in the current rapidly changing market.

[0336] Therefore, the construction method of the credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data in this application further proposes the construction steps of the dynamic monitoring and early warning model, such as Figure 3 As shown, it may include step S130: establishing data connection with a third-party data source to obtain a real-time risk assessment data set for SMEs, such as step S200; and using the real-time risk assessment data set to train a dynamic monitoring and early warning model to assess the real-time credit risk of SMEs.

[0337] The real-time risk assessment data set collected by this application may include the following real-time data: short-term liquidity indicator data set, business activity indicator data set, credit behavior indicator data set, market and public opinion indicator data set, industrial chain indicator data set, macroeconomic indicator data set and / or abnormal transaction indicator data set, all of which can be obtained in real time by connecting with third-party data sources.

[0338] In some embodiments, the real-time variables in the short-term liquidity indicator data set may include: the daily average bank deposit balance change rate, the cash inflow and outflow ratio in the past 7 days and / or the overdue accounts payable ratio; the real-time variables in the operating activity indicator data set include: the daily operating income change rate, the major customer order volume change rate and / or the capacity utilization rate.

[0339] In some embodiments, the real-time variables in the credit behavior indicator data set may include: bank loan repayment timeliness rate, supplier payment timeliness rate and / or tax payment timeliness rate.

[0340] In some embodiments, real-time variables in the market and public opinion indicators dataset may include: stock price volatility, number of negative news reports, and / or social media sentiment scores.

[0341] In some implementation schemes, real-time variables in the industry chain indicator data set may include: the on-time delivery rate of major suppliers, the on-time payment rate of core customers, the stock prices of upstream and downstream companies, and / or changes in the credit ratings of upstream and downstream companies.

[0342] In some embodiments, real-time variables in the macroeconomic indicator data set may include: the rate of change of the industry prosperity index, the volatility of related commodity prices and / or the volatility of exchange rates; and real-time variables in the abnormal transaction indicator data set include: the frequency of large-value capital transactions, the cross-border capital flow abnormality index and / or the change in the proportion of related-party transactions.

[0343] The present application can use the real-time risk assessment data set to train a dynamic monitoring and early warning model. Specifically, the dynamic monitoring and early warning model can be used to calculate the conditional dependency between each of the multiple real-time variables in the real-time risk assessment data set and the remaining real-time variables, as shown in Figure S210. For example, the dynamic monitoring and early warning model can calculate the conditional probability of each real-time variable occurring in real time based on the Bayesian network combined with the risk assessment data set and the current real-time risk assessment data set, and clarify the conditional dependency. The core formula of the Bayesian network is:

[0344]

[0345] Among them, A is related to a specific risk event, and B is the characteristic of the real-time variable.

[0346] Therefore, the dynamic monitoring and early warning model also includes a method for calculating the conditional probability of each real-time variable among multiple real-time variables, and based on the conditional probability of real-time variables that are highly correlated with the real-time credit risk events of SMEs, the occurrence probability of real-time credit risk events of SMEs at multiple risk levels is calculated respectively, such as step S220. In this way, the dynamic monitoring and early warning model can obtain the risk level with the highest probability of occurrence of real-time credit risk of SMEs, thereby providing real-time risk forecasts for SMEs and creditors, such as step S230.

[0347] Specifically, in some implementation schemes, the dynamic monitoring and early warning model can make probability predictions for low, medium and high risk levels by setting different real-time risk thresholds, and output the maximum risk level as a real-time credit risk early warning signal. For high-risk enterprises, the dynamic monitoring and early warning model can also recommend additional risk mitigation measures, such as increasing collateral requirements or reducing credit limits to help small and medium-sized enterprises survive high-risk difficult periods.

[0348] This application also proposes a system for constructing a credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data, such as Figure 4 As shown, it includes:

[0349] The multi-source data acquisition unit 300 acquires a multi-source data set for credit risk assessment of SMEs containing multiple variables, wherein the multi-source data set for credit risk assessment includes a government data set, an industry data set, an Internet data set, and a data set owned by SMEs;

[0350] The data feature engineering unit 310 extracts features from multiple variables in the credit risk assessment multi-source data set to obtain a first feature set; and obtains a second feature set by screening multiple features in the first feature set;

[0351] The multi-dimensional model building unit 320 divides the second feature set into a first feature subset, a second feature subset, and a third feature subset to respectively train a first evaluation model, a second evaluation model, and a third evaluation model; integrates the first evaluation model, the second evaluation model, and the third evaluation model to obtain a comprehensive credit scoring model to perform a multi-dimensional evaluation of the credit risk of small and medium-sized enterprises; and

[0352] The dynamic monitoring and early warning model building unit 330 establishes data connection with a third-party data source to obtain a real-time risk assessment data set for SMEs; and uses the real-time risk assessment data set to train a dynamic monitoring and early warning model to assess the real-time credit risk of SMEs.

