A Smart Identification Method for Systemic Financial Risk Based on Causal Networks and Multi-Model Optimization
By employing causal networks and multi-model optimization, the challenge of identifying dynamic interactions among institutions within the financial system was solved, enabling highly accurate prediction and interpretable analysis of systemic risks, and improving the model's transferability and prediction accuracy.
Patent Information
- Application Number
- CN202610170195.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-02
Smart Images
Figure CN122134454A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of financial technology and complex system risk management. Specifically, it relates to a systemic financial risk identification method that integrates a causal relationship network identification mechanism, a multi-model prediction fusion algorithm, and an intelligent optimization strategy. It can be widely applied to the construction of risk monitoring, regulatory technology, and systemic risk early warning platforms for financial institutions. Background Technology
[0002] Current mainstream financial risk measurement tools, such as Value at Risk (VaR), volatility analysis, and multi-factor models, primarily focus on the static risk exposure assessment of individual financial institutions. These methods are largely based on the assumption of marginal characteristics of return distribution, neglecting the dynamic interactions and interconnected structures between institutions within the financial system, making it difficult to identify cross-institutional risk contagion chains. For example, traditional VaR models only focus on the extreme losses of a single institution at a certain confidence level, without considering the impact of that institution's extreme state on other entities in the system, resulting in insufficient early warning in systemic risk scenarios.
[0003] Furthermore, in events such as the severe financial market volatility triggered by the global subprime mortgage crisis, risks exhibited a cascading propagation characteristic across markets and institutions. Numerous studies have shown that financial institutions have a high degree of risk interdependence and transmission paths, and the outbreak of systemic risk is often triggered by a few nodes, spreading rapidly through the financial network. However, traditional methods are generally based on the assumption of independence or static covariance structure, lacking causal identification mechanisms, and making it difficult to accurately capture the key paths and pivotal institutions in the risk contagion chain.
[0004] Furthermore, current methods for identifying the risk status of enterprises or financial instruments mostly employ fixed threshold classifications or static machine learning classifiers (such as SVM and decision trees), ignoring the heterogeneity of risk indicators across time and populations. In real-world markets, risk status boundaries often dynamically adjust with macroeconomic cycle fluctuations, changes in asset types, and other factors. Traditional static classification standards struggle to handle nonlinear transitions and structural drifts under complex environments. In addition, model parameters are often manually set or tuned empirically, lacking a systematic optimal search mechanism, which limits the model's transferability and predictive accuracy.
[0005] Therefore, there is an urgent need to develop a systematic risk identification method that integrates structural identification and intelligent prediction. This method should have the following core capabilities: (1) It can identify the dynamic causal structure between institutions within the financial system; (2) It can adaptively learn and classify the distribution of risk status; (3) It can integrate multiple model structures to capture the evolution law of time series risk; (4) It can use intelligent optimization algorithms to automatically adjust the parameters of the model and improve the performance ceiling. Summary of the Invention
[0006] To address the aforementioned issues, this invention provides an intelligent identification method for systemic financial risk based on causal networks and multi-model optimization, aiming to overcome the shortcomings of existing financial risk measurement methods in system structure identification, state classification, and prediction accuracy. This method integrates the Conditional Value at Risk (CoVaR) mechanism from financial engineering, Granger causal network modeling from statistical causal inference, Long Short-Term Memory (LSTM) neural networks and Gradient Boosting Tree (XGBoost) models from machine learning, and Particle Swarm Optimization (PSO) algorithm from evolutionary computation, forming a systematic closed-loop process of structure identification—label construction—intelligent prediction—global optimization.
[0007] Specifically, this method first calculates the CoVaR index at a specific confidence level based on the company's historical return and market return series, which measures individual vulnerability under extreme overall market conditions. To avoid manually setting classification thresholds, a Gaussian mixture model (GMM) is used for unsupervised clustering of CoVaR values, enabling automatic identification and label construction of high / low risk status for companies. This method can capture the heterogeneous structure in the index distribution, avoiding model bias caused by subjective partitioning.
[0008] Subsequently, the inventors constructed an inter-firm risk propagation network based on Granger causality. By combining a vector autoregression (VAR) model with lag order optimization criteria (AIC / BIC), they automatically identified directional contagion paths between institutions and extracted network topology features (such as in-degree and centrality) as input variables to enhance the model's ability to characterize structural risks.
