Dynamic credit atlas-based bill realization risk management and control method and system

Abstract

The invention relates to the technical field of bill credit evaluation, and discloses a bill realization risk management and control method and system based on a dynamic credit graph. According to the method, historical transaction data and real-time bill information of a bill holder are collected, and a bill credit feature set including bill discount frequency, cashing default records and related party credit rating is generated through a feature extraction engine. A dynamic credit graph is constructed based on the feature set, nodes of the dynamic credit graph represent bill holders and associated entities, and edges of the dynamic credit graph represent fund exchange relations and are bound with credit conduction strength labels. And traversing the atlas by adopting a stratified sampling algorithm, and identifying a high-risk bill cluster and a core conduction link formed by edges of which the credit conduction intensity exceeds a threshold value. And according to the cluster distribution density and the link topology structure, generating a credit hedging strategy set including discount amount dynamic allocation and risk reserve fund calculation and extraction proportion adjustment. And receiving an external market fluctuation signal, correcting the edge weight of the atlas through the credit conduction intensity attenuation model, and triggering iterative updating of the strategy set.

Description

Technical Field

[0001] This invention relates to the field of bill liquidation risk management technology, specifically to a bill liquidation risk management method and system based on dynamic credit graphs. Background Technology

[0002] As an important financial instrument, bills play a crucial role in short-term corporate financing and payment settlement. Credit risk assessment during the bill realization process is a core challenge for financial institutions. Traditional bill credit assessment methods primarily rely on static analysis of the bill's intrinsic elements, such as the drawer's qualifications, the bill's amount, and maturity date, combined with limited third-party credit rating information. This method struggles to comprehensively reflect the true creditworthiness of the bill holder, particularly failing to effectively capture the potential risks accumulated during the bill's circulation process due to multiple endorsements and discounting. With the increasing complexity of participants in the bill market and the continuous extension of transaction chains, the limitations of traditional assessment models are becoming increasingly apparent.

[0003] Most existing risk assessment systems employ statistical models based on historical data, pricing risk by analyzing macroeconomic indicators such as bill default rates and discount rates. These models lack sensitivity to systemic risk and cannot dynamically respond to rapid changes in market conditions. More importantly, traditional methods neglect the complex network relationships between related parties in bill transactions. In the real-world bill circulation ecosystem, risk often propagates and amplifies across multiple entities through related-party transactions such as guarantees, endorsements, and discounting, forming risk clusters. Current technologies lack the ability to effectively identify and quantify such risk transmission paths.

[0004] Regarding risk response strategies, current practices often employ broad-based management methods such as uniform discount rate increases or quota controls, lacking refined hedging solutions tailored to bills of different risk levels. Financial institutions struggle to identify clusters of high-risk bills in a timely manner, and are even less able to predict potential risk transmission paths, resulting in lagging and unpredictable risk management measures. Furthermore, existing systems typically isolate risk assessment from market dynamics, failing to establish a linkage mechanism between external market signals and internal credit conditions, causing risk assessment results to become disconnected from the real-time market environment. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for managing the liquidity risks of bills based on dynamic credit graphs, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for managing the liquidity risk of bills based on dynamic credit graphs, the method comprising: Historical transaction data and real-time bill information of bill holders are collected, and a bill credit feature set is generated through a feature extraction engine. The bill credit feature set includes bill discounting frequency, payment default records and related party credit rating. A dynamic credit graph is constructed based on the credit feature set of bills. The nodes of the dynamic credit graph represent bill holders and related entities, and the edges represent fund transfer relationships and are bound with credit transmission strength labels. A hierarchical sampling algorithm is used to traverse the dynamic credit graph to identify high-risk bill clusters and core transmission links. The core transmission links are composed of edges whose credit transmission strength exceeds a dynamic threshold. Based on the distribution density of high-risk bill clusters and the topology of the core transmission link, a set of credit hedging strategies is generated. The set of credit hedging strategies includes a dynamic allocation scheme for bill discounting quotas and a scheme for adjusting the risk reserve provision ratio. Receive external market fluctuation signals, correct the edge weights of the dynamic credit graph through the credit transmission strength decay model, and trigger iterative updates of the credit hedging strategy set; The updated credit hedging strategy set is executed, and the credit risk diffusion path in the process of bill realization is monitored in parallel. The feedback is then fed into the feature extraction engine for incremental learning of the bill credit feature set.

[0007] Preferably, the specific implementation of the feature extraction engine generating the bill credit feature set includes: The ratio of the number of discount applications by bill holders to the total number of bills held in the past preset period is used to generate bill discounting frequency characteristics. Extract the number of non-payment or delayed payment events from the historical payment records of bills, and combine them with the total face value of the bills involved in the events to generate payment default record characteristics; Call a third-party credit rating interface to obtain the latest rating results of related parties, convert the rating results into numerical credit scores, and use them as credit rating features of related parties; The characteristics of bill discounting frequency, default record, and related party credit rating are integrated according to preset weight coefficients to output a bill credit feature set.

[0008] Preferably, the construction process of the dynamic credit graph further includes: Based on the flow data of funds between the bill holders, the proportion of the amount of a single transaction to the total historical transaction amount of both parties is calculated, and the initial credit transmission strength is generated by combining the transaction occurrence time decay factor. The initial credit transmission strength of all related transactions of the same note holder is normalized, and the normalization result is bound to the corresponding edge as a credit transmission strength label. When real-time bill information is updated, the credit transmission strength label is dynamically adjusted, and the adjustment range is proportional to the rate of change of the bill credit feature set.

