Multi-industry adaptive enterprise credit scoring model based on modular stacking and pluggable data system and implementation method

Abstract

The application discloses a multi-industry adaptive enterprise credit scoring model based on a modular stacking and a pluggable data system and an implementation method, relates to the technical field of financial technology and enterprise credit evaluation, and the method comprises the following steps: industry adaptation determination and multi-dimensional feature system construction, data acquisition and preprocessing, feature engineering and fine binning processing, machine learning optimization and credit scoring calculation, credit rating and result output.In the application, by constructing an industry-specific model, fine feature binning and stacking integrated learning, the problems of poor industry adaptability of the existing model, unscientific feature processing, subjective weight distribution and limited practicality of the scoring result are solved, the evaluation accuracy is significantly improved, the feature scientificity is enhanced, the weight credibility is improved, and a multi-dimensional credit report is output, thereby supporting efficient credit decision-making of financial institutions and enterprise credit optimization.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and corporate credit assessment technology, and in particular to a multi-industry adaptable corporate credit scoring model and implementation method based on modular stacking and pluggable data system. Background Technology

[0002] With the continued advancement of inclusive finance policies, financial institutions have an increasingly urgent need for credit assessment of small and micro enterprises. As a core tool for credit decision-making, the accuracy and adaptability of corporate credit scoring models directly affect the efficiency of financial services and risk control capabilities. Existing corporate credit scoring technologies have undergone years of development, forming various technical paths such as rule engines and single machine learning algorithms. However, significant shortcomings remain in areas such as multi-industry adaptability, scientific feature processing, reasonable weight allocation, and the practicality of results, making it difficult to meet the refined needs of current inclusive finance scenarios.

[0003] Currently, most existing models adopt a unified assessment framework and fail to design differentiated indicator systems to address the differences in operating characteristics between manufacturing and non-manufacturing sectors, resulting in significant assessment biases. Feature binning often employs simple statistical methods such as equal frequency and equal intervals, without considering the business logic and risk correlation of enterprises to set grouping intervals, leading to features failing to accurately reflect the degree of influence of indicators on credit risk. Many models rely on a single algorithm to determine indicator weights or directly assign weights based on human experience, lacking a scientific method that combines quantitative and qualitative approaches. Furthermore, existing models mostly output only a single credit score, failing to provide score breakdowns for each dimension, analysis of the positive and negative impacts of indicators, and targeted risk warnings.

[0004] Therefore, this paper proposes a multi-industry adaptable enterprise credit scoring model and implementation method based on modular stacking and pluggable data system to solve the above problems. Summary of the Invention

[0005] The main objective of this invention is to provide a multi-industry adaptable enterprise credit scoring model and implementation method based on modular stacking and pluggable data system, so as to solve the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a multi-industry adaptable enterprise credit scoring method based on modular stacking and a pluggable data system, the method comprising the following steps: S1. Industry Adaptation Judgment and Multi-Dimensional Feature System Construction: Divide manufacturing and non-manufacturing industries, query the complete electricity consumption data of enterprises for the past 12 months, adapt a dedicated scoring model including an electricity consumption level module for manufacturing enterprises, and adapt dedicated scoring models for electricity data and no electricity data for non-manufacturing enterprises respectively; at the same time, build a multi-dimensional feature system that combines general modules and dedicated modules. S2. Data Acquisition and Preprocessing: Integrate public data from multiple departments and supplementary operational data submitted by enterprises themselves. Use a combination of real-time synchronization via API interface and batch acquisition via direct database connection to acquire data. Valid authorization from the enterprise must be obtained before acquisition. Perform multiple preprocessing operations on the acquired data. S3. Feature Engineering and Refined Binning: A hybrid binning strategy, driven by business logic and assisted by statistical algorithms, is used to process features in various dimensions. S4. Logistic regression modeling within the first-level module: For each of the six modules (basic information, operating status, development capability, performance risk, electricity consumption level, and adjustment items), features after S3 binning and assignment are extracted. Historical default labels are used as the target variable, with "good customer" = 1 and "bad customer" = 0. Independent logistic regression models are trained for each module. The model format for each module is as follows: ; in Assign bin values ​​to features within the module. This is the predicted probability of default for this module. This is the intercept term for the m-th module, reflecting the baseline risk level of that module. Let be the regression coefficient of the i-th feature in the m-th module. The module score is obtained through a linear transformation (such as multiplying by 100). As an intermediate output of this module, the logistic regression coefficients of each module record the direction and strength of the influence of internal features on the module score, which are used for subsequent indicator-level interpretation. If a module is deemed unusable due to missing data, it will not participate in the first-level modeling and will be skipped directly. S5. Logistic Regression Stacking Integration and Credit Score Calculation Between Second-Level Modules: Scoring each module output from the first layer (M is the actual number of available modules) as input features, and historical default labels as target variables, a logistic regression model is trained as a meta-model, with the model form as follows: ; in These are the meta-model coefficients, reflecting the impact of module scores on the final default risk. The probability of the final predicted business being a good customer. This is the score output by the m-th module in the first layer; a 5-fold stratified cross-validation is used to prevent overfitting, and the stratification is based on the industry category of the enterprise to ensure that the ratio of good customers to bad customers in each fold is consistent with the overall ratio.

