A financial processing method based on multi-level debt circulation

Abstract

The application discloses a kind of financial processing methods based on multi-level creditor's rights circulation, including S1, the behavior time series data of the associated entity of target creditor's rights is collected;S2, the behavior abnormality score is calculated;S3, the behavior abnormality score is as initial risk distribution, input PageRank algorithm is iteratively calculated, and network risk conduction stress value is obtained;S4, preset dynamic threshold is judged to network risk conduction stress value, and combined with whether risk source is core debtor, establish early warning of grading;S5, according to early warning level matching and executing corresponding grading intervention operation process.The application is coupled by behavior abnormality score and network risk conduction stress value Calculation, when risk has not completely broken out, signal is captured, then through individualized PageRank algorithm simulates the conduction and convergence process of risk along network edge, finally calculates the network risk conduction stress value borne by target creditor's rights, realizes risk forward-looking perception and early warning.

Description

Technical Field

[0001] This invention relates to the field of debt transfer processing, and in particular to a financial processing method based on multi-level debt transfer. Background Technology

[0002] In traditional debt risk management practices, financial institutions and asset holders primarily rely on static assessments and ex-post monitoring of the debtor's creditworthiness. Conventional methods typically focus on the direct debtor's financial indicators, repayment history, and the value of collateral, triggering risk alerts by setting single, rule-based thresholds. While this point-to-point management model may be effective in simple debt relationships, its inherent limitations are amplified dramatically in multi-level, complex debt transfer scenarios.

[0003] Traditional methods treat debt risk in isolation as an individual credit issue of the debtor, neglecting the complex guarantee networks, equity linkages, and supply chain dependencies among modern enterprises. When risk propagates between related entities, a break in one link of the guarantee chain allows the risk to spread rapidly along the network. Traditional risk control methods only detect the risk passively when it has already reached the direct debtor, missing the opportunity for proactive intervention. Furthermore, existing early warning systems are mostly based on past negative events or severely deteriorated financial data, making it difficult to identify potential risks from early abnormal behavior patterns. Therefore, this proposal suggests a financial processing method based on multi-level debt transfer to address these issues. Summary of the Invention

[0004] The purpose of this invention is to provide a financial processing method based on multi-level debt transfer to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a financial processing method based on multi-level debt transfer, comprising the following processing steps: S1, collect time-series data on the behavior of related entities of the target claim; S2, extract the behavioral time series data of each associated entity, aggregate them in a fixed time window to generate feature vectors, then input the feature vectors into the trained isolated forest model to output point anomaly scores, input the data of the same time window into the trained LSTM autoencoder, calculate the reconstruction error to obtain the sequence anomaly score, and use weighted geometric mean to fuse the point anomaly score and the sequence anomaly score to obtain the behavioral anomaly score. S3. Based on the relationships and strengths of each entity, a directed edge and transition probability are constructed. Then, the abnormality score of each node's behavior is used as the initial risk distribution. The PageRank algorithm with the target creditor node is input for iterative calculation to finally obtain the steady-state probability distribution. The probability values ​​of all nodes except the target node are weighted and averaged to obtain the network risk transmission pressure value. S4, preset dynamic thresholds to judge the pressure value of network risk transmission, and establish a graded early warning system based on whether the source of risk is a core debtor; S5 matches and executes the corresponding graded intervention procedures based on the warning level.

[0006] Preferably, the calculation and processing of the behavioral anomaly score includes the following steps: At least extract data on judicial litigation, equity pledge, personnel changes, administrative penalties, and public opinion sentiment for each related entity, and aggregate the data in a rolling manner according to a fixed time window to generate a feature vector for each entity in each time window; First, train an isolated forest model using historical normal period behavior data. Then, input the feature vector into the trained isolated forest model to calculate the anomaly score of the point where the current behavior deviates from the normal group distribution. The LSTM autoencoder is trained using historical normal behavior sequences. Then, the behavioral time series data within the same time window is input into the trained LSTM autoencoder. After calculating the reconstruction error, the sequence abnormality score, which represents the abnormality of the behavioral sequence pattern, is obtained. The point anomaly score and the sequence anomaly score are fused using a weighted geometric average to obtain the final behavioral anomaly score of each associated entity in the corresponding time window.

