Method for evaluating loan availability of small and micro enterprises

Abstract

The invention discloses a small and micro enterprise loan availability evaluation method, which belongs to the technical field of data processing, and integrates unstructured operation and semi-structured behavior data, establishes a dynamic data confidence evaluation matrix, and starts artificial recheck to guarantee data quality. And constructing a three-layer fuzzy membership function system, and generating a standardized score by using an improved interval type-2 fuzzy set and an alpha-cut algorithm. Defining an interest rate sensitive dimension, generating an initial interest rate hierarchy through an improved spectral clustering algorithm, and predicting operation fluctuation and adjusting parameters by using a double-layer LSTM (Long Short Term Memory). And through IPC cross validation, carrying out residual analysis on the fuzzy data and on-site collected data, generating a joint confidence interval through D-S evidence fusion, and evaluating loan availability and outputting a suggestion report on the basis of the joint confidence interval. According to the method, the existing evaluation data processing problem is effectively solved, evaluation scientificity, accuracy and comprehensiveness are improved, a reliable decision basis is provided for financial institutions, and accurate customer layering and product adaptation are assisted to be achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly relates to a method for evaluating the loan availability of small and micro enterprises. Background Art

[0002] In today's financial market environment, small and micro enterprises, as an important part of economic development, have an extremely urgent need for funds. However, due to their small scale, relatively low transparency of financial information, and relatively poor business stability, small and micro enterprises face many challenges in obtaining loans. Accurately evaluating the loan availability of small and micro enterprises has become a key issue for financial institutions and related fields.

[0003] Currently, in the process of evaluating the loan availability of small and micro enterprises, data integration is the basic step, and data is usually divided into explicit data and fuzzy data. Explicit data has certainty. For example, the public security blacklist data can confirm that a customer has a public security record once it is found. Such data can make customer information more transparent, and banking institutions generally use it, and business personnel are also more likely to accept it because it can be accurately verified. Fuzzy data, on the other hand, has a certain degree of discrimination and is generally data obtained after being processed by a big data model by a data company. Such data cannot directly determine whether a customer is of high quality or poor quality.

[0004] In the development process of hierarchical classification of fuzzy data, it has mainly gone through the following three stages:

[0005] The stage of credit rating methods mainly based on expert experience: This stage mainly relies on the financial factor analysis method. By analyzing various economic and financial indicators in the enterprise financial statements, static comparison and dynamic trend analysis are carried out to obtain the current operating conditions and debt repayment ability of the enterprise. The evaluation elements cover five aspects: the character, ability, capital, collateral, and conditions of the evaluation object, and attempt to comprehensively consider from dimensions such as the repayment willingness, repayment ability, financial status, whether there is collateral, and other external conditions (such as financial crises), and combine the analysis of experts to draw evaluation conclusions. However, this method overly relies on the subjective judgment of experts and lacks quantitative accuracy and objectivity. Different experts may have large deviations in evaluation results due to differences in experience and cognition, and it is difficult to meet the needs of accurately evaluating small and micro enterprises in the increasingly complex financial market.

[0006] Credit rating stage combining expert experience and mathematical models: For example, the Analytic Hierarchy Process combines quantitative and qualitative indicators, expressing subjective judgments in quantitative form, making the evaluation process more scientific and systematic. The fuzzy comprehensive evaluation method uses fuzzy mathematics and combines the Analytic Hierarchy Process to determine the weights of evaluation indicators, and obtains the comprehensive evaluation results by calculating the membership degrees of the evaluation objects. Although this method improves the accuracy and objectivity of the evaluation to a certain extent, there are still limitations. On the one hand, expert experience still has a great influence in determining indicator weights and other aspects, and it is difficult to completely eliminate the interference of subjective factors; on the other hand, some mathematical models themselves may also have problems of incomplete consideration of complex real situations, resulting in the evaluation results not being able to accurately reflect the true risk status of small and micro enterprises.

[0007] Rating method stage based on mathematical models: With the development of technology, rating methods based on mathematical models such as Logit models, KMV models, and CreditMetrics models have emerged. These models rely on the financial and operating data of enterprises for evaluation. However, the existing operating data in enterprises often lacks data related to economic cycle fluctuations, and economic fluctuations will have a significant impact on the rating quality. For example, during the economic boom period, some small and micro enterprises may perform well, but during the economic recession, their repayment ability may be severely impacted. Due to the lack of effective consideration of economic cycle fluctuations in the existing models, they cannot accurately evaluate this risk change, resulting in the rating results not matching the actual situation.

[0008] In addition, in the existing evaluation system, although the IPC model (an evaluation model that emphasizes on-site investigation and comprehensive analysis of enterprise financial and non-financial information) is introduced, business personnel have a relatively high acceptance of explicit data, but there are still deficiencies in the processing of fuzzy data. When fuzzy data is combined with IPC technology, its probability inference characteristics cannot be fully utilized, and accurate stratification and classification of customers and product adaptation cannot be achieved.

[0009] In summary, the existing evaluation methods for the loan availability of small and micro enterprises have many problems in dealing with fuzzy data, coping with economic cycle fluctuations, integrating multi-source data, and achieving accurate customer stratification and product adaptation. There is an urgent need for a more scientific, accurate, and comprehensive data processing and evaluation method to improve the accuracy and objectivity of small and micro enterprise loan evaluation in data processing and meet the development needs of the financial market. Summary of the Invention

[0010] In view of the deficiencies of the prior art, the present invention provides an evaluation method for the loan availability of small and micro enterprises, which effectively solves the data processing problems of the existing methods in dealing with fuzzy data, coping with economic cycle fluctuations, integrating multi-source data, and achieving accurate customer stratification and product adaptation. It significantly improves the scientificity, accuracy, and comprehensiveness of the evaluation data, providing a more reliable data decision-making basis for financial institutions.

[0011] To achieve the above objectives, the present invention is realized through the following technical solutions: An evaluation method for the loan availability of small and micro enterprises, comprising the steps of: Integrating unstructured business data and semi-structured behavioral data; Establishing a dynamic data confidence evaluation matrix, and initiating an artificial review process for fields with a missing rate greater than a threshold or an abnormal fluctuation exceeding the threshold; Constructing a three-layer fuzzy membership function system; Using an improved interval type-2 fuzzy set to process data uncertainty, and generating a 0-1 standardized score through the α-cut algorithm; Defining the interest rate-sensitive dimensions that affect loan availability: the interest rate-sensitive dimensions include the capital turnover elasticity coefficient, the leverage tolerance, and the price elasticity test value; Applying an improved spectral clustering algorithm to generate an initial interest rate level in the reduced-dimensional feature space to achieve the mapping of interest rate-sensitive classification and risk dimensions; Establishing a two-layer LSTM time memory unit to predict the business fluctuations in the next two quarters and dynamically adjust the membership function parameters; Performing residual analysis on the monthly average sales volume calculated from fuzzy data, the tax return forms and bank statements collected on-site by IPC. When the data deviation degree is greater than the threshold, a credibility attenuation function for data adjustment is triggered to reduce the weight of fuzzy data and increase the score of the manual verification item; Performing D-S evidence fusion on the preliminary classification results of the fuzzy data layer and the IPC verification conclusions to generate a joint confidence interval;

[0012] Evaluating the loan availability of small and micro enterprises based on the joint confidence interval, and outputting a recommendation report on whether to approve the enterprise loan.

[0013] Optionally, a dynamic data confidence evaluation matrix is constructed. The rows of the dynamic data confidence evaluation matrix represent different data fields, and the columns record the values of various confidence evaluation indicators related to the fields.

[0014] Optionally, the three-layer fuzzy membership function system includes a basic layer fuzzy membership function, a derivative layer fuzzy membership function, and a comprehensive layer fuzzy membership function.

[0015] Optionally, an improved interval type-2 fuzzy set is adopted to process data uncertainty, and 0-1 normalized scores are generated through the α-cut algorithm. Specifically: for the transaction frequency stability and transaction amount stability at the basic layer, the cash flow cycle matching degree at the derivative layer, and the industry prosperity sensitivity coefficient index at the comprehensive layer, corresponding improved interval type-2 fuzzy sets are constructed respectively; For each data point, calculate the membership degree interval in each fuzzy set; Through the α-cut algorithm, convert the improved interval type-2 fuzzy set into an ordinary set; then, generate 0-1 normalized scores according to the characteristics of the ordinary set.