[0007] The final default probability output by the metamodel Mapping to a preset credit score range (e.g., 300-850 points) can be achieved using the following formula: ; in , The calibration parameters were determined by fitting the correlation between credit score and default rate onto historical data; At the same time, the coefficients of the meta-model Normalization is performed to obtain the weights of each module, which are used to explain the degree of contribution of each module to the final credit score. It should be noted that, through the pluggable mechanism, if a module is removed due to missing data, the second-layer model will automatically be trained on the scores of the remaining modules, with the weights adaptively adjusted, without the need for manual intervention. S6. Credit Rating and Output: The K-means clustering algorithm is used to define a five-level credit rating range based on the needs of inclusive finance business. A corresponding risk level label and credit recommendation are set for each rating range. A credit report is generated and output, which includes the total credit score, credit rating, breakdown of scores for each module, positive and negative impact of indicators and targeted risk warnings. At the same time, the scoring results and feedback data are pushed to the machine learning engine. Preferably, the electricity consumption level module includes at least one of the following indicators: the enterprise's average monthly electricity consumption over the past 12 months, monthly electricity consumption fluctuation rate, peak-valley electricity consumption ratio, timely electricity payment rate, average monthly electricity load rate, and year-on-year growth rate of electricity consumption; the general module includes five modules: basic information, operating status, development capability, performance risk, and adjustment items; the dedicated module is a scoring model that is only applicable to the manufacturing industry and non-manufacturing industries with electricity data.

[0008] Preferably, the steps of the multi-step preprocessing operation in S2 are as follows: S21. For continuous features, the random forest regression model is used to predict and fill missing values. For categorical features, the mode method is used to fill missing values. Outliers of numerical indicators are identified by combining the IQR interquartile range with the reasonable range of industry operations. S22. Features with multicollinearity were removed by using the variance inflation factor (VIF) test, and all feature data were mapped to the [0,100] interval using the Min-Max standardization method.

[0009] Preferably, the missing value handling is set with threshold rules: when the missing rate of continuous features exceeds 30%, the review group first judges the correlation between the feature and credit risk in combination with business logic. If the correlation is low, the feature is directly removed; if the correlation is high, a random forest regression model is used to fill in the missing value. When the missing rate of categorical features exceeds 40%, the feature is directly removed. The outlier determination rule of the IQR interquartile range method is: values ​​less than and greater than The samples are outliers, among which The lower quartile, It is the upper quartile. The reasonable operating range for the industry is determined by the review panel based on the average operating data of the corresponding industry over the past three years. For example, samples whose average monthly electricity consumption in the manufacturing industry is 10% lower than the industry average are judged as outliers.

[0010] Preferably, the box splitting process in S3 includes the following steps: S31. Continuous indicators are first binned using the K-means clustering algorithm, and then the binning boundaries are adjusted based on the operating rules and risk correlation of the corresponding industry. Discrete indicators are binned directly according to category attributes. S32. Perform a monotonicity test on all binning results, calculate the WOE weighted evidence value and IV information value for each bin, and remove low-discrimination indicators with IV values ​​less than 0.02; assign values ​​to each bin using a basic score and risk coefficient calibration model, assign static indicators linearly according to interval gradient, assign dynamic indicators using segmented differentiated gradient, and set punitive binning rules for special risk indicators involving default and breach of trust.

[0011] Preferably, the monotonicity test adopts the Kendall rank correlation coefficient test. When the absolute value of the Kendall rank correlation coefficient between the binning result and credit risk is less than 0.3, the binning boundary is readjusted until the monotonicity requirement is met. The The interval division criteria are as follows: It is a low-discrimination index. It is a medium-to-low discrimination index. As a medium discrimination index, For indicators with high discrimination, only indicators with medium discrimination or higher will be retained for subsequent scoring.

[0012] Preferably, in S5, the Stacking integration strategy uses the logistic regression model within each module as the base model, takes the module risk probability output by each base model as the input feature, and uses the logistic regression model as the meta-model to train the final integrated model; a 5-fold stratified cross-validation method is used to prevent model overfitting, and the cross-validation is stratified according to the industry category of the enterprise to ensure data distribution consistency.

[0013] Preferably, the credit report also includes: a visual chart showing the score percentage of each dimension indicator, a list of the top 5 positive contribution indicators, a list of the top 5 negative drag indicators, and credit optimization path suggestions for enterprises, including specific and implementable measures to improve the timeliness of electricity payment, stabilize monthly operating data, and repair credit records.

[0014] A multi-industry adaptable enterprise credit scoring model based on modular stacking and a pluggable data system includes the following modules: Industry Adaptation Determination Module: Based on the classification of a company into manufacturing or non-manufacturing industries, this module queries the company's electricity consumption data for the past 12 months and matches a corresponding exclusive scoring model based on the industry category and the availability of electricity data. Feature system construction module: used to build a multi-dimensional feature system that combines general modules and special modules; Data acquisition and preprocessing module: used to integrate public data from multiple departments and data submitted by enterprises themselves, to collect data and process the collected data; Feature binning module: It adopts a hybrid binning strategy to process features of various dimensions, performs monotonicity test and IV value calculation on binning results, removes low discrimination indicators and assigns values ​​to each bin; Two-layer logistic regression modeling module: The first layer is used to train logistic regression within each module to obtain module scores. The second layer uses the module scores as features to train the logistic regression meta-model, outputs the final credit score, and automatically learns the module weights. Built-in model training unit and cross-validation unit support continuous model iteration and optimization. Rating output module: Used to define a five-level credit rating range using K-means clustering, generate and output a credit report, and push the results to the machine learning engine for model iteration; Data security module: used to perform masking and replacement desensitization processing on the collected sensitive enterprise data; Model Iteration Module: Used to collect updated information on credit feedback data from financial institutions, credit change data from enterprises, and industry operating data every quarter, retrain the two-layer logistic regression model, and adjust the feature binning boundaries.