[0009] In terms of predictive modeling, this invention employs a dual-channel fusion mechanism: on the one hand, it uses an LSTM model to process the multi-period time series characteristics of enterprises, learning the dynamic laws governing the evolution of risk status over time; on the other hand, it introduces an XGBoost model to model the nonlinear interaction relationships between cross-sectional variables of enterprises, and then weights and fuses the output probabilities of the two models to generate the final systemic risk score. To improve the model's generalization ability and accuracy, the inventors further introduce the PSO algorithm to perform global search optimization on model structural parameters (such as the number of LSTM units, dropout rate, XGBoost depth, etc.), overcoming the local optima and computational bottleneck problems of traditional grid parameter tuning.
[0010] Ultimately, the system outputs multi-period risk scores, enterprise risk index rankings, causal network graphs, and interpretability indicators (such as SHAP values). It is deployed as a modular service platform through a standardized RESTful API, suitable for application scenarios such as financial supervision, bank risk control, asset management, and enterprise credit risk early warning.
[0011] The specific plan is as follows:
[0012] A method for intelligent identification of systemic financial risk based on causal networks and multi-model optimization includes the following steps:
[0013] Step S1: Read and clean financial market and corporate financial data, and construct a quarterly logarithmic return feature series;
[0014] Step S2: Calculate the company-level risk exposure value based on the CoVaR mechanism, and binarize the risk status label of the enterprise using Gaussian Mixture Model (GMM) clustering.
[0015] Step S3: Construct a causal directed network among financial institutions using the Granger causality test method, and extract lag parameters, significance coefficients, and structural measures;
[0016] Step S4: Construct a risk prediction model based on LSTM neural network and XGBoost tree model, and jointly optimize the model parameters and structure using particle swarm optimization algorithm;
[0017] Step S5: Weight and fuse the output results of multiple models to generate a systematic risk score and state classification results, and deploy the interpretability and visualization service module to realize online application.
[0018] Furthermore, in step S2, the CoVaR value is defined by the following formula:
[0019]
[0020] in, For the company The logarithmic return series, To set the confidence level, CoVaR was calculated based on historical simulation.
[0021] Furthermore, in step S2, the Gaussian Mixture Model (GMM) is used for risk status identification. The GMM model is defined as follows:
[0022]
[0023] in, In this embodiment, the number of components in the Gaussian mixture model is defined as high-risk clusters and low-risk clusters. and through the mean Determine the risk category; This is an index of the mixture components in a Gaussian mixture model, used to distinguish different cluster components; it does not represent a risk level itself. and The first The mean and variance of a Gaussian distribution.
[0024] Furthermore, in step S3, the causal network construction is based on the Granger causality test, with the null hypothesis being:
[0025]
[0026] Among them, the null hypothesis For: Company All Lagged variables on the company The current rate of return has no significant impact, i.e., all lagged coefficients. ,in ;
[0027] The significance of the test was performed using the following F-statistic:
[0028]
[0029] in, , These are the sums of squared residuals for the restricted and unrestricted models, respectively. This represents the number of samples.
[0030] Furthermore, in step S3, the lag order is determined by minimizing the AIC and BIC criteria, where AIC is defined as follows:
[0031]
[0032] in For the number of model parameters, For the sample size, To predict the residual variance.
[0033] Furthermore, in step S4, the state update of the LSTM neural network model includes the following calculation process:
[0034] Cell status update:
[0035]
[0036] Output status:
[0037]
[0038] in , , These are the activation vectors for the forget gate, input gate, and output gate, respectively.
[0039] Furthermore, in step S4, the prediction function of the XGBoost tree model is defined as:
[0040]
[0041] in For the regression tree function space, the loss function includes a penalty term for the structural complexity of each tree; The total number of trees, Indicates the first The input variable vector for each sample.
[0042] Furthermore, in step S4, the particle swarm optimization (PSO) algorithm is used to jointly tune the prediction model. The optimization process is as follows: Let the first step be... In the nth iteration, the 1st Each particle represents a set of hyperparameter configurations, and its current position is... The speed is The best position reached by this particle in history is The known optimal position in the entire group is ;
[0043] Speed updates:
[0044]
[0045] Location update:
[0046]
[0047] in Inertia factor , The learning coefficient, .
