Bank credit risk assessment method, system and device based on big data
By constructing a causal knowledge graph and a time-series graph neural network, the problems of lag and interpretability in traditional bank credit risk assessment methods in the face of economic fluctuations and environmental changes are solved, and dynamic risk assessment and forward-looking early warning with causal transparency are realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-05
- Publication Date
- 2026-03-31
AI Technical Summary
Traditional bank credit risk assessment methods deteriorate sharply in the face of economic cycle fluctuations, lacking causal explanatory power and dynamic adaptability. This results in models exhibiting lag and interpretability gaps in risk identification and early warning, and failing to proactively adapt to changes in the external environment.
A knowledge graph based on causal relationships is constructed. Direct connections between financial variables are identified through causal discovery algorithms, generating a basic financial causal knowledge graph. Time-series data of credit entities are mapped onto the graph, and a time-series graph neural network model is trained to predict risk evolution paths. Combined with causal structure drift monitoring and adaptive correction modules, dynamic assessment and environmental adaptation are achieved.
It achieves causal transparency and dynamic foresight in risk assessment, enabling the prediction of the probability of risk events before they occur, proactively adapting to environmental changes, and improving the interpretability of risk management and the accuracy and timeliness of assessment.
Smart Images
Figure CN121767076A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of financial risk management, and more specifically, to a method, system, and apparatus for assessing bank credit risk based on big data. Background Technology
[0002] Bank lending is a core pillar of the modern financial system. Traditional risk assessment methods, such as scorecards based on expert rules or statistical models, are no longer adequate to meet current challenges. While these methods perform reasonably well during periods of stable business and macroeconomic conditions, their performance deteriorates sharply during economic cycles. Industry data shows that a standard scorecard model typically maintains a KS value of around 0.35 during periods of economic stability, but during economic recessions, the KS value can plummet below 0.22 due to the emergence of numerous previously unseen risk patterns, resulting in a severe decline in its risk differentiation capabilities.
[0003] With the advent of the big data era and the rapid development of artificial intelligence technology, traditional risk assessment methods based on expert rules or simple statistical models have been gradually replaced by complex models based on machine learning and deep learning. These modern risk control models improve the accuracy and coverage of risk identification to a certain extent by learning complex nonlinear relationships from massive amounts of user data.
[0004] However, current mainstream risk assessment technologies based on big data and machine learning still reveal inherent, deep-seated limitations in practice. The core of these models lies in uncovering statistical correlations between variables, rather than underlying causality. While models can identify patterns where "high debt ratios" and "high delinquency risk" coexist, they cannot explain the underlying transmission mechanisms, making the models themselves incomprehensible "black boxes." This neglect of causality renders the model's conclusions lacking sufficient interpretability, making it difficult for risk approvers to fully trust and effectively utilize the model results. Furthermore, prediction logic based on chance or indirect correlations inherently exhibits fragility in stability and reliability when faced with new sample distributions.
[0005] Furthermore, most existing risk assessment models offer a "static snapshot" evaluation of credit entities. Even when using users' time-series data, it's often compressed into several static statistics (such as "average income over the past six months") through feature engineering, inputting them into the model to derive a risk score for the current point in time. This approach ignores the evolving nature of risk as a dynamic process, failing to capture the continuous trajectory and development trend of a user's financial situation over time. Therefore, the models struggle to provide proactive risk warnings, often only identifying risks after they are about to occur or have already occurred, thus missing the opportunity for early risk management.
[0006] More importantly, the financial market itself is a complex, dynamic system. Macroeconomic cycles, regulatory policy adjustments, and even the emergence of new business models constantly reshape the interactions between financial variables. Currently prevalent technological paradigms, once their models are trained, have their internal parameters and structures fixed. Such static models cannot proactively adapt to changes in the external environment; the statistical laws upon which they are based may have subtly altered in the real world, causing model performance to continuously decline over time. The existing system lacks an automated mechanism to effectively monitor and respond to this "causal structure drift." Model updates and iterations often rely on manual intervention, which is time-consuming and costly, making risk assessment systems particularly passive and lagging in dynamic environments.
[0007] With the development of big data and artificial intelligence technologies, complex models based on machine learning (such as XGBoost) have improved in accuracy, but fundamental bottlenecks still exist: Post-loan early warning is delayed: Existing models are mostly static assessments and are not sensitive to the dynamic evolution of a customer's risk status. When risk signals appear, banks' post-loan early warnings are delayed by an average of one to two quarters, missing the best opportunity for intervention.
[0008] Poor data and model timeliness: In the real business context of banks, data is scattered across more than ten heterogeneous systems. A complete feature calculation requires an average of T+2 days to retrieve data from across systems, and the missing rate of key fields is often higher than 8%. At the same time, once a model is deployed, it tends to become fixed, and its performance will continue to degrade over time and with changes in the environment, lacking an effective automated update loop.