[0009] Preferably, the execution process of the stratified sampling algorithm includes: The nodes of the dynamic credit graph are divided into three levels: high, medium, and low, based on the scores of the bill credit feature set. A predetermined proportion of nodes are randomly selected from each level as sample nodes, and the flow of funds of the sample nodes is tracked within a predetermined backtracking period. If the credit transmission strength labels of three consecutive edges in the capital flow path of a sample node are all higher than the hierarchical average, then the path is marked as a core transmission link. The standard deviation of the frequency of bill discounting among sample nodes in each level is statistically analyzed, and node clusters that exceed twice the standard deviation are identified as high-risk bill clusters.

[0010] Preferably, the logic for generating the credit hedging strategy set includes: The risk diffusion index is calculated based on the ratio of the node density of high-risk bill clusters to the number of core transmission links. If the risk diffusion index exceeds the first critical value, a dynamic allocation scheme for bill discounting quotas will be generated. This scheme requires a tiered reduction in quotas for new discounting applications within high-risk bill clusters. If the risk diffusion index is lower than the first threshold but higher than the second threshold, a risk reserve provision ratio adjustment plan is generated. This plan is based on the sum of credit transmission strength labels of the core transmission link, and linearly increases the reserve provision ratio.

[0011] Preferably, the correction process for the credit transmission strength attenuation model includes: Analyze the magnitude of interest rate changes and the increase in industry default rates in external market volatility signals to generate a market risk coefficient; The weakened edge weight is obtained by multiplying the credit transmission strength label of each edge in the dynamic credit graph by the market risk coefficient. When the weight of an edge after decay is lower than a preset percentage of its original value, the edge is cut off and the distribution density of the high-risk bill cluster is recalculated.

[0012] Preferably, the iterative update triggering conditions for the credit hedging strategy set include: The system detected that more than a preset number of edge weights in the dynamic credit graph were attenuated. Or it may identify that the node density of newly added high-risk bill clusters has reached a historical peak; Or it may receive a manual intervention instruction requiring a forced policy refresh.

[0013] Preferably, the incremental learning is implemented as follows: The actual default events that occur during the bill realization process are compared with the predicted risk path to calculate the credit assessment deviation rate; The weighting coefficients of the bill discounting frequency feature and the payment default record feature in the feature extraction engine are adjusted according to the credit assessment deviation rate. When the deviation rate continues to exceed the tolerance threshold, the topology reconstruction of the dynamic credit graph is triggered.

[0014] Preferably, the specific implementation of the dynamic allocation scheme for bill discounting quotas includes: Seventy percent of the benchmark quota will be reserved for the first discount application within a high-risk bill cluster; If the holder does not have any new default records in the subsequent preset period, then 90% of the benchmark amount will be restored; Otherwise, the credit limit will be frozen and the relevant third-party interface for the credit rating feature will be notified.

[0015] Preferably, the present invention also includes a bill monetization risk management system based on dynamic credit graphs, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the bill monetization risk management method based on dynamic credit graphs described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a dynamic credit graph, enabling a networked and visualized representation of credit risk in negotiable instruments. Traditional methods assess negotiable instruments as isolated entities, while this method treats related entities such as the holder, endorser, and discounter as nodes, their financial transactions as edges, and quantifies the credit transmission strength of the edges. This clearly reveals the potential transmission paths and aggregation patterns of risk within the negotiable instrument network, providing an intuitive basis for risk tracing and precise intervention.

[0017] A stratified sampling algorithm was employed to identify high-risk clusters and core transmission links, improving the efficiency and accuracy of risk identification. Instead of traversing the entire risk map, this method uses intelligent sampling to focus on key paths and areas with high credit transmission intensity, quickly locating risk concentration points and critical transmission channels. This allows risk management resources to be prioritized for the most critical links, improving the targeting and efficiency of risk prevention and control.

[0018] By generating credit hedging strategies based on cluster density and link topology, the system achieves refined and differentiated risk response. It not only identifies risks but also automatically generates customized hedging solutions, including dynamic allocation of credit limits and adjustment of reserves, based on the specific distribution characteristics and transmission structure of those risks. This network-structure-based strategy generation mechanism makes risk management measures more closely aligned with actual risk conditions, enhancing the scientific rigor and effectiveness of risk management.

[0019] By introducing external market signals and dynamically correcting the credit graph using an attenuation model, the system's environmental adaptability and real-time response capabilities are enhanced. The system can sense market fluctuations and adjust the credit transmission strength weights of related edges accordingly, ensuring that the credit graph and risk assessment results keep pace with market changes. This avoids assessment biases caused by information lag and guarantees the timeliness of risk decisions.

[0020] Establishing a feedback loop from strategy execution and risk monitoring to incremental learning endows the system with the ability to continuously optimize. While executing hedging strategies, the system continuously monitors the actual risk diffusion path and feeds the monitoring results back to the feature extraction engine for model updates. This closed-loop learning mechanism enables the system to continuously accumulate experience from practice, self-improve its evaluation model and strategy library, and gradually enhance its risk assessment and control effectiveness in different market environments. Attached Figure Description