[0007] Preferably, the calculation and processing of the network risk transmission pressure value includes the following steps; Based on the dynamic creditor relationship graph, each related entity and the target creditor is used as a node. A directed edge set is constructed according to the equity control, guarantee chain, transaction relationship and association strength between entities. The basic transition probability between nodes is calculated based on the association strength to form an initial transition probability matrix. The abnormality scores of each node are transformed by the Sigmoid function and mapped to the initial risk value of network propagation. The node corresponding to the target claim is used as the convergence target of the personalized PageRank algorithm. The mapped initial risk value is used as input, a damping factor is set, and the transition relationship is defined by the transition probability matrix. Iterative calculation is performed to simulate the process of risk propagation and diffusion along the directed edge and finally converge to the target node until the probability distribution converges and the steady-state probability distribution is obtained. The weighted summation of the probability values ​​of all nodes except the target creditor node in the steady-state probability distribution is the network risk transmission pressure value.

[0008] Preferably, in the process of generating feature vectors for each time window in S2, data quantization is performed first, including: Extract the number of cases, the total amount involved, the proportion of cases in which the defendant / judgment debtor is located, and the distribution of case types from judicial litigation data; Extract the pledge ratio, number of pledges, volatility of pledged market value, and number of warnings triggered by the liquidation line from the equity pledge data; Extract the number of senior executives leaving, the turnover rate of key technical personnel, and the frequency of business registration changes from personnel change data; Extract a weighted sum of the number of penalties, the amount of penalties, and the severity level of penalties from the administrative penalty data; In the public opinion sentiment data, NLP sentiment analysis models are used to process news, reports, and social media information to generate negative sentiment index, public opinion heat and frequency of occurrence of default / litigation keywords; Then, a fixed time window is set and rolled over in days or weeks. The original time data in each window is aggregated according to entities to generate multidimensional time series panel data. For each entity in each time window, not only the current value of the quantified data is calculated, but also the rate of change of the quantified data relative to the previous window, the rolling statistics, and the percentile of the same industry are calculated. Finally, each entity is represented as a feature vector in each time window.

[0009] Preferably, the establishment of the tiered early warning system includes the following steps: Obtain the calculated network risk transmission pressure value; Based at least on preset dynamic thresholds A, B, and C, where A < B < C, the network risk transmission pressure value is judged, and it is identified whether the main source of risk is the core debtor. When the pressure value of network risk transmission is greater than C, or when a judicial freeze or loss of contact event is detected in a core debtor, the highest level three early warning is generated; When the network risk transmission pressure value is greater than B and not greater than C, or when the network risk transmission pressure value is greater than A and not greater than B, but the main source of risk is the core debtor, a secondary warning of intermediate level is generated; When the network risk transmission pressure value is greater than A but not greater than B, and the risk mainly originates from non-core debtors, the lowest level, Level 1 warning, is generated.

[0010] Preferably, after the tiered early warning is generated, the following intervention operations are performed: When a Level 1 warning is triggered, an account confirmation letter is automatically sent to the debtor, and a diagnostic work order is generated to assign an account manager to conduct informal communication and verification. The risk status is marked based on the verification results. When a Level 2 warning is triggered, a risk concern letter is sent to the debtor and guarantor, a stress test is conducted on the target debt, and negotiations are initiated with the debtor and guarantor based on the test results to reach an agreement on additional guarantees or to sign a repayment commitment. When a Level 3 warning is triggered, a crisis management team will be established to initiate an emergency investigation of the debtor's and guarantor's assets, and simultaneously review the completeness of the legal documents related to the transfer of rights of the target claim, and prepare for legal action.

[0011] Preferably, the verification of the integrity of legal documents in the rights transfer chain includes the following steps: When responding to a Level 3 warning, the original debt contract, all previous debt transfer agreements, and the debtor's notification documents corresponding to each transfer are retrieved from the collected data. Verify the connection between the transferor and the transferee in each transfer agreement, and verify the evidence that the debtor has been effectively notified of each debt transfer, to form a complete chain of evidence.

[0012] Preferably, after implementing the tiered intervention procedure, an observation period is preset, and during the observation period, the behavioral anomaly score of the associated entities, the network risk transmission pressure value, and new risk events are continuously tracked. At the end of the observation period, based on the final state of the claims, a structured case data package is generated, which includes initial characteristics, early warning actions, disposal strategies, stress reduction rate, and outcome labels. Using the structured case data package, the model parameters for calculating the behavioral anomaly score and the network risk transmission pressure value are iteratively optimized.

[0013] Preferably, the structured case data package includes at least the following information: The initial network risk transmission pressure value, the core debtor's abnormal behavior score, and the characteristics of related entities when the early warning is triggered; The graded early warning levels implemented and the corresponding complete response strategies and actions; The network risk transmission pressure value at the end of the observation period, and the pressure reduction rate calculated based on the initial value and the end-of-period value; The final status label of the target claim at the end of the observation period.