[0015] The present invention has the following beneficial effects: 1. In this invention, when conducting enterprise credit assessment, by determining industry suitability and constructing a multi-dimensional feature system, and taking into account the differences in operating characteristics between manufacturing and non-manufacturing industries, and combining the availability of electricity data, a scoring model containing a dedicated module for electricity consumption level or a general model without electricity data is adapted for different types of enterprises. This enables the system to overcome the assessment bias caused by the existing unified assessment framework and achieve a more accurate and differentiated assessment of the credit risk of enterprises in different industries that is more in line with the actual industry situation.

[0016] 2. In this invention, when processing credit scoring features and determining weights, a hybrid binning strategy is used to adjust binning boundaries based on business logic, and the Stacking algorithm is used to determine feature binning and indicator weights respectively. This enables the system to overcome the limitations of simple statistical binning and single algorithm or manual experience-based weighting, and to achieve accurate quantification of the degree of influence of features on credit risk and scientific allocation of weights, significantly improving the interpretability and evaluation accuracy of the credit scoring model.

[0017] 3. In this invention, when outputting credit scores and optimizing the model, a credit report is generated that includes the total credit score, rating, breakdown of scores for each module, analysis of the positive and negative impacts of indicators, and targeted optimization suggestions. A continuous learning mechanism including a data security module and a model iteration module is also established. This enables the system to overcome the limitation of existing models that only output a single score, providing users with in-depth insights and actionable suggestions. At the same time, through secure data processing and regular model retraining, the practicality, security, and long-term adaptive evolution capability of the scoring system are ensured. Attached Figure Description

[0018] Figure 1 This is a framework diagram of the multi-industry adaptable enterprise credit scoring method based on modular stacking and pluggable data system of the present invention; Figure 2 This is a framework diagram of the multi-industry adaptable enterprise credit scoring model based on modular stacking and pluggable data system of the present invention. Figure 3 This is a flowchart of the multi-industry adaptable enterprise credit scoring method based on modular stacking and pluggable data system of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1-3 This invention provides a technical solution: a multi-industry adaptable enterprise credit scoring method based on modular stacking and a pluggable data system, the method comprising the following steps: S1. Industry Adaptation Determination and Multi-Dimensional Feature System Construction: The system categorizes enterprises into manufacturing and non-manufacturing sectors, queries complete electricity consumption data for the past 12 months, and adapts a dedicated scoring model including an electricity consumption level module for manufacturing enterprises. The electricity consumption level module is retained for all manufacturing enterprises regardless of whether they have electricity data. For non-manufacturing enterprises, dedicated scoring models based on electricity data availability are provided, along with models for those without. This enables the pluggable and dynamic adaptation of the electricity consumption level module based on electricity data availability. Simultaneously, a multi-dimensional feature system combining general and dedicated modules is constructed, including the following steps: S11. Initiate the assessment, collect basic enterprise information, industry classification, and power data availability: Collect basic enterprise information: Obtain enterprise business registration information through the market supervision department's API interface, including enterprise name, unified social credit code, legal representative, establishment date, registered capital, business scope, etc. Industry classification: Based on the industry classification code for enterprise registration (GB / T 4754-2017), enterprises are classified into manufacturing (Category C) and non-manufacturing (other categories besides Category C). Electricity data availability query: Through the API interface of the power supply agency (such as State Grid and China Southern Power Grid), query the enterprise's electricity consumption data details for the past 12 months (including average monthly electricity consumption, peak and valley electricity consumption ratio, payment records, etc.) to determine whether the enterprise has valid electricity data (defined as having electricity consumption records for 6 or more consecutive months in the past 12 months). S12. Matching the corresponding scoring model: Three types of exclusive scoring models are matched: If the enterprise is a manufacturing enterprise, regardless of whether electricity data is available, the exclusive model for manufacturing enterprises (including the electricity consumption level module) is adapted; if the manufacturing enterprise does not have valid electricity data, the indicators of the electricity consumption level module are replaced by the industry average and used as normal feature input when modeling within the module; If the enterprise is not in the manufacturing sector but has valid electricity data, a dedicated model for non-manufacturing enterprises (with electricity data) (including an electricity consumption level module) will be adapted. Through the above industry adaptation judgment, the power consumption level module is dynamically adapted to be pluggable: for enterprises with power data, the power consumption level module is included in the scoring model; for enterprises without power data (non-manufacturing and without electricity), the module is automatically removed, and the weights are adaptively learned on the remaining modules by the second layer of logistic regression to ensure the integrity and stability of the scoring model.

[0021] Definition of pluggable mechanism: The modular structure of the pluggable index scoring model in this invention supports the dynamic addition or removal of specific modules, such as the electricity consumption module, based on the availability of enterprise data. When the core data required by a module is missing and cannot be effectively replaced, the module is automatically removed from the first-layer modeling and does not participate in the logistic regression training within the module; the second-layer Stacking ensemble model automatically retrains on the scores of the remaining available modules, with adaptive adjustment of module weights, requiring no manual intervention. This mechanism ensures the stability and applicability of the model under different data conditions.