[0021] Figure 1 The importance distribution of core features in bill credit assessment; Figure 2 A flowchart for generating a bill credit feature set for the feature extraction engine; Figure 3 A flowchart illustrating the execution of the stratified sampling algorithm; Figure 4 This is a feature adjustment graph for incremental learning. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1This invention provides a method and system for managing the risk of bill monetization based on a dynamic credit graph. The method includes integrated modules for data acquisition, feature analysis, graph construction, and dynamic adjustment to achieve accurate assessment and hedging of bill credit risk. Raw data, including transaction time, amount, and participant information, is collected from the historical transaction records and real-time bill flows of bill holders. This data is input into a feature extraction engine, which runs a preset algorithm to generate a bill credit feature set. The feature set covers multi-dimensional indicators such as bill discounting frequency, default records, and related party credit ratings, thus forming the basic data layer for risk assessment. A dynamic credit graph is constructed based on the bill credit feature set. The graph structure uses nodes to represent bill holders and related entities, and edges to represent fund transfer relationships. Each edge is bound to a credit transmission strength label to quantify risk propagation capability. The graph construction process is updated in real time to ensure consistency with market changes. Next, a stratified sampling algorithm is used to traverse the dynamic credit graph. This algorithm stratifies nodes by credit level and then randomly samples them to identify high-risk bill clusters and core transmission links. The core transmission links are formed by edges whose credit transmission strength exceeds a dynamic threshold, thereby locating high-risk areas. Based on the distribution density of high-risk bill clusters and the topology of the core transmission links, a credit hedging strategy set is generated. This set includes a dynamic allocation scheme for bill discounting quotas and a risk reserve provision ratio adjustment scheme. These schemes trigger different levels of response measures by calculating a risk diffusion index. The system receives external market fluctuation signals and corrects the edge weights of the dynamic credit graph using a credit transmission strength decay model. The model adjusts the credit transmission strength of the edges based on market risk coefficients, triggering iterative updates to the strategy set when the weight decays significantly. The updated credit hedging strategy set is executed, and the credit risk diffusion path is monitored simultaneously during bill monetization. Actual default data is fed back to the feature extraction engine for incremental learning of the bill credit feature set to optimize future assessment accuracy.

[0024] Example 1: See Figure 2The specific implementation process of the feature extraction engine generating a bill credit feature set involves the integration of a series of data processing steps. The engine obtains raw information from multiple sources through its built-in data acquisition module, including the historical transaction database of the bill holder and the real-time bill transaction log system. Historical transaction data covers all discounting operations and redemption records within the past preset period. Real-time bill information is dynamically retrieved from banks or trading platforms through a secure interface to ensure the comprehensiveness and timeliness of the data. After data input, the feature calculation module initiates the analysis process. For the generation of bill discounting frequency characteristics, the module calculates the ratio of the number of discounting applications by the holder within the past preset period to the total number of bills held. The total number of bills held is calculated cumulatively based on the face value of the bills. The ratio calculation uses a standardized algorithm to avoid scale bias, thereby deriving a numerical indicator reflecting the discounting activity. This indicator is smoothed to eliminate the impact of abnormal fluctuations. The extraction of default record characteristics is handled by the event analysis module. This module scans historical default records for events of non-payment or delayed payment. The number of events is combined with the total face value of the bills involved, and the severity of the default is calculated by a weighted average method. The weights are dynamically adjusted according to the time of the event, with more recent events given higher weights. Finally, a comprehensive default score is output to quantify the holder's reliability of payment.

[0025] The generation of related-party credit rating features begins with the activation of the external integration module. This module acts as a bridge between the system and the external credit information environment. Its workflow includes several standardized steps. First, the module generates a query list based on related-party identifiers identified in the dynamic credit graph. Query requests are sent via an encrypted channel to pre-integrated third-party credit rating interfaces. These interfaces may include databases of large credit reporting agencies or API gateways of professional financial data service providers. After the request is sent, the module enters a waiting state and initiates a timeout control mechanism. Rating results are typically returned in industry-standard letter grades, such as "AA" or "BBB" for a company. The module has a pre-stored mapping rule table that defines the conversion relationship from letter grades to numerical scores and embeds weighting coefficients for the authority of different rating agencies. For example, ratings from national-level credit reporting agencies may receive a weight of 1.0, while ratings from regional agencies may have a weight of 0.9, thus reconciling the quality differences between different information sources. The conversion process also verifies the validity period of the rating report, prioritizing data updated within the last three months to ensure that the scores reflect the latest credit status of the related parties.

[0026] After the numerical conversion is complete, the feature fusion module begins operation. This module receives three core feature vectors from different computational paths: bill discounting frequency feature, default payment record feature, and the newly generated related party credit rating feature. Preprocessing of each feature before fusion includes data normalization, scaling the original values ​​of different dimensions to a uniform numerical range, such as 0 to 100. The module then uses preset weight coefficients for linear weighted summation. These weight coefficients are not fixed but dynamically managed by a background machine learning model. This model periodically uses historical transaction and default data as a training set, evaluating the contribution of each feature to the final credit event prediction through regression analysis, thereby automatically adjusting the weight allocation. For example, if a default payment record feature is found to have a significantly enhanced correlation with actual risk in a certain period, the model may increase its weight from 0.5 to 0.6, while simultaneously fine-tuning the weights of other features accordingly. The weighted sum is processed by a standardization function to output a standardized credit feature set for the bills of exchange. This feature set exists in the form of a structured data object, containing the comprehensive credit score and individual feature values ​​for each bill holder. This feature set is then pushed to the shared data area, providing complete and quantitative input data for subsequent node attribute assignment in the dynamic credit graph and risk analysis models. The entire process, from data acquisition and transformation to fusion, forms an automated data processing chain with self-optimizing capabilities.