[0016] Optionally, the capital turnover elasticity coefficient is calculated by the ratio of the accounts receivable turnover days to the accounts payable turnover days, the leverage tolerance is calculated by the invisible debt coverage ratio, and the price elasticity test value is calculated by the historical interest rate adjustment response rate.

[0017] Optionally, an initial interest rate level is generated in the dimensionality-reduced feature space to achieve dimensionality reduction in the mapping between interest rate-sensitive classification and risk dimensions. The dimensionality reduction method can choose principal component analysis or linear discriminant analysis; The improved spectral clustering algorithm includes: In the dimensionality-reduced feature space, construct a similarity matrix according to the data characteristics of small and micro enterprises; Based on the constructed similarity matrix, calculate the Laplacian matrix; Perform eigenvalue decomposition on the Laplacian matrix, select the eigenvectors corresponding to the first k smallest non-zero eigenvalues, and map the data points to a k-dimensional low-dimensional space; In the low-dimensional embedding space, use the K-means algorithm to cluster and partition the data points; Correspond the clustering results to different initial interest rate levels.

[0018] Optionally, establish a two-layer LSTM time memory unit to predict the business fluctuations in the next two quarters. Dynamically adjusting the membership function parameters includes: First, input the time series data of the business data of small and micro enterprises into the first-layer LSTM unit in chronological order. The first-layer LSTM unit processes the data at each time step through the coordinated work of the input gate, forget gate, and output gate, and extracts the short-term fluctuation characteristics in the data; The feature information processed by the first-layer LSTM unit is passed as input to the second-layer LSTM unit; the second-layer LSTM unit further mines the long-term trends and complex patterns in the data based on these inputs; Through the layer-by-layer calculation and information transfer of the two-layer LSTM units, predict the business data at each future time step; Dynamically adjust the membership function parameters: Re-evaluate the enterprise's cash inflow and outflow cycles based on the predicted operating data for the next two quarters; Dynamically adjust the membership function parameters according to the re-evaluated cash flow cycle. The adjusted membership function parameters are specifically the parameters of the derivative layer fuzzy membership function.

[0019] Optionally, perform D-S evidence fusion on the preliminary classification results of the fuzzy data layer and the IPC verification conclusion to generate a combined confidence interval, specifically including: Compare the basic probability assignments of the preliminary classification results of the fuzzy data layer and the IPC verification conclusion to detect conflicts between the two; By calculating the conflict coefficient to quantify the degree of conflict, where A i and B j are different events in the fuzzy data and the IPC verification conclusion respectively; A i represents the event in the preliminary classification result of the fuzzy data, and B j represents the event in the IPC verification conclusion. When the intersection of A i and B j is an empty set, it indicates that these two events conflict with each other. Multiply and accumulate their corresponding basic probability assignment values. The larger the conflict coefficient k value, the more serious the conflict; When the conflict coefficient k is small, the Dempster combination rule can be directly used for fusion; for events with a non-empty intersection, the combined basic probability assignment is: ; Based on the combined basic probability assignment, determine the parameters of the combined confidence interval.

[0020] Optionally, evaluate the availability of small and micro enterprise loans based on the combined confidence interval, and output a recommendation report on whether to approve the enterprise loan, including: dividing the risk levels of small and micro enterprises into three levels: low risk, medium risk, and high risk according to the risk probability range determined by the combined confidence interval; generating a recommendation report including the risk assessment results.

[0021] This application also discloses an electronic device, which includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute one or more steps of the method for evaluating the availability of small and micro enterprise loans.

[0022] The present invention provides an evaluation method for the loan availability of small and micro enterprises, having the following beneficial effects:

[0023] The evaluation method for the loan availability of small and micro enterprises proposed in this application effectively solves the data processing problems of existing methods in dealing with fuzzy data, coping with economic cycle fluctuations, integrating multi-source data, and achieving accurate customer stratification and product adaptation. It significantly improves the scientificity, accuracy, and comprehensiveness of evaluation data, providing a more reliable data decision-making basis for financial institutions.

[0024] Specifically:

[0025] By integrating unstructured POS transaction records, supply chain bill images, and semi-structured e-commerce transaction logs and mobile payment record behavior data, a dynamic data confidence evaluation matrix is established. For fields with a missing rate greater than 30% or abnormal fluctuations exceeding the threshold, an artificial review process is initiated to ensure data quality. Furthermore, a three-layer fuzzy membership function system is constructed, and an improved interval type-2 fuzzy set and α-cut algorithm are used to convert fuzzy data into a 0-1 standardized score, accurately generating features for decision-making. For example, in the evaluation of the stability of transaction frequency, by calculating the ratio of the standard deviation to the mean of the number of transactions, the stability of the operation of small and micro enterprises can be accurately measured, providing a solid foundation for subsequent evaluations and solving the problem that fuzzy data in traditional methods is difficult to effectively utilize and quantify.

[0026] Compared with traditional fuzzy sets, the improved interval type-2 fuzzy set can better capture the uncertainty of data, and its membership interval setting is more in line with the actual situation of fuzzy data. When dealing with the stability data of transaction amounts, the influence of market fluctuations and seasonal factors is fully considered, and a standardized score is generated through the α-cut algorithm, enabling a more scientific quantification method for fuzzy data in the evaluation system. Compared with the traditional method that relies on expert experience to handle fuzzy data, it greatly reduces the interference of subjective factors and improves the objectivity and accuracy of the evaluation.

[0027] A two-layer LSTM time memory unit is established, and its powerful time series analysis ability is used to accurately predict the business fluctuations of small and micro enterprises in the next two quarters. Based on the prediction results, the parameters of the membership function are dynamically adjusted. For example, when the economy changes from prosperity to recession, LSTM can predict in advance the changes in business data such as enterprise sales and procurement costs, and timely adjust the evaluation of the cash flow cycle matching degree, making the evaluation results more in line with the true risk status of enterprises in different economic cycles, and making up for the defect that existing models lack effective consideration of economic cycle fluctuations.

[0028] In the stage of fuzzy data preprocessing and feature engineering, complex data from multiple sources are integrated and quantitatively evaluated by establishing a dynamic data confidence evaluation matrix. In the IPC cross-validation link, the monthly average sales amount calculated from fuzzy data is subjected to residual analysis with the tax return forms and bank statements collected on-site by IPC. The weights of fuzzy data and the scores of manual verification items are adjusted according to the data deviation degree, realizing the effective integration of multi-source data. It is ensured that the evaluation results comprehensively consider various aspects of data information, improving the comprehensiveness and reliability of the evaluation.

[0029] Define the interest rate sensitivity dimension including the capital turnover elasticity coefficient, leverage tolerance, and price elasticity test value, and use the improved spectral clustering algorithm to generate the initial interest rate hierarchy in the reduced-dimensional feature space, realizing the effective mapping of interest rate sensitivity classification and risk dimensions. Different interest rate hierarchies correspond to small and micro enterprises with different risk characteristics, providing a basis for financial institutions to design differentiated loan products according to the risk situation and helping to achieve precise product adaptation.

[0030] In the process of D-S evidence fusion of the preliminary classification results of the fuzzy data layer and the IPC verification conclusion, the organic combination of the fuzzy data results processed by machine intelligence and the IPC verification conclusion dominated by manual experience is realized. Through evidence conflict detection, conflict coefficient calculation, and Dempster combination rule operation, a joint confidence interval is generated. The evaluation results not only include the efficiency and objectivity of machine intelligence processing but also incorporate the advantages of manual experience in controlling the authenticity of data, finally forming an interpretable, traceable, and iterative evaluation data system for small and micro customer groups and achieving accurate customer stratification and product adaptation. Brief Description of the Drawings

[0031] Figure 1 It is a schematic flow chart of an evaluation method for the loan availability of small and micro enterprises according to the present invention;

[0032] Figure 2 It is a schematic structural diagram of an electronic device provided by the present invention. Detailed Embodiments

[0033] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0034] This application discloses an evaluation method for the loan availability of small and micro enterprises. Refer to Figure 1 , and in implementation, this method includes the steps: Fuzzy data preprocessing and feature engineering: S100 integrates unstructured business data (POS transactions, supply chain bill images) and semi-structured behavioral data (e-commerce transaction logs, mobile payment records); establishes a dynamic data confidence evaluation matrix, and initiates a manual review process for fields with a missing rate > 30% or abnormal fluctuations exceeding the threshold; In the implementation of this step, first, for the integrated unstructured business data (such as POS transactions, supply chain bill images) and semi-structured behavioral data (such as e-commerce transaction logs, mobile payment records), corresponding confidence evaluation indicators are created for each data field.