[0014] Preferably, the model parameters are iteratively optimized, specifically including: Using completed case data packages, the edge weight matrix representing the relationships between entities in the PageRank algorithm model is adjusted through reinforcement learning algorithms; By using the newly added normal and abnormal behavior samples in the case data package, the isolated forest model and LSTM autoencoder are incrementally learned to update the model parameters to adapt to the new behavior patterns.

[0015] The technical effects and advantages of this invention are as follows: (1) This invention first collects data on the rights transfer chain, the timing of entity behavior and multi-level related networks, and constructs a unified dynamic spatiotemporal map. By identifying and quantifying the potential paths and intensity of risk transmission through the complex connections of guarantee chains, equity control and supply chains, the timeliness and foresight of risk warning are improved. (2) This invention captures signals before risks fully erupt by coupling the abnormality score of behavior with the network risk transmission pressure value. Then, it simulates the transmission and convergence process of risk along the network edge through a personalized PageRank algorithm, and finally calculates the network risk transmission pressure value borne by the target claim. Thus, it issues an early warning before the debtor's own financial statements deteriorate, realizing the forward-looking perception and early warning of risks, and carrying out forward-looking risk control and handling. Attached Figure Description

[0016] Figure 1 This is the overall operation flowchart for this solution.

[0017] Figure 2 This is a flowchart illustrating the tiered early warning system of the present invention. Detailed Implementation

[0018] 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.

[0019] This invention provides, for example Figure 1 The financial processing method based on multi-level debt transfer shown includes the following processing steps: S1, collect time-series data on the behavior of related entities of the target claim; Specifically, the system collects time-series data on the behavior of entities associated with the target debt, dynamically and continuously capturing negative behavioral events of all parties related to the debt, especially debtors and key guarantors; this includes at least judicial litigation, equity pledges, personnel changes, administrative penalties, and public opinion sentiment indices. The data collection method begins by identifying a list of core entities to be monitored based on the rights transfer chain and guarantee relationships, including debtors, controlling shareholders, actual controllers, and major guarantors, using standardized entity names, social credit codes, or ID numbers for data integration; then, through a third-party data service provider, it periodically queries new cases where the entity is the defendant or judgment debtor, collecting information on the cause of action, amount in dispute, filing date, and latest progress, and connecting to public data interfaces from market supervision, environmental protection, and taxation departments to collect penalty decisions, reasons, amounts, and dates; subsequently, it monitors industrial and commercial equity pledge registration information, recording the pledged equity ratio, pledgee, guarantee amount, and status, and monitors changes in legal representatives, directors, supervisors, and senior management personnel; thus forming a risk signal time series. It should be noted that the data collection sources include corporate credit reports, business registration maps, supply chain data, and guarantee relationship data.

[0020] S2, extract the behavioral time series data of each associated entity, aggregate them in a fixed time window to generate feature vectors, then input the feature vectors into the trained isolated forest model to output point anomaly scores, input the data of the same time window into the trained LSTM autoencoder, calculate the reconstruction error to obtain the sequence anomaly score, and use weighted geometric mean to fuse the point anomaly score and the sequence anomaly score to obtain the behavioral anomaly score. Specifically, the calculation and processing of the behavioral abnormality score includes the following steps: At least extract data on judicial litigation, equity pledge, personnel changes, administrative penalties, and public opinion sentiment for each related entity, and aggregate the data in a rolling manner according to a fixed time window to generate a feature vector for each entity in each time window; Furthermore, in the process of generating feature vectors for each time window, data quantization is first performed, including: Extract the number of cases, the total amount involved, the proportion of cases in which the defendant / judgment debtor is located, and the distribution of case types from judicial litigation data; Extract the pledge ratio, number of pledges, volatility of pledged market value, and number of warnings triggered by the liquidation line from the equity pledge data; Extract the number of senior executives leaving, the turnover rate of key technical personnel, and the frequency of business registration changes from personnel change data; Extract a weighted sum of the number of penalties, the amount of penalties, and the severity level of penalties from the administrative penalty data; In the public opinion sentiment data, NLP sentiment analysis models are used to process news, reports, and social media information to generate negative sentiment index, public opinion heat and frequency of occurrence of default / litigation keywords; Then, a fixed time window is set and rolled over in days or weeks. The original time data in each window is aggregated according to entities to generate multidimensional time series panel data. For each entity in each time window, not only the current value of the quantified data is calculated, but also the rate of change of the quantified data relative to the previous window, the rolling statistics, and the percentile of the same industry are calculated. Finally, each entity is represented as a feature vector in each time window.