[0027] The construction of the dynamic credit graph follows immediately after the feature extraction stage. The graph initialization module calculates the proportion of a single transaction amount to the total historical transaction amount between the two parties based on the fund flow data between the bill holders. This proportion serves as the base value for the initial credit transmission strength. The calculation considers the transaction time factor, introducing a time decay factor; for example, recent transactions have a greater impact. The decay function uses an exponential model to smooth and adjust for the influence of historical data. The normalization module standardizes the initial credit transmission strength of all related transactions of the same holder. The method uses Min-Max scaling to map the strength values ​​to the range of 0 to 1, eliminating dimensional differences. The normalized results are bound to the corresponding edges as credit transmission strength labels. The label data is stored in the edge attributes of the graph database for easy querying and updating. When real-time bill information changes, such as a new transaction or a bill status update, the graph update module automatically triggers a dynamic adjustment of the credit transmission strength labels. The adjustment magnitude is proportional to the rate of change of the bill credit feature set, which is calculated by comparing the differences between the previous and current versions of the feature set. This ensures that the graph reflects market dynamics in real time. The adjustment process includes recalculating edge weights and updating node connections to maintain the accuracy and responsiveness of the graph.

[0028] The data flow forms a closed loop from acquisition to fusion and then to graph construction. The output of the feature extraction engine is directly fed into the initialization stage of the dynamic credit graph. The graph construction not only relies on static data but also adapts to changes through a real-time monitoring mechanism. For example, when the transaction behavior of a bill holder is abnormal, the feature set update is immediately propagated to the graph edge weights. This dynamism ensures the real-time nature of credit assessment. The graph's storage structure adopts a distributed graph database, supporting high-concurrency access and complex queries. Node attributes include holder identifiers and credit feature scores, while edge attributes include transaction timestamps and credit propagation strength. This design facilitates subsequent traversal and risk identification operations of the hierarchical sampling algorithm. The entire system achieves efficient collaboration through modular design, ensuring the consistency and reliability of the credit assessment process.

[0029] Example 2: See Figure 3 The stratified sampling algorithm, as a core component of dynamic credit graph analysis, aims to efficiently identify high-risk bill clusters and core transmission links. Its implementation is based on systematic stratification and sampling analysis of graph nodes. In the algorithm initialization phase, all nodes in the dynamic credit graph are first stratified according to the scores of the bill credit feature set. The scores are derived from standardized scores generated by the feature extraction engine. The stratification rules use a predefined threshold range; for example, high-scoring nodes are assigned to high-level strata, medium-scoring nodes to medium-level strata, and low-scoring nodes to low-level strata. The threshold setting is determined through historical data distribution analysis to ensure that the stratification reflects a continuous spectrum of credit risk. After stratification, the sampling module initiates a random sampling procedure, selecting sample nodes from each stratum according to a preset proportion. This preset proportion is dynamically adjusted through configuration parameters; for example, 30% from high-level strata, 20% from medium-level strata, and 10% from low-level strata. The sampling process uses a pseudo-random number generator to ensure randomness and avoid duplicate selection.

[0030] After sampling, the path tracing module performs a deep scan of the fund flow path of each sample node within a preset backtesting period. The backtesting period length is set according to business needs, such as six months or one year. Path scanning is achieved by traversing the graph edge sequence. The algorithm records the transaction time, amount, and credit transmission strength label on each path. Path analysis focuses on detecting whether there are three consecutive edges whose credit transmission strength labels are all higher than the mean level of that layer. The layer mean is obtained by calculating the average strength of the relevant edges of all sample nodes in that layer. If the condition is met, the path is marked as a core transmission link. Link information includes node sequence and strength value, stored in a dedicated data structure for risk diffusion analysis. At the same time, the algorithm counts the frequency data of sample nodes involved in bill discounting in each layer. The frequency calculation is based on the output of the feature extraction engine, and the statistics module calculates the standard deviation of the frequency to measure the dispersion of discounting behavior among nodes. If the standard deviation of the frequency of a node cluster exceeds twice the layer mean, the cluster is determined to be a high-risk bill cluster. The determination result is further verified in combination with the graph topology to ensure the accuracy of risk identification.

[0031] The stratified sampling algorithm incorporates a comprehensive optimization mechanism to handle complex data environments. When the sample nodes obtained by the system through the initial sampling ratio show significant differences in key feature distribution compared to the overall graph nodes—for example, if the sampling ratio of high-credit-score nodes is too low, resulting in insufficient representativeness of that stratum—the system will trigger an automatic adjustment procedure. The adjustment is based on a real-time calculated coverage index, which is quantified by comparing the distribution similarity between the sample node set and the full node set in key features such as bill discounting frequency and default payment records. If the coverage rate is lower than a preset threshold, the system will gradually increase the sampling ratio or re-divide the stratification boundaries based on feature distribution until the samples can better represent the overall characteristics. The path tracing stage employs a breadth-first search algorithm to systematically explore fund flow paths. This algorithm starts from each sample node, visiting its directly connected neighbor nodes layer by layer, and recording the credit transmission strength labels of all traversed edges. This search strategy avoids the path omission problem that may be caused by depth-first search and is particularly suitable for discovering short-term, high-frequency fund flow relationships. Faced with a large dynamic credit graph containing tens of thousands of nodes, the algorithm uses a distributed computing framework to divide the graph data into multiple computing nodes to process the path search task in parallel. Each computing node is responsible for the path exploration of a subgraph, and the results are finally aggregated, which significantly shortens the time required for full graph traversal.