[0035] Then, a dynamic data confidence evaluation matrix is constructed. The rows of the matrix represent different data fields, and the columns record the values of various confidence evaluation indicators related to that field. For example, for the POS transaction data field, its columns may include the reputation score of the data source, the stability score of the data collection frequency, and the matching degree score with the cash inflow data in the enterprise's financial statements during the same period.

[0036] Next, thresholds for data missing rate and abnormal fluctuations are set. In this solution, it is determined that a missing rate > 30% or abnormal fluctuations exceeding the threshold are the conditions for initiating a manual review. When calculating the data missing rate, the proportion of the number of missing data in each data field to the total data volume of that field is statistically calculated. For abnormal fluctuations, a reasonable fluctuation range is determined as the threshold by calculating the mean and standard deviation statistics of the data field over a period of time. For example, if the daily transaction amount data in an e-commerce transaction log has a recent standard deviation exceeding 20% of the mean in the past month, it is determined that the abnormal fluctuations exceed the threshold.

[0037] When the data missing rate or abnormal fluctuation situation triggers the manual review condition, the manual review process is initiated. First, the manual review personnel will carefully check the data collection process to confirm whether there are data missing or incorrect collection caused by technical failures. For example, check whether the hardware connection of the POS machine is stable and whether there are loopholes in the recording program of the e-commerce transaction log.

[0038] Then, the reviewer will compare multiple data sources and try to supplement the missing data or explain the reasons for abnormal fluctuations. Taking the missing supply chain bill image data as an example, the reviewer can contact the supplier to obtain a copy of the paper bill and compare it with the enterprise's internal purchase records to supplement the missing bill information.

[0039] During the manual review process, the relevant indicators in the data confidence evaluation matrix will be corrected. If it is found through the review that the data missing is due to a temporary data transmission problem and the data has been successfully supplemented, the reliability score of the data source for that data field can be appropriately increased. If the abnormal fluctuation is caused by reasonable business activities, the fluctuation stability score of that field can be adjusted.

[0040] Finally, according to the results after manual review, re-evaluate the confidence level of the data and update the dynamic data confidence level evaluation matrix.

[0041] S101 Construct a three-layer fuzzy membership function system, which includes a basic-layer fuzzy membership function (transaction frequency / amount stability), a derivative-layer fuzzy membership function (cash flow cycle matching degree), and a comprehensive-layer fuzzy membership function (industry prosperity sensitivity coefficient); In the implementation of this step, constructing a three-layer fuzzy membership function system includes: The basic-layer fuzzy membership function (transaction frequency / amount stability) specifically refers to: Transaction frequency stability: Count the number of transactions of small and micro enterprises within a certain period (such as the past year), and measure the fluctuation of the transaction frequency by calculating the ratio of the standard deviation to the mean of the number of transactions. If the ratio is small, it means the transaction frequency is relatively stable, and the membership function value approaches 1; if the ratio is large, it indicates that the transaction frequency fluctuates greatly, and the membership function value approaches 0. For example, if a company's monthly transaction times are relatively fixed, the membership degree of its transaction frequency stability may be as high as 0.8; while for another company with fluctuating transaction times, this value may be only 0.3.

[0042] Transaction amount stability: Calculate the deviation degree of each transaction amount from the average transaction amount within this period. A small deviation degree means high transaction amount stability, and the membership function value is close to 1; a large deviation degree means the membership function value is close to 0. Taking the POS transaction data of a small and micro enterprise as an example, if most transaction amounts fluctuate within ±10% of the average amount, the membership degree of its transaction amount stability can be set to 0.7; if the fluctuation range exceeds ±30%, the membership degree may be 0.4.

[0043] The derivative-layer fuzzy membership function (cash flow cycle matching degree) specifically refers to: First, determine the cash inflow cycle and cash outflow cycle of the enterprise. For cash inflow, analyze the collection cycle of the enterprise's main income sources, such as the time of receiving payment after selling products; for cash outflow, pay attention to the cycle of the enterprise's main expenditures such as paying for goods and employees' salaries.

[0044] Then calculate the cash flow cycle matching degree. If the cash inflow cycle is similar to the cash outflow cycle and within a reasonable business cycle range, it indicates a high cash flow cycle matching degree, and the membership function value is close to 1. For example, if the average cash inflow cycle of an enterprise is 45 days and the average cash outflow cycle is 40 days, the membership degree of its cash flow cycle matching degree may be 0.9; if the inflow cycle is 90 days and the outflow cycle is 30 days, with a large difference between the two, the membership function value may be only 0.3. This membership degree reflects the health of the enterprise's cash flow and has an important impact on evaluating the availability of loans.

[0045] The comprehensive layer fuzzy membership function (industry prosperity sensitivity coefficient) specifically refers to: Collect industry data, including indicators such as the overall sales growth rate, profit growth rate, and market demand changes in the industry.

[0046] Analyze the degree of correlation between the prosperity of the industry where the small and micro enterprises are located and the operating conditions of the enterprises. If the enterprise's operating performance is significantly improved when the industry is prosperous and the enterprise's operation is greatly impacted when the industry is not prosperous, it indicates that the enterprise is highly sensitive to industry prosperity and the membership function value is close to 1; on the contrary, if the enterprise's operating conditions are less affected by industry prosperity, the membership function value is close to 0. For example, in the tourism industry, during the peak tourist season (industry prosperity), the occupancy rate and revenue of a hotel increase significantly, and its membership of the industry prosperity sensitivity coefficient may be 0.8; while for some enterprises producing daily necessities, the demand is relatively stable regardless of industry prosperity, and the membership may be 0.2. Through the membership function of this layer, comprehensively consider the impact of industry factors on the loan availability of enterprises.

[0047] S102 uses an improved interval type-2 fuzzy set to handle data uncertainty and generates a 0-1 standardized score through the α-cut algorithm; In the implementation of this step, using an improved interval type-2 fuzzy set to handle data uncertainty and generating a 0-1 standardized score specifically includes: Introduction of the improved interval type-2 fuzzy set: When evaluating the loan availability of small and micro enterprises, there is a lot of uncertainty in fuzzy data. For example, when analyzing transaction frequency, transaction amount, cash flow cycle, and industry prosperity factors, the data may be interfered by various complex factors, making it difficult to guarantee the accuracy of the data. Traditional fuzzy set methods have certain limitations in dealing with this uncertainty, while the improved interval type-2 fuzzy set can better capture and handle these uncertainties.

[0048] Compared with the traditional fuzzy set, the membership degree of the improved interval type-2 fuzzy set is no longer an exact value but an interval. This is because in practical applications, it is often difficult to accurately determine the degree to which an element belongs to a fuzzy set, and there is a certain fuzzy range. For example, for the stability of the transaction amount of a small and micro enterprise, due to market fluctuations and seasonal factors, it is difficult to accurately say that the membership degree of the transaction amount stability is a fixed value, and it is more reasonable to represent it with an interval.

[0049] The process of handling data uncertainty: Constructing fuzzy sets: For the transaction frequency stability and transaction amount stability at the basic layer, the cash flow cycle matching degree at the derivative layer, and the industry prosperity sensitivity coefficient index at the comprehensive layer, corresponding improved interval type-2 fuzzy sets are constructed respectively. Taking the transaction frequency stability as an example, according to the ratio of the standard deviation to the mean of the number of transactions, it is divided into different fuzzy levels, such as "high stability", "medium stability", "low stability", etc., and each level corresponds to an improved interval type-2 fuzzy set.

