Credit customer risk identification method and system based on multi-dimensional financial assets

Abstract

The invention relates to the technical field of financial science and technology and data processing, and discloses a credit customer risk identification method and system based on multi-dimensional financial assets, and the method comprises the following steps: firstly, aligning external unstructured government affair data with internal structured financial data through a unique identifier; carrying out discretization and normalization processing to construct a standardized multi-dimensional feature vector; then, a comprehensive risk score is calculated in parallel based on the metadata rule base and the statistical anomaly detection model; if the score exceeds a threshold value, generating a risk clue, controlling the clue to flow among states of to-be-checked, handling, archived and the like by using a finite-state machine, and executing timeout monitoring; and finally, in response to clue termination, dynamically adjusting a rule weight based on misinformation data, and encrypting and synchronizing a credit decision result to an external data source. According to the invention, through multi-source data fusion, double-track scoring and state machine closed-loop management, the accuracy of credit risk identification, the standardization of process management and control and the adaptive ability of the model are effectively improved.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and data processing technology, specifically to a method and system for identifying credit customer risk based on multidimensional financial assets. Background Technology

[0002] With the rapid development of financial technology, risk management in credit business has become a key component of the core competitiveness of commercial banks and financial institutions. In actual credit risk control scenarios, accurately identifying potential customer risks not only relies on the internal fund transaction records and account status of financial institutions, but also requires extensive integration of external social credit data such as government affairs, judicial data, and public opinion data to construct a comprehensive risk profile.

[0003] However, existing credit risk identification technologies have limitations when facing the deep integration of multi-source heterogeneous data. On the one hand, internal bank data is usually high-frequency, structured time-series data, while external government credit data is mostly in the form of unstructured text reports, PDF documents, or discrete events. Traditional technologies often struggle to effectively align these two types of data on the same dimension, typically relying on manual review of external reports to aid decision-making or simply performing keyword matching. There is a lack of effective mechanisms to transform unstructured text semantics into quantitative indicators that can be jointly calculated with internal time-series features, resulting in internal and external data being fragmented in risk models and failing to form a cohesive force.

[0004] Furthermore, in the risk assessment and calculation phase, existing risk control systems mostly employ static rule engines or black-box machine learning models. Static rules are often hard-coded into the system logic, lacking abstraction of business metadata, resulting in inflexible rule adjustments and difficulty in responding to rapidly changing risk situations. While single statistical models can detect anomalies, they lack business interpretability. Simultaneously, current systems generally lack adaptive evaluation mechanisms for rule effectiveness. When certain rules generate a large number of false alarms due to environmental changes, the system cannot automatically reduce their weight based on business feedback, often requiring regular manual review and parameter adjustments, resulting in high maintenance costs and significant delays.

[0005] In terms of managing the handling of risk leads after they are generated, existing technologies largely rely on general workflow software or manual ledgers, lacking strict status control logic. This makes the risk handling process susceptible to human interference, such as unauthorized jumps in process status, arbitrary rollbacks of handling actions, or prolonged stagnation at a certain stage without system-level monitoring. This loose control model cannot guarantee the certainty and irreversibility of the risk handling process, and fails to meet the stringent requirements of financial regulation for full-process traceability. Finally, due to the lack of cross-platform data synchronization mechanisms, the handling results of internal credit systems often form data silos, unable to be fed back to external data sources in a timely manner, hindering the closed-loop circulation and value enhancement of social credit system data. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for identifying credit customer risks based on multidimensional financial assets. This solves the problems in existing credit risk identification technologies, such as the difficulty in effectively integrating heterogeneous government and financial data, the static and rigid risk assessment rules lacking an adaptive adjustment mechanism, and the lack of closed-loop management throughout the entire risk disposal process.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for identifying credit customer risk based on multidimensional financial assets.

[0008] This method first performs the collection and alignment of multi-source heterogeneous data. The system obtains unstructured government report data of the target customer from a first external data source and structured financial data of the target customer from a second internal data source. The system uses the Unified Social Credit Code as a unique identifier to associate the unstructured government report data and structured financial data. Specifically, the system parses the two types of data, extracts the Unified Social Credit Code, and constructs a wide data table with the code as the primary key. The system sets a valid time window based on the current system date, retains government records occurring within the window, extracts the corresponding financial time-series data, and generates a raw dataset containing complete dimensions through inner join operations.

[0009] Building upon this foundation, the method performs feature engineering and vector space mapping. The system loads a pre-built risk semantic mapping dictionary and uses a regular expression matching algorithm to parse unstructured government report data, converting textual descriptions into discrete numerical features. Simultaneously, the system calculates the statistical features of structured financial data within the effective time window, including capital volatility features and balance trend slope features. The system further obtains historical extreme value parameters for each feature dimension and uses an extreme value normalization algorithm to map discrete numerical features and statistical features to a closed interval between zero and one, thereby constructing a standardized multidimensional feature vector and eliminating dimensional differences between different data sources.