[0032] The identification of high-risk bill clusters is a multi-factor comprehensive decision-making process. Besides checking whether the standard deviation of node discounting frequency exceeds a threshold, the system also assesses the spatial structural attributes of the cluster. Spatial density is measured by calculating the geographical or topological clustering of nodes within the cluster in the graph, while connectivity tightness analyzes the number and weight of edges between nodes within the cluster. These auxiliary indicators are calculated using ensemble clustering algorithms, such as density peak detection methods based on node distance, to help identify node groups that may not be statistically prominent but are structurally closely related and have strong risk transmission capabilities, thereby reducing the probability of misjudgment and omission. The entire stratified sampling algorithm maintains close coordination with the dynamic credit graph update mechanism. Once the node attributes or edge weights of the graph change due to real-time transaction data or external market signals, the system immediately detects this topological or attribute-level change. This change acts as an event signal, triggering the automatic re-initialization of the sampling analysis process. The algorithm then re-stratifies, samples, and identifies risks based on the updated graph data, ensuring that the risk analysis results reflect the latest market dynamics in near real-time, maintaining the timeliness and accuracy of risk monitoring. The algorithm outputs a core transmission link list and high-risk bill cluster identifiers. This data is directly fed into the credit hedging strategy generation module. The link list is used to analyze risk propagation paths, and the cluster identifiers are used to locate high-risk areas. The algorithm also integrates logging and error handling mechanisms. The logs record sampling parameters, path scanning results, and judgment details for easy auditing and debugging. Error handling addresses anomalies such as missing data or computation timeouts by employing retry or degradation strategies to ensure system stability. Through this hierarchical sampling method, the system can accurately locate key risk points in massive graph data, providing reliable input for subsequent strategy formulation. Simultaneously, the algorithm design balances efficiency and scalability, adapting to data environments of varying scales.

[0033] Example 3: The generation logic of the credit hedging strategy set and the correction process of the credit transmission strength decay model are closely linked, forming the system's quantitative response mechanism to dynamic risks. The generation of the credit hedging strategy set begins with a deep analysis of the output results of the stratified sampling algorithm. The system receives the distribution density data of high-risk bill clusters and the identification results of core transmission links. The distribution density is obtained by calculating the ratio of the number of nodes within the cluster to the total number of nodes in the graph, while the number of core transmission links is the statistical sum of all marked paths. The risk calculation module uses these two inputs to calculate a key indicator, the risk diffusion index. The calculation logic of this index is that the ratio of cluster distribution density to the number of core transmission links is normalized, and its value range is set between 0 and 1 to facilitate subsequent threshold comparisons. The strategy generation module presets two critical thresholds, for example, the first critical value is set to 0.7 and the second critical value is set to 0.4, to classify different risk response levels.

[0034] When the calculated risk diffusion index exceeds the first threshold, it indicates a high probability of rapid risk diffusion. At this point, the module generates a dynamic allocation scheme for bill discounting quotas. The core of this scheme is to implement tiered quota management for bill discounting applications within high-risk clusters. For example, for the first discounting application from a holder within the cluster, the available basic quota will be adjusted to 70% of the original benchmark quota. If the holder does not generate any new default records in a subsequent monitoring period, the quota can gradually recover to 90% of the benchmark level; otherwise, a quota freeze process is triggered. If the risk diffusion index is between the first and second thresholds, it indicates a moderate level of risk. The system generates a risk reserve provision ratio adjustment scheme. This scheme sums the credit transmission strength labels of all edges on the core transmission link to obtain a total strength value, and linearly adjusts the reserve provision ratio based on this value. For example, for every unit increase in the total strength value, the provision ratio increases by 5%. This linear relationship makes the risk reserve provision more risk-sensitive.

[0035] The correction process of the credit transmission strength decay model is a dynamic adjustment mechanism that responds to changes in the external market. The model continuously monitors market fluctuation signals from external data sources, including key information such as the magnitude of changes in interbank market interest rates and the increase in default rates in specific industries. The signal analysis module cleans and standardizes this raw data, extracting the absolute value of interest rate changes and the year-on-year change in default rates. The model uses this purified data to generate a comprehensive market risk coefficient, the calculation of which ensures that it comprehensively reflects the overall risk volatility of the market. The core operation of the decay calculation is to multiply the credit transmission strength label value bound to each edge in the dynamic credit graph with this market risk coefficient, thereby obtaining the new edge weights after decay. This multiplication operation is expressed as: , Where: symbol This represents the new weight value of an edge *e* in the dynamic credit graph after correction using the decay model, reflecting the current strength of the credit association represented by that edge after considering market fluctuations. (Symbol) This represents the original credit transmission strength label value bound to edge e before the attenuation calculation, derived from the previous graph construction and normalization process. (Symbol) This represents the market risk coefficient calculated from external market volatility signals. Its value is greater than zero and is used to quantify the general attenuation or amplification effect of the current market environment on the strength of credit transmission. The weight checking submodule then systematically scans the attenuated edge weights. When it finds that the new weight of an edge has dropped below a preset percentage (e.g., 30%) of its original weight value, the system will perform an edge severing operation, temporarily removing the edge's connection in the graph. Significant attenuation of edge weights and potential severing operations will change the graph's topology. Therefore, the system will immediately trigger a recalculation of the distribution density of high-risk bill clusters, ensuring the timeliness and accuracy of the risk map and providing the latest data for the iterative updates of the credit hedging strategy set.

[0036] Example 4: The system utilizes the synergistic operation of iterative update triggering conditions and incremental learning mechanisms for credit hedging strategy sets. Through multi-path monitoring and feedback loops, the system achieves adaptive optimization of the evaluation model. In a specific business scenario, assuming the system is monitoring a large-scale bill trading network with thousands of nodes, the edge weights of the dynamic credit graph continuously change due to market fluctuations. The update triggering module continuously scans the edge weight modification status in the graph. When, within a certain monitoring period (e.g., 6 consecutive hours), it detects that more than a preset number (e.g., 5% of the total) of edge weights have decreased by more than a certain margin due to decay model correction, the module immediately generates a strategy update signal. For example, if the graph originally has 10,000 edges, and it detects that the weight values ​​of 600 edges have decayed by more than 20% of the historical average, the system determines that network connectivity has changed significantly. At this point, the triggering condition is met, and the update signal is sent to the strategy generation engine.