[0353] The system for constructing the SME credit risk assessment model proposed in the present application can be a system for executing the method for constructing the SME credit risk assessment model. Those skilled in the art can understand that the specific functions and operations of each module in the system for constructing the SME credit risk assessment model have been introduced in detail in the description of the method for constructing the SME credit risk assessment model, and therefore, repeated description thereof will be omitted.

[0354] As described above, the construction system of the SME credit risk assessment model of the present application can be implemented in various terminal devices, for example, for constructing a comprehensive credit assessment model or one or two or more of the three assessment models, and further, for constructing a server for a comprehensive credit assessment model based on industrial and financial big data. In some embodiments, the construction system of the SME credit risk assessment model described in the present application can be integrated into the terminal device as a software module and / or hardware module. For example, the construction system of the SME credit risk assessment model can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; the construction system of the SME credit risk assessment model can also be one of the many hardware modules of the terminal device.

[0355] Alternatively, in other embodiments, the system for constructing the SME credit risk assessment model and the terminal device may be separate devices, and the system for constructing the SME credit risk assessment model may be connected to the terminal device via a wired and / or wireless network, and transmit interactive information, such as various types of industrial and financial big data, in accordance with an agreed data format.

[0356] In the present application, an electronic device is also proposed, and the electronic device may include:

[0357] Processor; and

[0358] A memory stores computer program instructions, which, when executed by a processor, enable the processor to execute the method for constructing a credit risk assessment model for small and medium-sized enterprises described in the present application.

[0359] In the present application, the processor may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions.

[0360] The memory may include one or more computer program products, and the computer program product may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor may run the program instructions to implement the method for constructing the SME credit risk assessment model described in this application and / or other desired functions. Various contents such as a multi-source data set for SME credit risk assessment, a first feature set, a second feature set, and a real-time risk assessment data set may also be stored in the computer-readable storage medium.

[0361] In some embodiments, the electronic device may further include: an input device and an output device, and these components are interconnected through a bus system and / or other forms of connection mechanisms. The input device may include, for example, a keyboard, a mouse, etc. The output device may output various information to the outside, including a trained comprehensive credit scoring model, etc. The output device may include, for example, a display, a speaker, a printer, and a communication network and a remote output device connected thereto, etc.

[0362] This application also proposes a credit risk assessment method for small and medium-sized enterprises based on industrial and financial big data. It uses the comprehensive credit scoring model and dynamic monitoring and early warning model constructed by the method for constructing a credit risk model for small and medium-sized enterprises based on industrial and financial big data proposed in this application to assess the current comprehensive credit risk and current real-time credit risk of small and medium-sized enterprises respectively.

[0363] Specifically, the method for assessing the credit risk of small and medium-sized enterprises based on industrial and financial big data may include: a data collection step, obtaining a multi-source data set for assessing the credit risk of small and medium-sized enterprises containing multiple variables, the multi-source data set for assessing the credit risk of small and medium-sized enterprises including a government data set, an industrial data set, an Internet data set, and a data set owned by small and medium-sized enterprises. In particular, the multi-source data set for assessing the credit risk of small and medium-sized enterprises obtained in the data collection step may be the same as the multi-source data set for assessing the credit risk of small and medium-sized enterprises obtained by the method for constructing a credit risk model for small and medium-sized enterprises based on industrial and financial big data.

[0364] The credit risk assessment method for small and medium-sized enterprises based on industrial and financial big data may also include: a data processing step, performing feature extraction and feature screening on multiple variables in the credit risk assessment multi-source data set to obtain a risk assessment feature set. Specifically, an initial feature set may be extracted from multiple variables, and the features in the initial feature set may be screened for relevance, credit risk information value, importance, principal component, and feature fusion to obtain a risk assessment feature set. It can be seen that the initial feature set may be equivalent to the first feature set obtained by the method for constructing a credit risk model for small and medium-sized enterprises based on industrial and financial big data, and the risk assessment feature set may be equivalent to the second feature set obtained by the method for constructing a credit risk model for small and medium-sized enterprises based on industrial and financial big data.

[0365] In particular, the processing processes of correlation screening, credit risk information value screening, importance screening, principal component screening and feature fusion of the features in the initial feature set can also be consistent with the feature extraction, dimensionality reduction and fusion methods in the feature screening method of the construction method of the small and medium-sized enterprise credit risk model based on industrial and financial big data, so as to ensure the effective learning of multiple features in the risk assessment feature set by the comprehensive credit score.