[0050] Determining the membership degree interval: For each data point, calculate its membership degree interval in each fuzzy set. Taking the transaction frequency data of a small and micro enterprise as an example, by calculating the ratio of the standard deviation to the mean of its number of transactions and combining historical data and industry experience, determine the membership degree interval of the enterprise's transaction frequency stability in fuzzy sets such as "high stability", "medium stability", "low stability", etc. For example, the membership degree interval of the enterprise's transaction frequency stability in the "high stability" fuzzy set may be [0.7, 0.9], indicating that its degree of belonging to "high stability" is between 0.7 and 0.9.

[0051] Handling uncertainty: The improved interval type-2 fuzzy set can handle the uncertainty in data, mainly reflected in the following aspects. First, it can take into account the errors and noises in the data collection process. For example, when collecting POS transaction data, there may be certain errors in the data due to system failures and data transmission problems, and the improved interval type-2 fuzzy set can reflect this uncertainty through the membership degree interval. Second, it can handle the fuzziness of business rules. For example, for the judgment of the cash flow cycle matching degree, different business scenarios and industry characteristics may lead to certain fuzziness in the judgment criteria, and the improved interval type-2 fuzzy set can better adapt to this fuzziness through the interval membership degree.

[0052] Application of the α-cut algorithm: Definition of α-cut: α-cut is a method of converting a fuzzy set into a normal set. For an improved interval type-2 fuzzy set, given an α value (0 ≤ α ≤ 1), the α-cut is the normal set composed of the elements in this fuzzy set whose membership degree is greater than or equal to α. For example, for the "high stability" fuzzy set of transaction frequency stability, when α = 0.8, its α-cut is the normal set composed of the elements in this fuzzy set whose membership degree is greater than or equal to 0.8.

[0053] Generating 0-1 standardized scores: Through the α-cut algorithm, convert the improved interval type-2 fuzzy set into a series of normal sets. Then, generate 0-1 standardized scores according to the characteristics of these normal sets. The specific steps are as follows: Determining multiple α values: For example, select multiple values of α = 0.2, 0.4, 0.6, 0.8.

[0054] Calculate the α - cut corresponding to each α value: For each α value, calculate the α - cut of the improved interval type - 2 fuzzy set.

[0055] Comprehensive evaluation: According to the situation of each α - cut, comprehensively evaluate the membership degree of this data point in this fuzzy set. For example, the membership degree of this data point in this fuzzy set can be comprehensively evaluated by calculating the area and the number of elements index of each α - cut.

[0056] Generate a 0 - 1 normalized score: Normalize the result of the comprehensive evaluation to generate a 0 - 1 normalized score. For example, map the result of the comprehensive evaluation to the interval [0, 1] to obtain the 0 - 1 normalized score of this data point in this fuzzy set.

[0057] The data processed by the improved interval type - 2 fuzzy set is accurate and complete data obtained after fuzzy data pre - processing. During the fuzzy data pre - processing process, by establishing a dynamic data confidence evaluation matrix and manual review methods, the quality of the data is ensured, providing a reliable basis for the improved interval type - 2 fuzzy set to handle data uncertainty.

[0058] The 0 - 1 normalized scores generated by the α - cut algorithm will be important inputs for constructing the multi - dimensional feature space. These scores will be used to define the interest - rate - sensitive dimensions that affect loan availability, such as the capital turnover elasticity coefficient, leverage tolerance, and price elasticity test value, providing data support for the subsequent spectral clustering algorithm and the generation of the initial interest - rate hierarchy. The generated 0 - 1 normalized scores will participate in the IPC cross - validation process and be compared and fused with the data collected on - site by IPC.

[0059] Construction of multi - dimensional feature space: S103 Define the interest - rate - sensitive dimensions that affect loan availability: The interest - rate - sensitive dimensions include the capital turnover elasticity coefficient, leverage tolerance, and price elasticity test value; in the implementation of this step, for the capital turnover elasticity coefficient: Definition and calculation: The capital turnover elasticity coefficient is calculated by the ratio of the days sales outstanding (DSO) to the days payable outstanding (DPO). The days sales outstanding reflects the average time required for an enterprise to collect payments after selling products, and the calculation method is: days sales outstanding = (average accounts receivable balance × number of days in the calculation period) ÷ sales revenue. The days payable outstanding reflects the average time for an enterprise to pay for goods after purchasing, and the calculation method is: days payable outstanding = (average accounts payable balance × number of days in the calculation period) ÷ cost of sales. For example, in the past year, the average accounts receivable balance of a small micro-enterprise was 500,000 yuan, the sales revenue was 5 million yuan, and calculated based on 360 days in a year, its days sales outstanding was (50×360)÷500 = 36 days; the average accounts payable balance was 300,000 yuan, and the cost of sales was 3 million yuan, its days payable outstanding was (30×360)÷300 = 36 days. Then the capital turnover elasticity coefficient of this enterprise was 36÷36 = 1.

[0060] Impact on loan availability: The capital turnover elasticity coefficient reflects the flexibility of an enterprise in capital turnover. When the coefficient is close to 1, it means that the enterprise's collection cycle and payment cycle are relatively matched, the rhythm of capital inflow and outflow is stable, and when facing interest rate fluctuations, the enterprise has better capital allocation ability to repay loans on time, and the loan availability is relatively high. If the coefficient is much greater than 1, it indicates that the collection of accounts receivable is slower than the payment of accounts payable, the enterprise may face capital shortage, is more sensitive to interest rate changes, the loan risk increases, and the loan availability decreases; conversely, if the coefficient is much less than 1, although the enterprise has strong payment ability, there may be idle funds, which will also affect its performance in interest rate-sensitive businesses and the evaluation of loan availability.

[0061] Specifically for leverage tolerance: Definition and calculation: Leverage tolerance is calculated by the hidden debt coverage ratio. Hidden debt refers to those debts that are not clearly reflected in the enterprise's financial statements but may actually affect the enterprise's debt repayment ability, such as external guarantees and potential debts arising from pending litigation. Hidden debt coverage ratio = (total hidden debt + total explicit debt) ÷ total assets. Suppose the explicit debt of a small micro-enterprise is 2 million yuan, and after investigation, it is found that its potential debt from external guarantees is 500,000 yuan, and the total assets of the enterprise are 5 million yuan. Then the hidden debt coverage ratio of this enterprise is (200 + 50)÷500 = 0.5.

[0062] Impact on loan availability: The leverage tolerance reflects an enterprise's ability to bear the debt burden. A lower implicit debt coverage ratio means that the enterprise has a relatively lighter debt burden and is more capable of withstanding the increased repayment pressure caused by rising interest rates, resulting in a higher loan availability. Conversely, if the implicit debt coverage ratio is too high, the enterprise's debt repayment pressure is large, and even a slight fluctuation in interest rates may plunge the enterprise into financial difficulties. Banks and financial institutions will consider its loan risk to be high, thereby reducing its loan availability.

[0063] Specifically for the price elasticity test value: Definition and calculation: The price elasticity test value is determined by the historical interest rate adjustment response rate. Specifically in the calculation, observe the changes in the enterprise's business volume (such as sales volume, order volume) during the past interest rate adjustment cycle. For example, when the market interest rate was last raised by 10%, the enterprise's sales volume decreased by 15%, then its historical interest rate adjustment response rate is 15%÷10% = 1.5; if the interest rate is lowered by 10% and the enterprise's sales volume increases by 12%, the response rate is calculated in the same way. In practical applications, the response rate data for multiple interest rate adjustment cycles are integrated, and the average value is taken as the final price elasticity test value.

[0064] Impact on loan availability: The price elasticity test value reflects the sensitivity of an enterprise's business to interest rate changes. If an enterprise has a high price elasticity test value, it indicates that its business is greatly affected by interest rate fluctuations. When interest rates rise, it may face operational difficulties due to factors such as decreased market demand and increased costs, threatening its repayment ability and reducing loan availability. On the contrary, an enterprise with a lower price elasticity test value is more adaptable to interest rate changes and can maintain a relatively stable operating condition in different interest rate environments, resulting in a higher loan availability.