[0037] Referring to Table 1, the risk cluster monitoring module works in parallel. This module periodically recalculates the distribution density of high-risk bill clusters in the graph, based on a node spatial clustering algorithm. Assume that in the most recent calculation, a new high-risk bill cluster was identified, consisting of nodes A, B, C, and D. Its node density is calculated as the ratio of the number of nodes in the cluster to the total number of nodes in the graph, with a value of 0.15 (i.e., 225 nodes out of 1500 nodes). The system compares this density value with historical peak records, which are stored in a dedicated log. For example, the highest density recorded in the past 30 days is 0.12. If the current density of 0.15 exceeds the historical peak, the module immediately generates a high-risk alert. This alert triggers a comprehensive iterative update of the credit hedging strategy set as another path. The system provides a manual intervention interface. When a risk administrator observes extreme market events not included in the model through the graphical interface, they can manually click the forced refresh button. After authentication, the command directly interrupts the automatic logic and immediately initiates the strategy recalculation process.

[0038] Table 1: Credit Assessment Deviation Rate and Feature Weight Adjustment Records

[0039] The incremental learning process is implemented based on continuous data comparison and parameter optimization as shown in the table above. At the end of each monitoring period (e.g., weekly), the data comparison module counts the actual number of bill default events that occurred during that period and compares it with the number of risk paths predicted by the system through the core transmission link during the same period. The credit assessment deviation rate is calculated using a specific algorithm, and its value reflects the accuracy of the prediction. As shown in the table, from period T1 to T4, the number of actual default events gradually increases, while the number of predicted risk paths remains relatively stable, causing the deviation rate to decrease significantly from 0.45 to 0.12. The weight adjustment module dynamically fine-tunes the weight coefficients of two key features in the feature extraction engine based on the calculated deviation rate. The direction and magnitude of the adjustment depend on the sign and magnitude of the deviation rate. For example, when the deviation rate continues to decrease, the system may determine that the current model is too sensitive to certain risk factors, thereby correspondingly reducing the weight of historical punitive features such as "payment default record" and appropriately increasing the weight of behavioral features such as "bill discounting frequency" to optimize the model's judgment of future risks.

[0040] The system sets a tolerance threshold for deviation rate (e.g., 10%). When the deviation rate remains below this threshold for several consecutive periods (e.g., T3 and T4), it indicates that the model's predictions are quite accurate, requiring only minor adjustments to the weights. However, if the deviation rate continuously exceeds the tolerance threshold, the incremental learning mechanism will initiate a more thorough model reconstruction process. For example, assuming the deviation rate remains above 15% in subsequent periods T5 and T6, the system determines that the current dynamic credit graph's topology can no longer effectively map the true credit risk relationships. At this point, the system triggers topology reconstruction, a process that includes, but is not limited to: redefining node stratification thresholds, recalculating the credit transmission strength of all edges based on the latest transaction data, and even reassessing the relationship definitions between related entities. The reconstructed graph will serve as the basis for a new round of stratified sampling and risk identification, enabling the entire credit assessment system to learn and evolve from prediction deviations, rather than remaining stuck with the initial model. This design, which deeply couples strategy iteration updates with incremental model learning, ensures the long-term effectiveness and robustness of the method in the face of complex and volatile market environments.

[0041] See Figure 4This diagram visually illustrates the dynamic optimization process of core feature weights within the "incremental learning" mechanism of the bill credit assessment system. The upper sub-chart shows the weight adjustment trend of the "bill discounting frequency" feature: as the learning steps progress from 1 to 6, the weight adjustment of this feature continuously increases. This is because, during iteration, the system compares "actual default events" with "predicted risk paths" and discovers that the contribution of bill discounting frequency to the prediction of real-time risk transmission gradually increases. Therefore, its weight is dynamically increased, allowing the model to more sensitively capture the potential risks behind trading activity. The lower sub-chart shows the weight adjustment trend of the "payment default record" feature: the weight of this feature continuously decreases because the system finds that over-reliance on historical default records leads to a lag in response to dynamic market risks. Therefore, its weight is gradually reduced, making the model more focused on characterizing the risks of real-time trading behavior. This dynamic adjustment of feature weights is the core manifestation of the feedback loop "from strategy execution and risk monitoring to incremental learning," ensuring that the credit assessment model can continuously evolve with market changes and improve the accuracy and timeliness of identifying credit risks in bill monetization.

[0042] Example 5: Specific application and execution details of the dynamic allocation scheme for bill discounting quotas in a real business scenario. As a key component of the credit hedging strategy set, this scheme directly affects the transaction entities within the high-risk bill cluster. Consider an example: A large trading company, "Company A," experienced a payment default by a supplier with whom it had close financial ties. This resulted in Company A's node in the dynamic credit graph being assigned to a newly identified high-risk bill cluster. When Company A submitted a bank acceptance bill discounting application for 5 million yuan through the electronic bill platform, the dynamic quota allocation scheme was immediately activated. The first step of the scheme execution is initialization. The system retrieves Company A's benchmark discounting quota of 3 million yuan, calculated based on its average transaction volume and good credit record over the past year. Since Company A is currently in a high-risk cluster, the system deducts the quota for this first discounting application according to preset rules, retaining only 70% of the benchmark quota, i.e., 2.1 million yuan, available for use. After the application is submitted, the status is marked as "Pending Review - Quota Restricted," triggering a 30-day monitoring period. During the monitoring period, the system's default detection module becomes highly active. This module not only connects to internal transaction records but also queries the central bank's credit reporting system and related commercial credit reporting platforms in real time, aiming to comprehensively track changes in the credit status of Company A and its key related parties (such as the suppliers mentioned above). The module automatically runs a scanning program daily to check if any new negative events such as non-payment, delayed payment, or judicial freezes have been recorded. The scan results are summarized to generate a daily credit status report. Assuming that no new default records are generated by Company A and its related parties during the 30-day monitoring period, the system automatically triggers a credit limit adjustment process at the end of the period. Company A's available discount limit is restored from 2.1 million yuan to 90% of the benchmark limit, i.e., 2.7 million yuan. This status change is notified to Company A's finance personnel through the platform message center and recorded in the audit log.