[0022] S2. Data Acquisition and Preprocessing: Integrate public data from multiple departments and supplementary operational data submitted by enterprises themselves. Use a combination of real-time synchronization via API interface and batch acquisition via direct database connection to acquire data. Valid authorization from the enterprise must be obtained before acquisition. Perform multiple preprocessing operations on the acquired data. Multi-source data collection: Integrating public data from multiple departments such as market supervision, taxation, social security, electricity, justice, and real estate, as well as data self-reported by enterprises, using a combination of real-time synchronization via API interfaces and batch collection via direct database connections. Real-time data collection: Dynamic data such as data self-reported by enterprises and records of judicial dishonesty are synchronized in real time using HTTPS API interface, and data transmission is secured using SSL / TLS 1.3 encryption protocol; Batch data collection: For static data such as business registration information, annual tax report data, and social security payment data, batch data is collected every day at midnight using a direct database connection. Data Authorization: Written authorization from the enterprise is required before any public data is collected. The authorization process is completed through an electronic signature system to ensure the compliance of data collection. S3. Feature Engineering and Refined Binning: A hybrid binning strategy, driven by business logic and assisted by statistical algorithms, is used to process features across various dimensions. The binning process includes the following steps: S31. Continuous indicators are first binned using the K-means clustering algorithm, and then the binning boundaries are adjusted based on the corresponding industry's operating patterns and risk correlation. Discrete indicators are binned directly according to category attributes, including the following steps: First, initial binning is performed using the K-means clustering algorithm: the number of clusters K is determined by the elbow rule, which involves trying different K values ​​(usually 3-8), calculating the sum of squared errors (SSE) of each cluster, and selecting the K value corresponding to the obvious inflection point (elbow) of the SSE decline trend as the optimal number of clusters. After obtaining the initial binning, the binning boundaries are manually adjusted based on industry operating rules and expert experience. For example, for the monthly electricity consumption index of manufacturing enterprises, after the initial binning by K-means clustering, enterprises with monthly electricity consumption ≥50,000 kWh are classified into a separate category (corresponding to large manufacturing enterprises), and enterprises with monthly electricity consumption <1,000 kWh are classified into another category (corresponding to small workshop-style enterprises), ensuring that the binning results conform to the business logic. Discrete indicators are binned: binned directly according to category attributes. For example, judicial dishonesty records are divided into three categories: no dishonesty record, 1 dishonesty record, and 2 or more dishonesty records. Electricity payment timeliness rate is divided into three categories: ≥95%, 90%-95%, and <90%. S32. Perform a monotonicity test on all binning results, calculate the WOE weighted evidence value and IV information value for each bin, and remove low-discrimination indicators with IV values ​​less than 0.02: Monotonicity test: Calculate the WOE value for each bin and use Spearman's rank correlation coefficient to test the monotonicity (increasing or decreasing) of the WOE value with the bin number. If the absolute value of the correlation coefficient is less than 0.3 and the p value is greater than 0.05, the monotonicity is considered not significant and the bin boundaries need to be adjusted. Discrimination test: Calculate the discrimination of each feature. value, The standard formula for calculating (information value) is: ; In this model, a good customer is defined as a company that has not defaulted (label=1), and a bad customer is defined as a company that has defaulted (label=0). The WOE (Warranty of Default) is calculated using the following formula: a higher WOE value indicates a higher proportion of good customers in that bin, i.e., lower risk. It represents the proportion of good customers in the i-th sub-box to the total number of good customers. It represents the proportion of bad customers in the i-th bin to the total number of bad customers. The standard formula for calculating (weight of evidence) is: ; Since the WOE (Warranty of Ownership) direction aligns with the proportion of good customers, the IV (Indicator Value) reflects the feature's ability to distinguish between good and bad customers; a higher value indicates a stronger ability of the feature to differentiate credit risk. The evaluation criteria are as follows: (No distinguishability, discard); (Weak discrimination); (Medium discrimination); (Strong differentiation), usually excluded based on business needs. Indicators with values ​​less than 0.02; S33. A basic score and risk coefficient calibration mode is used to assign values ​​to each bin. Static indicators are assigned values ​​linearly according to interval gradients, while dynamic indicators are assigned values ​​using segmented differentiated gradients. Punitive bin rules are set for special risk indicators involving default and breach of trust. Linear mapping of base scores: Each sub-box is linearly mapped to a base score range according to its business performance (e.g., electricity consumption from low to high, credit record from none to some). For example, for positive indicators (the higher the value, the better), the best sub-box is mapped to 85-100 points, and the worst sub-box is mapped to 0-40 points. Risk coefficient Calibration: The base score is fine-tuned based on the WOE value of each bin, using the following formula: ,in It is a scaling factor (such as 10) used to scale The logarithmic odds effect is converted into a fractional adjustment. A positive value adds points, and a negative value subtracts points. Final score range control: Extreme value processing is performed on the above calibrated scores to ensure that the final scores of all sub-boxes fall within a uniform preset range (such as 0-100 points); Punitive binning rules: For special risk indicators (such as judicial dishonesty records, loan overdue records), punitive binning rules are set: After completing the above basic score and risk coefficient calibration, an additional fixed deduction (such as -20 points or -15 points) is directly applied to specific high-risk bins (such as having 2 or more dishonesty records, loan overdue for ≥30 days). The deduction is significantly greater than the regular calibration range to reflect the principle of veto or key warning in business.