[0009] Explainability gap: Complex machine learning models are often “black boxes”, making it difficult to clearly explain the decision-making basis to business people and regulators, and do not fully meet the Basel Accords’ requirements for model risk management.
[0010] Therefore, there is an urgent need for a completely new technological paradigm that not only predicts risks, but also explains their causes and can dynamically adapt to environmental changes. Summary of the Invention
[0011] The purpose of this invention is to provide a method, system, and apparatus for assessing bank credit risk based on big data, which can not only predict risks but also explain the causes of risks and dynamically adapt to environmental changes.
[0012] The above-mentioned technical objective of this invention is achieved through the following technical solution: a bank credit risk assessment method based on big data, comprising the following steps: S1. Based on historical credit data, construct a basic financial causal knowledge graph that represents the causal relationships between multiple financial variables; S2. Map the time series data of the specified credit subject to the basic financial causal knowledge graph to generate a dynamic risk evolution graph of the subject that represents the dynamic changes in the financial status of the credit subject. S3. By using the dynamic risk evolution map of the main body, a predictive model is trained to generate a prediction model for predicting the risk evolution path of the credit entity within a future time window; S4. Continuously monitor structural changes in causal relationships and revise the basic financial causal knowledge graph and prediction model based on the monitoring results.
[0013] As a preferred embodiment of the present invention, S1 includes: A causal discovery algorithm is used to perform a series of conditional independence tests on historical credit data to identify direct causal relationships between variables and generate an initial causal structure. Using prior knowledge as a hard constraint, the initial causal structure is calibrated and the edges are oriented to generate a directed acyclic graph, namely the basic financial causal knowledge graph.
[0014] As a preferred technical solution of the present invention, S2 includes: acquiring time-series data of a specified credit entity within consecutive time steps in the past, and instantiating the specific data values at each time step as node attributes onto the topological structure of a basic financial causal knowledge graph to form a time-series graph sequence, namely, a dynamic risk evolution graph of the entity; each graph in the time-series graph sequence represents a snapshot of the financial state of the credit entity at the corresponding time, and the time-series graph sequence as a whole constitutes a dynamic risk evolution graph of the entity used to fully describe the dynamic change process of the financial state of the credit entity.
[0015] As a preferred technical solution of the present invention, in S3, the time series neural network model is trained by the subject dynamic risk evolution map. During the training, the causal chain propagation path related to the changes in financial variables and the impact path on downstream variables are captured by the message passing mechanism, and the parameters of the mechanism are updated to output the risk evolution path. The risk evolution path output by the time-series graph neural network model is a risk state transition probability matrix, which is used to characterize the probability that the risk level of a credit entity will change between preset states within a preset time window in the future.
[0016] As a preferred embodiment of the present invention, in S3, the node state update process in the temporal graph neural network model is represented as follows: , represents the state vector of node v at time t, which is a mathematical abstraction of risk characteristics; UPDATE is the state update function; AGGREGATE is the neighbor information aggregation function, which is used to gather the state information of all direct cause nodes at the previous time step; Let v be the external input feature of node v at time t.
[0017] As a preferred embodiment of the present invention, in S4, S1 is continuously re-executed in the latest rolling time window data to generate a causal graph for the current time window period and calculate the difference value between it and the currently used basic financial causal knowledge graph. When the difference value exceeds a preset threshold, a correction operation is triggered. The correction operation includes: correcting the current basic financial causal knowledge graph based on the new causal relationships in the causal graph for the current time window period; and triggering the prediction model to retrain on the updated basic financial causal knowledge graph and the latest data.
[0018] As a preferred embodiment of the present invention, the method for calculating the difference between the causal graph of the current time window period and the currently used basic financial causal knowledge graph is as follows: A structured comparison is made between the causal graph of the current time window and the basic financial causal knowledge graph currently in use; The causal outcome offset between the causal graph of the current time window and the currently used basic financial causal knowledge graph is calculated using structured Hamming distance. ,in, For the calculated drift score, For the structured Hamming distance calculation function, It is a causal knowledge graph for basic finance. This is a causal graph of the window period.
[0019] A big data-based bank credit risk assessment system includes: The basic financial causal knowledge graph construction module is used to construct a basic financial causal knowledge graph representing the causal relationships between multiple financial variables based on historical credit data. The entity dynamic risk evolution graph generation module is used to map the time series data of a specified credit entity to a basic financial causal knowledge graph to generate an entity dynamic risk evolution graph that represents the dynamic changes in the financial status of the credit entity. The risk evolution path prediction module is used to train and generate a prediction model for predicting the risk evolution path of a credit entity within a future time window by using the entity's dynamic risk evolution map. The adaptive correction module is used to continuously monitor structural changes in causal relationships and correct the basic financial causal knowledge graph and prediction model based on the monitoring results.