[0043] If the default detection module detects a new risk event during the monitoring period, such as on day 15, when another partner of Company A, "Company B," which has frequent transactions with Company A, is disclosed to have defaulted on a commercial draft, and this transaction relationship is a strong connection edge in the graph, the system will immediately assess the risk transmission effect of this event. Upon confirmation of the assessment, the implementation process shifts to stringent control measures, immediately terminating Company A's credit limit recovery process for the current monitoring period and triggering a credit limit freeze order. Under this order, all of Company A's unprocessed discount applications are suspended, and its available credit limit is reduced to zero. Simultaneously, the notification module is activated. This module generates risk warning information according to a preset template, including Company A's basic information and a summary of the risk event that triggered the freeze. This information is then sent via a secure application programming interface to the third-party credit data service provider upon which the related party's credit rating characteristics rely. This action aims to alert external institutions to the potential systemic risks of this cluster and potentially affect the external rating model's score for the relevant companies.

[0044] The credit limit management mechanism incorporates a dynamically adjusted retrospective process, which provides enterprises with a remedial channel after automatically freezing credit limits, mitigating risk. When an enterprise's credit limit is frozen due to a related-party risk event, the system generates a detailed explanation of the freezing reasons and a list of unfreezing conditions, pushing this information to the enterprise's registered account via the enterprise's interface. If the enterprise intends to restore its credit limit, it needs to prepare and upload a series of supporting documents within a specified subsequent period (e.g., six months). These documents aim to prove that it has effectively isolated the source of risk. Examples include legal documents formally terminating business cooperation with the defaulting related party, relevant judgments issued by courts or arbitration institutions, or a special audit report issued by a qualified third-party auditing institution. The report must detail the specific content and effectiveness of the risk isolation measures. After the enterprise submits a review application and relevant supporting documents online, the system automatically creates a pending work order and assigns it to the risk administrator's to-do list. After receiving the task in the console interface, the risk administrator will first review all of the company's behavioral records during the freeze period and throughout its history through the system-integrated data panel. This includes, but is not limited to, all subsequent bill transaction records, changes in relationships with other nodes in the data graph, and any new positive or negative credit information. The administrator needs to cross-verify the written evidence submitted by the company with the objective behavioral data recorded by the system. For example, even if the company provides an agreement to terminate cooperation, the system still needs to verify whether there have indeed been no new financial transactions with the defaulting party after the freeze.

[0045] After successful verification, the risk administrator begins a comprehensive assessment on a dedicated interface provided by the system. This interface centrally displays all relevant data about the company. The core of the assessment is not simply confirming the technical facts of risk isolation, but rather quantifying the potential impact of remaining risks. The administrator needs to analyze whether the company's connections with other economic entities have truly weakened in the data graph. Even with legal isolation, transactional inertia or implicit guarantees may still lead to risk transmission. Simultaneously, the administrator must examine the company's overall credit status trend since the freeze, including analyzing changes in its bill discounting frequency, whether new financing channels have emerged, and whether the stability of its upstream and downstream partners has improved. The system interface visualizes these key indicators in trend charts to assist the administrator in making judgments.

[0046] Administrator accounts are pre-set with tiered handling permissions to accommodate different risk assessment conclusions. If the assessment deems the evidence provided by the company sufficient and strong, and system monitoring data indicates that its operations are independent of the original risk source and are in a healthy state, the administrator can choose the "Complete Unfreeze" option. This operation will completely remove the company's credit limit freeze, allowing it to re-enter the normal credit limit calculation and approval process, restoring the full discount limit commensurate with its credit rating. If the assessment believes that the risk has been controlled, but some uncertainties still require time to observe, such as the creditworthiness of a new partner still being established, the administrator can choose "Set Trial Period Credit Limit." In this mode, the system allows the administrator to input a specific credit limit value and observation period length (e.g., three months). This limit is usually set between 50% and 80% of the normal benchmark limit, higher than the zero limit in a frozen state but lower than the completely normal level. Correspondingly, the system will automatically activate an enhanced monitoring mode for the company, increasing the frequency of its transaction data checks during the trial period.

[0047] Regardless of the administrator's final choice of action, the system interface will automatically pop up a box prompting the administrator to provide a detailed explanation of the rationale behind their decision. The entire set of electronic materials for the review application, the administrator's operation logs for reviewing various data views during the review process, and the final approval result and rationale will be automatically packaged into a complete digital file. This file is archived using encryption and timestamp technology, forming an unalterable audit trail that facilitates future regulatory oversight or internal case review.