[0010] Subsequently, the method performs dynamic risk calculation based on metadata. The system performs parallel calculations on multi-dimensional feature vectors based on a pre-built metadata rule base and a statistical anomaly detection model to obtain a comprehensive risk score for the target customer. During the calculation, the system parses the metadata rule base to construct a logical decision tree, traverses the decision tree using an interpreter pattern to determine the triggered risk monitoring rules, and calculates the sum of the risk deduction weights for all triggered rules. Simultaneously, the system inputs the multi-dimensional feature vectors into a pre-trained isolated forest model or a single-class support vector machine model to obtain normalized anomaly probability scores. The final comprehensive risk score is obtained by superimposing the sum of weights with the anomaly probability score corrected by an adjustment coefficient.

[0011] When the comprehensive risk score exceeds a preset global early warning threshold, the method generates a unique risk clue object and sets its initial state to "pending verification." Subsequently, the system uses a finite state machine to manage the risk clue object throughout its entire lifecycle. The finite state machine strictly controls the flow of the risk clue object through a set of states based on predefined input events and a state transition matrix. This set of states includes "pending verification," "in progress," "pending confirmation," "archived," and "ignored."

[0012] The specific state transition logic includes: in response to a verification command submitted by the verifier confirming the verification is true, the state transitions from "Pending Verification" to "In Progress"; in response to a handling command submitted by the handling personnel, the state transitions from "In Progress" to "Pending Confirmation"; and in response to a confirmation command submitted by the confirming personnel, the state transitions from "Pending Confirmation" to "Archived". The system determines that the lifecycle of a risk clue object terminates only when the object is in the "Archived" or "Ignored" state. Furthermore, this method includes a timeout control mechanism. The system starts a timer to monitor the duration of a risk clue object in the "Pending Verification" state. Once a preset timeout threshold is exceeded, a timeout alarm event is automatically triggered, and a follow-up command is sent to prevent process blockage.

[0013] Finally, the method performs closed-loop feedback processing and cross-platform synchronization. In response to a risk lead object transitioning to a terminated state, the system dynamically adjusts the weight parameters of the metadata rule base based on false positive data. Specifically, it periodically analyzes false positive samples that have migrated to the ignored state, calculates the false positive rate for each rule, and automatically reduces the weight of that rule based on attenuation logic if the false positive rate exceeds the tolerance threshold. Simultaneously, when a risk lead object migrates to the archived state, the system extracts the target customer's final business operation records from the internal credit system, converts them into standard data metadata defined by the first external data source using a data dictionary mapping engine, digitally signs and encapsulates the code using hybrid encryption, and synchronizes it to the first external data source via a secure transmission protocol, achieving cross-platform sharing of credit decision data.

[0014] A second aspect of this invention provides a credit customer risk identification system based on multidimensional financial assets.

[0015] The system includes a data integration and standardization module, a dynamic rule configuration and calculation module, a risk management module for the entire process, and an interface interaction and feedback module.

[0016] The data integration and standardization module is configured to acquire data from a first external data source and a second internal data source, respectively. It performs entity alignment using unique identifiers, and then discretizes and normalizes the data, outputting a standardized multidimensional feature vector. This module solves the technical problems of heterogeneous data fusion and numerical processing.

[0017] The dynamic rule configuration and calculation module is configured to calculate multi-dimensional feature vectors based on a metadata rule base and a statistical anomaly detection model, outputting a comprehensive risk score. This module achieves comprehensive quantification of known risk patterns and unknown anomaly patterns through the combination of rule-based judgment and statistical models.

[0018] The risk management module is configured to generate risk clue objects when the comprehensive risk score exceeds the global early warning threshold, and uses a finite state machine to control the flow of risk clue objects between the pending verification state, the handling state, the pending confirmation state, the archived state, and the ignored state. This module ensures the determinism, irreversibility, and traceability of the risk handling process.

[0019] The interface interaction and feedback module is configured to dynamically adjust the weight parameters of the metadata rule base based on false alarm data in response to the risk clue object's transition to the terminated state, and to encrypt and synchronize the credit decision results to the first external data source. This module endows the system with adaptive optimization capabilities and establishes a closed loop for internal and external data feedback.

[0020] This invention provides a method and system for identifying credit customer risk based on multidimensional financial assets. It has the following beneficial effects: 1. This invention uses the unified social credit code as the primary key to align unstructured government reporting data with structured financial data, and performs discretization mapping and normalization to construct a standardized multidimensional feature vector. This technique effectively eliminates the dimensional differences and dimensional influences between heterogeneous data sources, transforming qualitative text descriptions into computer-processable quantitative indicators. This achieves deep integration of government credit information and internal financial transaction behavior, improving the computational accuracy and coverage of the risk identification model in the multidimensional feature space.

[0021] 2. This invention utilizes finite state machine technology to manage the entire lifecycle of generated risk clue objects. It strictly controls the flow of clues between states such as pending verification, in progress, and archived, based on predefined input events and state transition matrices, and incorporates a timeout alarm mechanism. This technical solution mandates that risk handling must go through irreversible logical steps such as verification, handling, and confirmation. From the system's underlying logic, it avoids the problems of illegal jumps, handling omissions, or delayed responses that occur in traditional manual processes, ensuring the certainty, standardization, and full traceability of the credit risk management process.