[0366] Therefore, the credit risk assessment method for SMEs based on industrial and financial big data can include: a credit risk assessment step, inputting multiple features in the risk assessment feature set into a comprehensive credit scoring model, and then correctly using the comprehensive credit scoring model to assess the credit risk of SMEs.

[0367] The credit risk assessment method for SMEs based on industrial and financial big data may also include: a dynamic monitoring and early warning step, establishing data connection with a third-party data source to obtain a real-time risk assessment data set for SMEs containing multiple real-time variables; inputting multiple real-time variables in the real-time risk assessment data set into a dynamic monitoring and early warning model to assess the real-time credit risk of SMEs. In particular, the real-time risk assessment data set for SMEs obtained by the dynamic monitoring and early warning step may be the same as the real-time risk assessment data set for SMEs obtained by the method for constructing a credit risk model for SMEs based on industrial and financial big data, thereby ensuring that the dynamic monitoring and early warning model effectively learns multiple real-time features in the real-time risk assessment data set for SMEs.

[0368] This application also proposes a credit risk assessment system for small and medium-sized enterprises based on industrial and financial big data, such as Figure 5 As shown, it may include:

[0369] The data collection unit 400 obtains a multi-source data set for credit risk assessment of small and medium-sized enterprises containing multiple variables, wherein the multi-source data set for credit risk assessment includes a government data set, an industry data set, an Internet data set, and a data set owned by small and medium-sized enterprises;

[0370] The data processing unit 410 performs feature extraction and feature screening on the multiple variables in the credit risk assessment multi-source data set to obtain a risk assessment feature set;

[0371] A credit risk assessment unit 420 inputs a plurality of features in the risk assessment feature set into a comprehensive credit scoring model to assess the credit risk of the SME; and

[0372] The dynamic monitoring and early warning unit 430 establishes data connection with a third-party data source to obtain a real-time risk assessment data set of SMEs containing multiple real-time variables; and inputs the multiple real-time variables in the real-time risk assessment data set into a dynamic monitoring and early warning model to assess the real-time credit risk of SMEs.

[0373] The SME credit risk assessment system proposed in the present application can be a system for executing the SME credit risk assessment method. Those skilled in the art can understand that the specific functions and operations of each module in the SME credit risk assessment system have been introduced in detail in the description of the SME credit risk assessment method, and therefore, their repeated description will be omitted.

[0374] As described above, the SME credit risk assessment system of the present application can be implemented in various terminal devices, for example, to construct a comprehensive credit assessment model or one or two or more of the three assessment models, and further, to construct a server for a comprehensive credit assessment model based on industrial and financial big data. In some embodiments, the SME credit risk assessment system described in the present application can be integrated into a terminal device as a software module and / or a hardware module. For example, the SME credit risk assessment system can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; the SME credit risk assessment system can also be one of the many hardware modules of the terminal device.

[0375] Alternatively, in other embodiments, the SME credit risk assessment system and the terminal device may be separate devices, and the SME credit risk assessment system may be connected to the terminal device via a wired and / or wireless network, and transmit interactive information, such as various types of industrial and financial big data, in accordance with an agreed data format.

[0376] In the present application, an electronic device is also proposed, and the electronic device may include:

[0377] Processor; and

[0378] A memory stores computer program instructions, which, when executed by a processor, enable the processor to execute the credit risk assessment method for small and medium-sized enterprises described in the present application.

[0379] In the present application, the processor may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions.

[0380] The memory may include one or more computer program products, and the computer program product may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor may run the program instructions to implement the SME credit risk assessment method described in this application and / or other desired functions. Various contents such as SME credit risk assessment multi-source data set, initial feature set, risk assessment feature set, and real-time risk assessment data set may also be stored in the computer-readable storage medium.

[0381] In some embodiments, the electronic device may further include: an input device and an output device, and these components are interconnected through a bus system and / or other forms of connection mechanisms. The input device may include, for example, a keyboard, a mouse, etc. The output device may output various information to the outside, including a trained comprehensive credit scoring model, etc. The output device may include, for example, a display, a speaker, a printer, and a communication network and a remote output device connected thereto, etc.

[0382] Most existing technologies rely on single financial data, such as annual reports and quarterly reports. These data are updated infrequently and with lags, making it difficult to reflect the real-time credit status of the company. This application integrates government data, industry data, Internet data, and the company's own data to establish a multi-dimensional data system covering corporate finances, market performance, operational compliance, and public opinion dynamics, providing comprehensive information support for credit risk assessment. Traditional credit risk assessments usually rely on annual or quarterly data, which has a large lag and cannot reflect changes in the company's credit status in a timely manner. This application achieves real-time monitoring and early warning of corporate credit risks through real-time data collection and analysis combined with a dynamic early warning mechanism.