[0048] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for managing the liquidity risk of bills based on dynamic credit graphs, characterized in that, Includes the following steps: Historical transaction data and real-time bill information of bill holders are collected, and a bill credit feature set is generated through a feature extraction engine. The bill credit feature set includes bill discounting frequency, payment default records and related party credit rating. A dynamic credit graph is constructed based on the credit feature set of bills. The nodes of the dynamic credit graph represent bill holders and related entities, and the edges represent fund transfer relationships and are bound with credit transmission strength labels. A hierarchical sampling algorithm is used to traverse the dynamic credit graph to identify high-risk bill clusters and core transmission links. The core transmission links are composed of edges whose credit transmission strength exceeds a dynamic threshold. Based on the distribution density of high-risk bill clusters and the topology of the core transmission link, a set of credit hedging strategies is generated. The set of credit hedging strategies includes a dynamic allocation scheme for bill discounting quotas and a scheme for adjusting the risk reserve provision ratio. Receive external market fluctuation signals, correct the edge weights of the dynamic credit graph through the credit transmission strength decay model, and trigger iterative updates of the credit hedging strategy set; The updated credit hedging strategy set is executed, and the credit risk diffusion path in the process of bill realization is monitored in parallel. The feedback is then fed into the feature extraction engine for incremental learning of the bill credit feature set.

2. The method for managing the liquidity risk of bills based on dynamic credit graphs according to claim 1, characterized in that, The specific implementation of the feature extraction engine in generating the bill credit feature set includes: The ratio of the number of discount applications by bill holders to the total number of bills held in the past preset period is used to generate bill discounting frequency characteristics. Extract the number of non-payment or delayed payment events from the historical payment records of bills, and combine them with the total face value of the bills involved in the events to generate payment default record characteristics; Call a third-party credit rating interface to obtain the latest rating results of related parties, convert the rating results into numerical credit scores, and use them as credit rating features of related parties; The characteristics of bill discounting frequency, default record, and related party credit rating are integrated according to preset weight coefficients to output a bill credit feature set.

3. The method for managing the liquidity risk of bills based on dynamic credit graphs according to claim 2, characterized in that, The construction process of the dynamic credit graph also includes: Based on the flow data of funds between the bill holders, the proportion of the amount of a single transaction to the total historical transaction amount of both parties is calculated, and the initial credit transmission strength is generated by combining the transaction occurrence time decay factor. The initial credit transmission strength of all related transactions of the same note holder is normalized, and the normalization result is bound to the corresponding edge as a credit transmission strength label. When real-time bill information is updated, the credit transmission strength label is dynamically adjusted, and the adjustment range is proportional to the rate of change of the bill credit feature set.

4. The method for managing the liquidity risk of bills based on dynamic credit graphs according to claim 3, characterized in that, The execution process of the stratified sampling algorithm includes: The nodes of the dynamic credit graph are divided into three levels: high, medium, and low, based on the scores of the bill credit feature set. A predetermined proportion of nodes are randomly selected from each level as sample nodes, and the flow of funds of the sample nodes is tracked within a predetermined backtracking period. If the credit transmission strength labels of three consecutive edges in the capital flow path of a sample node are all higher than the hierarchical average, then the path is marked as a core transmission link. The standard deviation of the frequency of bill discounting among sample nodes in each level is statistically analyzed, and node clusters that exceed twice the standard deviation are identified as high-risk bill clusters.

5. The method for managing the liquidity risk of bills based on dynamic credit graphs according to claim 4, characterized in that, The logic for generating the credit hedging strategy set includes: The risk diffusion index is calculated based on the ratio of the node density of high-risk bill clusters to the number of core transmission links. If the risk diffusion index exceeds the first critical value, a dynamic allocation scheme for bill discounting quotas will be generated. This scheme requires a tiered reduction in quotas for new discounting applications within high-risk bill clusters. If the risk diffusion index is lower than the first threshold but higher than the second threshold, a risk reserve provision ratio adjustment plan is generated. This plan is based on the sum of credit transmission strength labels of the core transmission link, and linearly increases the reserve provision ratio.

6. The method for managing the liquidity risk of bills based on dynamic credit graphs according to claim 5, characterized in that, The correction process for the credit transmission strength attenuation model includes: Analyze the magnitude of interest rate changes and the increase in industry default rates in external market volatility signals to generate a market risk coefficient; The weakened edge weight is obtained by multiplying the credit transmission strength label of each edge in the dynamic credit graph by the market risk coefficient. When the weight of an edge after decay is lower than a preset percentage of its original value, the edge is cut off and the distribution density of the high-risk bill cluster is recalculated.

7. The method for managing the liquidity risk of bills based on dynamic credit graphs according to claim 6, characterized in that, The iterative update triggering conditions for the credit hedging strategy set include: The system detected that more than a preset number of edge weights in the dynamic credit graph were attenuated. Or it may identify that the node density of newly added high-risk bill clusters has reached a historical peak; Or it may receive a manual intervention instruction requiring a forced policy refresh.

8. The method for managing the liquidity risk of bills based on dynamic credit graphs according to claim 7, characterized in that, The incremental learning is implemented as follows: The actual default events that occur during the bill realization process are compared with the predicted risk path to calculate the credit assessment deviation rate; The weighting coefficients of the bill discounting frequency feature and the payment default record feature in the feature extraction engine are adjusted according to the credit assessment deviation rate. When the deviation rate continues to exceed the tolerance threshold, the topology reconstruction of the dynamic credit graph is triggered.

9. The method for managing the liquidity risk of bills based on dynamic credit graphs according to claim 8, characterized in that, The specific implementation of the dynamic allocation scheme for bill discounting quotas includes: Seventy percent of the benchmark quota will be reserved for the first discount application within a high-risk bill cluster; If the holder does not have any new default records in the subsequent preset period, then 90% of the benchmark amount will be restored; Otherwise, the credit limit will be frozen and the relevant third-party interface for the credit rating feature will be notified.

10. A bill realization risk management system based on dynamic credit graph, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the bill realization risk management method based on dynamic credit graph as described in any one of claims 1 to 9.

Images