[0021] The Isolation Forest model is first trained using historical data from normal periods. Data from historically recognized, risk-free periods is selected, and feature vectors from all entities are extracted as training samples. This sample set is then used to train the Isolation Forest model. The model constructs isolation trees by randomly partitioning the feature space. Normal points, due to their dense distribution, require more partitions to be isolated, while anomalous points require less. Based on the training results, a basic anomalous score threshold is determined, such as the 95th percentile of the sample scores. The feature vectors are then input into the trained Isolation Forest model for calculation. The model calculates the anomalous score of points whose current behavior deviates from the normal group distribution based on the average path length of the feature vectors across all isolation trees. The closer the anomalous score is to 1, the more anomalous the feature combination of the entity is in the current window, and the easier it is to isolate. For each entity, the feature vectors of multiple consecutive time windows are arranged in chronological order to form a behavior sequence. The LSTM autoencoder is trained using the historical normal behavior sequence. An encoder-decoder structure is used. The encoder compresses the input sequence into a low-dimensional context vector, which represents the normal pattern of the sequence. The decoder attempts to reconstruct the original input sequence from this vector. The training objective is to minimize the reconstruction error, forcing the model to learn the ability to compress and reconstruct normal behavior sequences. Then, the behavioral time series data within the same time window are input into the trained LSTM autoencoder. After calculating the reconstruction error, the sequence anomaly score, which represents the abnormality of the behavioral sequence pattern, is obtained. The sequence anomaly score is the mean square error or mean absolute error between the original sequence and the reconstructed sequence. The larger the error, the greater the difference between the current behavioral sequence pattern and the historical normal pattern, and the higher the degree of sequence anomaly. The point anomaly score and the sequence anomaly score are fused using a weighted geometric mean to obtain the final behavioral anomaly score of each associated entity within the corresponding time window. Specifically, the point anomaly scores are normalized to the [0,1] interval to eliminate the influence of dimensions, and then calculated using a weighted geometric mean. The calculation is expressed as follows: (1); Where M represents the weighted average sum, U represents the point outlier score, and V represents the sequence outlier score; and These are the weights corresponding to the point anomaly score and the sequence anomaly score, respectively. , and The value of β can be optimized through grid search or based on the performance of the validation set, i.e., the historical risk event samples. If historical validation finds that sequence anomalies are more predictive of medium- to long-term risks, then the weight of β should be increased.

[0022] S3. Based on the relationships and strengths of each entity, a directed edge and transition probability are constructed. Then, the abnormality score of each node's behavior is used as the initial risk distribution. The PageRank algorithm with the target creditor node is input for iterative calculation to finally obtain the steady-state probability distribution. The probability values ​​of all nodes except the target node are weighted and averaged to obtain the network risk transmission pressure value. Specifically, the calculation and processing of network risk transmission pressure values ​​includes the following steps; Based on the dynamic creditor relationship graph, each related entity and the target creditor is used as a node. A directed edge set is constructed based on the equity control, guarantee chain, transaction relationship and relationship strength between entities. The basic transition probability between nodes is calculated based on the relationship strength to form an initial transition probability matrix. The relationship strength includes the shareholding ratio, guarantee amount and transaction amount ratio. The abnormality scores of each node are transformed using the Sigmoid function and mapped to an initial risk value for network propagation. The node corresponding to the target claim is used as the convergence target of the personalized PageRank algorithm, and the mapped initial risk value is taken as input. The transformation using the Sigmoid function is expressed as follows: (2); Where s represents the original behavioral anomaly score, s0 represents the risk initiation threshold, and k represents the curve steepness factor. The score of each node is converted into an initial risk value. The parameters s0 and k need to be determined by statistical calibration based on the anomaly distribution of nodes before the occurrence of historical risk events. Set a damping factor and define the transition relationship with the transition probability matrix. Perform iterative calculations to simulate the process of risk propagation and diffusion along the directed edge and finally converge to the target node until the probability distribution converges and the steady-state probability distribution is obtained. The weighted summation of the probability values ​​of all nodes except the target creditor node in the steady-state probability distribution yields the network risk transmission pressure value.