[0020] As a preferred technical solution of the present invention, the basic financial causal knowledge graph construction module adopts a preset causal discovery algorithm to perform a series of conditional independence tests on historical credit data to identify the direct causal relationship between variables and generate an initial causal structure; using prior knowledge as a hard constraint, the initial causal structure is calibrated and the edges are oriented to generate a directed acyclic graph, i.e., the basic financial causal knowledge graph. The entity dynamic risk evolution graph generation module is used to obtain time series data of a specified credit entity within a continuous time step in the past. The specific data value at each time step is used as a node attribute and instantiated onto the topological structure of the basic financial causal knowledge graph to form a time series graph sequence, i.e., the entity dynamic risk evolution graph. The risk evolution path prediction module is used to train the time-series graph neural network model through the dynamic risk evolution map of the subject. During training, the module captures the causal chain propagation path related to changes in financial variables and the path of influence on downstream variables through a message passing mechanism, updates the parameters of the mechanism, and outputs the risk evolution path. The risk evolution path output by the time-series graph neural network model is a risk state transition probability matrix, which is used to characterize the probability that the risk level of the credit subject will change between preset states within a preset time window in the future. The adaptive correction module continuously regenerates the causal graph for the current time window based on the latest rolling time window data, and calculates the difference between it and the currently used basic financial causal knowledge graph. When the difference exceeds a preset threshold, a correction operation is triggered. The correction operation includes: correcting the current basic financial causal knowledge graph based on the new causal relationships in the causal graph for the current time window; and triggering the prediction model to retrain on the updated basic financial causal knowledge graph and the latest data.
[0021] A bank credit risk assessment device based on big data includes a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor implements the above-described method when executing the computer program.
[0022] In summary, this invention offers the following advantages: Unlike traditional "black box" models, its core lies in constructing a fundamental financial causal knowledge graph. This ensures that any assessment conclusion is built upon a logically clear and well-defined causal network. When the model outputs a risk warning, approvers can intuitively trace the basis of their judgment, clearly seeing which initial variables(s) changed and through what risk transmission path affected the final result. This inherent transparency transforms the model from a mere predictive tool into a trustworthy analytical system that assists human understanding and decision-making.
[0023] This invention captures the continuous trajectory and development trend of risk over time, rather than merely assessing a static time slice. This allows the model to predict the probability of a risk event occurring at a future point in time, before the event actually happens. It changes the lag inherent in traditional risk control, shifting risk management from a passive "post-event response" to a proactive "pre-event prediction."
[0024] This invention designs a closed-loop system incorporating causal structure drift monitoring and adaptive correction. This enables the system to proactively and periodically detect the impact of changes in the external macroeconomic and market environment on the intrinsic logic between financial variables. Once a significant deviation in the causal relationship in the real world is detected, the system automatically triggers knowledge graph updates and model retraining, ensuring that its evaluation logic remains up-to-date. This overcomes the inherent problem of performance degradation over time in traditional static models, guaranteeing the model's continued effectiveness and accuracy in long-term applications.
[0025] This invention provides a systematic insight into the mechanisms of risk formation by constructing and applying causal knowledge graphs. It organizes complex financial data into a structured network of causal relationships, intuitively demonstrating the complete chain of mutual influence and transmission among different risk factors. This helps financial institutions understand the nature of risk from a more macro and fundamental perspective, providing in-depth theoretical support for formulating more scientific credit policies and risk management strategies.
[0026] This invention generates a unique dynamic risk evolution map for each assessment subject, making the assessment process highly personalized. The model moves beyond a broad, coarse matching based on general group characteristics, delving into the dynamic changes in individual data to analyze the risk evolution path of a specific subject. This allows the system to accurately identify the most critical risk points and the most likely transmission pathways for that particular subject, achieving a far more refined and targeted dynamic assessment than traditional methods, significantly improving the granularity of risk differentiation. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of the functional modules of the system of the present invention; Figure 2 This is an overall flowchart of the risk assessment and adaptive correction method of the present invention; Figure 3 This is an example schematic diagram of the basic financial causal knowledge graph of the present invention. Detailed Implementation
[0028] The present invention will be further described in detail below with reference to the accompanying drawings.
[0029] This invention provides a big data-based method and system for assessing bank credit risk. The system includes: a basic financial causal knowledge graph construction module, a dynamic risk evolution graph generation module, a risk evolution path prediction module, and an adaptive correction module. This system can discover and utilize the causal relationships between financial variables to dynamically and proactively assess the risk of credit entities, and has the ability to adaptively correct for environmental changes. The system aims to achieve dynamic and proactive risk assessment of credit entities by deeply mining the causal logic hidden in the data and combining it with time-series inference, while also possessing the ability to adapt to changes in the macro environment.
[0030] The following are the steps executed through each module in the system: S1. Executed through the basic financial causal knowledge graph construction module: Based on historical credit data, construct a basic financial causal knowledge graph representing the causal relationships between multiple financial variables; The core function of this module is to automatically construct a knowledge network from large-scale historical credit data that reflects the inherent causal relationships between various financial variables.