[0065] S104 uses an improved spectral clustering algorithm to generate an initial interest rate hierarchy in the reduced-dimensional feature space to achieve the mapping of interest rate sensitivity classification and risk dimensions; In the implementation of this step, it should be noted that before applying the improved spectral clustering algorithm, it is necessary to first reduce the dimension of the feature space composed of interest rate-sensitive dimensions (capital turnover elasticity coefficient, leverage tolerance, price elasticity test value) and multi-dimensional data of 0-1 standardized scores generated by the improved interval type-2 fuzzy set and α-cut algorithm. This is because the original high-dimensional data is not only computationally complex but may also have data redundancy and noise interference, affecting the clustering effect. The dimensionality reduction method can be selected as principal component analysis (PCA) or linear discriminant analysis (LDA), etc. Taking PCA as an example, its principle is to project the original data into a new coordinate space through linear transformation, so that the variance of the data is maximized as much as possible in the new dimension, thereby extracting the main feature components of the data. In actual operation, the number of principal components to be retained is determined according to the eigenvalue contribution rate. For example, select the principal components that can explain more than 80% of the data variance, compress the high-dimensional data into a low-dimensional space, which not only retains the key information but also reduces the computational complexity, providing a more efficient and concise data basis for the subsequent spectral clustering algorithm.

[0066] Implementation process of the improved spectral clustering algorithm: In the reduced-dimensional feature space, construct a similarity matrix for the data features of small and micro enterprises.

[0067] For the data of the interest rate-sensitive dimension, distance measurement methods such as Euclidean distance and cosine similarity are used to measure the similarity degree between the data points of different small and micro enterprises. For example, for two small and micro enterprises with similar capital turnover elasticity coefficients, leverage tolerances, and price elasticity test values, the corresponding element values in the similarity matrix will be higher, indicating a strong similarity between the two. For the 0-1 standardized score data generated by the improved interval type-2 fuzzy set and α-cut algorithm, the similarity is also determined based on the closeness of the scores. When constructing the similarity matrix, considering that the importance of data in different dimensions may be different, the data in each dimension can be weighted. The determination of the weights can be combined with expert experience and the analysis of historical data to ensure that more important features account for a larger proportion in the similarity measurement.

[0068] Calculate the Laplacian matrix: Based on the constructed similarity matrix, calculate the Laplacian matrix. The Laplacian matrix can reflect the connection relationship and local structure between data points. Common forms of the Laplacian matrix include the unnormalized Laplacian matrix and the normalized Laplacian matrix. In this application, the normalized Laplacian matrix with better adaptability to data distribution is selected, such as the symmetric normalized Laplacian matrix. By calculating the eigenvalues and eigenvectors of the Laplacian matrix, it prepares for the subsequent clustering operation.

[0069] Eigenvalue decomposition and low-dimensional embedding: Perform eigenvalue decomposition on the Laplacian matrix, select the eigenvectors corresponding to the first k smallest non-zero eigenvalues, and map the data points into a k-dimensional low-dimensional space. The value of k here is determined according to the actual business requirements and data characteristics. For example, by trying different k values and combining clustering evaluation metrics (such as silhouette coefficient, Calinski-Harabasz index), determine a k value that can achieve the best clustering effect. In this low-dimensional space, the distribution of data points can better reflect their inherent clustering structure, providing a clearer pattern for subsequent clustering division.

[0070] Clustering division: In the low-dimensional embedding space, use the K-means algorithm to perform clustering division on the data points. The K-means algorithm randomly selects k initial clustering centers, and then continuously iteratively calculates the distances from each data point to each clustering center, assigns the data points to the clusters where the nearest clustering centers are located, and updates the clustering centers until the clustering centers no longer change significantly or reach the preset number of iterations. In this application, map the clustering results to different initial interest rate levels. For example, classify small and micro enterprises in a clustering cluster into the same interest rate level, and different clustering clusters correspond to different interest rate levels.

[0071] During the clustering process, strictly control the within-class distance and between-class distance. For the within-class distance, by continuously adjusting the clustering centers and the assignment of data points, ensure that the distance (such as Euclidean distance) between the data points of small and micro enterprises within the same clustering cluster is less than 0.15. This means that small and micro enterprises within the same interest rate level have high similarity in terms of interest rate-sensitive dimension features, and their response patterns and risk characteristics to interest rates are relatively close. For the between-class distance, ensure that the distance between different clustering clusters (such as the minimum Euclidean distance) is greater than 0.35, so that small and micro enterprises in different interest rate levels have obvious differences in features, facilitating the formulation of differentiated credit strategies according to different interest rate levels in the future. If it is found during the clustering process that the within-class distance or between-class distance does not meet the requirements, the clustering parameters (such as the selection of initial clustering centers, the number of iterations of the clustering algorithm) can be readjusted, or the data can be further preprocessed (such as outlier handling, data standardization) until the conditions of within-class distance < 0.15 and between-class distance > 0.35 are met, thereby generating reasonable and effective initial interest rate levels.

[0072] S105 Establish a two-layer LSTM time memory unit to predict the business fluctuations in the next two quarters and dynamically adjust the membership function parameters; In the implementation of this step, the structure of the two-layer LSTM time memory unit includes: The first layer of LSTM units: Input Gate: Responsible for controlling the entry of new input data into the unit. In the scenario of predicting small and micro enterprise operation data, for example, when facing newly input monthly sales data, the input gate will decide whether to allow the sales data to enter the unit for processing based on the current state of the unit and the characteristics of the input data. If the enterprise's business model has been relatively stable recently, the input gate allows the new data to enter because these data are of great significance for reflecting the current operating conditions of the enterprise.

[0073] Forget Gate: Decides which information in the unit state at the previous moment to retain or discard. For example, if a small and micro enterprise had an abnormal peak in sales due to seasonal promotion activities before, but has now entered the regular sales period, the forget gate will reduce the retention degree of the state information related to the abnormal sales at the previous moment to avoid its interference with the judgment of the current normal operating conditions.

[0074] Output Gate: Based on the current unit state and input data, determines the information finally output to the next layer or the next moment. For the first-layer LSTM unit, the information it outputs is the characteristics of the small and micro enterprise operation data after preliminary processing, and these characteristics contain the short-term fluctuation information of the recent operation data, such as the change trends of sales and procurement costs between months.

[0075] Memory Cell: The memory cell is like a long-term storage warehouse that can store important information over multiple time steps. In the processing of small and micro enterprise operation data, the memory cell can remember the operation trends of the enterprise over a past period of time, such as the continuous growth of sales or the gradual increase in costs over several consecutive months. Even if there are small fluctuations in the data in the short term, the memory cell can maintain the memory of the overall trend.

[0076] Second-layer LSTM Unit: Input Gate, Forget Gate, and Output Gate: Similar to the first layer, but the input they process is the output of the first-layer LSTM unit. The input gate of the second layer will decide which information processed by the first layer to allow to enter this layer for further processing based on the characteristics output by the first layer and the state of its own unit. The forget gate judges which parts of the long-term state information passed from the first layer to retain or discard according to the current situation. The output gate finally determines the more comprehensive and regular information to be output after two-layer processing.

[0077] Memory cells: The memory cells in the second layer integrate the information transmitted by the memory cells in the first layer and the operation results of the input gate and forget gate in this layer, and can capture the longer-term and more macroscopic trends in the business operation data of small and micro enterprises. For example, it can identify the overall business situation of an enterprise in the past year or even longer, showing the overall business situation as the market environment changes and the industry cycle fluctuates. Such as the long-term trend of the enterprise's steadily increasing sales and good cost control during the industry's upward period, or the business difficulties faced during the industry's downward period and the signs of gradual recovery after adjusting strategies.

[0078] Application process of the double-layer LSTM time memory unit: First, the business operation data of small and micro enterprises in the past few years, such as the time series data of sales, procurement costs, inventory quantities, and the number of employees, are sequentially input into the first-layer LSTM unit in chronological order. The first-layer LSTM unit processes the data at each time step through the coordinated work of the input gate, forget gate, and output gate, and extracts the short-term fluctuation characteristics in the data. For example, it can capture the subtle changes in monthly sales and the short-term correlations between such changes and factors such as procurement costs and marketing activities.

[0079] The feature information processed by the first-layer LSTM unit is passed as input to the second-layer LSTM unit. Based on these inputs, the second-layer LSTM unit further explores the long-term trends and complex patterns in the data. For example, by comprehensively analyzing the sales fluctuation characteristics of consecutive months output by the first layer, the second layer can identify the annual cycle law of the enterprise's sales changing with seasons, or discover the gradually changing trend presented by the business operation data under the long-term strategies of the enterprise to explore new markets and launch new products.

