Credit risk early warning strategy automatic evaluation method and device driven by AI agent, control device and computer readable storage medium

Abstract

The application relates to an AI Agent-driven credit risk early warning strategy automatic evaluation method and device, a control equipment and a computer readable storage medium, which comprises the following steps: acquiring structured financial data, unstructured text data and dynamic time series data of an enterprise through multiple data sources, and performing standardization processing to output a standardized risk data set; based on the standardized risk data set, static financial features, dynamic derived features and correlation network features are extracted, feature screening and dimension reduction optimization are performed, and a high-discrimination feature set is generated; a multi-model integration strategy is adopted, the high-discrimination feature set is input into a neural network, a Stacking integrated model and a graph neural network, and enterprise risk scores and grades are output; and based on the enterprise risk scores and grades, early warning rules are automatically matched, and real-time monitoring and early warning are performed on enterprise credit according to the early warning rules.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of credit management, in particular to an AI Agent driven credit risk early warning strategy automatic evaluation method and device, a control equipment and a computer readable storage medium. BACKGROUND

[0002] Under the background of the intensification of global economic fluctuations and the rise of market uncertainty, the field of financial credit and enterprise risk management is facing the severe challenges of accelerated credit risk transmission speed and complex risk patterns. Accurate assessment of enterprise credit risk is a core link for preventing the spread of systemic financial risks and ensuring the safety and stable operation of financial institutions, and its importance is increasingly prominent.

[0003] Traditional technical means mainly rely on single financial indicators or simple statistical models for risk assessment, which has significant limitations: first, the static assessment mode cannot capture the dynamic changes of enterprises, such as ignoring the real-time impact of time series characteristics such as revenue volatility and frequent transaction frequency on debt servicing ability; second, isolated analysis of a single enterprise subject fails to effectively integrate the risk transmission effects of associated networks such as supply chains and guarantee circles, resulting in delayed identification of associated default risks. These defects make it difficult for traditional methods to meet the needs of financial institutions for early identification, early warning and early disposal of risks in complex market environments.

[0004] Based on this, the application provides an AI Agent driven credit risk early warning strategy automatic evaluation method, device, control equipment and computer readable storage medium. SUMMARY

[0005] In order to improve the problem that the traditional method is difficult to meet the needs of financial institutions for early identification, early warning and early disposal of risks in complex market environments, the application provides an AI Agent driven credit risk early warning strategy automatic evaluation method, device, control equipment and computer readable storage medium.

[0006] In a first aspect, the application provides an AI Agent driven credit risk early warning strategy automatic evaluation method, which adopts the following technical solution: comprising:

[0007] Obtain structured financial data, unstructured text data and dynamic time series data of enterprises through multi-source data, and perform standardization processing to output a standardized risk data set;

[0008] Based on the standardized risk data set, extract static financial features, dynamic derived features and associated network features, and generate a high-discrimination feature set through feature screening and dimensionality reduction optimization;

[0009] Adopting a multi-model integration strategy, inputting the high-differentiation feature set into a neural network, a Stacking integrated model and a graph neural network, and outputting enterprise risk scores and grades;

[0010] Based on the enterprise risk score and grade, automatically matching early warning rules, and monitoring and warning enterprise credit in real time according to the early warning rules.

[0011] Preferably, the structured financial data, unstructured text data and dynamic time series data of the enterprise are obtained through multi-source data fusion technology, and standardized processing is performed, and a standardized risk data set is output, including:

[0012] Connecting enterprise ERP systems and credit investigation platform APIs to collect balance sheets, profit and loss statements, cash flow statements and historical default records, and using OCR technology to identify financial reports to obtain the structured financial data;

[0013] Based on a large language model, the implicit risk signals of the enterprise annual report are analyzed to obtain the unstructured text data;

[0014] Using federated learning technology to collect enterprise real-time transaction records, Internet of Things production data and economic indicators, and completing data encryption fusion without sharing original data to obtain the dynamic time series data;

[0015] Constructing a generative adversarial network, using an LSTM time series prediction model to fill in short-term missing values, and marking long-term missing data as data anomaly features;

[0016] Based on a pre-set standardized rule library, the financial indicators are quantile standardized, and the time series data is uniformly converted to the same timestamp format to output the standardized risk data set.

[0017] Preferably, based on the standardized risk data set, static financial features, dynamic derived features and associated network features are extracted, and high-differentiation feature sets are generated through feature screening and dimensionality reduction optimization, including:

[0018] From the standardized risk data set, indicators reflecting the basic financial condition of the enterprise are extracted, including solvency indicators, profitability indicators and operational capacity indicators, forming the static financial features;

[0019] Based on the dynamic time series data in the standardized risk data set, the rate of change, volatility and trend slope of the indicators are calculated through time series analysis, and the dynamic derived features are generated in combination with the sliding window statistical method;

[0020] According to the enterprise correlation relationship data, a network topology structure is constructed, node degree centrality, intermediate centrality and edge weight features are extracted, and the associated network features are generated in combination with a network propagation algorithm.