[0023] Ultimately, the abnormal behavior score of the core debtor is combined with the stress value, and a preset dynamic threshold is used for judgment to output a graded early warning. refer to Figure 2 As shown in Figure S4, a preset dynamic threshold is used to determine the network risk transmission pressure value, and a tiered early warning system is established based on whether the source of the risk is a core debtor; specifically, the steps for generating the tiered early warning output include: Obtain the behavioral abnormality score of the core debtor and calculate the network risk transmission pressure value; Based at least on preset dynamic thresholds A, B, and C, where the values ​​of A, B, and C are all in the range of 0 to 1, and A < B < C, the network risk transmission pressure value is judged, and it is identified whether the main source of risk is the core debtor. When the pressure value of network risk transmission is greater than C, or when a judicial freeze or loss of contact event is detected in a core debtor, the highest level three early warning is generated; When the network risk transmission pressure value is greater than B and not greater than C, or when the network risk transmission pressure value is greater than A and not greater than B, but the main source of risk is the core debtor, a secondary warning of intermediate level is generated; When the network risk transmission pressure value is greater than A but not greater than B, and the risk mainly originates from non-core debtors, the lowest level, Level 1 warning, is generated.

[0024] S5, match and execute the corresponding tiered intervention procedure according to the warning level; specifically, after the tiered warning is generated, the following intervention operations are executed: When a Level 1 warning is triggered, an account confirmation letter is automatically sent to the debtor, and a diagnostic work order is generated to assign an account manager to conduct informal communication verification. The risk status is marked based on the verification results. Specifically, an automated process sends a confirmation letter verifying relevant invoice information and generates a diagnostic work order requiring the account manager to complete the informal communication verification within 2-3 business days. This is used for early verification, confirming the authenticity of risk signals, and conducting informal communication. After a Level 1 warning is triggered, an email or encrypted message is automatically sent to the debtor or their designated contact person, containing a standardized account information confirmation letter. The letter lists the basic information of the target debt, including the contract number, amount, and due date, and attaches the most recent invoice, requesting confirmation or feedback within 10-15 business days if there are any objections. The alert is simultaneously assigned to the account manager responsible for the debt to communicate with the debtor. Within 2-3 business days, the account manager is required to communicate with the debtor's finance or business manager through informal channels, including telephone or WeChat, to verify whether operations are normal, whether there are recent payment difficulties, and whether they are aware of relevant risk signals. The account manager records the communication results, including the other party's response, explanation, and expected payment arrangements. Based on the communication results, the account manager marks the corresponding debtor or designated contact person as being in a normal situation (false alarm), confirming short-term difficulties but with a strong willingness to repay, or having questionable circumstances, and recommends escalation. When marked as being in a normal situation (false alarm), this information will be used for subsequent model threshold optimization. When marked as having questionable circumstances and recommending escalation, the process transitions to manual escalation, with human collaboration to increase the accuracy of the alert. When a Level 2 warning is triggered, a risk concern letter is sent to the debtor and guarantor, a stress test is conducted on the target debt, and negotiations with the debtor and guarantor are initiated based on the test results to reach an agreement on additional guarantees or a repayment commitment. Specifically, a concern message is sent, a stress test is conducted on the creditor, and the process of negotiating additional guarantees or signing a repayment commitment is initiated to proactively exert pressure, activate risk mitigation measures, and prepare for potential risks. After the Level 2 warning response, the asset preservation specialist drafts a risk concern letter, clearly indicating the detected risk signals, requiring the debtor to provide a written explanation within 5-7 working days, and formally copying the guarantor. Simultaneously, the risk analyst... Based on the core debtor's involvement in litigation, setting compensation for losses or enforcement, a stress test is conducted on the cash flow for recovering the target debt. Following the test, a stress test report is generated, outlining the expected recovery rate, recovery time, and loss amount under the worst-case scenario. Based on the stress test results, the account manager or asset preservation specialist immediately negotiates with the debtor and guarantor to propose risk mitigation measures, including additional guarantees or signing repayment commitments. Additional guarantees involve negotiating additional collateral, pledges, or introducing new guarantors, while signing repayment commitments requires the debtor to provide a phased, specific written repayment plan. Once an agreement is reached, relevant legal documents are immediately prepared and signed, and registration is completed. When a Level 3 warning is triggered, a crisis management team is established to initiate an emergency investigation of the debtor's and guarantor's assets, and simultaneously review the completeness of the legal documents related to the transfer of rights of the target claim, prepare legal actions for comprehensive emergency response, and protect assets to the greatest extent possible. Furthermore, the review of the integrity of legal documents in the rights transfer chain includes the following steps: When responding to a Level 3 warning, the original debt contract, all previous debt transfer agreements, and the debtor's notification documents corresponding to each transfer are retrieved from the collected data. Verify the connection between the transferor and the transferee in each transfer agreement, and verify the evidence that the debtor has been effectively notified of each debt transfer, to form a complete chain of evidence; Retrieve and verify all previous debt transfer agreements and corresponding debtor confirmation documents, and verify the continuity and validity of the evidence chain for notifying the debtor. Upon a Level 3 early warning response, the risk director immediately leads a crisis management team, including legal, asset preservation, external lawyers, and finance personnel, to hold an emergency meeting. Then, lawyers or investigation agencies are commissioned to conduct a comprehensive and urgent investigation of the debtor's and guarantor's real estate, movable property, bank accounts, equity investments, and accounts receivable, including those under the names of natural persons. Based on the investigation results, the necessity and feasibility of immediately applying for pre-litigation property preservation are assessed. Legal personnel retrieve the complete rights transfer document package of the target debt from the data collected by S1, including the original debt contract, previous debt transfer agreements, debt transfer notices corresponding to each transfer stage, and delivery vouchers, and verify them one by one to ensure that the transferor and transferee are connected in the previous and subsequent agreements without any gaps. It is also ensured that the debtor has been effectively notified of each transfer, i.e., there is a signed receipt or registration certificate. Special note: If it is agreed that no subsequent notices are required, it must be verified whether the initial notice contains a clause allowing subsequent transfers.