[0023] S4. Logistic regression modeling within the first-level module: For each of the six modules (basic information, operating status, development capability, performance risk, electricity consumption level, and adjustment items), features after S3 binning and assignment are extracted. Using historical default labels (good customer = 1, bad customer = 0) as the target variable, independent logistic regression models are trained. The model form for each module is as follows: ; in Assign bin values ​​to features within the module. This is the predicted probability of default for this module. This is the intercept term for the m-th module, reflecting the baseline risk level of that module. Let be the regression coefficient of the i-th feature in the m-th module; The module score is obtained through a linear transformation (such as multiplying by 100). As an intermediate output of this module, the logistic regression coefficients of each module record the direction and strength of the influence of internal features on the module score, which are used for subsequent indicator-level interpretation. If a module is deemed unusable due to missing data, it will not participate in the first-level modeling and will be skipped.

[0024] S5. Logistic Regression Stacking Integration and Credit Score Calculation Between Second-Level Modules: Scoring each module output from the first layer Using M (where M is the actual number of available modules) as input features and historical default labels as the target variable, a logistic regression model is trained as the meta-model. The model form is: ; in These are the meta-model coefficients, reflecting the impact of module scores on the final default risk. The probability of the final predicted business being a good customer. This is the score output by the m-th module in the first layer; a 5-fold stratified cross-validation is used to prevent overfitting, and the stratification is based on the industry category of the enterprise to ensure that the ratio of good customers to bad customers in each fold is consistent with the overall ratio. The final default probability output by the metamodel Mapping to a preset credit score range (e.g., 300-850 points), the mapping formula is as follows: ; in , The calibration parameters were determined by fitting the correlation between credit score and default rate onto historical data; At the same time, the coefficients of the meta-model Normalization is performed to obtain the weights of each module, which are used to explain the contribution of each module to the final credit score. It should be noted that, through a pluggable mechanism, if a module is removed due to missing data, the second-layer model will automatically train on the scores of the remaining modules, and the weights will be adaptively adjusted without manual intervention.

[0025] S6. Credit Rating and Output: A five-level credit rating range is defined using the K-means clustering algorithm combined with the needs of inclusive finance. A corresponding risk level label and credit recommendation are set for each rating range. A credit report is generated, including the total credit score, credit rating, breakdown of scores for each module, the positive and negative impact of indicators, and targeted risk warnings. (The credit report also includes: a visualization chart of the score percentage for each dimension indicator, a list of the top 5 positive contributing indicators, a list of the top 5 negative drag indicators, and credit optimization path suggestions for enterprises, including specific and implementable measures such as improving electricity payment timeliness, stabilizing monthly operating data, and repairing credit records.) The scoring results and feedback data are then pushed to the machine learning engine, including the following steps: K-means clustering was used in conjunction with business requirements to define a five-level credit rating range: Clustering input feature construction: The final total credit score of historical enterprises is used as the main clustering feature. To further enhance the risk differentiation ability of the rating, the coefficient of variation of the core module score (standard deviation / mean of each module score) can be added as the second dimension feature to capture the stability of the enterprise credit structure. Clustering process and boundary determination: The clustering number K=5 was set, and historical corporate credit scores data from the most recent 12 months (sample size ≥1000) were used for clustering. Based on the clustering results, the median of adjacent cluster centers was used as the initial boundary of the rating intervals. Subsequently, considering the risk tolerance requirements of inclusive finance (e.g., setting the proportion of high-risk customers to no more than 3%), the boundaries of each interval were fine-tuned to ensure that the rating distribution met business expectations. Rating labels and credit recommendations mapping: The five ranges are mapped from high to low as: AAA (excellent), AA (good), A (average), B (watch list), and C (risky), and correspond to credit strategy recommendations of active credit granting, recommended credit granting, moderate credit granting, cautious credit granting, and restricted credit granting, respectively. Generate a multi-dimensional credit report: Generate a credit report that includes the following: Key information: Company name, Unified Social Credit Code, Total Credit Score, Credit Rating, Risk Level Label, Credit Recommendation; Dimensional score breakdown: The scores of each module (such as basic information, operating status, electricity consumption level, etc.) and the percentile ranking among companies in the same industry (e.g., the operating status module score is 92 points, which is higher than 90% of companies in the industry). Risk Warnings and Optimization Suggestions: Specific and quantifiable risk warnings and optimization suggestions are provided for negative impact indicators (e.g., if your electricity volatility reaches 25%, which is 10% higher than the industry average, there may be a risk of production instability; it is recommended to optimize the production plan, which is expected to reduce volatility to 15% and improve credit score by 8-10 points). The expected score improvement is calculated based on the logistic regression model score corresponding to the score increment for each preset sub-level improvement of this indicator. Output reports and push them to application scenarios; the results are fed back to the machine learning engine to optimize the model. Report Output: Output credit reports in formats such as PDF and HTML, and push them to inclusive financial service platforms, financial institutions, or corporate users via API interfaces; Results Feedback and Model Iteration: The final credit decision results and loan performance (such as whether the loan is overdue and the number of overdue days) of the enterprise are used as new label data and fed back to the machine learning engine along with the corresponding feature data. The system establishes two model iteration mechanisms: regular (such as quarterly) and triggered (such as when the bad debt rate of a certain industry fluctuates significantly). Regular iteration: Collect new data every quarter, first re-perform the monotonicity test and IV value calculation of feature binning, and adjust the binning boundaries if necessary; then, retrain the Stacking ensemble model using all available data. Model Validation and Deployment: The new model after iteration needs to be tested on an independent validation set. Its performance (such as KS value, AUC) is not lower than that of the original model and it passes the business logic verification before it can be deployed to replace the old model and complete the closed-loop optimization.