[0021] Performing feature importance evaluation and multicollinearity test on the static financial feature subset, dynamic derived feature subset and associated network feature subset, removing redundant features, and then mapping the high-dimensional feature space to a low-dimensional space through dimension reduction algorithm to generate a high-discrimination feature set.

[0022] Preferably, the high-discrimination feature set is input into a neural network, a Stacking integrated model and a graph neural network using a multi-model integration strategy, and an enterprise risk score and grade are output.

[0023] The static financial features and the dynamic derived features are input into a neural network, and the associated network features are input into a graph neural network to obtain meta-features as a Stacking integrated model.

[0024] The neural network is trained using regularization and early stopping mechanism, the graph neural network is optimized using neighborhood aggregation algorithm to propagate node features, the Stacking integrated model is combined by screening base classifiers and meta-classifiers through cross-validation, and the model training result is compared with a preset validation set to obtain a validation set risk prediction error.

[0025] Based on the validation set risk prediction error, the weights of each model are dynamically adjusted, and based on the meta-features, the base risk score output by the neural network and the associated conduction score output by the graph neural network are fused by weighted voting to generate input features for the Stacking integrated model.

[0026] Based on the input features, a comprehensive risk score is output by the Stacking integrated model, and the comprehensive risk score is mapped to a risk grade by a preset risk grade division rule, and the risk grade includes at least three levels of low risk, medium risk and high risk.

[0027] Preferably, the comparison of the model training result with the preset validation set to obtain the validation set risk prediction error comprises:

[0028] The original sample data is divided into a training set, a validation set and a test set, and the validation set is independent of the model training process.

[0029] Based on the training set, the neural network, the Stacking integrated model and the graph neural network are trained, and the trained neural network and graph neural network are used to predict the risk of the validation set, respectively. The predicted risk score is compared with the true risk label of the sample, and the validation set risk prediction error is calculated by a preset error index.

[0030] Preferably, the enterprise risk score and grade are automatically matched with the early warning rules, and the enterprise credit is monitored and warned in real time according to the early warning rules.

[0031] The preset multi-dimensional early warning rule includes a risk level threshold rule, a score fluctuation rule, and a feature anomaly rule;

[0032] The enterprise risk score and level are input into a rule library, and corresponding early warning rules are automatically associated through a matching algorithm to generate early warning instructions including early warning levels, trigger conditions, and disposal suggestions;

[0033] Real-time collection of enterprise credit data and associated features, updating risk scores and levels according to a preset period;

[0034] If the updated score or level meets the trigger condition of the matched early warning rule, an early warning signal is pushed to the credit management system through an interface, and the risk control personnel are notified synchronously;

[0035] Record the early warning response results and feed back the disposal effects to the rule library.

[0036] Preferably, after recording the early warning response results and feeding back the disposal effects to the rule library, it further includes:

[0037] Collect the early warning response results and the actual credit risk performance of the corresponding enterprise to generate a feedback data set;

[0038] Update the standardized risk data set based on the feedback data set, and optimize the importance evaluation threshold value through an incremental learning mechanism to dynamically adjust the weight proportion of the static financial features, the dynamic derivative features, and the associated network features;

[0039] Based on the updated high-differentiation feature set, re-execute the multi-model training process, and use the actual risk results in the feedback data set as a new verification set;

[0040] Through the verification set risk prediction error correction model weight distribution strategy, realize the adaptive iteration of neural networks, Stacking integrated models, and graph neural networks;

[0041] According to the actual trigger effectiveness of the early warning rules in the feedback data set, adjust the threshold value parameters of the early warning rule library.

[0042] In a second aspect, an AI Agent-driven credit risk early warning strategy automatic evaluation device is provided, which adopts the following technical scheme, comprising:

[0043] A data processing module acquires structured financial data, unstructured text data, and dynamic time series data of an enterprise through multiple data sources, and performs standardized processing to output a standardized risk data set;

[0044] The feature extraction module is configured for extracting static financial features, dynamic derived features and associated network features based on the standardized risk data set, and generating a high-discrimination feature set through feature screening and dimension reduction optimization.

[0045] The model training module is configured for inputting the high-discrimination feature set into a neural network, a Stacking integrated model and a graph neural network by using a multi-model integrated strategy, and outputting an enterprise risk score and a grade.

[0046] The risk early warning module is configured for automatically matching a warning rule based on the enterprise risk score and the grade, and performing real-time monitoring and early warning on enterprise credit according to the warning rule.

[0047] In a third aspect, the present application further provides a control device, which comprises:

[0048] The control device comprises a memory and a processor, and the memory stores a computer program capable of being loaded and executed by the processor to implement the AI Agent-driven credit risk early warning strategy automated assessment method.

[0049] In a fourth aspect, the present application further provides a computer readable storage medium storing a computer program capable of being loaded and executed by a processor to implement the AI Agent-driven credit risk early warning strategy automated assessment method.