[0383] In addition, in credit risk assessment, dealing with complex nonlinear relationships and high-dimensional data is a key challenge. Traditional methods such as linear regression or scoring cards usually perform poorly, especially when faced with dynamically changing and complex data environments. This application significantly improves the prediction accuracy and robustness of the model by introducing advanced algorithms such as GBDT (gradient boosted decision tree), Logistic regression, and AHP (analytic hierarchy process). Most traditional credit scoring systems rely on single-dimensional scoring models, which are difficult to fully reflect the multi-dimensional risks of enterprises; the multi-dimensional credit scoring system designed in this application, through independent modeling of three aspects of business operations, market competitiveness, and business owner risks, ultimately generates the credit score of the enterprise through integrated learning, supporting personalized and differentiated credit decisions.

[0384] Finally, the method of this application can rely on corresponding systems and electronic devices, and can achieve full-process automated risk assessment and decision support by introducing automated data processing, real-time monitoring and intelligent scoring mechanisms. The system and electronic devices can autonomously collect and process multi-source data, and automatically generate credit recommendations based on real-time analysis results, significantly reducing the need for manual intervention.

[0385] Example

[0386] Example 1 Dataset acquisition and preprocessing

[0387] Obtain raw data from four major data sources (government affairs, industry, Internet, and enterprise-owned) and pre-process them. The specific processing process is as follows:

[0388] Government data collection: Through API connection with the tax bureau, social security bureau and industrial and commercial departments, the tax declaration, social security payment and business compliance data of enterprises are captured once a day. Specific government data types include: enterprise business registration information, including registered capital, establishment time, business scope and other information; tax payment information, including monthly value-added tax, corporate income tax and other information; social security payment information, including employee information, social security payment, etc.; environmental compliance information, including environmental impact assessment reports, pollution information, etc.; and administrative penalty information, including reasons for penalties, penalty content, etc.

[0389] In total, basic government data of approximately 1 million small and medium-sized enterprises were collected, with an average number of data fields for each enterprise being 50.

[0390] Industry data collection: Market share information, upstream and downstream supply chain relationships, industry development and other indicator data are obtained from industry associations and a research and consulting Internet company every quarter. Specific industry data include: market share data, including enterprise market share, industry ranking, etc.; upstream and downstream supply chain information, including major suppliers, customer concentration, etc.; industry prosperity index, including PMI, capacity utilization, etc.; product price index, including raw material prices, product factory information, etc.; industry technology innovation indicators, including industry patent applications, R&D investment, etc.

[0391] The industrial data of 50 major industries covers a total of about 1 million small and medium-sized enterprises, and the average industrial data of each industry contains 1,000 variables.

[0392] Internet data collection: Scrapy crawler technology is used to continuously collect enterprise-related information from public social media, news and e-commerce platforms, which is collected and updated every hour, and sentiment analysis and keyword extraction are performed through natural language processing technology. Specific Internet data include: social media data, including enterprise-related discussion texts on platforms such as Weibo and WeChat public accounts; news media reporting data, including mainstream media reports on enterprises; e-commerce platform data, including sales and evaluation texts of enterprise products; recruitment information, including recruitment positions and salary levels released by enterprises; and enterprise official website and public account data, including business information released by enterprises.

[0393] In total, approximately 100GB of unstructured related text was collected every day.

[0394] Enterprise-owned data collection: Sales, customer management and financial data are regularly exported through the enterprise's internal ERP system and CRM system, and collected and updated once a month. Specific enterprise-owned data include: detailed financial statement data, including balance sheet, income statement, cash flow statement, etc.; sales data, including product sales, customer composition, sales area distribution, etc.; production data, including production costs, capacity utilization information, inventory turnover, etc.; and enterprise R&D data, including R&D investment, patent application status, new product development progress, etc.

[0395] In total, the company-owned data covers about 10,000 small and medium-sized enterprises, and the average company-owned data for each enterprise contains 200 variables.

[0396] Data cleaning: The raw data needs to be preprocessed after collection. The specific operations are as follows:

[0397] Missing value processing: Use interpolation, mean filling and other methods to fill missing values ​​in the data. For example, in financial data, if the sales data of a certain month is missing, linear interpolation can be performed through the sales trends of the adjacent months. Among them, linear interpolation can be implemented by filling missing values ​​in multiple variables in the implementation manual.

[0398] Outlier detection and processing: For extreme data points, Z-Score analysis and box plot (IQR) are used to detect and eliminate them to ensure the rationality of data distribution. Among them, Z-Score analysis can be achieved by eliminating abnormal data in multiple variables as described in the implementation manual.