[0048] Furthermore, in step S5, the final output of the fused prediction probability is:
[0049]
[0050] in The fusion coefficient is adjustable. The first given by the LSTM model Risk probability of the sample (company-quarterly portfolio) The risk probability output by the XGBoost model;
[0051] Constructing a risk index:
[0052]
[0053] Used to uniformly quantify and rank the risk levels of all companies in the same quarter;
[0054] The model prediction results are interpreted using the SHAP method and satisfy the following form:
[0055]
[0056] in, For the model to sample The predicted risk probability; The global baseline prediction of the model represents the average output when there is no input information. Representation of features The marginal contribution of the prediction results to the explanation.
[0057] Furthermore, the method of the present invention is deployed in a modular microservice architecture, including at least: a data processing module, a tag construction module, a causal graph construction module, a predictive modeling module, a parameter optimization module, an interpretation and display module, and a predictive service interface module, supporting RESTful API format request returns.
[0058] The beneficial effects of this invention include:
[0059] (1) It can achieve structural identification and path-level tracking of systemic financial risks;
[0060] (2) Provides automated and highly accurate risk state prediction capabilities;
[0061] (3) Support interpretability analysis to enhance model transparency and regulatory compliance;
[0062] (4) Provide a general deployment solution and have the ability to promote and apply it in multiple types of financial data. Attached Figure Description
[0063] Figure 1 This diagram illustrates the overall framework of the systematic risk intelligent identification method based on causal networks and multi-model optimization, as presented in this invention. The diagram shows the overall process of data input, causal network construction, multi-model prediction, optimization and fusion, and risk index output.
[0064] Figure 2 This is a detailed flowchart of an embodiment of the present invention. The diagram further illustrates that, within the overall framework, multi-model prediction is performed by combining LSTM and XGBoost models, hyperparameter optimization is achieved through the PSO algorithm, and interpretability analysis is implemented using the SHAP module, ultimately outputting risk identification results.
[0065] Figure 3 This diagram compares the workflow of traditional methods. It illustrates the general process of existing statistical modeling methods, which only includes data collection, preprocessing, model training, and prediction, lacking the causal network and multi-model optimization mechanism proposed in this invention.
[0066] Figure 4This is a schematic diagram of the loss curve during the training process of the method of the present invention. The figure shows that the loss of the training set and the validation set gradually decreases with the number of iterations, verifying that the model has good convergence.
[0067] Figure 5 This is a schematic diagram showing the AUC change curve of the method of the present invention during the training process. The figure shows that as iterations proceed, the AUC of both the training set and the validation set show an upward trend, indicating that the model performance continues to improve.
[0068] Figure 6 This is a schematic diagram of the ROC curve of the method of the present invention on the test set. The figure shows that the method of the present invention has a high classification accuracy, with an AUC value of 0.98.
[0069] Figure 7 This is a schematic diagram of the precision-recall curves of the method of this invention under ten-fold cross-validation. The figure illustrates the robustness and generalization ability of the model under different folds.
[0070] Figure 8 This is a schematic diagram of the final precision-recall curve of the method of the present invention on the overall test set. This figure further demonstrates the effectiveness of the method of the present invention in intelligent identification of systemic financial risks. Detailed Implementation
[0071] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0072] As shown in the figure, this invention provides a method for intelligent identification of systemic financial risk based on causal networks and multi-model optimization, characterized by the following steps:
[0073] S1. Data Loading & Preprocessing
[0074] This step aims to standardize the input data format, remove outlier data, and construct standardized input variables, laying the foundation for subsequent causal identification and model building.
[0075] Step S1.1: Data Reading and Date Format Standardization
[0076] In the data preprocessing stage, this invention first reads financial data and market price data from the quarterly level of enterprises, in a standardized CSV file format. Each record contains three main fields: a time field, a company code field, and several financial or price indicator fields. The time field, named "Date," identifies the accounting period corresponding to the record; the company code field, named "ID," uniquely identifies each enterprise; and the remaining fields represent the company's relevant financial or market variables for that quarter.
[0077] To ensure the integrity and validity of the time series data in subsequent modeling, this invention first performs unified format identification and standardization on the time field, converting the original time string into a parsable standard timestamp format. During this process, any records with format errors or parsing failures in the time field, such as null values, invalid dates, or garbled content, are considered abnormal records. These abnormal records are automatically identified and removed, ensuring that all data ultimately used in the calculation has a valid and comparable time identifier.
[0078] Through the above steps, the system establishes a high-quality time series dataset with time as the main axis and companies as the objects, laying a solid data foundation for subsequent return calculation, risk identification, and sequence modeling.