[0022] 3. This invention constructs an adaptive optimization and cross-platform synchronization mechanism based on closed-loop feedback. This mechanism automatically and dynamically decays the weight parameters of inefficient rules in the rule base based on historical false alarm data, and encrypts and sends the final credit decision results back to the external data source. This mechanism enables the system to perform self-iteration and parameter tuning based on post-hoc business data, effectively reducing the false alarm rate generated by the model over time. Simultaneously, it breaks down data barriers between the internal credit system and the external government affairs platform, achieving cross-platform sharing and verification of risk disposal results. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart illustrating the overall process of the credit customer risk identification method of the present invention. Detailed Implementation

[0024] The technical solutions in 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.

[0025] See attached document Figure 1 This invention provides a credit customer risk identification system 100 based on multidimensional financial assets. The credit customer risk identification system 100 is deployed in a computer server cluster and is connected to a first external data source 200 and a second internal data source 300 through network communication interfaces. The first external data source 200 is a service platform providing government credit data, and the second internal data source 300 is a financial management system providing internal bank credit flow and accounting data.

[0026] The credit customer risk identification system 100 includes: a data integration and standardization module 10, a dynamic rule configuration and calculation module 20, a risk full-process management module 30, an interface interaction and feedback module 40, and a storage module 50.

[0027] The data integration and standardization module 10 is configured to perform the collection, entity alignment, and feature vectorization processing of multi-source heterogeneous data. This module obtains unstructured government reporting data, including business, tax, judicial, and environmental protection dimensions, from the first external data source 200 through an application programming interface (API) or file transfer protocol, and obtains structured financial data, including account transaction records, deposit and loan balances, and related transactions, from the second internal data source 300.

[0028] The data integration and standardization module 10 uses the unified social credit code as the primary key to align and correlate government reporting data and financial data of the same customer entity in a spatiotemporal manner. For numerical data, this module directly extracts feature values; for unstructured text data, this module uses a pre-built mapping table to convert text descriptions into corresponding numerical scores. The data integration and standardization module 10 further performs normalization processing on all feature data to eliminate differences in units and generate standardized multidimensional feature vectors.

[0029] Set up a customer The original feature set is , of which Each feature component is The data integration and standardization module 10 calculates the normalized eigenvalues. The formula is as follows: ; In the formula, Represents the normalized i-th Each feature component; Represents the original eigenvalues; Indicates the first sample in the sample set The minimum value of each characteristic component; Indicates the first sample in the sample set The maximum value of each feature component. The final normalized feature vector is represented as: .

[0030] The dynamic rule configuration and calculation module 20 is configured to build a risk monitoring model and calculate risk scores based on a metadata-driven approach. This module predefines standardized feature vectors. The module displays atomic indices corresponding to each component and receives user-configured logical rules. The dynamic rule configuration and calculation module 20 includes a rule parser, which compiles the rule templates configured on the front end into executable logical judgment scripts.

[0031] The dynamic rule configuration and calculation module 20 processes the standardized feature vectors according to a preset comprehensive scoring algorithm. Perform calculations and output customer data. The comprehensive risk score is calculated using a combination of rule-based decision-making and statistical models. The formula is as follows: ; In the formula, Indicates customer Comprehensive risk score; This indicates the total number of currently active risk monitoring rules; Indicates the first The judgment result of the rule is that when the standardized feature vector Satisfy the first The value is 1 if the condition set by the rule is met, and 0 otherwise. Indicates the first The risk deduction weights corresponding to each rule; This represents the anomaly score output by the anomaly detection model based on statistical learning; This represents the adjustment coefficient for the statistical model score.

[0032] The risk management module 30 is configured to manage risk clues throughout their entire lifecycle based on finite state machine logic. This is done when the dynamic rule configuration and calculation module 20 calculates the comprehensive risk score. When the preset threshold is exceeded, the risk management module 30 generates a unique risk clue object and controls the flow of the object according to a predefined finite set of states and state transition functions.

[0033] The risk end-to-end management module defines the set of states. This includes the following statuses: generated, pending verification, in progress, pending confirmation, archived, and ignored. This module monitors input events and performs status transitions to ensure that risk leads must undergo verification, processing, and confirmation before reaching a terminated state. The risk end-to-end management module 30 also has a built-in timer; when a risk lead remains in the pending verification state for longer than a preset time limit, a timeout event is automatically triggered, and an escalation push operation is executed.

[0034] The interface interaction and feedback module 40 is configured for performance data feedback optimization and cross-platform synchronization. This module acquires clue data that is ultimately marked as ignored in the risk end-process management module 30, calculates the false positive rate, and dynamically updates the rule weights in the dynamic rule configuration and calculation module 20 based on the false positive rate. .

[0035] The formula for updating the weights in the Interface Interaction and Feedback Module 40 is as follows: ; In the formula, Indicates the updated number Rule weights; Indicates the number before the update Rule weights; This indicates the preset learning rate parameter; Indicates the first The historical false alarm rate of a rule is defined as the proportion of clues triggered by that rule that are ultimately marked as ignored.