[0399] Statistical feature variable acquisition: For numerical variables, common statistical methods are used to generate statistical feature variables and feature values, such as mean, standard deviation, maximum value, minimum value, etc. Afterwards, these new variables are included in feature selection and weighted combination.

[0400] Text data preprocessing: Text information in Internet data is processed through NLP technology, including operations such as word segmentation, removal of stop words, word vectorization, and the BERT model is used for context embedding generation. For the obtained text data, TF-IDF and BERT embedding are used to generate text feature vectors. After that, these new variables are also included in the feature selection and weighted combination. In addition, sentiment analysis is performed on text data, such as social media comments, to generate positive and negative sentiment scores. For example, the BERT model is used to analyze the public opinion information of the company to obtain a sentiment score to indicate the market's overall attitude towards the company.

[0401] Example 2 Feature Engineering and Dimensionality Reduction

[0402] Feature extraction is performed based on the multiple enterprise credit risk related variables preprocessed in Example 1, including the following steps:

[0403] Feature Dimensionality Reduction: Reduce the dimensionality of high-dimensional features by extracting principal components to reduce redundant features. Specifically, in the feature extraction process of one batch, among more than 100 initial features to be selected, the top 10 principal components with the largest contribution were extracted, accounting for 85% of the total variance.

[0404] Feature selection and weight calculation: LASSO regression is used for feature selection to screen out key features that have a greater impact on the model. Then, weights are assigned to different features, and weighted features are generated based on consistency tests (CR value < 0.1).

[0405] Among them, the main features of the principal components after dimensionality reduction and their explanations are as follows:

[0406] Principal component 1 (explained variance rate 30.5%): mainly reflects the overall financial health of the enterprise and is highly correlated with the operating cash flow / total debt ratio, quick ratio, debt-to-asset ratio, etc.

[0407] Principal component 2 (15.2% explained variance): mainly reflects the operational efficiency of the enterprise and is highly correlated with accounts receivable turnover rate, inventory turnover rate, etc.

[0408] Principal component 3 (explained variance rate 11.8%): mainly reflects the growth of the enterprise and is highly correlated with the net profit growth rate, operating income growth rate, etc.

[0409] Principal component 4 (explained variance rate 8.7%): mainly reflects the characteristics of the enterprise's industrial chain and is highly correlated with the upstream supplier concentration and downstream customer stability index.

[0410] Principal component 5 (7.3% explained variance): mainly reflects the characteristics of the enterprise's transaction behavior and is highly correlated with cash flow volatility, transaction frequency anomaly index, etc.

[0411] Principal component 6 (5.9% explained variance): mainly reflects the external evaluation of the enterprise and is highly correlated with social media sentiment scores, industry prosperity index, etc.

[0412] Principal component 7 (4.2% explained variance): mainly reflects the innovation capability of the enterprise and is highly correlated with the intensity of R&D investment.

[0413] Principal component 8 (explained variance rate 3.5%): mainly reflects the profitability of the enterprise and is highly correlated with gross profit margin, return on net assets, etc.

[0414] Principal component 9 (explained variance 2.8%): mainly reflects the asset utilization efficiency of the enterprise and is highly correlated with total asset turnover rate, etc.

[0415] Principal component 10 (explained variance 2.3%): mainly reflects the position of the enterprise in the industrial chain and is highly correlated with the industrial chain health index.

[0416] Principal component 11 (1.9% explained variance): mainly reflects the business history of the enterprise and is highly correlated with the age of the enterprise.

[0417] Principal component 12 (explained variance rate 1.6%): mainly reflects the short-term debt repayment ability of the enterprise and is highly correlated with the current ratio, etc.

[0418] Example 3 Model training and optimization

[0419] The third evaluation model was constructed by using GBDT and Logistic regression models with an integrated learning method to train multidimensional data and predict the main risks. The specific operations are as follows:

[0420] GBDT model training: GBDT improves the prediction accuracy of the model by optimizing the residuals round by round. In each round of iteration, the system generates a new decision tree to fit the prediction error of the previous round:

[0421] F m (x) = F m-1 (x)+γ m h m (x)

[0422] (Formula 11)

[0423] Among them, γ m is the learning rate, h m (x) is the decision tree output of the mth round.

[0424] Logistic regression: For binary classification problems, such as whether a company has a default risk, independent modeling is performed. The loss function of the Logistic regression model selects logarithmic loss:

[0425]

[0426] Ensemble learning and model fusion: The results of GBDT and Logistic regression are integrated, and the final prediction of the third evaluation model is generated by weighted average. The model parameters are optimized through grid search and cross-validation to ensure the robustness of the model in different scenarios.