[0050] In summary, the present application obtains enterprise structured financial data, unstructured text data and dynamic time series data through multi-source data fusion, constructs a risk data set through standardization processing, extracts static financial features, dynamic derived features and associated network features based on the data set, generates a high-discrimination feature set through screening and dimension reduction, adopts a multi-model integrated strategy of a neural network, a Stacking integrated model and a graph neural network to output an enterprise risk score and a grade, matches a warning rule based on the score and the grade to realize real-time monitoring and early warning, and optimizes the whole process through early warning disposal effect feedback. This scheme improves data integrity through multi-dimensional data fusion, enhances risk characterization capability through multi-feature cooperation, improves prediction accuracy through multi-model integration, realizes adaptive optimization of features, models and rules through a feedback mechanism, effectively improves the real-time performance, accuracy and dynamic adjustment capability of credit risk early warning, and provides a credit risk management scheme. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 FIG. 1 is a flowchart of an AI Agent-driven credit risk early warning strategy automated assessment method.

[0052] Figure 2 FIG. 1 is a structural block diagram of an AI Agent-driven credit risk early warning strategy automated assessment device. DETAILED DESCRIPTION

[0053] The following description is provided in connection with Figure 1 - Figure 2 The application is further described in detail.

[0054] In the context of the rapid development of current financial technology, traditional credit risk assessment methods face challenges such as single data source, limited feature dimension, insufficient model adaptability, and delayed early warning response, making it difficult to meet the needs of financial institutions for dynamic risk monitoring.

[0055] To solve the above problems, the present patent proposes an AI Agent-driven credit risk early warning strategy automatic evaluation method, which integrates structured financial data, unstructured text data and dynamic time series data through multi-source data fusion technology, and constructs a standardized risk data set. Based on the data set, static financial features, dynamic derived features and associated network features are extracted, and a high-discrimination feature set is generated after screening and optimization. A multi-model integration strategy of neural network, Stacking integrated model and graph neural network is used to output enterprise risk score and grade, and automatic matching of early warning rules is realized to achieve real-time monitoring and early warning. At the same time, through the feedback of early warning disposal effect, a whole-process self-adaptive optimization mechanism is formed.

[0056] With reference to Figure 1 The embodiments of the present application at least include steps S10 to S40.

[0057] S10, structured financial data, unstructured text data and dynamic time series data of enterprises are obtained through multi-source data, and standardized processing is performed to output a standardized risk data set.

[0058] S20, based on the standardized risk data set, static financial features, dynamic derived features and associated network features are extracted, and a high-discrimination feature set is generated through feature screening and dimensionality reduction optimization.

[0059] S30, a multi-model integration strategy is adopted, and the high-discrimination feature set is input into neural network, Stacking integrated model and graph neural network to output enterprise risk score and grade.

[0060] S40, based on the enterprise risk score and grade, the early warning rules are automatically matched, and the enterprise credit is monitored and warned in real time according to the early warning rules.

[0061] Specifically, in the bank credit approval scenario, the system obtains structured financial data such as balance sheet by interfacing with enterprise ERP system, uses NLP to analyze annual report text to mine implicit risks, uses federated learning to encrypt and fuse real-time transaction water and other dynamic data, and eliminates data heterogeneity through standardization processing to lay a high-quality data foundation for risk assessment; extract static features such as solvency from standardized data, calculate dynamic features such as volatility based on time series data, extract centrality features from enterprise correlation network, remove redundancy through screening and dimensionality reduction, and form a high-discrimination feature set that can accurately depict enterprise risks; input the features into neural networks and graph neural networks to process financial and correlation data respectively, and output risk scores and grades through a Stacking integrated model weighted fusion, and multiple models are used to improve prediction accuracy; automatically match early warning rules according to the scoring level, monitor enterprise credit data in real time, and push early warning signals and notify risk control personnel when the risk exceeds the standard, effectively reducing the bad debt rate.

[0062] In some embodiments, step S10 specifically includes the following steps: interfacing with enterprise ERP system and credit investigation platform API, collecting balance sheet, profit table, cash flow table and historical default records, and using OCR technology to identify financial reports to obtain structured financial data; analyzing the implicit risk signals of enterprise annual reports based on a large language model to obtain unstructured text data; using federated learning technology to collect enterprise real-time transaction water, Internet of Things production data and economic indicators, and completing data encryption fusion without sharing original data to obtain dynamic time series data; constructing a generative adversarial network, using an LSTM time series prediction model to fill in short-term missing values, and marking long-term missing data as data anomaly features; based on a pre-set standardization rule library, quantile standardizing financial indicators, and uniformly converting time series data into the same timestamp format to output standardized risk data sets.

[0063] Specifically, by directly connecting to enterprise ERP system and credit investigation platform API to obtain structured data such as balance sheet, and using OCR technology to digitize financial information from paper financial reports, the problem of low data collection efficiency is solved; using a large language model to analyze annual report management discussion and other texts to mine implicit risk signals such as market competition, making up for the lack of structured data; using federated learning technology to encrypt and fuse transaction water, Internet of Things production data and macroeconomic indicators without obtaining enterprise original data, balancing data privacy and multi-source data value; filling in short-term data gaps and marking long-term missing data as abnormal features through a generative adversarial network and an LSTM model to improve data integrity; finally, quantile normalizing financial indicators and unifying time stamps for time series data according to a standardization rule library to eliminate data heterogeneity and lay a high-quality data foundation for subsequent risk feature extraction.