[0031] Specifically, S1 includes: Using a pre-defined causal discovery algorithm, such as a constraint-based PC algorithm or a variant of the FCI algorithm, a series of conditional independence tests are performed on historical credit data to identify direct causal relationships between variables and generate an initial causal structure. By integrating well-known prior knowledge in the financial field, such as "income is the cause of consumption" rather than the other, as a hard constraint, the initial causal structure is calibrated and the edges are oriented to generate a structurally stable directed acyclic graph with clear directionality, namely the basic financial causal knowledge graph.
[0032] S2. Executed through the subject dynamic risk evolution graph generation module: The time series data of the specified credit subject is mapped to the basic financial causal knowledge graph to generate a subject dynamic risk evolution graph that represents the dynamic changes in the financial status of the credit subject. This module is connected to the basic financial causal knowledge graph construction module, and its function is to apply general causal knowledge maps to specific credit entities.
[0033] Specifically, S2 includes: The system acquires time-series data of a specified credit entity over a continuous period of time, such as income, debt, and consumption data for several consecutive months. Then, it instantiates the specific data values at each time step as node attributes onto the topology of a basic financial causal knowledge graph, thus forming a time-series graph sequence, i.e., a dynamic risk evolution graph of the entity. Each graph in the time-series sequence represents a snapshot of the credit entity's financial state at the corresponding moment. The entire time-series sequence constitutes a dynamic risk evolution graph of the entity, used to fully describe the dynamic changes in the financial state of the credit entity.
[0034] S3. Executed through the risk evolution path prediction module: By using the entity's dynamic risk evolution map, a prediction model is trained and generated to predict the risk evolution path of the credit entity within a future time window; This module connects to the entity's dynamic risk evolution map generation module, aiming to predict the entity's future risk trends by learning the evolution patterns of risk states. Specifically, S3 includes: training a time-series graph neural network model using a dynamic risk evolution graph of the subject. This model can effectively process graph-structured time-series data. During training, a message passing mechanism is used to capture how changes in financial variables propagate along causal chains and affect downstream variables. Through training on a large number of historical evolution graphs, the model can learn the parameters of this update mechanism. Finally, the risk evolution path output by this module can be represented as a risk state transition probability matrix. This matrix clearly shows the probability that the risk level of a credit subject will transition between multiple preset states (low, medium, and high) within a future preset time window.
[0035] In S3, the node state update process in the time-series graph neural network model is represented as follows: , represents the state vector of node v at time t, which is a mathematical abstraction of risk characteristics; UPDATE is the state update function, usually a gated recurrent unit or other neural network that can learn time-series dependencies; AGGREGATE is the neighbor information aggregation function, which is used to gather the state information of all direct cause nodes at the previous time step. Let v be the external input feature of node v at time t.
[0036] S4. Executed through the adaptive correction module: Continuously monitor structural changes in causal relationships and correct the basic financial causal knowledge graph and prediction model based on the monitoring results.
[0037] This module constitutes a key feedback loop in the system, ensuring that the system can adapt to changes in the macro environment.
[0038] Specifically, in S4, the causal discovery process (S1) is continuously rerun on the latest rolling time window data to generate a temporary window-period causal graph that reflects recent market logic. The difference between this graph and the currently used basic financial causal knowledge graph is calculated. When the difference exceeds a preset threshold, the system determines that a significant causal structure drift has occurred and triggers a correction operation. The correction operation includes: 1. Integrating or updating the validated and stable new causal relationships in the window-period causal graph to the basic financial causal knowledge graph based on the new causal relationships in the current time window; 2. Triggering the prediction model to retrain on the updated basic financial causal knowledge graph and the latest data, thereby completing the system's self-adaptation.
[0039] The method for calculating the difference between the causal graph of the current time window and the currently used basic financial causal knowledge graph is as follows: A structured comparison is made between the causal graph of the current time window and the basic financial causal knowledge graph currently in use; The causal outcome offset between the causal graph of the current time window and the currently used basic financial causal knowledge graph is calculated using structured Hamming distance. ,in, For the calculated drift score, For the structured Hamming distance calculation function, It is a causal knowledge graph for basic finance. This is a causal graph of the window period.
[0040] Corresponding to the methods and systems described above, the present invention also provides a bank credit risk assessment device based on big data, comprising: a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor implements the methods described in S1-S4 when executing the computer program.
[0041] Compared to traditional risk assessment models that rely on static and correlational features, the technical solution of this invention can identify deep-seated causal transmission paths between financial variables. It innovatively designs a closed-loop system that includes causal knowledge graph construction, time-series neural network prediction, and adaptive correction of causal structure drift, possessing the following irreplaceable advantages: The "auditability" brought about by causal path explanation: By constructing a basic financial causal knowledge graph, the system no longer treats variables as flat features, but organizes them into a structured network with internal logic. This makes risk assessment results (such as pre-loan access decisions) not only probability values, but also traceable causal attribution paths, greatly improving the interpretability and auditability of the model.