[0025] It should be noted that after the early warning response at each level is completed, the response actions, time taken, and results are recorded, including the amount recovered, losses, and changes in legal status. The early warning signal, transmission analysis, the entire response process, and the results are packaged into a structured case and archived in the training library of the computational model. Furthermore, the precision rate and false alarm rate of each early warning level are statistically analyzed. The precision rate is the percentage of early warning levels that are ultimately confirmed as real risks. The thresholds A, B, and C are dynamically adjusted according to the following rules: if the false alarm rate of a certain level continues to be higher than the set upper limit, the threshold corresponding to that level is appropriately increased; if a serious missed alarm occurs and a high-risk event quickly breaks out after a low-level early warning, the threshold is decreased.

[0026] Finally, data on the early warning and handling process and results are collected to generate case data packages. The case data packages are used to optimize the associated edge weights of the PageRank algorithm and optimize the parameters of the isolated forest model and LSTM autoencoder. After the hierarchical intervention operation process is executed, an observation period is preset, and during the observation period, the abnormality score of the behavior of associated entities, the network risk transmission pressure value and new risk events are continuously tracked. At the end of the observation period, based on the final state of the claims, a structured case data package is generated, which includes initial characteristics, early warning actions, disposal strategies, stress reduction rate, and outcome labels. Using structured case data packages, the model parameters for calculating behavioral anomaly scores and network risk transmission pressure values ​​are iteratively optimized; Specifically, a structured case data package should include at least the following information: The initial network risk transmission pressure value, the core debtor's abnormal behavior score, and the characteristics of related entities when the early warning is triggered; The graded early warning levels implemented and the corresponding complete response strategies and actions; The network risk transmission pressure value at the end of the observation period, and the pressure reduction rate calculated based on the initial value and the end-of-period value; The final status label of the target claim at the end of the observation period.

[0027] Specifically, the model parameters are iteratively optimized, including: Using completed case data packages, the edge weight matrix representing the relationships between entities in the PageRank algorithm model is adjusted through reinforcement learning algorithms; By using the newly added normal and abnormal behavior samples in the case data package, the isolated forest model and LSTM autoencoder are incrementally learned to update the model parameters to adapt to the new behavior patterns.

[0028] Based on the latest entity behavior data, the abnormality score of the related entities in this case and the network risk transmission pressure value are recalculated, forming two time-series curves. Any new risk events related to the debtor or guarantor are recorded, including new litigation and equity freezes, or positive events such as repayment completion and new financing. At the end of the observation period, or when the debt has been pre-settled or written off, the asset manager marks the final status of the debt in the system. Status labels include normal settlement, restructured settlement, partial recovery, total loss, and still under observation. Normal settlement indicates that the debtor has repaid the full amount on schedule; restructured settlement indicates settlement after extension or reduction; partial recovery indicates partial recovery through collection and disposal of collateral; total loss indicates write-off without recovery; still under observation indicates that the risk has not been eliminated but no loss has yet occurred at the end of the observation period. When optimizing the associated edge weights using reinforcement learning, a batch of completed case data packages is collected periodically. A policy gradient-based reinforcement learning algorithm is used to adjust the edge weight matrix in the PageRank transmission model with the goal of maximizing the cumulative reward of all cases. The anomaly score related calculation model update involves extracting new normal samples and new anomalous samples from the data packet. For LSTM autoencoders, incremental training is performed using the new normal samples to adjust model parameters, enabling the model to better reconstruct the normal patterns of the current period and thus be more sensitive to deviations. For the isolation forest model, it is periodically retrained using the full dataset combined with new samples, or an incremental update isolation forest algorithm is used to identify new anomalous patterns.