[0064] In some embodiments, step S20 specifically comprises the following steps: extracting indicators reflecting the basic financial status of the enterprise from the standardized risk data set, including solvency indicators, profitability indicators and operational capacity indicators, to form static financial features; based on the dynamic time series data in the standardized risk data set, calculating the change rate, volatility and trend slope of the indicators through time series analysis, and generating dynamic derivative features in combination with the sliding window statistical method; constructing a network topology according to the enterprise correlation data, extracting node degree centrality, intermediate centrality and edge weight features, and generating correlation network features in combination with the network propagation algorithm; performing feature importance evaluation and multicollinearity test on the static financial feature subset, the dynamic derivative feature subset and the correlation network feature subset, removing redundant features, and then mapping the high-dimensional feature space to the low-dimensional space through the dimension reduction algorithm to generate a high-discrimination feature set.

[0065] Specifically, the system extracts static financial features such as liquidity ratio and gross profit margin from standardized data, which intuitively reflects the current debt paying, profitability and operational capacity of the enterprise; based on dynamic time series data such as transaction flow, dynamic derivative features such as revenue growth rate volatility and inventory turnover trend slope are calculated to capture the change rule of the enterprise's operating status; a network topology is constructed according to the enterprise equity association and supply chain relationship, and correlation network features such as core enterprise node degree centrality and guarantee chain intermediate centrality are extracted to identify the risk transmission path; through feature importance evaluation, redundant indicators such as accounts receivable turnover rate are removed, and after multicollinearity test and PCA dimension reduction, the high-dimensional features are mapped to the low-dimensional space to generate a feature set that takes into account comprehensiveness and discrimination, providing accurate input for subsequent risk prediction models and effectively improving the accuracy and robustness of credit risk assessment.

[0066] In some embodiments, step S30 specifically comprises the following steps: inputting static financial features and dynamic derivative features into a neural network, and inputting correlation network features into a graph neural network to obtain meta-features as a Stacking ensemble model; training the neural network using regularization and early stopping mechanism, optimizing node feature propagation of the graph neural network using neighborhood aggregation algorithm, screening the base classifier and meta-classifier combination of the Stacking ensemble model through cross-validation, and comparing the model training result with the preset validation set to obtain the validation set risk prediction error; based on the validation set risk prediction error, dynamically adjusting the weights of each model, based on the meta-features, combining the base risk score output by the neural network and the correlation transmission score output by the graph neural network through weighted voting to generate input features of the Stacking ensemble model; based on the input features, outputting a comprehensive risk score through the Stacking ensemble model, and mapping the comprehensive risk score to a risk level through a preset risk level division rule, the risk level at least including three levels of low risk, medium risk and high risk.

[0067] Specifically, the system inputs static financial and dynamic derivative features into a neural network to process time series changes, and inputs network feature into a graph neural network to capture risk transmission between enterprises. The outputs of the two are used as meta-features of a Stacking integrated model. Regularization and early stopping mechanism are used to avoid overfitting of the neural network, and a neighborhood aggregation algorithm is used to optimize the propagation of node features in the graph neural network. Cross-validation is used to screen the combination of base classifiers and meta-classifiers of Stacking. Comparative validation set error is used to dynamically adjust the model weight. The basic risk score and the associated transmission score are fused to generate input features. Finally, the comprehensive risk score is output by Stacking and mapped to low, medium and high levels. Multiple models are used to improve the prediction accuracy and provide accurate risk level reference for credit approval.

[0068] In some embodiments, step S40 specifically includes the following steps: dividing the original sample data into a training set, a validation set, and a test set, and the validation set is independent of the model training process; training the neural network, the Stacking integrated model, and the graph neural network based on the training set, and using the trained neural network and the graph neural network to predict the risk of the validation set, comparing the predicted risk score with the real risk label of the sample, and calculating the risk prediction error of the validation set by a preset error indicator.

[0069] Specifically, in the bank credit risk control scenario, historical credit data is divided into a training set (70% of the samples), a validation set (15%), and a test set (15%). The validation set is independent of the model training to ensure objectivity of the evaluation. The neural network, the graph neural network, and the Stacking integrated model are trained based on the training set. The trained neural network and the graph neural network are used to predict the risk of the validation set. The output risk score is compared with the real default label of the enterprise. The prediction error of the validation set is calculated by RMSE (Root Mean Square Error). This error is used to dynamically adjust the weight of each model (such as reducing the weight of the high-error model) to avoid overfitting and optimize the integrated strategy, ensuring that the final risk score has strong generalization ability in the actual credit scenario and provides a reliable model output basis for subsequent early warning rule matching.

[0070] In some embodiments, step S50 specifically includes the following steps: presetting multi-dimensional early warning rules, including risk level threshold rules, score fluctuation rules, and feature anomaly rules; inputting the enterprise risk score and level into the rule library, and automatically associating the corresponding early warning rules through a matching algorithm to generate early warning instructions containing early warning levels, trigger conditions, and disposal suggestions; real-time collection of enterprise credit data and associated features, updating of risk scores and levels according to a preset period; if the updated score or level meets the trigger condition of the matched early warning rule, an early warning signal is pushed to the credit management system through an interface, and the risk control personnel are notified synchronously; the early warning response result is recorded, and the disposal effect is fed back to the rule library.