[0036] The interface interaction and feedback module 40 is also configured to extract credit decision result information, including credit limit adjustment results and contract status, after the risk clue object reaches the archived state, and to encrypt and encapsulate the information and send it to the first external data source 200 through the network communication interface.

[0037] Storage module 50 is configured to provide the data persistence environment required for system operation. Storage module 50 includes a relational database, an unstructured object repository, and a memory cache database. The relational database stores business status data, rule metadata, and structured financial data; the unstructured object repository stores raw PDF-format government report files obtained from the first external data source 200; and the memory cache database stores frequently accessed real-time feature vector data. Data integration and standardization module 10, dynamic rule configuration and calculation module 20, risk management module 30, and interface interaction and feedback module 40 are all connected to storage module 50 for data read and write operations.

[0038] See attached document Figure 2 This embodiment details the logical steps of risk identification and closed-loop processing based on multidimensional financial asset data.

[0039] This invention provides a method for identifying credit customer risk, which is automatically executed by a computer system and specifically includes the following steps: Step S1: Triggering and aligning the acquisition of multi-source heterogeneous data with entity association The system is configured with timed task scheduling or event trigger listeners. When preset trigger conditions are met (including the arrival of the daily batch processing time or the receipt of a new credit application event), the system starts the data acquisition process in parallel.

[0040] The system first calls the application programming interface (API) of the first external data source through a pre-configured encrypted channel to request and download the target customer's government credit report (PDF or JSON format), which includes business registration change records, administrative penalty information, tax rating information, and judicial litigation records. Simultaneously, the system accesses a second internal data source through an internal database connection, executes structured query statements (SQL), and extracts the target customer's bank account transaction history, average daily deposit balance, loan contract status, and related counterparty information within a preset time window.

[0041] After acquiring the two types of data mentioned above, the system performs entity alignment. The system reads the Unified Social Credit Code field from each record as a unique identifier. The system constructs a hash table in memory, using the Unified Social Credit Code as the primary key, and matches and merges unstructured data objects from the first external data source with structured data rows from the second internal data source to form the original dataset containing the customer's full financial asset information.

[0042] Step S2: Feature Engineering and Vector Space Mapping The system parses and extracts features from the raw dataset generated in step S1, converting it into standardized numerical vectors for subsequent computation modules to process.

[0043] For unstructured government data, the system uses a pre-built text mapping dictionary for discretization. For example, when scanning the environmental protection penalty field, if a serious violation keyword is matched, it is mapped to a preset high-risk score; if a minor violation keyword is matched, it is mapped to a preset low-risk score; if the field is empty or normal, it is mapped to zero. For structured financial data, the system performs time-series aggregation operations to calculate statistical characteristic values ​​within a specified period, including the ratio of capital inflows and outflows, the monthly average balance change rate, and the proportion of related transaction amounts.

[0044] Subsequently, the system performs a normalization operation on all extracted numerical features. The system calls the normalization formula defined in the aforementioned Example 1 to linearly map the feature values ​​of each dimension to the closed interval [0,1], eliminating dimensional differences between different data domains, and ultimately generating the standardized feature vector for the client. .

[0045] Step S3: Dynamic rule parsing and comprehensive score calculation based on metadata The system loads the metadata for the currently active risk rule configuration. This metadata defines the logical decision tree structure, where leaf nodes correspond to standardized feature vectors. The specific components in the array have intermediate nodes corresponding to logical operators (AND, OR, NOT) and comparison operators (greater than, less than, equal to).

[0046] The system uses the interpreter pattern to traverse the logic decision tree and standardize the feature vectors. The system performs calculations by substituting the rules into the model. First, it determines whether each hard rule has been triggered (i.e., outputs a truth value of 1). Simultaneously, the system calls a pre-trained statistical anomaly detection model, inputting standardized feature vectors. This yields the anomaly probability score output by the model.

[0047] Based on this, the system performs a weighted summation operation. The system reads the dynamic weight parameters and model adjustment coefficients corresponding to each rule stored in the database, and applies the comprehensive risk scoring formula detailed in Example 1. The customer's current risk score is calculated.

[0048] Step S4: Instantiation of risk clue objects and initialization of the state machine The system compares the comprehensive risk score calculated in step S3 with a preset global early warning threshold. If the comprehensive risk score exceeds the threshold, the system determines that there is potential credit risk and immediately instantiates a new risk clue object in the database.

[0049] This risk lead contains a unique serial number (UUID), the associated customer's unified social credit code, a list of specific triggering rules, a snapshot of the relevant original voucher (such as a transaction record or screenshot of a government report), and a current status flag. The system initially marks the object as pending verification and, based on the customer's organizational hierarchy, queries the institution-user mapping table to route the processing authority for this lead to the task queue of the corresponding account manager or risk management personnel.

[0050] Step S5: Full lifecycle flow control based on finite state machine The system initiates a finite state machine controller to continuously monitor the generated risk clue objects. The controller only responds to predefined input events (including submission for verification, risk confirmation, risk elimination, and timeout) and updates the object state strictly according to the state transition matrix.