[0029] Statistical analysis was conducted on the stress reduction rate and final recovery rate of each disposal strategy under different early warning scenarios. Based on the effectiveness of the quantitative strategies, the recommended priority and weight of each strategy under different scenarios were adjusted. The stress reduction rate is the ratio of the initial stress value minus the stress value at the end of the observation period to the initial stress value. The stress reduction rate is used to measure the degree of relief of risk stress. The final recovery rate is the ratio of the total cash recovered to the total principal of the debt.

[0030] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A financial processing method based on multi-level debt transfer, characterized in that, The following processing steps are included: S1, collect time-series data on the behavior of related entities of the target claim; S2, extract the behavioral time series data of each associated entity, aggregate them in a fixed time window to generate feature vectors, then input the feature vectors into the trained isolated forest model to output point anomaly scores, input the data of the same time window into the trained LSTM autoencoder, calculate the reconstruction error to obtain the sequence anomaly score, and use weighted geometric mean to fuse the point anomaly score and the sequence anomaly score to obtain the behavioral anomaly score. S3. Based on the relationships and strengths of each entity, a directed edge and transition probability are constructed. Then, the abnormality score of each node's behavior is used as the initial risk distribution. The PageRank algorithm with the target creditor node is input for iterative calculation to finally obtain the steady-state probability distribution. The probability values ​​of all nodes except the target node are weighted and averaged to obtain the network risk transmission pressure value. S4, preset dynamic thresholds to judge the pressure value of network risk transmission, and establish a graded early warning system based on whether the source of risk is a core debtor; S5 matches and executes the corresponding graded intervention procedures based on the warning level.

2. The financial processing method based on multi-level debt transfer according to claim 1, characterized in that, The calculation and processing of the behavioral anomaly score includes the following steps: At least extract data on judicial litigation, equity pledge, personnel changes, administrative penalties, and public opinion sentiment for each related entity, and aggregate the data in a rolling manner according to a fixed time window to generate a feature vector for each entity in each time window; First, train an isolated forest model using historical normal period behavior data. Then, input the feature vector into the trained isolated forest model to calculate the anomaly score of the point where the current behavior deviates from the normal group distribution. The LSTM autoencoder is trained using historical normal behavior sequences. Then, the behavioral time series data within the same time window is input into the trained LSTM autoencoder. After calculating the reconstruction error, the sequence abnormality score, which represents the abnormality of the behavioral sequence pattern, is obtained. The point anomaly score and the sequence anomaly score are fused using a weighted geometric average to obtain the final behavioral anomaly score of each associated entity in the corresponding time window.

3. The financial processing method based on multi-level debt transfer according to claim 1, characterized in that, The calculation and processing of the network risk transmission pressure value includes the following steps; Based on the dynamic creditor relationship graph, each related entity and the target creditor is used as a node. A directed edge set is constructed according to the equity control, guarantee chain, transaction relationship and association strength between entities. The basic transition probability between nodes is calculated based on the association strength to form an initial transition probability matrix. The abnormality scores of each node are transformed by the Sigmoid function and mapped to the initial risk value of network propagation. The node corresponding to the target claim is used as the convergence target of the personalized PageRank algorithm. The mapped initial risk value is used as input, a damping factor is set, and the transition relationship is defined by the transition probability matrix. Iterative calculation is performed to simulate the process of risk propagation and diffusion along the directed edge and finally converge to the target node until the probability distribution converges and the steady-state probability distribution is obtained. The weighted summation of the probability values ​​of all nodes except the target creditor node in the steady-state probability distribution is the network risk transmission pressure value.