[0071] Specifically, the preset multi-dimensional early warning rules include risk level threshold rules (e.g., a high risk level directly triggers early warning), score fluctuation rules (e.g., a risk score increases by 50% within 1 month), and feature anomaly rules (e.g., a liquidity ratio suddenly drops to less than 50% of the industry average); the real-time risk score and level of an enterprise are input into the rule library, corresponding rules are automatically associated through a fuzzy matching algorithm, instructions containing early warning levels, triggering conditions (e.g., a revenue fluctuation rate exceeding a threshold for 3 consecutive months), and disposal suggestions (e.g., suspending new credit granting) are generated; enterprise credit data is collected daily to update the score, and if the triggering condition is met, an early warning signal is pushed to a credit management system through an API, and a risk control personnel is notified through a short message and an email; disposal results (e.g., the enterprise has been required to supplement a mortgage) are recorded, and indicators such as a false alarm rate and disposal time efficiency are fed back to the rule library, the rule threshold is continuously optimized, and the risk response efficiency is improved.

[0072] In some embodiments, step S50 further includes the following steps: collecting early warning response results and actual credit risk performance of corresponding enterprises to generate a feedback data set; updating the standardized risk data set based on the feedback data set, and optimizing the importance evaluation threshold through an incremental learning mechanism to dynamically adjust the weight proportion of static financial features, dynamic derived features, and associated network features; based on the updated high-discrimination feature set, re-executing the multi-model training process, taking the actual risk results in the feedback data set as a new validation set; through the validation set risk prediction error correction model weight distribution strategy, realizing adaptive iteration of the neural network, the Stacking integrated model, and the graph neural network; and according to the actual triggering effectiveness of the early warning rules in the feedback data set, adjusting the threshold parameters of the early warning rule library.

[0073] Specifically, the system generates a feedback data set by collecting early warning response results (e.g., customer defaults 3 months after early warning) and actual risk performance of enterprises (delinquency rate, non-performing loan rate); the standardized risk data set is updated with this data set, and the feature weights are dynamically adjusted through incremental learning (the importance proportion of the associated enterprise guarantee default times feature is increased); the actual default cases in the feedback data are taken as a new validation set, the neural network and the graph neural network are retrained, the weight distribution strategy of the Stacking integrated model is corrected through the validation set error, and the threshold of the early warning is adjusted according to the actual triggering effectiveness (e.g., the default prediction accuracy of the liquidity ratio sudden drop rule) to continuously improve the accuracy and timeliness of risk early warning and reduce credit asset losses.

[0074] The implementation principle of an AI Agent driven credit risk early warning strategy automatic evaluation method according to an embodiment of the present application is as follows: structured financial data, unstructured text data and dynamic time series data of an enterprise are obtained through multi-source data fusion, and a risk data set is constructed through standardization processing; static financial features, dynamic derived features and associated network features are extracted based on the data set, and a high-discrimination feature set is generated through screening and dimension reduction; a multi-model integration strategy of neural network, Stacking integrated model and graph neural network is adopted to output an enterprise risk score and grade; real-time monitoring and early warning are realized based on the matching of the score and grade, and the whole process is optimized through feedback of early warning disposal effect. The scheme improves data integrity through multi-dimensional data fusion, enhances risk characterization capability through multi-feature cooperation, improves prediction accuracy through multi-model integration, realizes adaptive optimization of features, models and rules through a feedback mechanism, effectively improves the real-time performance, accuracy and dynamic adjustment capability of credit risk early warning, and provides a credit risk management scheme.

[0075] Figure 1 FIG. 1 is a flowchart of an AI Agent driven credit risk early warning strategy automatic evaluation method according to an embodiment of the present application. It should be understood that although each step in the flowchart is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow; unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences; and Figure 1 at least part of the steps in the flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of the sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps. Figure 1

[0076] Based on the same technical concept, referring to Figure 2 the present application also provides an AI Agent driven credit risk early warning strategy automatic evaluation device, which adopts the following technical scheme, and the device comprises:

[0077] a data processing module, which obtains structured financial data, unstructured text data and dynamic time series data of an enterprise through multi-source data, and performs standardization processing to output a standardized risk data set;

[0078] a feature extraction module, which is configured to extract static financial features, dynamic derived features and associated network features based on the standardized risk data set, and generate a high-discrimination feature set through feature screening and dimension reduction optimization;

[0079] ​a model training module configured to input the high-discrimination feature set into a neural network, a stacking integrated model, and a graph neural network using a multi-model integrated strategy, and output an enterprise risk score and a grade;

[0080] a risk early warning module configured to automatically match early warning rules based on the enterprise risk score and the grade, and perform real-time monitoring and early warning on enterprise credit according to the early warning rules.