[0427] In addition, the following optimization process is performed based on the third evaluation model obtained above:

[0428] Initial parameter range determination: Based on domain knowledge and preliminary experiments, the initial parameter range is determined for the model. For example, the number of trees of the GBDT model is set to [100, 500, 1000] and the maximum depth is set to [3, 5, 7, 9].

[0429] Cross-validation and grid search: 5-fold cross-validation is used in combination with the grid search method to perform a coarse-grained search in the initial parameter space, and then a fine-grained search in the parameter interval with better performance.

[0430] Bayesian optimization: Introduce Bayesian optimization method to effectively explore high-dimensional parameter space: use Gaussian process as the surrogate model, adopt expected improvement as the acquisition function, set 100 iterations, and check convergence every 25 times.

[0431] Model fusion weight optimization: A genetic algorithm-based method was used to optimize the model fusion weight, with the population size set to 100 and 50 iterations. Finally, the parameter tuning results of the two models were obtained, as shown below:

[0432] The number of decision trees of the GBDT model is 500, and the maximum depth of the GBDT model is 5 to enable GBDT to achieve optimal prediction; the learning rate of the GBDT model is 0.05 to balance its training cost and training effect; and the sampling ratio of the feature of the GBDT model is 0.8 to divide the training features and the test features.

[0433] The systematic parameter optimization process of this embodiment not only improves the prediction performance of the model, but also provides an in-depth understanding of the impact of each parameter on the model, laying a foundation and a feasible solution for the practical application of the model and subsequent continuous optimization.

[0434] Example 4 Dynamic Monitoring and Early Warning

[0435] In order to achieve real-time monitoring and early warning of corporate credit risk, a dynamic monitoring and early warning model is constructed based on the streaming data processing of a five-year-old technology manufacturing enterprise (hereinafter referred to as Enterprise A) in Example 1 to monitor the real-time credit risk changes of Enterprise A. The following are the specific construction steps:

[0436] Streaming data processing: Data streams are transmitted through Kafka, and Flink is used to process the company's operating data, market dynamics, and public opinion information in real time. For example, when a company's sales revenue drops abnormally, this change can be captured within minutes and the cause can be automatically analyzed.

[0437] Warning rule setting: Based on the dynamic adjustment mechanism of the Bayesian network, the system automatically issues a warning signal when it detects that the fluctuation of key indicators (such as cash flow and market sentiment) exceeds the set threshold. The Bayesian network can be implemented by calculating the conditional dependency relationship between each of the multiple real-time variables in the real-time risk assessment data set and the remaining real-time variables as described in the specification.

[0438] Among them, the daily monitoring data of key financial and market indicators of Enterprise A in August 2024 are shown in the following table:

[0439] Table 1 Dynamic monitoring index values ​​of enterprise A at different time points in August 2024

[0440]

[0441]

[0442] As can be seen above, the cash flow and sales revenue of Company A have continued to decline since August 10, while the tax compliance score and public opinion change index have shown a downward trend. These changes triggered a real-time early warning mechanism, that is, the model analyzes the streaming data in real time and recommends that financial institutions re-evaluate the credit line of Company A and suspend new credit issuance to strengthen real-time regulation of the company's financial status.

[0443] Example 5 Credit score generation and verification

[0444] Finally, the third assessment model described in Example 3 is integrated with the pre-built first assessment model and the second assessment model to generate a comprehensive credit score for the enterprise, and risk grading prediction and development strategy recommendations are made for the enterprise based on the comprehensive credit score results, specifically including:

[0445] Comprehensive credit score calculation, verification and analysis: The scores of each dimension are combined using the weighted average method as described in the instructions. The effectiveness of the model is verified by comparing the model's predicted score with the company's actual credit performance, such as the company's actual default rate and capital liquidity.

[0446] Finally, after the same model multiple rounds of iterative optimization and parameter tuning method as in Example 3, the optimal model parameters of the first evaluation model are obtained as follows: the autoregressive coefficient of ARIMA is 2, the difference order of ARIMA is 1; and the moving average order of ARIMA is 2. The optimal weights of the features obtained by AHP for the second evaluation model are as follows: the total weight of market share and market share growth rate is 0.30, the total weight of the number of patents and patent quality score is 0.25, the total weight of customer growth rate and customer quality score is 0.20, the weight of the proportion of core technical personnel is 0.15, and the total weight of the proportion of new product revenue and the ratio of R&D expenses to operating income is 0.10.

[0447] The test data set is randomly sampled from the feature set. The performance of the model on the test data set is as follows:

[0448] Table 2 Performance test results of each model in the test data set

[0449]

[0450] From the above results, it can be seen that the integrated comprehensive credit scoring model performs well in multiple indicators, especially in the AUC value of 0.958, indicating that the model can better distinguish between high-risk and low-risk enterprises under different thresholds.