[0051] When a verification feedback event is received from an investigator, the system verifies the integrity of the data. If the feedback indicates a genuine risk, the controller moves the status from "Pending Verification" to "In Progress"; if the feedback indicates a false alarm, it moves the status to "Ignored." When a handling action event (such as a record of a frozen credit limit operation) is received from a handling personnel, the controller moves the status to "Pending Confirmation." Finally, the status moves to "Archived" only after the confirming personnel submit a deconfirmation event.

[0052] During the state transition, the system's background daemon periodically scans all clue objects in the pending verification state. It calculates the difference between the current time and the state entry time. If the difference exceeds the system's preset time limit threshold (e.g., 24 hours), the daemon automatically triggers a timeout event, forcibly executes the upgrade transition logic, pushes the clue to the superior management account, and records the timeout exception in the log.

[0053] Step S6: Closed-loop feedback optimization and cross-platform data synchronization Once a risk clue object reaches a terminated state (archived or ignored), the system executes a post-processing procedure.

[0054] On one hand, the system extracts the clue samples that are ultimately marked as ignored and stores their feature vectors and corresponding false alarm labels in the sample database. The system periodically triggers a weight update algorithm, automatically reducing the weight parameters of rules with high false alarm rates according to the weight update formula in Example 1. This enables adaptive optimization of the model.

[0055] On the other hand, for leads marked as archived and confirmed as genuine risks, the system reads the customer's final handling result in the credit system (including credit limit adjustment value and loan classification status change) through an internal interface. The system encapsulates the above handling results into standard data messages, encrypts them using an asymmetric encryption algorithm, and sends them to the first external data source through a secure transmission protocol, completing the cross-platform synchronization of credit risk data.

[0056] This embodiment describes in detail the specific technical steps of the data integration and standardization module, focusing on the implementation mechanism of converting unstructured text data and structured time-series data into a unified feature vector.

[0057] In this embodiment of the invention, the data processing flow first performs parallel acquisition of multi-source data. The system establishes connection channels with a first external data source and a second internal data source through a pre-set ETL (Extract, Transform, Load) job scheduler. For the first external data source, the system sends a request message with an authentication token via HTTPS to obtain the target customer's government credit information data packet. This data packet is a nested JSON format or Base64 encoded PDF file stream, containing unstructured text fields for four dimensions: tax credit rating, administrative penalty records, environmental assessment results, and court enforcement status. For the second internal data source, the system directly accesses a read-only copy database of the core banking system through a JDBC connection pool to lock and extract the target customer's past... Daily deposit balance sequence, daily transaction details, and credit account status change logs for a period of one month (e.g., 12 months).

[0058] Subsequently, the system performs entity alignment and spatiotemporal association based on the primary key. The system parses all the acquired data records and extracts the Unified Social Credit Code field as the unique primary key for entity alignment. The system constructs a wide data table indexed by this primary key within an in-memory computing framework. For the time dimension, the system defines a valid time window based on the current system date. For discrete event data (such as administrative penalties) from the first external data source, records are retained if their occurrence dates fall within the time window; for time-series data from the second internal data source, the system extracts continuous sequences within the time window. The system performs an inner join operation, discarding entities that exist only in one data source, ensuring that each row of the generated wide table contains a complete view of both internal and external data.

[0059] Next, the system performs discretization mapping on the unstructured text data. The system loads a predefined risk semantic mapping dictionary, which contains a three-level key-value pair structure from field name to keyword to risk score. The system iterates through the text fields in the wide table, using a regular expression matching algorithm to find keywords in the dictionary.

[0060] Define text mapping function Taking the tax credit rating field as an example, the system executes the following mapping logic: if the field value is A, it is mapped to a score of 0; if it is B, it is mapped to a score of 20; if it is M, it is mapped to a score of 50; if it is C or D, it is mapped to a score of 100. Taking the environmental assessment result field as an example, if the text contains keywords such as "severe pollution" or "production shutdown and rectification," it is mapped to a score of 100; if it contains "warning" or "rectification within a specified period," it is mapped to a score of 40; otherwise, it is mapped to 0. Through this step, all qualitative textual descriptions are converted into quantitative discrete numerical features. .

[0061] Simultaneously, the system performs statistical feature extraction on structured time-series data. For bank transaction data, the system does not directly use the raw values, but instead calculates its statistical moment characteristics to characterize the stability of fund fluctuations. The system calculates the fund volatility characteristics within a specified time window. The calculation formula is as follows: ; In the formula, The total number of trading days within the time window; For the first Total transaction amount for the day (sum of absolute values ​​of inflows and outflows); This represents the average daily transaction amount within that time window. In addition, the system also calculates the monthly average balance showing a downward trend. The slope of the balance sequence is fitted using the least squares method. Through this step, the high-dimensional time series data is reduced to a set of key statistical indicators.

[0062] Finally, the system performs feature vector construction and standardization. The system then converts the discrete numerical features obtained from the above transformations... Characteristics of capital volatility Characteristics of the downward trend in average monthly balance By concatenating these features, the original feature vector is formed. To eliminate differences in numerical magnitude between different features (e.g., tax scores range from 0 to 100, while capital volatility can reach millions), the system independently performs Min-Max normalization on each component of the vector. The system reads the historical extreme value parameters (maximum values) of each feature dimension stored in the metadata. and minimum value By applying the normalization formula detailed in Example 1, all eigenvalues ​​are linearly compressed to the [0,1] interval.