[0080] After learning and feature extraction from historical data, the double-layer LSTM time memory unit constructs a model that can reflect the internal laws of the business operation data of small and micro enterprises. When predicting the business fluctuations in the next two quarters, the latest historical data up to the current time is input into the model.

[0081] Based on the historical trends and patterns that have been learned, the model predicts the business operation data at each future time step (such as each month) through the layer-by-layer calculation and information transmission of the two-layer LSTM unit. For example, predicting the changes in key indicators such as sales, procurement costs, and profits in the next two quarters. During the prediction process, the first-layer LSTM unit makes a preliminary prediction based on the short-term fluctuation patterns of recent data, and the second-layer LSTM unit adjusts and corrects the prediction results of the first layer by combining long-term trends, so that the final prediction results are more in line with the long-term business laws of the enterprise and the possible influence of macro factors.

[0082] Then dynamically adjust the membership function parameters: Re-evaluate the company's cash inflow and outflow cycles based on the forecasted operating data for the next two quarters. For the cash inflow cycle, consider the forecasted sales growth or decline, potential changes in customer payment habits, and expected fluctuations in market demand. For example, if sales are forecast to increase significantly and the company strengthens accounts receivable management, the cash inflow cycle may be shortened; conversely, if market demand declines and customer payments are delayed, the cash inflow cycle may be extended.

[0083] For the cash outflow cycle, analyze the predicted changes in procurement costs, adjustments to suppliers' payment policies, and changes in the company's internal cost control strategies. If raw material prices are expected to rise, the company may need to pay for goods in advance, thereby shortening the cash outflow cycle; if the company plans to optimize its cost structure and delay some expenses, the cash outflow cycle may be extended.

[0084] According to the re-evaluated cash flow cycle, the membership function parameters are dynamically adjusted. The adjusted membership function parameters are specifically the parameters of the derivative layer fuzzy membership function. If the cash inflow and outflow cycles are closer and in line with the reasonable operating cycle range of the enterprise's industry, the parameter value representing high matching in the cash flow cycle matching membership function is increased; otherwise, the parameter value is reduced. For example, the high matching parameter value in the original membership function is 0.8. If the cash flow cycle matching is significantly improved after adjustment, the parameter value can be increased to 0.9; if the matching becomes worse, it is reduced to 0.7. By continuously adjusting the parameters, the membership function can more accurately reflect the health of the enterprise's future cash flow, providing a more reliable basis for the subsequent loan availability assessment.

[0085] IPC cross validation and joint evaluation: S106 conducts residual analysis on the average monthly sales volume estimated from fuzzy data and the tax return forms and bank statements collected on-site by IPC. When the data deviation is >15%, the credibility decay function is triggered to reduce the weight of fuzzy data and increase the score of manual verification items; In the implementation of this step, it should be noted that in the IPC cross-validation phase, first ensure that comprehensive and accurate monthly average sales data calculated from fuzzy data is obtained. This requires reviewing the various data sources involved in the fuzzy data preprocessing and feature engineering stages, such as POS flow, e-commerce transaction logs, and mobile payment records, summarizing and calculating these data, and deriving the monthly average sales value. At the same time, IPC on-site collection personnel obtain tax returns and bank statements from the company. The tax return should cover all taxable income details of the company during the corresponding time period, and the bank statement must contain a complete record of account funds in and out, including detailed information on each income and expenditure, the time of occurrence, and the other party's account information. Organize these data into a unified format to facilitate subsequent residual analysis.

[0086] Residual analysis process: Perform a residual analysis on the monthly average sales volume calculated from fuzzy data and the sales volume data in the tax returns and bank statements collected on-site by IPC. Taking the tax return as an example, extract the taxable sales volume data for each month from the tax return and match them one by one with the monthly average sales volume calculated from the fuzzy data according to the month. Calculate the difference between the two, which is the residual. For example, if the monthly average sales volume of a small and micro enterprise calculated from fuzzy data in March is 500,000 yuan, and the taxable sales volume shown in the tax return for that month is 450,000 yuan, then the residual is 500,000 - 450,000 = 50,000 yuan. Apply the same method to the sales volume data in the bank statement. After summarizing the monthly sales income recorded in the bank statement, compare it with the value calculated from the fuzzy data to calculate the residual. Then, calculate the ratio of the residual to the data collected on-site by IPC (taking the tax return or bank statement data as the standard), and use this as an indicator to measure the degree of data deviation. Assuming in the above example, taking the tax return data as the benchmark, the data deviation degree is 5÷45×100%≈11.1%.

[0087] When the calculation result of the data deviation degree is greater than 15%, trigger the credibility attenuation function. The role of this credibility attenuation function is to reduce the weight of the fuzzy data in the subsequent evaluation. For example, originally the weight of the fuzzy data in the comprehensive evaluation system is set to 0.6. After triggering the credibility attenuation function, according to the pre-set attenuation rule, the weight is reduced to 0.4. At the same time, increase the score of the manual verification item. The manual verification item covers the details of the on-site inspection of the enterprise's business premises, the content of the interviews with the enterprise's responsible person and employees, and the operating conditions of the enterprise's production equipment. Originally, the proportion of these manual verification items in the total evaluation score is 0.3. After triggering the credibility attenuation function, the proportion of their scores is increased to 0.5. This means that in the subsequent joint evaluation, the information obtained through manual verification will play a greater role in the overall evaluation to make up for the inaccurate impact that may be brought by the excessive deviation of the fuzzy data. The staff will re-comb the previous manual verification records and re-verify the issues that were not explored in depth before, such as asking about the specific uses and relevant vouchers for the recent large expenditures of the enterprise, to ensure that the evaluation of the enterprise's operating conditions is more accurate and reliable.

[0088] The fuzzy data after adjusting the weight by the credibility attenuation function and the manual verification item with increased scores will jointly participate in the construction of the subsequent joint evaluation matrix and the D-S evidence fusion process. In the D-S evidence fusion stage, the evidence strength carried by the preliminary classification result of the original fuzzy data weakens due to the reduction of the weight, and the evidence strength provided by the manual verification relatively increases. The joint confidence interval generated after the two are fused will be more biased towards the real situation of the enterprise reflected by the on-site verification, thereby improving the accuracy and reliability of the evaluation of the loan availability of the entire small and micro enterprise.

[0089] The three - layer fuzzy membership function system is a tool for multi - dimensional quantitative assessment of an enterprise's business conditions. The relevant assessment results generated by this system can be used for in - depth analysis and verification of business data such as monthly average sales. In the IPC cross - verification link, residual analysis is carried out on the monthly average sales calculated from fuzzy data and the tax returns and bank statements collected on - site by IPC. Among them, the monthly average sales calculated from fuzzy data is obtained after processing multi - source data through the three - layer fuzzy membership function system method. Through comparative analysis, the accuracy and reliability of the monthly average sales data can be further evaluated. When the data deviation degree is greater than the threshold value, corresponding processing mechanisms such as the credibility attenuation function will be triggered, reducing the weight of fuzzy data and increasing the score of manual verification items to ensure the scientificity and accuracy of the assessment. The monthly average sales and the three - layer fuzzy membership function system are interdependent. The business data of the monthly average sales lays the foundation for constructing the membership function system, while the assessment results of the membership function system can provide strong support for the analysis and verification of data such as the monthly average sales.

[0090] Definition and explanation of the credibility attenuation function: The credibility attenuation function is a function specifically designed by the present invention to address the deviation problem between fuzzy data and the data collected on - site by IPC during the assessment of the loan availability of small and micro enterprises. When residual analysis is carried out on the monthly average sales calculated from fuzzy data and the tax returns and bank statements collected on - site by IPC, and the data deviation degree is greater than the set threshold value (such as the data deviation degree > 15% mentioned in the text), this function will be triggered. Its main role is to reduce the weight of fuzzy data in subsequent assessments and at the same time increase the score of manual verification items. For example, originally the weight of fuzzy data in the comprehensive assessment system is 0.6. After triggering the credibility attenuation function, according to the pre - set attenuation rule, its weight may be reduced to 0.4; while originally the proportion of the manual verification item in the total assessment score is 0.3, and at this time it will be increased to 0.5. Through such adjustments, the information obtained from manual verification can play a greater role in the overall assessment in subsequent joint assessments, to make up for the possible inaccurate assessment caused by the excessive deviation of fuzzy data, and improve the accuracy and reliability of the entire assessment of the loan availability of small and micro enterprises.