[0079] Step S1.2: Calculation of Logarithmic Returns on Price Data
[0080] In this invention, asset price time series data from financial markets are used. Calculate the logarithmic return over continuous time. This metric measures the relative change of an asset between two adjacent points in time, and is defined as follows:
[0081]
[0082] in, Indicates the time of an asset The price at that time This indicates the price of the asset at the previous point in time. This represents the natural logarithm function.
[0083] The calculation of logarithmic returns helps to transform price changes into standardized risk factor inputs, making the return series more stable and comparable in statistical modeling. This indicator is not only applicable to traditional volatility analysis and covariance matrix construction, but is also widely used in financial risk measurement models, such as the estimation of VaR (Value at Risk) and CoVaR (Conditional Value at Risk).
[0084] In practice, logarithmic returns are calculated at the company level in chronological order, and this transformation is uniformly applied to all price-related indicators to ensure that the time dynamics of risk factors can accurately reflect the changing trends of asset values.
[0085] Step S1.3: Remove invalid observations and outliers
[0086] It should be noted that in actual financial market data, some asset prices may exhibit outliers less than or equal to zero, such as due to trading suspensions, unusual quotes, or missing data. Such price data lacks economic significance and cannot be used for efficient calculation of logarithmic returns because the logarithmic function is undefined in the non-positive domain.
[0087] Therefore, this invention pre-screens all asset price variables before calculating the logarithmic return. If the price value at a certain point in time satisfies... If the value is less than or equal to zero, the record is considered an invalid observation and is removed from the return calculation sample. This removal rule applies to all companies and all quarterly data sets to ensure that the final return series used for modeling originates from valid observations with positive prices.
[0088] This cleaning process can effectively eliminate the accumulation of errors caused by data anomalies, providing high-quality raw input for subsequent risk modeling and causal network identification.
[0089] Step S1.4: Data Normalization
[0090] To eliminate the influence of differences in measurement units among different indicators and ensure that all types of variables participate in model training and analysis on the same numerical scale, this invention uses standardization and normalization methods to uniformly process the original input data.
[0091] Specifically, for financial and market indicator data with significant differences in magnitude or units, the following two types of data transformation methods are used:
[0092] (1) Z-score standardization (standard deviation transformation): This method transforms the original feature variable into a standard normal distribution centered at 0 with a standard deviation of 1 by subtracting the sample mean and dividing by the standard deviation. The transformation formula is as follows:
[0093]
[0094] in, Represents the original variable. The sample mean. Let be the sample standard deviation. This method is suitable for scenarios where the original data roughly follows a normal distribution.
[0095] (2) Min-Max Normalization (Interval Scaling): This method linearly maps data to a specified interval proportionally, typically as follows: The transformation formula is as follows:
[0096]
[0097] in, and These are the minimum and maximum values of the original variable, respectively. This method is suitable for indicators that do not follow a normal distribution but require a uniform numerical range, and it works particularly well when the input needs to meet specific limits (such as the input range of a neural network activation function).
[0098] In the implementation of this invention, the most suitable standardization method is selected for different types of indicators (such as ratios, totals, structures, etc.) so that each variable is modeled on the same scale, thereby improving the convergence efficiency and prediction stability of the model.
[0099] Step S1.5: Handling Missing Values
[0100] In practical financial data processing, there are often cases where some indicators have missing values. For example, a company may not have disclosed specific financial items in a certain quarter, or the records may be incomplete due to differences in data acquisition channels. To ensure the completeness of variables and the reliability of predictions during the modeling process, this invention introduces a missing value imputation method based on the K-Nearest Neighbors (KNN) principle.
[0101] Specifically, let the original feature matrix be... ,in Indicates the number of samples. This indicates the number of features. For any cell containing a missing value... This invention finds the sample that is most similar across the remaining feature dimensions. Neighboring samples The mean is then used to interpolate the feature using the known values of neighboring samples. The interpolation formula is as follows:
[0102]
[0103] in, This is the estimated value. Indicates the observed value in the neighboring samples. This represents the number of nearest neighbors. In this invention, Generally, a value of 5 to 10 is used, but the specific value can be dynamically adjusted according to the sample density and variable characteristics.
[0104] Furthermore, variables with a high missing value ratio—that is, features with more than 40% missing values in all samples—will be directly removed. This selection rule aims to prevent variables with severely missing information from interfering with the model training process, thereby ensuring that the features ultimately used in modeling have sufficient data foundation and representativeness.