[0042] The "Dynamic Individual Profile" achieved through time-series graph modeling: By mapping the time-series data of credit entities onto a causal graph and using a time-series graph neural network for evolutionary path prediction, this invention transforms risk assessment from a "static snapshot" judgment to a dynamic, forward-looking "video" inference. This enables the system to capture the trajectory of risk evolution and is effectively applied to alerting deviations in loan behavior.
[0043] The "Adaptive Update Closed Loop" constructed by causal drift correction: This invention endows the system with the ability to self-evolve by establishing a causal structure drift monitoring module and an adaptive correction module. Once a significant drift is identified, the system can automatically update the knowledge graph and trigger model retraining for adaptive adjustment of post-loan risk thresholds and feature sets, ensuring the long-term effectiveness of the model.
[0044] The following is an embodiment of the present invention, such as Figures 1-3 As shown, the system of the present invention includes a basic financial causal knowledge graph construction module, a subject dynamic risk evolution graph generation module, a risk evolution path prediction module, a causal structure drift monitoring module, and an adaptive correction module.
[0045] These modules are deployed on one or more servers and communicate and interact with data via an internal data bus or network. The server hardware environment includes processors, memory, and storage, while the software environment can be built on an operating system such as Linux, using distributed computing frameworks such as Spark to process massive amounts of data, and using deep learning frameworks such as PyTorch and TensorFlow to implement complex model calculations.
[0046] The basic financial causal knowledge graph construction module, as the logical cornerstone of the entire system, has the core task of automatically identifying and constructing a knowledge network that reflects the inherent causal relationships between various financial variables from massive amounts of anonymous historical credit data.
[0047] The basic financial causal knowledge graph construction module first retrieves data from the data warehouse, covering multiple dimensions such as user attributes, behaviors, and transactions, and then abstracts it into a set of well-defined financial variables. These variables include, for example, "total monthly income," "credit debt ratio," "frequency of high-frequency online purchases," and "employment stability."
[0048] Next, the basic financial causal knowledge graph construction module employs constraint-based causal discovery algorithms, such as the PC algorithm, to systematically explore the causal structure among these variables. The process begins with a fully connected graph where all variable nodes are connected by undirected edges. Subsequently, the algorithm iteratively performs a series of conditional independence checks, gradually eliminating edges in the graph that do not meet the conditions. For any pair of variables... and The algorithm checks whether they are independent given their adjacent subset Z, i.e., it checks... Whether the condition of independence holds. If this condition of independence holds, then the connection... and Edges will be removed. This process continues until no more edges can be removed from the graph, forming a stable undirected skeleton graph.
[0049] After obtaining the skeleton graph, the basic financial causal knowledge graph construction module uses specific orientation rules to determine the direction of edges. For example, by identifying V-shaped structures, some edges are assigned causal directions. To further improve the accuracy of the graph, the module also uses well-known prior knowledge in the financial field, such as "age" being one of the causes of "income level," rather than the other way around, as a hard constraint to calibrate and supplement the orientation of the graph structure discovered by the algorithm. After the above steps, a directed acyclic graph is finally output, namely the basic financial causal knowledge graph: ,in, Let be the set of directed edges representing causal relationships.
[0050] The module for generating a dynamic risk evolution graph for the main entity is connected to the module for constructing a basic financial causal knowledge graph. It is responsible for applying a general causal knowledge map to each specific credit entity in a personalized way. When a credit application enters the system, this module retrieves the applicant's personal time-series data over a continuous time step, such as the past 12 months. Then, the specific data values at each time step are "attached" as node attributes to the basic financial causal knowledge graph. Based on the topological structure, the main dynamic risk evolution map generation module generates a time series diagram sequence for the main entity. Each of the images Each represents a snapshot of the financial state of the subject at a specific time t, and the entire sequence constitutes a dynamic risk evolution map of the subject that can fully describe its state transition process.
[0051] The risk evolution path prediction module receives a dynamic risk evolution graph generated by the module itself. Its goal is to learn and predict how risk states propagate and evolve over time on this causal network. The module employs a time-series graph neural network model. At its core is a recursive state update mechanism that captures information from the causal nodes at each time step and updates the state of the current node.
[0052] In this embodiment, the time-series graph neural network model used in the risk evolution path prediction module has a core recursive state update mechanism that can be characterized by the following formula: ;in, It is the hidden state vector of node v at time t. This vector is a high-dimensional mathematical representation of the risk characteristics of the financial variable represented by this node at the current time.
[0053] Represents the basic financial causal knowledge graph The set of all neighboring nodes that point to node v, that is, all direct cause nodes of node v.
[0054] AGGREGATE is a neighbor information aggregation function that aggregates information from all cause nodes. The state information at the previous time t-1 This aggregation function can employ methods such as mean pooling, max pooling, or attention-based weighted summation to effectively integrate upstream risk signals.
[0055] For node v, the external input feature at time t is the actual observed value of the financial variable for that credit entity at that time.