[0063] After the above processing, the system outputs the final standardized multidimensional feature vector. The vector is a vector of length . A one-dimensional array, where the first Each component corresponds to a risk quantification value for government data, and then... Each component corresponds to a statistical feature value of the internal financial data. This vector data is serialized and written to an in-memory cache database for direct use by the subsequent dynamic rule engine module.

[0064] This embodiment elaborates on the internal operation mechanism of the dynamic rule configuration and calculation module 20, especially the technical details of how to abstract business rules into metadata and realize risk quantification through a dual-track algorithm.

[0065] In this embodiment, the system first constructs an atomic indicator library in memory to achieve decoupling mapping between feature vectors and business semantics. The system defines a set of atomic indicators. Each atomic index Each contains a unique identifier, a data type definition, and a pointer to a standardized feature vector. A mid-dimensional index pointer. For example, the atomic indicator tax credit risk points to a vector. The Each component. When the eigenvector When the data is transmitted from the data integration and standardization module in the aforementioned embodiment three to this module, the system quickly locates and reads the corresponding value based on the index pointer.

[0066] Building upon this, the system utilizes the interpreter pattern to dynamically parse the rules. The system will process each risk monitoring rule... The metadata structure is stored as an Abstract Syntax Tree (AST) in JSON or XML format. In this syntax tree, leaf nodes represent specific atomic indices compared to thresholds (such as...). Non-leaf nodes represent logical operators (AND, OR, NOT). During computation, the system recursively traverses this syntax tree.

[0067] For the Each rule defines the system's logical discrimination function. Assume this rule is... The atomic conditions are composed of logical operations, and the system performs the following Boolean operation: ; In the formula, This represents the logical AND operation; Indicates the first The feature vector index corresponding to each condition; For eigenvalues; This is the weighting coefficient for this condition; The preset judgment threshold; This is an indicator function that outputs 1 if the inequality within the parentheses is true, and 0 otherwise. The system iterates through all active rules and generates a vector of Boolean results.

[0068] Simultaneously, the system performs parallel anomaly detection computation based on statistical models. The system incorporates a lightweight anomaly detection algorithm (such as Isolation Forest or Single-Class Support Vector Machine), which is pre-trained based on historical unlabeled samples. The system then uses the complete feature vector... This is fed into the model as input. The model then computes the standardized feature vectors. The distance or path length to the high-density cluster in the feature space is used to output a normalized anomaly probability value. This value does not depend on rules used to discover unknown risk patterns outside the rule base's coverage.

[0069] Subsequently, the system performs an aggregation calculation of the comprehensive score. The system invokes the comprehensive risk scoring formula detailed in Example 1. The above logical discriminant function The output result (0 or 1) and the corresponding rule weight Multiply and sum to obtain the rule scoring part; then output the statistical model. With adjustment coefficient Multiplying the two components yields the model score. Adding the two together gives the final comprehensive risk score. .

[0070] Finally, the system performs a risk trigger determination. The system reads the global warning threshold from the configuration parameters. The system will calculate the results. and Perform a numerical comparison. If The system determines that a risk warning has been triggered and immediately calls the interface of the risk end-to-end management module, transmitting the customer ID and a list of triggered rule IDs. And the current overall score, to initialize the risk clue object; if The system will write the calculation results to the log database for archiving, without triggering subsequent processing procedures.

[0071] This embodiment describes in detail how the risk management module 30 utilizes deterministic automata theory to conduct rigorous logical control over the risk clue objects generated by the aforementioned calculation module, ensuring the irreversibility and traceability of the credit risk disposal process.

[0072] In this embodiment, the system first defines the state space of the risk cues in memory or persistent storage. The system then constructs a finite state machine model. It contains a set of mutually exclusive states. This set It should include at least the following discrete states: Initial generated state ( ): This indicates that the clue has just been generated by the rules engine and has not yet been assigned to a specific person for processing.

[0073] Pending verification status ( This indicates that the clue has been routed to the task queue of the person in charge, awaiting manual intervention to verify the authenticity of the data.

[0074] Execution status of disposal ( ): This indicates that the clue has been verified as a real risk and is in the stage of formulating and implementing risk control measures (such as reducing the credit limit or freezing the account).

[0075] Result confirmed ( ): This indicates that the action has been entered into the system and is awaiting review and approval by the supervisor or risk manager.

[0076] Archived status ( ): This indicates that the risk has been eliminated or resolved, the process has ended normally, and it is in the AcceptingState.

[0077] False alarm ignore state ( ): This indicates that after verification, the risk was determined to be non-substantive, the process was terminated, and it is also considered an accepted status.

[0078] The system is configured with a state transition controller, which maintains a state transition matrix that defines the legal transition paths between states. ,in This is the set of input events. The controller only responds to predefined business operation events or system-triggered events. Any migration request not defined in the matrix will be intercepted and an exception will be thrown to prevent process jumps (e.g., prohibiting direct jumps from the pending verification state to the archived state).