[0451] In addition, according to some other specific comparison results, the scoring results of the model in different scenarios are also highly correlated with the actual performance. For example, for a company with a high credit score (850 points), its loan default probability is only 2.3%, proving that the scoring results of the model are highly consistent with the actual default risk; in the scoring verification of another company, its actual default rate is 7.5%, while the default risk predicted by the model is 6.8%, and the deviation is controlled within 1%, indicating that the model has a high accuracy.

[0452] In particular, three representative small and medium-sized enterprises were selected and their credit risk was evaluated to further verify the accuracy of the model; the three enterprises were Enterprise A, a traditional trading enterprise that has been in operation for more than ten years (hereinafter referred to as Enterprise B), and an Internet service enterprise that has been established for two years (hereinafter referred to as Enterprise C). For Enterprise A, Enterprise B and Enterprise C, the third evaluation model GBDT was first used to independently predict the credit scores of Enterprise A, B, and C and compare them with the actual default rate, and the results were obtained as shown in the following table:

[0453] Table 3 Prediction results of credit scores of companies A, B, and C using GBDT of the third evaluation model

[0454]

[0455] It is understandable that the credit score here does not integrate the results of other evaluation models, but is an independent verification of the performance of the third evaluation model. From the results, it can be seen that the score of the trained GBDT on enterprise C (87 points) is highly consistent with the actual low default rate (2.5%), indicating that it can effectively capture the true credit status of the enterprise in a multidimensional data scenario. At the same time, the predicted score for enterprise A is 68, the actual default rate is 7.5%, and the model deviation is small, which shows that even in enterprises with large risk fluctuations, the trained GBDT model also shows good stability.

[0456] Finally, the comprehensive credit assessment model was used to predict the comprehensive credit scores of the three companies, and corresponding credit strategy recommendations were provided based on the scoring results. The specific contents are as follows:

[0457] Table 4 Prediction results of the comprehensive credit scores of enterprises A, B, and C using the comprehensive credit assessment model

[0458]

[0459] According to the scoring results, it is recommended to give priority to the credit of enterprise C and increase the credit limit, while a more conservative strategy is adopted for the credit of enterprise B. The functional structure diagram of the credit risk assessment method for small and medium-sized enterprises in Examples 1 to 5 is shown in FIG. Figure 1As shown, the differentiated strategies provided for different enterprises through evaluation methods can not only optimize capital allocation, but also reduce potential default risks and ensure the safety and yield of financial assets.

[0460] Example 6 Performance comparison between the credit risk assessment method for small and medium-sized enterprises and existing methods

[0461] Comparing the data processing time of the data processing methods of Example 1 and Example 5 with the traditional manual review process, the obtained multi-source data processing efficiency is as follows:

[0462] Table 5 Performance comparison between SME credit risk assessment methods and existing data analysis methods

[0463] Processing Flow Traditional method (hours) Evaluation Method for this Application (minutes) Data collection and cleaning 72 15 Data analysis and modeling 96 20 Credit decision making 48 5 Total processing time 216 40

[0464] It can be seen that the overall processing time of the evaluation method and the evaluation system constructed based on this application is reduced from 216 hours to 40 minutes, and the processing efficiency is improved by more than 300 times that of the traditional method. At the same time, the evaluation method and evaluation system of this application process data through standardized processes and consistency checks, which can ensure the stability and consistency of the evaluation results and avoid evaluation bias caused by differences in manual judgment.

[0465] The evaluation method, evaluation system and electronic equipment of this application have broad application prospects in credit approval, supply chain management, government supervision and other scenarios of financial institutions. Through real-time monitoring and dynamic early warning, the system can flexibly adjust credit strategies in a changing market environment and provide accurate decision-making support for financial institutions. In the future, it can be further expanded to cross-border trade, supply chain finance and other fields, using more data sources and optimization algorithms to improve the intelligence level of risk management.

Claims

1. A method for constructing a credit risk assessment model for small and medium-sized enterprises based on industrial and financial big data, which includes: Acquire a multi-source data set for credit risk assessment of small and medium-sized enterprises containing multiple variables, wherein the multi-source data set for credit risk assessment includes a government data set, an industry data set, an Internet data set, and a data set owned by small and medium-sized enterprises; Extracting features from the multiple variables in the credit risk assessment multi-source data set to obtain a first feature set; and obtaining a second feature set by screening multiple features in the first feature set; Dividing the second feature set into a first feature subset, a second feature subset and a third feature subset to respectively train a first assessment model, a second assessment model and a third assessment model; integrating the first assessment model, the second assessment model and the third assessment model to obtain a comprehensive credit scoring model to perform a multi-dimensional assessment of the credit risk of small and medium-sized enterprises; as well as Establish data connection with third-party data sources to obtain real-time risk assessment data sets for SMEs; Using the real-time risk assessment data set to train a dynamic monitoring and early warning model to assess the real-time credit risk of small and medium-sized enterprises; Among them, the first feature subset includes the recent operating status characteristics of SMEs, the second feature subset includes the market competitiveness characteristics of SMEs, and the third feature subset includes the main credit risk characteristics of SMEs.