[0026] The electricity consumption level module includes at least one of the following indicators: the enterprise's average monthly electricity consumption over the past 12 months, monthly electricity consumption volatility, peak-valley electricity consumption ratio, timely electricity payment rate, average monthly electricity load rate, and year-on-year growth rate of electricity consumption; the general module includes five modules: basic information, operating status, development capability, performance risk, and adjustment items; the exclusive module is a scoring model that is only applicable to the manufacturing industry and non-manufacturing industries with electricity data. General module characteristics: Basic information module: Features such as the company's establishment year, paid-in capital ratio, shareholder background (whether it is a state-owned enterprise / listed company), and credit record of the legal representative; Operating performance module: Features such as revenue growth rate, net profit margin, debt-to-equity ratio, cash flow turnover rate, and tax payment growth rate over the past 3 years; Development Capability Module: Characteristics of R&D investment ratio, employee growth rate, market share growth rate, and new business revenue ratio; Performance Risk Module: Judicial Dishonesty Records, Number of Contract Defaults, Loan Overdue Records, Supplier Evaluation Characteristics; Adjustment Modules: Subsidies affected by the pandemic, eligibility for policy support, and characteristics of sudden risk events (such as fires and environmental penalties); Dedicated module features (electricity consumption level module): Only compatible with manufacturing and non-manufacturing (with electricity data) models, including average monthly electricity consumption over the past 12 months, electricity consumption volatility, peak-valley electricity consumption ratio, timely electricity payment rate, and electricity consumption growth rate features.

[0027] Missing value handling threshold rules: When the missing rate of continuous features exceeds 30%, the review group first judges the correlation between the feature and credit risk based on business logic. If the correlation is low, the feature is directly removed. If the correlation is high, a random forest regression model is used to fill in the missing value. When the missing rate of categorical features exceeds 40%, the feature is directly removed. The outlier detection rule for the IQR interquartile range method is: values ​​less than... and greater than The samples are outliers, among which The lower quartile, It is the upper quartile. The reasonable operating range for an industry is determined by the review panel based on the industry's average operating data over the past three years. For example, samples whose average monthly electricity consumption in the manufacturing sector is 10% lower than the industry average are considered outliers.

[0028] The steps of the multi-step preprocessing operation in S2 are as follows: S21. For continuous features, missing values ​​are predicted and imputed using a random forest regression model; for categorical features, missing values ​​are imputed using the mode method; and outliers of numerical indicators are identified using the IQR interquartile range combined with reasonable operating ranges for the industry. This includes the following steps: Missing value handling: For continuous features (such as revenue, net profit, and average monthly electricity consumption), random forest regression is used for prediction and imputation. Specifically, the feature is used as the target variable. 3-5 features that are strongly correlated with the target variable in business logic and statistically correlated (such as Pearson correlation coefficient absolute value > 0.3) are selected from the dataset as input variables (e.g., tax amount, number of social security contributors, electricity consumption level, etc. when revenue is missing). A random forest regression model is trained on non-missing data, and then the trained model is used to predict missing values. For categorical features (such as industry category and judicial dishonesty records), the mode is used for imputation. If the mode is not unique, the category with the highest business share is used for imputation. Outlier Handling: Numerical indicators are identified by using the IQR (Interquartile Range) method combined with a reasonable business range. First, the upper quartile of the indicator is calculated. Lower quartiles Interquartile range Exceeding or The values ​​are marked as potential outliers; then a second judgment is made based on the reasonable business range, which is determined by the review team according to industry white papers and historical data distribution. Percentiles or clear industry regulatory standards are used to define the range; for example, if a company's average monthly electricity consumption exceeds the range but it is in its peak production season (such as manufacturing companies before the Spring Festival), it will not be marked as an outlier; outliers are replaced by the predicted values ​​obtained from the random forest regression model trained on the feature, rather than being directly deleted, using the data of the sample on other non-outlier features. S22. Features exhibiting multicollinearity are eliminated using the variance inflation factor (VIF) test. Min-Max standardization is then applied to map all feature data to the [0, 100] interval, including the following steps: Multicollinearity removal: Multicollinearity is removed using the VIF (variance inflation factor) test. The correct formula for calculating VIF is: ; in It is the first The coefficient of determination obtained by performing linear regression on a feature as the dependent variable and all other features is... If a feature is found to have severe multicollinearity with other features, it will be removed; when removing features, features with clear business meaning will be retained first. Data standardization: Min-Max standardization is used to map all features to the interval [0, 100]. The calculation formula is as follows: ; in These are the original eigenvalues. The minimum value of the feature. The maximum value of the feature is used. The maximum and minimum values ​​here are calculated from the dataset after outliers have been removed to prevent extreme values ​​from distorting the standardization results.