[0056] UPDATE is a state update function that combines the node's own historical state. The aggregated upstream cause information, and the external input at the current moment. To calculate and generate the new state vector of the node at the current time t. Specifically, the update function can employ neural network units with excellent time-series information processing capabilities, such as gated recurrent units (GRUs) or long short-term memory networks (LSTMs).
[0057] To obtain a model with accurate predictive capabilities, the risk evolution path prediction module needs to undergo a training process. This process uses a large amount of historical dynamic risk evolution maps of credit entities and their known actual risk outcomes as training data. The goal of training is to minimize the difference between the risk state distribution predicted by the model and the actual risk state. In one embodiment, the cross-entropy loss function can be used as the objective function, and all learnable parameters in the time-series graph neural network model can be iteratively adjusted using a backpropagation algorithm and a gradient descent optimizer (such as Adam) until the model converges.
[0058] After training is complete, the final output of the risk evolution path prediction module is a clear risk state transition probability matrix. This matrix clearly shows the probability distribution of a credit entity's risk level transitioning between multiple preset states such as "low," "medium," and "high" within one or more preset time windows in the future, providing a dynamic and quantitative forward-looking basis for credit decisions.
[0059] In this invention, to ensure that the system can maintain the accuracy and timeliness of its assessments in the face of constantly changing macroeconomic environments and user behavior patterns, the system further includes a causal structure drift monitoring module and an adaptive correction module that form an adaptive monitoring and correction closed loop.
[0060] The core function of the causal structure drift monitoring module is to continuously and automatically "check up" the underlying financial causal knowledge graph upon which the system relies. Does it still conform to the latest market logic? The causal structure drift monitoring module runs on a preset period (e.g., quarterly). In each period, it selects data from the latest rolling time window and re-executes the causal discovery process in the basic financial causal knowledge graph construction module on this data, thereby generating a temporary causal graph that only reflects the logic of recent data within a specific window. .
[0061] Then, the module will With the system currently in use A rigorous structured comparison is performed to quantify the differences between the two. This difference is measured by calculating the causal structural drift score (DriftScore), which can be calculated using the Structural Hamming Distance (SHD): DriftScore = SHD The SHD function calculates the transformation of a graph structure. Transform into another graph structure The minimum number of edge operations (including adding, deleting, or reversing edges) required. A higher Drift Score indicates a significant change in the causal relationships between variables in the real world.
[0062] The adaptive correction module, closely linked to the causal structure drift monitoring module, is the execution unit for the system's self-evolution. The causal structure drift monitoring module sets a warning threshold. This threshold can be determined based on statistical distribution analysis of historical drift scores. When the DriftScore calculated by the module exceeds this threshold... At that time, the system determined that a significant causal structure drift had occurred.
[0063] Once the drift is confirmed, the adaptive correction module will automatically trigger a series of correction operations. First, analysis... New causal relationships that emerge and remain stable across multiple consecutive windows are identified, and these validated new structures are updated or integrated into the basic financial causal knowledge graph based on a pre-defined fusion strategy (e.g., introducing a confidence scoring mechanism for edges). Next, the module triggers the time-series graph neural network model used by the risk evolution path prediction module to undergo a complete retraining process based on the updated graph structure that better reflects the current reality and the latest data. Through this closed-loop process, the system described in this invention can achieve adaptive response to changes in the external environment, ensuring the continuous output of high-quality risk assessment results.
[0064] This invention also provides a big data-based method for assessing bank credit risk. By introducing a deep integration of causal inference and temporal graph neural networks, and constructing an adaptive closed-loop correction mechanism, it brings significant beneficial effects. This includes: Step 1: Construction of a basic financial causal knowledge graph. This step is completed by the basic financial causal knowledge graph construction module.
[0065] (1) Structure learning: Continuous optimization algorithms such as NOTEARS are used to analyze the conditional independence relationships between financial variables in historical credit data and generate an initial causal structure. At the same time, prior knowledge in the financial field (such as "customer age" is the cause of "income level", and the reverse is not true) is integrated as constraints to orient the edges of the causal graph.
[0066] (2) Parameter learning and counterfactual verification: Maximum likelihood estimation (MLE) is used to quantify the strength of causal relationships. Counterfactual analysis y(x←x′) is performed on the critical path. For example, to evaluate "If a customer's average daily deposit balance increases by 10%, how will their credit score change?", after passing the significance test (α=0.05), stable causal relationships are stored in the graph.
[0067] Step 2: Generation of the Entity's Dynamic Risk Evolution Map. This step is completed by the entity's dynamic risk evolution map generation module. It maps the time-series data of the specified credit entity (such as transaction records for the most recent 30 / 60 / 90 days, APP behavior, etc.) onto the base map, generating a highly personalized, time-varying dynamic map. The node types of this map... Includes {customers, accounts, guarantors, enterprises}, edge types This includes {transfer of funds, joint debt, guarantee, and joint control}.