[0081] In some embodiments, the data processing module is specifically configured to obtain structured financial data, unstructured text data, and dynamic time series data of the enterprise through multi-source data, and perform standardized processing to output a standardized risk data set;

[0082] Based on the standardized risk data set, static financial features, dynamic derived features, and associated network features are extracted, and the high-discrimination feature set is generated through feature screening and dimensionality reduction optimization;

[0083] The multi-model integrated strategy is used to input the high-discrimination feature set into a neural network, a stacking integrated model, and a graph neural network, and output an enterprise risk score and a grade;

[0084] The risk early warning module is configured to automatically match early warning rules based on the enterprise risk score and the grade, and perform real-time monitoring and early warning on enterprise credit according to the early warning rules.

[0085] In some embodiments, the feature extraction module is specifically configured to interface with an enterprise ERP system and a credit investigation platform API, collect balance sheets, profit statements, cash flow statements, and historical default records, and use OCR technology to identify financial reports to obtain structured financial data;

[0086] Based on a large language model, the implicit risk signals of the enterprise annual report are analyzed to obtain unstructured text data;

[0087] Using federated learning technology, real-time transaction records, Internet of Things production data, and economic indicators of the enterprise are collected, data encryption and fusion are completed without sharing original data, and dynamic time series data is obtained;

[0088] A generative adversarial network is constructed, an LSTM time series prediction model is used to fill in short-term missing values, and long-term missing data is marked as data anomaly features;

[0089] Based on a pre-set standardized rule library, the financial indicators are quantile standardized, and the time series data is uniformly converted to the same timestamp format to output a standardized risk data set.

[0090] In some embodiments, the model training module is specifically configured to extract indicators reflecting the basic financial condition of the enterprise from the standardized risk data set, including solvency indicators, profitability indicators, and operational capacity indicators, to form static financial features;

[0091] Based on the dynamic timing data in the standardized risk dataset, the rate of change, the volatility rate and the trend slope of the indicators are calculated by time series analysis, and the dynamic derivative features are generated by combining the sliding window statistical method;

[0092] According to the enterprise correlation data, the network topology structure is constructed, the node degree centrality, the intermediate centrality and the edge weight features are extracted, and the correlation network features are generated by combining the network propagation algorithm;

[0093] The feature importance evaluation and multicollinearity test are performed on the static financial feature subset, the dynamic derivative feature subset and the correlation network feature subset. After removing the redundant features, the high-dimensional feature space is mapped to the low-dimensional space by the dimension reduction algorithm to generate the high-discrimination feature set.

[0094] In some embodiments, the risk warning module is specifically configured to input the static financial features and the dynamic derivative features into a neural network, and input the correlation network features into a graph neural network to obtain meta-features as a Stacking ensemble model;

[0095] The neural network is trained using regularization and early stopping mechanism, the graph neural network is optimized using neighborhood aggregation algorithm to propagate node features, the Stacking ensemble model is combined by cross-validation to screen base classifiers and meta-classifiers, and the model training result is compared with the preset validation set to obtain the validation set risk prediction error;

[0096] Based on the validation set risk prediction error, the weights of each model are dynamically adjusted, based on the meta-features, the base risk score output by the neural network and the correlation conduction score output by the graph neural network are fused by weighted voting to generate the input features of the Stacking ensemble model;

[0097] Based on the input features, the comprehensive risk score is output by the Stacking ensemble model, and the comprehensive risk score is mapped to the risk level by the preset risk level division rule, and the risk level includes at least three levels of low risk, medium risk and high risk.

[0098] In some embodiments, the risk warning module is specifically configured to divide the original sample data into a training set, a validation set and a test set, and the validation set is independent of the model training process;

[0099] Based on the training set, the neural network, the Stacking ensemble model and the graph neural network are trained, and the risk prediction of the validation set is performed by the trained neural network and graph neural network respectively, and the predicted output risk score is compared with the sample true risk label, and the validation set risk prediction error is calculated by the preset error index.

[0100] In some embodiments, the risk early warning module is specifically configured to preset multi-dimensional early warning rules, the early warning rules including risk level threshold rules, score fluctuation rules and feature anomaly rules;

[0101] The enterprise risk score and level are input into the rule library, and corresponding early warning rules are automatically associated through a matching algorithm to generate early warning instructions including early warning levels, trigger conditions and disposal suggestions;

[0102] The enterprise credit data and associated features are collected in real time, and the risk score and level are updated according to a preset period;

[0103] If the updated score or level meets the trigger condition of the matched early warning rule, an early warning signal is pushed to the credit management system through an interface, and the risk control personnel are notified synchronously;

[0104] The early warning response results are recorded, and the disposal effects are fed back to the rule library.

[0105] In some embodiments, the risk early warning module is specifically configured to collect early warning response results and actual credit risk performance of corresponding enterprises to generate a feedback data set;

[0106] The standardized risk data set is updated based on the feedback data set, and the importance evaluation threshold is optimized through an incremental learning mechanism to dynamically adjust the weight proportion of static financial features, dynamic derived features and associated network features;

[0107] Based on the updated high-discrimination feature set, the multi-model training process is re-executed, and the actual risk results in the feedback data set are taken as a new validation set;

[0108] The model weight distribution strategy is corrected through the validation set risk prediction error to realize adaptive iteration of neural networks, Stacking integrated models and graph neural networks;

[0109] According to the actual trigger effectiveness of the early warning rules in the feedback data set, the threshold parameters of the early warning rule library are adjusted.