[0091] Relationship between the credibility attenuation function and the three - layer fuzzy membership function system: Data association relationship: The fuzzy data processed by the credibility attenuation function contains the data information obtained after being processed and analyzed by the three - layer fuzzy membership function system.

[0092] Interaction relationship: On the one hand, the evaluation results of the three-layer fuzzy membership function system will affect whether the credibility attenuation function is triggered. If the evaluation results of the three-layer fuzzy membership function system for an enterprise lead to a large deviation between the monthly average sales volume calculated from fuzzy data and the actual situation (referring to the data collected on-site by IPC), the credibility attenuation function may be triggered. On the other hand, the triggering and adjustment of the credibility attenuation function (reducing the weight of fuzzy data and increasing the score of manual verification items) will also affect the subsequent evaluation based on fuzzy data, including the re-application and analysis of the three-layer fuzzy membership function system.

[0093] The credibility attenuation function is used to adjust the weight of fuzzy data and the score of manual verification items when the data deviation degree is greater than the threshold: For example, let the weight of fuzzy data be wf, with an initial value of wf0; the score of manual verification items be wh, with an initial value of wh0; the data deviation degree be d; the set data deviation threshold be T (15% in the text); the attenuation coefficient be k1 (0 < k1 < 1, used to control the attenuation degree of the weight of fuzzy data), and the promotion coefficient be k2 (k2 > 1, used to control the promotion degree of the score of manual verification items).

[0094] Then the credibility attenuation function can be expressed as: When d ≤ T, wf = wf0; wh = wh0; When d > T, wf = k1 × wf0; wh = k2 × wh0.

[0095] S107 adopts the D-S evidence combination rule to perform D-S evidence fusion on the preliminary classification results of the fuzzy data layer and the IPC verification conclusion to generate a joint confidence interval; In the implementation of this step, it should be noted that in the fuzzy data preprocessing and feature engineering stage, multi-source data such as POS transactions, supply chain bill images, e-commerce transaction logs, and mobile payment records are integrated. By constructing a three-layer fuzzy membership function system, the operating conditions of enterprises are quantitatively evaluated from multiple dimensions, from the basic layer (transaction frequency / amount stability), derivative layer (cash flow cycle matching degree) to the comprehensive layer (industry prosperity sensitivity coefficient). For example, the membership degree of the transaction frequency stability of a small micro-enterprise is calculated to be 0.8, and this value reflects the stability characteristics of the enterprise in terms of transaction frequency.

[0096] Improved interval type-2 fuzzy sets are used to handle data uncertainty, and the α-cut algorithm is employed to generate 0-1 normalized scores. These scores are used in the construction of the multi-dimensional feature space to define the interest rate-sensitive dimensions (fund turnover elasticity coefficient, leverage tolerance, price elasticity test value) that affect loan availability, and an improved spectral clustering algorithm is used to generate the initial interest rate levels. Based on these results, basic probability assignments are generated for different evaluation categories (such as different risk levels, interest rate levels). Suppose that according to the fuzzy data evaluation, the basic probability assignment for the enterprise being in the low-risk level is m1(A1)=0.6 (where A1 represents the low-risk level), which indicates that from the perspective of fuzzy data, 60% of the "evidence" supports the enterprise being in the low-risk state. Here, m1 represents the basic probability assignment function based on fuzzy data, and A1 is the event of the low-risk level obtained from the fuzzy data evaluation.

[0097] In the IPC cross-validation process, IPC on-site collectors obtain the enterprise's tax returns and bank statements, and carefully verify fixed assets (such as factory buildings and equipment conditions) and current assets (inventory, accounts receivable), and calculate the asset conversion rate in combination with business data, while considering external credit data. For example, through the verification of the enterprise's fixed assets, it is found that its equipment is in good operation and the production capacity is stable, which provides strong support for the enterprise's business stability.

[0098] Perform residual analysis on the monthly average sales volume calculated from fuzzy data and the data collected on-site by IPC. When the data deviation > 15%, trigger the credibility attenuation function to adjust the weights of fuzzy data and manual verification items. Based on the comprehensive verification results, basic probability assignments are generated for different evaluation states of the enterprise. For example, after IPC verification, the basic probability assignment for the enterprise being in the medium-risk level is m2(A2)=0.5 (where A2 represents the medium-risk level), which reflects the judgment of the enterprise's risk status by the IPC verification conclusion. Here, m2 is the basic probability assignment function based on IPC verification, and A2 is the event of the medium-risk level obtained from IPC verification.

[0099] D-S evidence fusion process: Evidence conflict detection: Compare the basic probability assignments of the preliminary classification results of the fuzzy data layer and the IPC verification conclusion to detect the conflict between the two. For example, if the fuzzy data indicates that the enterprise is in the low-risk level (m1(A1)=0.6), while the IPC verification conclusion believes that the enterprise is in the medium-risk level (m2(A2)=0.5), there is an obvious conflict.

[0100] By calculating the conflict coefficient to quantify the degree of conflict, where A i and B j are different events in the fuzzy data and the IPC verification conclusion respectively. In this formula, A iRepresents the event in the preliminary classification result of fuzzy data, B j Represents the event in the IPC verification conclusion. When A i and B j The intersection of and is an empty set, indicating that these two events conflict with each other. Multiply and accumulate the corresponding basic probability assignment values, and the larger the obtained conflict coefficient k value, the more serious the conflict.

[0101] When the conflict coefficient k is small, the Dempster combination rule can be directly used for fusion. For events with a non-empty intersection, such as A1 ∩ A3 = A4 (assuming A3 is another event in the IPC verification conclusion and A4 is the fused event), the basic probability assignment after fusion is . In this formula, the numerator part represents multiplying and accumulating the basic probability assignment values of all event pairs with an intersection of A4 in the fuzzy data and the IPC verification conclusion. The denominator 1 - k is to correct the conflict to ensure that the fused probability value is within a reasonable range. If the conflict coefficient k is large, it indicates that the two conflict severely. At this time, the data needs to be further reviewed or the basic probability assignment needs to be adjusted before fusion.

[0102] Based on the fused basic probability assignment, determine the parameters of the joint confidence interval. For example, for the risk assessment of enterprise loans, a risk probability range can be set as the joint confidence interval according to the fused probabilities of different risk levels. Assume that the probability range of the enterprise being in the low - risk level after fusion is [0.4, 0.7], which is a preliminary joint confidence interval. Here, the lower limit 0.4 of the confidence interval represents the minimum probability estimate of the enterprise being in the low - risk level under the current assessment system, and the upper limit 0.7 represents the maximum probability estimate. This interval reflects the credibility range of the judgment that the enterprise is in the low - risk level.

[0103] S108 evaluates the availability of small - micro enterprise loans based on the joint confidence interval and outputs a recommendation report on whether to approve the enterprise loan; In the implementation of this step, according to the risk probability range determined by the joint confidence interval, the risk levels of small - micro enterprises are divided into three levels: low - risk, medium - risk, and high - risk. Specifically, if the probability that the joint confidence interval shows the enterprise is in the low - risk level is greater than or equal to 0.6, then the enterprise is determined to be in the low - risk level; if the probability of being in the low - risk level is between 0.3 and 0.6 (excluding 0.6), then it is in the medium - risk level; if the probability of being in the low - risk level is less than 0.3, then it is determined to be in the high - risk level.

[0104] Generate a recommendation report based on the previous evaluation results and enterprise data. Automatically fill in the report according to the enterprise's risk level, including the risk level evaluation result, loan availability analysis, loan recommendation and reasons, etc.

[0105] It is understandable that the evaluation method for the loan availability of small and micro enterprises proposed in this application effectively solves the data processing problems of existing methods in dealing with fuzzy data, coping with economic cycle fluctuations, integrating multi-source data, and achieving accurate customer stratification and product adaptation. It significantly improves the scientificity, accuracy, and comprehensiveness of the evaluation data, providing a more reliable data decision-making basis for financial institutions.