4. The financial processing method based on multi-level debt transfer according to claim 2, characterized in that, In the process of generating feature vectors for each time window in S2, data quantization is first performed, including: Extract the number of cases, the total amount involved, the proportion of cases in which the defendant / judgment debtor is located, and the distribution of case types from judicial litigation data; Extract the pledge ratio, number of pledges, volatility of pledged market value, and number of warnings triggered by the liquidation line from the equity pledge data; Extract the number of senior executives leaving, the turnover rate of key technical personnel, and the frequency of business registration changes from personnel change data; Extract a weighted sum of the number of penalties, the amount of penalties, and the severity level of penalties from the administrative penalty data; In the public opinion sentiment data, NLP sentiment analysis models are used to process news, reports, and social media information to generate negative sentiment index, public opinion heat and frequency of occurrence of default / litigation keywords; Then, a fixed time window is set and rolled over in days or weeks. The original time data in each window is aggregated according to entities to generate multidimensional time series panel data. For each entity in each time window, not only the current value of the quantified data is calculated, but also the rate of change of the quantified data relative to the previous window, the rolling statistics, and the percentile of the same industry are calculated. Finally, each entity is represented as a feature vector in each time window.

5. A financial processing method based on multi-level debt transfer according to claim 1, characterized in that, The establishment of the tiered early warning system includes the following steps: Obtain the calculated network risk transmission pressure value; Based at least on preset dynamic thresholds A, B, and C, where A < B < C, the network risk transmission pressure value is judged, and it is identified whether the main source of risk is the core debtor. When the pressure value of network risk transmission is greater than C, or when a judicial freeze or loss of contact event is detected in a core debtor, the highest level three early warning is generated; When the network risk transmission pressure value is greater than B and not greater than C, or when the network risk transmission pressure value is greater than A and not greater than B, but the main source of risk is the core debtor, a secondary warning of intermediate level is generated; When the network risk transmission pressure value is greater than A but not greater than B, and the risk mainly originates from non-core debtors, the lowest level, Level 1 warning, is generated.

6. A financial processing method based on multi-level debt transfer according to claim 5, characterized in that, After the tiered early warning is generated, the following intervention operations should be performed: When a Level 1 warning is triggered, an account confirmation letter is automatically sent to the debtor, and a diagnostic work order is generated to assign an account manager to conduct informal communication and verification. The risk status is marked based on the verification results. When a Level 2 warning is triggered, a risk concern letter is sent to the debtor and guarantor, a stress test is conducted on the target debt, and negotiations are initiated with the debtor and guarantor based on the test results to reach an agreement on additional guarantees or to sign a repayment commitment. When a Level 3 warning is triggered, a crisis management team will be established to initiate an emergency investigation of the debtor's and guarantor's assets, and simultaneously review the completeness of the legal documents related to the transfer of rights of the target claim, and prepare for legal action.

7. A financial processing method based on multi-level debt transfer according to claim 6, characterized in that, The review of the integrity of legal documents in the rights transfer chain includes the following steps: When responding to a Level 3 warning, the original debt contract, all previous debt transfer agreements, and the debtor's notification documents corresponding to each transfer are retrieved from the collected data. Verify the connection between the transferor and the transferee in each transfer agreement, and verify the evidence that the debtor has been effectively notified of each debt transfer, to form a complete chain of evidence.

8. A financial processing method based on multi-level debt transfer according to claim 1, characterized in that, After implementing the tiered intervention procedure, a pre-set observation period is established, during which the behavioral anomaly score of related entities, network risk transmission pressure value, and new risk events are continuously tracked. At the end of the observation period, based on the final state of the claims, a structured case data package is generated, which includes initial characteristics, early warning actions, disposal strategies, stress reduction rate, and outcome labels. Using the structured case data package, the model parameters for calculating the behavioral anomaly score and the network risk transmission pressure value are iteratively optimized.

9. A financial processing method based on multi-level debt transfer according to claim 8, characterized in that, The structured case data package includes at least the following information: The initial network risk transmission pressure value, the core debtor's abnormal behavior score, and the characteristics of related entities when the early warning is triggered; The graded early warning levels implemented and the corresponding complete response strategies and actions; The network risk transmission pressure value at the end of the observation period, and the pressure reduction rate calculated based on the initial value and the end-of-period value; The final status label of the target claim at the end of the observation period.

10. A financial processing method based on multi-level debt transfer according to claim 8, characterized in that, The model parameters are iteratively optimized, specifically including: Using completed case data packages, the edge weight matrix representing the relationships between entities in the PageRank algorithm model is adjusted through reinforcement learning algorithms; By using the newly added normal and abnormal behavior samples in the case data package, the isolated forest model and LSTM autoencoder are incrementally learned to update the model parameters to adapt to the new behavior patterns.

Images