[0110] Embodiments of the present application also disclose a control device.

[0111] Specifically, the control device includes a memory and a processor, and the memory stores a computer program capable of being loaded and executed by the processor to implement the AI Agent-driven credit risk early warning strategy automatic evaluation method described above.

[0112] Embodiments of the present application also disclose a computer readable storage medium.

[0113] Specifically, the computer readable storage medium, which stores the computer program capable of being loaded and executed by the processor, such as the AIAgent driven credit risk early warning strategy automatic evaluation method, the computer readable storage medium includes, for example, a variety of storage program codes, such as a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk.

[0114] The above are preferred embodiments of the present application, not limited to the protection scope of the present application, therefore: all equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. An AI Agent-driven credit risk early warning strategy automatic evaluation method, characterized in that, The application comprises the following steps: Obtaining structured financial data, unstructured text data and dynamic time series data of enterprises through multi-source data, and performing standardization processing to output a standardized risk data set; Based on the standardized risk data set, static financial features, dynamic derived features and correlation network features are extracted, and high-discrimination feature sets are generated through feature screening and dimensionality reduction optimization; Using a multi-model integration strategy, the high-discrimination feature sets are input into neural networks, Stacking integrated models and graph neural networks to output enterprise risk scores and grades; Based on the enterprise risk scores and grades, pre-warning rules are automatically matched, and the enterprise credit is monitored and pre-warned in real time according to the pre-warning rules; The multi-model integration strategy includes the following steps: The static financial features and dynamic derived features are input into neural networks, and the correlation network features are input into graph neural networks to obtain meta-features as Stacking integrated models; The neural networks are trained using regularization and early stopping mechanism, the graph neural networks are optimized using neighborhood aggregation algorithm to propagate node features, the Stacking integrated models are combined by screening base classifiers and meta-classifiers through cross-validation, and the model training results are compared with the preset validation set to obtain the validation set risk prediction error; Based on the validation set risk prediction error, the weights of each model are dynamically adjusted, and based on the meta-features, the base risk scores output by the neural networks and the correlation conduction scores output by the graph neural networks are fused by weighted voting to generate input features of the Stacking integrated model; Based on the input features, the comprehensive risk score is output by the Stacking integrated model, and the comprehensive risk score is mapped to a risk grade by a preset risk grade division rule, and the risk grade at least includes three levels of low risk, medium risk and high risk; The multi-source data includes the following steps: Connecting enterprise ERP systems and credit investigation platform APIs to collect balance sheets, profit tables, cash flow tables and historical default records, and using OCR technology to identify financial reports to obtain the structured financial data; Based on the large language model, the implicit risk signals of the annual report of the enterprise are analyzed to obtain the unstructured text data; Using federated learning technology to collect real-time transaction records, Internet of Things production data and economic indicators of enterprises, and completing data encryption fusion without sharing original data to obtain the dynamic time series data; Constructing a generative adversarial network, using an LSTM time series prediction model to fill in short-term missing values, and marking long-term missing data as data anomaly features; Based on the preset standardization rule library, the financial indicators are quantile standardized, and the time series data is converted to the same timestamp format to output the standardized risk data set; The standardized risk data set is used to extract static financial features, dynamic derived features and associated network features, high-discrimination feature sets are generated through feature screening and dimension reduction optimization, and the high-discrimination feature sets include: Indexes reflecting the basic financial status of the enterprise are extracted from the standardized risk data set, including solvency indexes, profitability indexes and operation capability indexes, to form the static financial features; Based on the dynamic time series data in the standardized risk data set, the change rate, volatility rate and trend slope of the indexes are calculated through time series analysis, and the dynamic derived features are generated by combining the sliding window statistical method; A network topology structure is constructed according to the enterprise association relationship data, node degree centrality, intermediate centrality and edge weight features are extracted, and the associated network features are generated by combining the network propagation algorithm; The feature importance evaluation and multicollinearity test are performed on the static financial feature subset, dynamic derived feature subset and associated network feature subset, and after removing the redundant features, the high-dimensional feature space is mapped to a low-dimensional space through a dimension reduction algorithm to generate high-discrimination feature sets.

2. The AI Agent-driven credit risk early warning strategy automated assessment method according to claim 1, characterized in that, The model training result is compared with the preset validation set to obtain the validation set risk prediction error, including: The original sample data is divided into a training set, a validation set and a test set, and the validation set is independent of the model training process; The neural network, Stacking integrated model and graph neural network are trained based on the training set, and the trained neural network and graph neural network are used to predict the risk of the validation set, respectively. The predicted risk score is compared with the actual risk label of the sample, and the validation set risk prediction error is calculated through a preset error index.