[0029] A multi-industry adaptable enterprise credit scoring model based on modular stacking and a pluggable data system includes the following modules: Industry Adaptability Determination Module: This module is used to query the electricity consumption data of enterprises over the past 12 months based on whether the enterprise belongs to the manufacturing or non-manufacturing industry. It matches the corresponding exclusive scoring model according to the industry category and the availability of electricity data. This module has a built-in GB / T 4754-2017 industry classification standard library, which can automatically identify the industry category of the enterprise and realize real-time data interaction with the API interface of the power supply agency. Feature system construction module: used to build a multi-dimensional feature system that combines general modules and special modules. This module has a built-in feature dimension library for various industries and can automatically select corresponding features according to the matching scoring model. Data Acquisition and Preprocessing Module: Used to integrate public data from multiple departments and data submitted by enterprises themselves, collect data and process the collected data, supports both real-time acquisition and batch acquisition modes, and has built-in SSL encrypted transmission and AES encrypted storage functions; Feature binning module: It adopts a hybrid binning strategy to process features of various dimensions, performs monotonicity test and IV value calculation on the binning results, removes low discrimination indicators and assigns values ​​to each bin. This module has built-in K-means clustering algorithm, WOE / IV value calculation unit and monotonicity test unit, and can automatically generate bin boundaries and perform validity test. Module weight generation module: used to normalize the regression coefficients of the second-level meta-model to obtain the contribution weight of each module to the final credit score, supporting module-level risk interpretation; Machine Learning Optimization Module: This module is used to build a framework for parallel implementation of logistic regression and random forest models. It adopts a stacking ensemble strategy to fuse the models, optimizes parameters through Bayesian optimization and performs 5-fold hierarchical cross-validation, and calculates the total credit score of enterprises by combining a weight system. This module has built-in model training unit, Bayesian parameter tuning unit, and cross-validation unit to support continuous iteration and optimization of the model. Rating Output Module: This module is used to define a five-level credit rating range using K-means clustering, generate a credit report and output it. At the same time, the results are pushed to the machine learning engine for model iteration. This module has built-in K-means clustering unit, report generation unit, and API push unit, and the report template and push method can be customized. Data security module: Used to perform masking and replacement desensitization processing on the collected sensitive enterprise data; this module includes a data encryption unit, a desensitization processing unit, and an authorization management unit to ensure the compliance and security of the entire process of data collection, storage, and transmission, and to meet the requirements of laws and regulations such as the Personal Information Protection Law and the Data Security Law; Model Iteration Module: Used to collect updated information on credit feedback data from financial institutions, credit change data from enterprises, and industry operating data every quarter, retrain the machine learning model, and adjust the feature binning boundaries.

[0030] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for implementing multi-industry adaptable enterprise credit scoring based on modular stacking and a pluggable data system, characterized in that: The method includes the following steps: S1. Industry Adaptation Judgment and Multi-Dimensional Feature System Construction: Divide manufacturing and non-manufacturing industries, query the complete electricity consumption data of enterprises for the past 12 months, adapt a dedicated scoring model including an electricity consumption level module for manufacturing enterprises, and adapt dedicated scoring models for electricity data and no electricity data for non-manufacturing enterprises respectively; at the same time, build a multi-dimensional feature system that combines general modules and dedicated modules. S2. Data Acquisition and Preprocessing: Integrate public data from multiple departments and supplementary operational data submitted by enterprises themselves. Use a combination of real-time synchronization via API interface and batch acquisition via direct database connection to acquire data. Valid authorization from the enterprise must be obtained before acquisition. Perform multiple preprocessing operations on the acquired data. S3. Feature Engineering and Refined Binning: A hybrid binning strategy, driven by business logic and assisted by statistical algorithms, is used to process features in various dimensions. S4. Logistic regression modeling within the first-level module: For the six modules, including basic information, operating status, development capability, performance risk, electricity consumption level, and adjustment items, features after S3 binning and assignment are extracted from each module. Historical default labels are used as the target variable, with "good customer" = 1 and "bad customer" = 0. Independent logistic regression models are trained for each module, with the model format as follows: ； in Assign bin values ​​to features within the module. This is the predicted probability of default for this module. This is the intercept term for the m-th module, reflecting the baseline risk level of that module. Let be the regression coefficient of the i-th feature in the m-th module; S5. Logistic Regression Stacking Integration and Credit Score Calculation Between Second-Level Modules: Scoring each module output from the first layer M represents the number of actual available modules, which are used as input features. A logistic regression model is trained as the meta-model, with historical default labels as the target variable. The model form is: ； in These are the meta-model coefficients, reflecting the impact of module scores on the final default risk. This represents the probability that the predicted company will be a good customer, with a value ranging from 0 to 1. This is the score output by the m-th module in the first layer; The final default probability output by the metamodel Mapping to a preset credit score range, the mapping formula is as follows: ； in , The calibration parameters were determined by fitting the correlation between credit score and default rate onto historical data; At the same time, the coefficients of the meta-model Normalization is performed to obtain the weights of each module, which are used to explain the degree of contribution of each module to the final credit score. S6. Credit Rating and Output: The K-means clustering algorithm is used to define a five-level credit rating range based on the needs of inclusive finance business. A corresponding risk level label and credit recommendation are set for each rating range. A credit report is generated and output, which includes the total credit score, credit rating, breakdown of scores for each module, positive and negative impact of indicators and targeted risk warnings. At the same time, the scoring results and feedback data are pushed to the machine learning engine.

2. The method for implementing multi-industry adaptable enterprise credit scoring based on modular stacking and a pluggable data system as described in claim 1, characterized in that: The electricity consumption level module includes at least one of the following indicators: the enterprise's average monthly electricity consumption over the past 12 months, monthly electricity consumption volatility, peak-valley electricity consumption ratio, timely electricity payment rate, average monthly electricity load rate, and year-on-year growth rate of electricity consumption; the general module includes five modules: basic information, operating status, development capability, performance risk, and adjustment items; the exclusive module is a scoring model that is only applicable to the manufacturing industry and non-manufacturing industries with electricity data.