[0079] Specifically, once a risk clue object is instantiated, the system places it in... Subsequently, the system automatically triggered an assignment event based on the organization code field carried in the clue. Transition the state to In this state, if the auditor submits a false verification command via the client, the controller executes a migration. The system marks clues as negative samples and stores them in the sample database; if the submitted verification command is true, the controller executes the migration. .

[0080] During the handling and confirmation phase, the system incorporates a rollback mechanism to ensure handling quality. When a lead is in... If the confirming personnel submit a rejection order... The controller forces a reverse migration. The personnel in charge are required to revise the procedures. Only upon receiving instructions ( The controller only performs the migration when ( ). This triggers the subsequent data synchronization interface.

[0081] Furthermore, this embodiment also includes a time-based automatic triggering mechanism to address the issue of lag in manual processing. The system triggers each non-terminating state (i.e., except...) and Bind a separate timer function to the state other than the state itself. A daemon runs in the background that periodically (e.g., every 15 minutes) scans the list of active risk leads.

[0082] For those in Any clue object in state The daemon reads the timestamp of when it entered that state. and get the current system time. The daemon process calculates its dwell time. and compared with the preset timeout threshold (For example, 24 hours) for comparison.

[0083] If the conditions are met The daemon process automatically generates timeout alarm events. The controller responds to this event by executing exception handling logic. In this embodiment, this logic does not alter the main state transition but instead triggers a side action: elevating the priority field of the lead object to urgent and sending a supervisory instruction to the superior management account of the user to whom the lead belongs. Through this mechanism, the system ensures that every risk lead receives a response within a specified time, preventing the accumulation and spread of credit risk due to human negligence.

[0084] This embodiment details the specific technical implementation of the interface interaction and feedback module 40, focusing on how the system uses post-processing data to reverse-calibrate the risk control model and how it securely transmits internal credit decision results back to the external government affairs platform to achieve closed-loop data flow. In this embodiment, the system first executes a model weight adaptive update process based on feedback data. This process is automatically triggered by the system's background batch job scheduler at a preset cycle (e.g., monthly or quarterly).

[0085] When the operation starts, the system first accesses the historical database of the risk management module to retrieve all risk clue records whose status has terminated in the previous period. The system then divides these records into two sets based on their final status markers: one set is the set of valid early warnings (final status: archived). One type represents the real risk, and the other type is the set of false alarms (the final state of which is ignored). (This represents a false alarm).

[0086] Subsequently, the system performs a performance evaluation at the rule level. The system iterates through the clue records in each false alarm set, reading the list of trigger rule IDs recorded in its metadata. For the first rule in the rule base... The system counts the total number of times a rule is triggered in the previous period. And the number of times it was judged as a false alarm. The system calculates the current false alarm rate of the rule based on the false alarm rate calculation logic defined in Implementation Example 1. .

[0087] Next, the system performs dynamic weight decay or boosting operations. The system reads the current weight value of the rule stored in the database. If the calculated false alarm rate If the weights exceed a preset tolerance threshold (e.g., 30%), the system invokes a weight update algorithm to perform a weight reduction process. Specifically, the system updates the current weights and learning rate parameters. and false alarm rate Substituting into the weight update formula, the new weight values ​​are calculated. The new weight value is written to the rule metadata table and takes effect immediately in the next round of real-time risk calculation. Through this mechanism, the system automatically reduces the influence of frequently triggered but low-hit noise rules on the overall score, without requiring manual parameter intervention.

Claims

1. A method for identifying credit customer risk based on multidimensional financial assets, characterized in that, Includes the following steps: S1. Obtain unstructured government report data of the target customer from the first external data source and structured financial data of the target customer from the second internal data source, and align the unstructured government report data with the structured financial data using a unique identifier. S2. Discretize the aligned unstructured government report data into numerical maps, extract statistical features from the structured financial data, normalize all features, and construct a standardized multidimensional feature vector. S3. Based on a pre-set metadata rule base and statistical anomaly detection model, the multi-dimensional feature vector is calculated in parallel to obtain the comprehensive risk score of the target customer; S4. When the comprehensive risk score exceeds the preset global early warning threshold, a unique risk clue object is generated, and the initial state of the risk clue object is set to the pending verification state. S5. Use a finite state machine to manage the risk clue object throughout its entire lifecycle, and control the risk clue object to flow through a set of states including pending verification, in progress, pending confirmation, archived, and ignored. S6. In response to the risk clue object being transferred to the termination state, perform closed-loop feedback processing, including dynamically adjusting the weight parameters of the metadata rule base based on false alarm data, and encrypting and synchronizing the credit decision results to the first external data source.

2. The method for identifying credit customer risk based on multidimensional financial assets according to claim 1, characterized in that, In step S1, the specific process of aligning entities using unique identifiers includes: The unstructured government report data and the structured financial data are analyzed to extract the unified social credit code as the unique identifier; Construct a wide data table with the unified social credit code as the primary key, and set a valid time window based on the current system date; The unstructured government report data records occurring within the effective time window are retained, and the structured financial data sequence within the effective time window is extracted. An inner join operation is performed based on the unified social credit code to generate the original dataset.