3. The AI Agent-driven credit risk early warning strategy automated assessment method according to claim 2, characterized in that, The enterprise risk score and grade are automatically matched with the early warning rules, and the enterprise credit is monitored and warned in real time according to the early warning rules, including: The early warning rules are preset in multiple dimensions, including risk grade threshold rules, score fluctuation rules and feature anomaly rules; The enterprise risk score and grade are input into the rule library, and the corresponding early warning rules are automatically associated through a matching algorithm to generate early warning instructions containing early warning levels, trigger conditions and disposal suggestions; The enterprise credit data and associated features are collected in real time, and the risk score and grade are updated at a preset period; If the updated score or grade meets the trigger condition of the matched early warning rule, an early warning signal is pushed to the credit management system through an interface, and the risk control personnel are notified synchronously; The early warning response results are recorded, and the disposal effects are fed back to the rule library.

4. The AI Agent-driven credit risk early warning strategy automated assessment method according to claim 3, characterized in that, After recording the early warning response results and feeding back the disposal effects to the rule library, it further includes: The early warning response results and the actual credit risk performance of the corresponding enterprise are collected to generate a feedback data set; The standardized risk data set is updated based on the feedback data set, and the importance evaluation threshold is optimized through an incremental learning mechanism to dynamically adjust the weight proportion of the static financial features, the dynamic derived features and the associated network features; Based on the updated high-discrimination feature set, the multi-model training process is re-executed, and the actual risk results in the feedback data set are used as a new validation set. The verification set risk prediction error correction model weight distribution strategy realizes self-adaptive iteration of a neural network, a stacking integrated model and a graph neural network; According to the actual triggering effectiveness of the early warning rules in the feedback data set, the threshold parameters of the early warning rule library are adjusted.

5. An AI Agent-driven credit risk early warning strategy automatic evaluation device, characterized in that, The device comprises: A data processing module acquires structured financial data, unstructured text data and dynamic time series data of an enterprise through multiple data sources, and performs standardization processing to output a standardized risk data set; A feature extraction module extracts static financial features, dynamic derived features and correlation network features based on the standardized risk data set, and generates a high-discrimination feature set through feature screening and dimensionality reduction optimization; A model training module inputs the high-discrimination feature set into a neural network, a stacking integrated model and a graph neural network using a multi-model integration strategy, and outputs an enterprise risk score and grade; A risk early warning module automatically matches early warning rules based on the enterprise risk score and grade, and performs real-time monitoring and early warning on enterprise credit according to the early warning rules; The model training module is specifically configured to input the static financial features and the dynamic derived features into the neural network, and input the correlation network features into the graph neural network to obtain meta-features of the stacking integrated model; The neural network is trained using a regularization and early stopping mechanism, the graph neural network is optimized using a neighborhood aggregation algorithm to propagate node features, the stacking integrated model is combined by screening base classifiers and meta-classifiers through cross-validation, and the model training result is compared with a preset verification set to obtain a verification set risk prediction error; Based on the verification set risk prediction error, the weights of each model are dynamically adjusted, and based on the meta-features, a basic risk score output by the neural network and a correlation conduction score output by the graph neural network are fused by weighted voting to generate input features of the stacking integrated model; Based on the input features, a comprehensive risk score is output by the stacking integrated model, and the comprehensive risk score is mapped to a risk grade by a preset risk grade division rule, and the risk grade at least includes three levels of low risk, medium risk and high risk. The data processing module is specifically configured to interface with an enterprise ERP system and a credit investigation platform API, collect balance sheets, profit tables, cash flow tables and historical default records, and acquire the structured financial data by recognizing financial reports using OCR technology; Based on a large language model, the implicit risk signals of the annual report of the enterprise are analyzed to obtain the unstructured text data; The dynamic time series data is acquired by collecting real-time transaction records, Internet of Things production data and economic indicators of the enterprise using federated learning technology without sharing the original data; A generative adversarial network is constructed, an LSTM time series prediction model is used to fill in short-term missing values, and long-term missing data is marked as data anomaly features; Based on a preset standardization rule library, the financial indicators are quantile standardized, and the time series data is uniformly converted to the same timestamp format to output the standardized risk data set. The feature extraction module is specifically configured to extract indexes reflecting the basic financial status of an enterprise from the standardized risk dataset, including solvency indexes, profitability indexes and operation capability indexes, to form the static financial features. Based on the dynamic time series data in the standardized risk dataset, the change rate, the fluctuation rate and the trend slope of indexes are calculated through time series analysis, and the dynamic derived features are generated in combination with a sliding window statistical method; A network topology structure is constructed according to enterprise correlation relationship data, node degree centrality, intermediate centrality and edge weight features are extracted, and the correlation network features are generated in combination with a network propagation algorithm; After feature importance evaluation and multicollinearity test are performed on the static financial feature subset, the dynamic derived feature subset and the correlation network feature subset, and redundant features are removed, high-dimensional feature space is mapped to low-dimensional space through a dimension reduction algorithm to generate a high-discrimination feature set.

6. A control device characterized by comprising: The device comprises: The device comprises:

7. A computer-readable storage medium, characterized in that, The device comprises: The device comprises:

Images