[0106] Figure 2 FIG. shows a schematic structural diagram of an electronic device 40 that can be used to implement an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as, for example, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, for example, personal digital processors, cellular phones, smart phones, wearable devices (such as helmets, glasses, watches, etc.) and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0107] As Figure 2 shown, the electronic device 40 includes at least one processor 41, and a memory communicatively connected to the at least one processor 41, such as a read-only memory (ROM) 42, a random access memory (RAM) 43, etc. Among them, the memory stores a computer program executable by the at least one processor. The processor 41 can execute various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 42 or the computer program loaded from the storage unit 48 into the random access memory (RAM) 43. In the RAM 43, various programs and data required for the operation of the electronic device 40 can also be stored. The processor 41, the ROM 42, and the RAM 43 are connected to each other through a bus 44. The input / output (I / O) interface 45 is also connected to the bus 44.

[0108] A plurality of components in the electronic device 40 are connected to the I / O interface 45, including: an input unit 46, such as a keyboard, a mouse, etc.; an output unit 47, such as various types of displays, speakers, etc.; a storage unit 48, such as a magnetic disk, an optical disc, etc.; and a communication unit 49, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 49 allows the electronic device 40 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0109] The processor 41 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 41 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The processor 41 executes the various methods and processes described above, such as an evaluation method for the availability of small and micro enterprise loans or a part of the steps of this method.

[0110] In some embodiments, an evaluation method for the availability of small and micro enterprise loans or a part of the steps of this method can be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as the storage unit 48. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 40 via the ROM 42 and / or the communication unit 49. When the computer program is loaded into the RAM 43 and executed by the processor 41, one or more steps of an evaluation method for the availability of small and micro enterprise loans or a part of the steps of this method described above can be executed.

[0111] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An evaluation method for the loan availability of small and micro enterprises, characterized in that: Including the steps: Integrate unstructured business data and semi-structured behavioral data; Establish a dynamic data confidence evaluation matrix, and initiate a manual review process for fields with a missing rate greater than the threshold or an abnormal fluctuation exceeding the threshold; Construct a three-layer fuzzy membership function system; Adopt an improved interval type-2 fuzzy set to handle data uncertainty, and generate a 0-1 standardized score through the α-cut algorithm; Define the interest rate sensitive dimensions that affect loan availability: the interest rate sensitive dimensions include the capital turnover elasticity coefficient, the leverage tolerance, and the price elasticity test value; Apply the improved spectral clustering algorithm to generate an initial interest rate hierarchy in the reduced-dimensional feature space to achieve the mapping of interest rate sensitivity classification and risk dimensions; Establish a two-layer LSTM time memory unit to predict the business fluctuations in the next two quarters and dynamically adjust the membership function parameters; Perform residual analysis on the average monthly sales amount calculated from fuzzy data, the tax return forms collected on-site by IPC, and the bank statements. When the data deviation degree is greater than the threshold, trigger a credibility attenuation function for data adjustment, reduce the weight of fuzzy data, and increase the score of the manual verification item; Conduct D-S evidence fusion on the preliminary classification results of the fuzzy data layer and the IPC verification conclusion to generate a joint confidence interval; Evaluate the availability of small and micro enterprise loans based on the joint confidence interval, and output a recommendation report on whether to approve the enterprise loan.

2. The evaluation method for the loan availability of small and micro enterprises according to claim 1, characterized in that: Construct a dynamic data confidence evaluation matrix. The rows of the dynamic data confidence evaluation matrix represent different data fields, and the columns record the values of various confidence evaluation indicators related to the fields.

3. The evaluation method for the loan availability of small and micro enterprises according to claim 1, wherein: The three-layer fuzzy membership function system includes a basic layer fuzzy membership function, a derivative layer fuzzy membership function, and a comprehensive layer fuzzy membership function.

4. The evaluation method for the loan availability of small and micro enterprises according to claim 1, characterized in that: Adopt an improved interval type-2 fuzzy set to handle data uncertainty, and generate a 0-1 standardized score through the α-cut algorithm. Specifically: for the transaction frequency stability and transaction amount stability in the basic layer, the cash flow cycle matching degree in the derivative layer, and the industry prosperity sensitivity coefficient indicator in the comprehensive layer, construct corresponding improved interval type-2 fuzzy sets respectively; For each data point, calculate its membership degree interval in each fuzzy set; Through the α-cut algorithm, transform the improved interval type-2 fuzzy set into an ordinary set; then, generate a 0-1 standardized score according to the characteristics of the ordinary set.

5. The evaluation method for the loan availability of small and micro enterprises according to claim 1, wherein: The capital turnover elasticity coefficient is calculated by the ratio of the accounts receivable turnover days to the accounts payable turnover days, the leverage tolerance is calculated by the invisible debt coverage ratio, and the price elasticity test value is calculated by the historical interest rate adjustment response rate.

6. The evaluation method for the loan availability of small and micro enterprises according to claim 1, wherein: The dimensionality reduction method for generating the initial interest rate hierarchy in the reduced-dimensional feature space can select principal component analysis or linear discriminant analysis; The improved spectral clustering algorithm includes: In the reduced-dimensional feature space, construct a similarity matrix according to the data characteristics of small and micro enterprises; Based on the constructed similarity matrix, calculate the Laplacian matrix; Perform eigenvalue decomposition on the Laplacian matrix, select the eigenvectors corresponding to the first k smallest non-zero eigenvalues, and map the data points to a k-dimensional low-dimensional space; In the low-dimensional embedding space, use the K-means algorithm to cluster and divide the data points; Map the clustering results to different initial interest rate levels.

7. An evaluation method for the loan availability of small and micro enterprises according to claim 3, characterized in that: Build a two-layer LSTM time memory unit to predict the business fluctuations in the next two quarters and dynamically adjust the membership function parameters, including: First, input the time series data of the business operation data of small and micro enterprises into the first-layer LSTM unit in chronological order. The first-layer LSTM unit processes the data at each time step through the coordinated work of the input gate, forget gate, and output gate, and extracts the short-term fluctuation characteristics in the data; The feature information processed by the first-layer LSTM unit is passed as input to the second-layer LSTM unit; the second-layer LSTM unit further mines the long-term trends and complex patterns in the data based on these inputs; Through the layer-by-layer calculation and information transfer of the two-layer LSTM units, predict the business data at each future time step; Dynamically adjust the membership function parameters: According to the predicted business data in the next two quarters, re-evaluate the cash inflow and outflow cycles of the enterprise; Based on the re-evaluated cash flow cycle, dynamically adjust the membership function parameters, and the adjusted membership function parameters are specifically the parameters of the derivative layer fuzzy membership function.

8. The evaluation method for the loan availability of small and micro enterprises according to claim 1, characterized in that: Perform D-S evidence fusion on the preliminary classification results of the fuzzy data layer and the IPC verification conclusions to generate a joint confidence interval, specifically including: Compare the basic probability assignments of the preliminary classification results of the fuzzy data layer and the IPC verification conclusions to detect the conflicts between them; By calculating the conflict coefficient to quantify the degree of conflict, where A i and B j are different events in the fuzzy data and the IPC verification conclusion respectively; A i represents the event in the preliminary classification result of the fuzzy data, and B j represents the event in the IPC verification conclusion. When the intersection of A i and B j is an empty set, it indicates that these two events conflict with each other. Multiply and accumulate their corresponding basic probability assignment values. The larger the conflict coefficient k value, the more serious the conflict; When the conflict coefficient k is small, the Dempster combination rule can be directly used for fusion; for events with a non-empty intersection, the fused basic probability assignment is: ； Based on the fused basic probability assignment, determine the parameters of the joint confidence interval.

9. The evaluation method for the loan availability of small and micro enterprises according to claim 1, characterized in that: Evaluate the availability of small and micro enterprise loans based on the joint confidence interval, and output a recommendation report on whether to approve the enterprise loan, including: according to the risk probability range determined by the joint confidence interval, classify the risk levels of small and micro enterprises into three levels: low risk, medium risk, and high risk; generate a recommendation report including the risk assessment results.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute one or more steps of an evaluation method for the availability of small and micro enterprise loans according to any one of claims 1-9.

Images