3. The method for implementing multi-industry adaptable enterprise credit scoring based on modular stacking and a pluggable data system as described in claim 1, characterized in that: The steps of the multi-step preprocessing operation in S2 are as follows: S21. For continuous features, the random forest regression model is used to predict and fill missing values. For categorical features, the mode method is used to fill missing values. Outliers of numerical indicators are identified by combining the IQR interquartile range with the reasonable range of industry operations. S22. Features with multicollinearity were removed by using the variance inflation factor (VIF) test, and all feature data were mapped to the [0,100] interval using the Min-Max standardization method.

4. The method for implementing multi-industry adaptable enterprise credit scoring based on modular stacking and a pluggable data system according to claim 3, characterized in that: The missing value handling settings include the following threshold rules: when the missing rate of a continuous feature exceeds 30%, the review group first determines the correlation between the feature and credit risk based on business logic. If the correlation is low, the feature is directly removed; if the correlation is high, a random forest regression model is used to fill in the missing value. When the missing rate of a categorical feature exceeds 40%, the feature is directly removed. The outlier determination rule of the IQR interquartile range method is: values ​​less than and greater than The samples are outliers, among which The lower quartile, It is the upper quartile. The reasonable operating range for the industry is determined by the review panel based on the average operating data of the corresponding industry over the past three years.

5. The method for implementing multi-industry adaptable enterprise credit scoring based on modular stacking and a pluggable data system according to claim 1, characterized in that: The box splitting process in S3 includes the following steps: S31. Continuous indicators are first binned using the K-means clustering algorithm, and then the binning boundaries are adjusted based on the operating rules and risk correlation of the corresponding industry. Discrete indicators are binned directly according to category attributes. S32. Perform a monotonicity test on all binning results, calculate the WOE weighted evidence value and IV information value for each bin, and remove low discrimination indicators with an IV value less than 0.

02. S33. A basic score and risk coefficient calibration model is used to assign values ​​to each bin. Static indicators are assigned values ​​linearly according to interval gradients, while dynamic indicators are assigned values ​​using segmented differentiated gradients. Punitive bin rules are set for special risk indicators involving default and breach of trust.

6. The method for implementing multi-industry adaptable enterprise credit scoring based on modular stacking and a pluggable data system according to claim 5, characterized in that: The monotonicity test employs the Kendall rank correlation coefficient test. When the absolute value of the Kendall rank correlation coefficient between the binning results and credit risk is less than 0.3, the binning boundaries are readjusted until the monotonicity requirement is met. The interval division criteria are as follows: It is a low-discrimination index. It is a medium-to-low discrimination index. As a medium discrimination index, For indicators with high discrimination, only indicators with medium discrimination or higher will be retained for subsequent scoring.

7. The method for implementing multi-industry adaptable enterprise credit scoring based on modular stacking and a pluggable data system according to claim 1, characterized in that: The Stacking integration strategy in S5 uses the logistic regression model within each module as the base model, takes the module risk probability output by each base model as the input feature, and uses the logistic regression model as the meta-model to train the final integration model. A 5-fold stratified cross-validation method is used to prevent model overfitting, and the cross-validation is stratified according to the industry category of the enterprise to ensure the consistency of data distribution.

8. The method for implementing multi-industry adaptable enterprise credit scoring based on modular stacking and a pluggable data system according to claim 1, characterized in that: The credit report also includes: a visual chart showing the score percentage of each dimension indicator, a list of the top 5 positive contribution indicators, a list of the top 5 negative drag indicators, and credit optimization path suggestions for enterprises, including specific and implementable measures to improve the timeliness of electricity payment, stabilize monthly operating data, and repair credit records.

9. A multi-industry adaptable enterprise credit scoring model based on modular stacking and a pluggable data system, comprising the multi-industry adaptable enterprise credit scoring implementation method based on modular stacking and a pluggable data system as described in any one of claims 1-8, characterized in that... Includes the following modules: Industry Adaptation Determination Module: Based on the classification of a company into manufacturing or non-manufacturing industries, this module queries the company's electricity consumption data for the past 12 months and matches a corresponding exclusive scoring model based on the industry category and the availability of electricity data. Feature system construction module: used to build a multi-dimensional feature system that combines general modules and special modules; Data acquisition and preprocessing module: used to integrate public data from multiple departments and data submitted by enterprises themselves, to collect data and process the collected data; Feature binning module: It adopts a hybrid binning strategy to process features of various dimensions, performs monotonicity test and IV value calculation on binning results, removes low discrimination indicators and assigns values ​​to each bin; Two-layer logistic regression modeling module: The first layer is used to train logistic regression within each module to obtain module scores. The second layer uses the module scores as features to train the logistic regression meta-model, outputs the final credit score, and automatically learns the module weights. Built-in model training unit and cross-validation unit support continuous model iteration and optimization. Rating output module: Used to define a five-level credit rating range using K-means clustering, generate and output a credit report, and push the results to the machine learning engine for model iteration; Data security module: used to perform masking and replacement desensitization processing on the collected sensitive enterprise data; Model Iteration Module: Used to collect updated information on credit feedback data from financial institutions, credit change data from enterprises, and industry operating data every quarter, retrain the two-layer logistic regression model, and adjust the feature binning boundaries.

Images