3. The method for identifying credit customer risk based on multidimensional financial assets according to claim 2, characterized in that, Step S2 specifically includes: Load a risk semantic mapping dictionary containing field names, keywords, and risk score correspondences, and use a regular expression matching algorithm to convert the text descriptions in the unstructured government affairs report data into discrete numerical features; Calculate the volatility characteristics of funds and the slope characteristics of the balance trend of the structured financial data within the effective time window; Obtain the historical maximum and minimum values ​​of each feature dimension, and use the extreme value normalization algorithm to map the discrete numerical features, the capital volatility features, and the balance trend slope features to a closed interval from zero to one, and combine them to form the multidimensional feature vector.

4. The method for identifying credit customer risk based on multidimensional financial assets according to claim 1, characterized in that, Step S3, the process of calculating based on the preset metadata rule base, includes: Parse the metadata rule base and construct a logical judgment tree. The leaf nodes of the logical judgment tree correspond to the specific feature components and preset thresholds in the multidimensional feature vector, and the non-leaf nodes of the logical judgment tree correspond to logical operators. The logic decision tree is traversed using the interpreter pattern, and the multi-dimensional feature vector is substituted into the calculation to determine whether each risk monitoring rule is triggered.

5. The method for identifying credit customer risk based on multidimensional financial assets according to claim 4, characterized in that, In step S3, the calculation logic for obtaining the comprehensive risk score of the target customer is as follows: Obtain the risk deduction weights corresponding to each triggered risk monitoring rule, and calculate the sum of the weights of all triggered rules; The multidimensional feature vector is input into a pre-trained isolated forest model or a single-class support vector machine model to obtain a normalized anomaly probability score. The product of the anomaly probability score and a preset adjustment coefficient is calculated. The sum of the weights is added to the product to obtain the comprehensive risk score.

6. The method for identifying credit customer risk based on multidimensional financial assets according to claim 1, characterized in that, In step S5, the state transition logic of the finite state machine includes: In response to the verification instruction submitted by the verification personnel that the verification is true, the status of the risk clue object is changed from the pending verification status to the processing status; In response to the disposal action instructions submitted by the disposal personnel, the status of the risk clue object is transferred from the disposal status to the pending confirmation status; In response to the deconfirmation instruction submitted by the confirming personnel, the status of the risk clue object is transferred from the pending confirmation status to the archived status; The lifecycle of a risk clue object is determined to have ended only when the object is in the archived state or the ignored state.

7. The method for identifying credit customer risk based on multidimensional financial assets according to claim 1, characterized in that, Step S5 also includes a timeout control step: Start a timer bound to the risk clue object to monitor the duration of the risk clue object's stay in the pending verification state; When the dwell time exceeds the preset time limit threshold, an overtime alarm event is automatically triggered, and a supervision instruction is sent to the associated management account.

8. The method for identifying credit customer risk based on multidimensional financial assets according to claim 1, characterized in that, In step S6, dynamically adjusting the weight parameters of the metadata rule base based on false alarm data specifically includes: Periodically analyze the risk clue objects whose final state has migrated to the ignored state, and mark the risk clue objects as false alarm samples; For each risk monitoring rule in the metadata rule base, the proportion of clues triggered by the risk monitoring rule that ultimately become false alarm samples is calculated to obtain the false alarm rate. When the false alarm rate exceeds the preset tolerance threshold, the risk deduction weight corresponding to the risk monitoring rule is automatically reduced according to the preset attenuation formula.

9. The method for identifying credit customer risk based on multidimensional financial assets according to claim 1, characterized in that, In step S6, encrypting and synchronizing the credit decision results to the first external data source specifically includes: When the status of the risk clue object is transferred to the archived status, the final business operation record of the target customer in the internal credit system is extracted. The final business operation record is converted into standard data metacode defined by the first external data source using a data dictionary mapping engine. The standard data metadata is digitally signed and encapsulated with hybrid encryption, and then sent to the first external data source via a secure transmission protocol.

10. A credit customer risk identification system based on multidimensional financial assets, characterized in that, The credit customer risk identification method based on multidimensional financial assets, as described in any one of claims 1-9, includes: The data integration and standardization module is configured to acquire data from a first external data source and a second internal data source, respectively, complete entity alignment through a unique identifier, and perform discretization mapping and normalization processing on the data to output a standardized multidimensional feature vector. The dynamic rule configuration and calculation module is configured to calculate the multidimensional feature vector based on the metadata rule base and the statistical anomaly detection model, and output a comprehensive risk score; The risk management module is configured to generate a risk clue object when the comprehensive risk score exceeds the global early warning threshold, and to use a finite state machine to control the flow of the risk clue object between the pending verification state, the handling state, the pending confirmation state, the archived state, and the ignored state. The interface interaction and feedback module is configured to dynamically adjust the weight parameters of the metadata rule base based on false alarm data in response to the risk clue object's transition to the termination state, and to encrypt and synchronize the credit decision results to the first external data source.

Images