[0068] Step 3: Risk Evolution Path Prediction. This step is completed by the risk evolution path prediction module. A temporal graphical neural network (Temporal GNN) model (number of layers L∈, dimension d∈) is used to learn the evolution pattern of the dynamic graph of the subject and predict its future risk state transition probability matrix.
[0069] Step 4: Adaptive correction. This step is completed by the adaptive correction module and is the core closed loop of the system.
[0070] (1) Drift detection: The system periodically relearns a window-period causal graph on new data and compares it with the base graph. At the same time, it continuously monitors the overall stability index (PSI) and concept drift index (CSI) of the input features and the model output.
[0071] (2) Triggering and Response: Set clear thresholds and strategies. For example, when PSI>0.2 or CSI>0.1, trigger the update process. The response is divided into three levels: L1 (cold start retraining), L2 (threshold recalibration), and L3 (feature subset replacement).
[0072] (3) Gray release: The updated model is first allocated 10% of the traffic for a 14-day A / B test. After passing the significance test (p<0.05), it is then fully switched to ensure the stability and effectiveness of the update.
[0073] Experimental evaluation and operational results: To demonstrate the significant technical effects of this invention, we conducted a multi-dimensional evaluation and compared it with a baseline (scoring card / GBDT).
[0074] A comparison table of model evaluation metrics is shown in Table 1 below: ; This invention can serve as the core foundational patent for the "Intelligent Credit Audit Assistant" project. Through extensions, the system can receive "audit material sets" (such as due diligence reports) as input; it can also link with the built-in "audit field label / template library"; ultimately outputting an "intelligent audit scoring report" that includes field evidence backlinks and responsibility chain annotations, seamlessly integrating with the project report engine.
[0075] Compared to traditional risk assessment models that rely on static, correlational features, the system and method of this invention can reveal deep-seated causal transmission paths between financial variables. By constructing a basic financial causal knowledge graph, the system no longer treats numerous variables as a flat set of features, but rather organizes them into a structured network with inherent logic. This makes the risk assessment result not just a simple probability value, but also allows for tracing its causes, thereby greatly improving the interpretability of the assessment model and helping risk managers understand and trust the model's output.
[0076] Furthermore, by mapping the time-series data of credit entities onto a causal graph and using a time-series graph neural network for evolutionary path prediction, this invention transforms risk assessment from a static, snapshot-like judgment at a static point in time into a dynamic, forward-looking, video-like inference. The system can capture the subtle trajectory and development trend of risk status over time, thereby identifying potential risks that are difficult for traditional models to detect, and achieving more forward-looking risk warnings.
[0077] More importantly, this invention endows the entire system with the ability to self-evolve and adapt to environmental changes by establishing a causal structure drift monitoring module and an adaptive correction module. This is achieved by periodically quantifying changes in real-world causal logic, i.e., calculating the DriftScore and comparing it with a threshold. By comparison, the system can automatically identify the risk of model failure caused by macroeconomic fluctuations, regulatory policy adjustments, or changes in user behavior patterns. Once a significant drift is identified, the system can automatically update its underlying causal knowledge graph and trigger model retraining, thereby ensuring the robustness of the system and the continued effectiveness of the evaluation results during long-term operation, overcoming the inherent defect of traditional models that are trained once and used for a long time until they fail.
[0078] System Architecture and Data Pipeline (SLA): This system adopts a unified stream and batch architecture. Data is collected in real-time from the source system via CDC / Kafka, and real-time feature computation is performed by Flink, ensuring an end-to-end latency P95 ≤ 15 minutes. Feature and graph data are stored in the Neo4j / Tiger Graph graph database. The model is provided through g RPC / REST API, with a system SLA target of ≥ 99.9%. The entire data pipeline has comprehensive lineage tracing and quality monitoring, with a null value rate of ≤ 1% for key fields and a consistency check pass rate of ≥ 99.5%.
[0079] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A big data-based method for assessing bank credit risk, characterized by: Includes the following steps: S1. Based on historical credit data, construct a basic financial causal knowledge graph that represents the causal relationships between multiple financial variables; S2. Map the time series data of the specified credit subject to the basic financial causal knowledge graph to generate a dynamic risk evolution graph of the subject that represents the dynamic changes in the financial status of the credit subject. S3. By using the dynamic risk evolution map of the main body, a predictive model is trained to generate a prediction model for predicting the risk evolution path of the credit entity within a future time window; S4. Continuously monitor structural changes in causal relationships and revise the basic financial causal knowledge graph and prediction model based on the monitoring results.
2. The bank credit risk assessment method based on big data according to claim 1, characterized in that: S1 includes: A causal discovery algorithm is used to perform a series of conditional independence tests on historical credit data to identify direct causal relationships between variables and generate an initial causal structure. Using prior knowledge as a hard constraint, the initial causal structure is calibrated and the edges are oriented to generate a directed acyclic graph, namely the basic financial causal knowledge graph.