2. The method according to claim 1, wherein: The government data set includes: business registration information variables, tax payment record variables, water and electricity payment variables, social security payment variables, environmental protection compliance variables and / or administrative penalty variables; The industry data set includes: industry trend prosperity index variables, market share variables, upstream and downstream supply chain variables, product price index variables and / or industry technology innovation indicator variables; The Internet data set includes: social media public opinion variables, e-commerce platform transaction variables, platform evaluation variables, news media report variables, recruitment information analysis variables, enterprise official website push variables and / or enterprise public account push variables; and The SME's own data set includes: financial statement variables, SME internal sales variables, customer relationship management variables, internal management change variables and / or R&D investment variables.

3. The method according to claim 1, wherein: The step of obtaining a multi-source data set for credit risk assessment of small and medium-sized enterprises further includes: preprocessing the multiple variables in the multi-source data set for credit risk assessment; wherein the preprocessing includes: calculating a plurality of statistical feature variables of the plurality of variables to be incorporated into the plurality of variables; filling missing values ​​in the plurality of variables; Eliminating abnormal data from the plurality of variables; and The duplicate data in the multiple variables are merged.

4. The method according to claim 3, wherein: The calculating of the plurality of statistical characteristic variables of the plurality of variables comprises: A center metric variable, a dispersion metric variable, a distribution metric variable, a time series variable and / or a location metric variable of the plurality of variables are calculated.

5. The method according to claim 3 or 4, wherein: Filling the missing values ​​in the plurality of variables comprises: calculating a first credit assessment quality value or a second credit assessment quality value for each variable of the plurality of variables; In response to the variable whose first credit assessment quality value or the second credit assessment quality value among the multiple variables is less than a first threshold or a second threshold, a first interpolation variable or a second interpolation variable of the multiple variables is obtained, and the variable whose first credit assessment quality value is less than the first threshold is replaced by the first interpolation variable or the variable whose second credit assessment quality value is less than the second threshold is replaced by the second interpolation variable.

6. The method according to claim 3 or 4, wherein: The removing of abnormal data from the plurality of variables comprises: Calculating the mean values ​​and standard deviations of the multiple variables, and obtaining abnormal distribution intervals based on the mean values ​​and standard deviations of the multiple variables; In response to the variable in the plurality of variables being located in the abnormal interval, an a% quantile of the plurality of variables is obtained, and the variable in the abnormal distribution interval is replaced by the a% quantile.

7. The method according to any one of claims 1 to 4, wherein: Before extracting features from the plurality of variables in the credit risk assessment multi-source data set, the method includes: sorting the plurality of variables based on the magnitude of the plurality of variables; Obtaining a first variable and a last variable of the plurality of variables; and A standardized value of each of the plurality of variables is calculated based on the first variable and the last variable to replace a value of each of the plurality of variables.

8. The method according to any one of claims 1 to 4, wherein: Before extracting features from the plurality of variables in the credit risk assessment multi-source data set, the method includes: A logarithm is taken of the value of each variable in the plurality of variables to replace the value of each variable in the plurality of variables.

9. The method according to any one of claims 1 to 4, wherein: Before extracting features from the plurality of variables in the credit risk assessment multi-source data set, the method includes: calculating the mean and standard deviation of the plurality of variables; A standard deviation normalized value of each of the plurality of variables is calculated based on the mean value and the standard deviation of the plurality of variables to replace the value of each of the plurality of variables.

10. The method according to any one of claims 1 to 4, wherein: The first feature set includes: basic financial features, industrial chain features, transaction behavior features, text data features and key indicator time series features, among which: The basic financial characteristics include: profitability indicators, debt-paying ability indicators, operating efficiency indicators and / or growth indicators; The industrial chain characteristics include: upstream concentration index, downstream customer stability index and / or supply chain health index; The transaction behavior characteristics include: cash flow volatility index, abnormal transaction frequency index and / or large transaction index; The text data features include: sentiment score, TF-IDF keywords and / or topic information; and The key indicator time series characteristics include: a key indicator moving average, a key indicator year-on-year growth rate, a key indicator month-on-month growth rate and / or a key indicator forecast trend.

Images