3. The bank credit risk assessment method based on big data according to claim 1, characterized in that: S2 includes: acquiring time-series data of a specified credit entity over consecutive time steps in the past, and instantiating the specific data values at each time step as node attributes onto the topological structure of the basic financial causal knowledge graph to form a time-series graph sequence, namely, the entity's dynamic risk evolution graph; each graph in the time-series graph sequence represents a snapshot of the credit entity's financial state at the corresponding moment, and the time-series graph sequence as a whole constitutes the entity's dynamic risk evolution graph used to fully describe the dynamic changes in the financial state of the credit entity.
4. The bank credit risk assessment method based on big data according to claim 1, characterized in that: In S3, the time series neural network model is trained by the subject dynamic risk evolution map. During training, the causal chain propagation path related to changes in financial variables and the impact path on downstream variables are captured through the message passing mechanism, and the parameters of the mechanism are updated to output the risk evolution path. The risk evolution path output by the time-series graph neural network model is a risk state transition probability matrix, which is used to characterize the probability that the risk level of a credit entity will change between preset states within a preset time window in the future.
5. The bank credit risk assessment method based on big data according to claim 4, characterized in that: In S3, the node state update process in the time-series graph neural network model is represented as follows: , The state vector of node v at time t represents a mathematical abstraction of the risk characteristics; UPDATE is the state update function; AGGREGATE is the neighbor information aggregation function, used to gather the state information of all direct cause nodes at the previous time step; Let v be the external input feature of node v at time t.
6. The bank credit risk assessment method based on big data according to claim 1, characterized in that: In S4, S1 is continuously re-executed in the latest rolling time window data to generate the causal graph of the current time window period and calculate the difference value between it and the currently used basic financial causal knowledge graph. When the difference value exceeds the preset threshold, a correction operation is triggered. The correction operations include: correcting the current basic financial causal knowledge graph based on the new causal relationships in the causal graph of the current time window; The predictive model is then triggered to be retrained on the updated underlying financial causal knowledge graph and the latest data.
7. The bank credit risk assessment method based on big data according to claim 6, characterized in that: The method for calculating the difference between the causal graph of the current time window and the currently used basic financial causal knowledge graph is as follows: A structured comparison is made between the causal graph of the current time window and the basic financial causal knowledge graph currently in use; The causal outcome offset between the causal graph of the current time window and the currently used basic financial causal knowledge graph is calculated using structured Hamming distance. ,in, For the calculated drift score, For the structured Hamming distance calculation function, It is a causal knowledge graph for basic finance. This is a causal graph of the window period.
8. A bank credit risk assessment system based on big data, characterized by: include: The basic financial causal knowledge graph construction module is used to construct a basic financial causal knowledge graph representing the causal relationships between multiple financial variables based on historical credit data. The entity dynamic risk evolution graph generation module is used to map the time series data of a specified credit entity to a basic financial causal knowledge graph to generate an entity dynamic risk evolution graph that represents the dynamic changes in the financial status of the credit entity. The risk evolution path prediction module is used to train and generate a prediction model for predicting the risk evolution path of a credit entity within a future time window by using the entity's dynamic risk evolution map. The adaptive correction module is used to continuously monitor structural changes in causal relationships and correct the basic financial causal knowledge graph and prediction model based on the monitoring results.
9. A bank credit risk assessment system based on big data according to claim 1, characterized in that: The basic financial causal knowledge graph construction module uses a pre-set causal discovery algorithm to perform a series of conditional independence tests on historical credit data to identify direct causal relationships between variables and generate an initial causal structure. Using prior knowledge as a hard constraint, the module calibrates the initial causal structure and orients the edges to generate a directed acyclic graph, i.e., the basic financial causal knowledge graph. The entity dynamic risk evolution graph generation module is used to obtain time series data of a specified credit entity within a continuous time step in the past. The specific data value at each time step is used as a node attribute and instantiated onto the topological structure of the basic financial causal knowledge graph to form a time series graph sequence, i.e., the entity dynamic risk evolution graph. The risk evolution path prediction module is used to train the time series neural network model through the subject's dynamic risk evolution map. During training, it captures the causal chain propagation path related to changes in financial variables and the impact path on downstream variables through a message passing mechanism, updates the parameters of the mechanism, and outputs the risk evolution path. The risk evolution path output by the time-series graph neural network model is a risk state transition probability matrix, which is used to characterize the probability that the risk level of a credit entity will change between preset states within a preset time window in the future. The adaptive correction module is used to continuously regenerate the causal graph of the current time window period in the latest rolling time window data, and calculate the difference value between it and the currently used basic financial causal knowledge graph. When the difference value exceeds a preset threshold, a correction operation is triggered. The correction operations include: correcting the current basic financial causal knowledge graph based on the new causal relationships in the causal graph of the current time window; The predictive model is then triggered to be retrained on the updated underlying financial causal knowledge graph and the latest data.
10. A bank credit risk assessment device based on big data, characterized by: include: A processor and a memory, the memory storing a computer program executable by the processor, wherein the processor, when executing the computer program, implements the method of any one of claims 1-7.