[0105] By employing the aforementioned missing value handling strategy, this invention effectively improves data integrity and model stability while preserving the original information to the maximum extent, providing high-quality input for subsequent risk prediction models.
[0106] This invention introduces the concept of Conditional Value at Risk (CoVaR) into the risk status identification module as an important indicator for identifying the vulnerability of a single financial institution in the context of systemic risk. CoVaR is a systemic financial risk measurement framework proposed by Adrian and Brunnermeier in 2016, used to measure the maximum loss that another institution or the entire financial system may suffer when a financial institution is in its own extreme risk state.
[0107] set up Indicates the first Family organization in time The rate of return, This represents the rate of return of the entire financial system at that moment. Therefore, CoVaR is defined as follows:
[0108]
[0109] in, This indicates the confidence level, typically set at 5%. This formula expresses the confidence level within the company. When in its own VaR risk state, the system return The probability of being less than or equal to a certain threshold is This threshold is CoVaR.
[0110] In a specific implementation of this invention, the company's rate of return will be... Compared with market returns Treating them as two random variables in a joint distribution, a nonparametric estimation is performed using historical simulation. That is, for the company... exist Set of return observations at each time section:
[0111]
[0112] Its CoVaR value is defined as the sequence at a confidence level. The percentile value below is specifically expressed as:
[0113]
[0114] In this way, CoVaR can directly quantify the vulnerability level of each company under extreme market conditions, providing a basis for subsequent risk labeling and systemic risk identification. Compared with traditional VaR models, this method focuses more on the "contribution risk" of individual institutions in systemic events, and has stronger systemic explanatory power and practical adaptability.
[0115] Step S2.2: GMM cluster analysis of CoVaR distribution
[0116] To further explore the risk state structure implied by CoVaR values, this invention introduces a Gaussian Mixture Model (GMM) to perform unsupervised clustering analysis on the CoVaR indices of enterprises. This method aims to transform the originally continuously distributed CoVaR risk factors into discrete risk labels that can be used for classification and prediction, thereby achieving structural identification of systemic risk states.
[0117] Suppose that the estimated CoVaR value for a certain company in each period is [value missing]. The overall sample set is:
[0118]
[0119] Assuming the CoVaR sample set follows a bimodal Gaussian mixture distribution, meaning there are two potential sub-distributions corresponding to the "high-risk" and "low-risk" states respectively, its probability density function can be expressed as:
[0120]
[0121] in, , representing the weights of the two mixture components; and The first The mean and variance of a Gaussian distribution.
[0122] In practice, this invention first standardizes all CoVaR values to eliminate the influence of dimensions and distribution skewness. Then, a Gaussian Mixture Model (GMM) is used to fit the values, identifying the two most representative cluster centers. After clustering, the risk state is defined based on the mean of each cluster.
[0123] If a certain CoVaR value Assigned to a school with a smaller mean In the clustering, if a node is classified as "high-risk," its corresponding label is set to 1; otherwise, it is considered "low-risk," and its corresponding label is set to 0. This classification logic can be formally represented as follows:
[0124]
[0125] Compared to traditional classification methods based on fixed thresholds, this method has stronger adaptability and distribution adaptability, and can automatically adjust the risk identification boundary according to different market environments, thereby improving the robustness and rationality of label construction.
[0126] The generated binary risk labels will serve as target variables for subsequent machine learning prediction models, providing clear and interpretable output basis for modeling systemic risk states.
[0127] Step S2.3: Risk Label Time Series Output and Visualization
[0128] After completing the clustering identification of CoVaR indicators based on Gaussian Mixture Model (GMM), this invention further extends the static clustering results into a label sequence with a time structure. Specifically, for each quarter... For each company (iii), construct a binary risk label variable. This is used to characterize whether the company is in a systemically high-risk state during the quarter. The marked variable is defined as follows: when the company... In the quarter When identified as being in a state of high systemic risk, ;otherwise In this context, State 1 indicates that the company was identified as belonging to a high-risk cluster during the quarter, while State 0 indicates that it is in a relatively low-risk state.
[0129] In this way, the present invention constructs a complete set of time series data on risk states:
[0130]
[0131] This sequence covers the entire Home institutions and The quarterly rate is an important target variable in the subsequent machine learning modeling stage, used to train the model to identify future systemic risk states.
[0132] To visualize and verify the distribution pattern of the clustering results, this invention generates a distribution map of all CoVaR values for each quarter and overlays a kernel density estimation curve to determine whether the current market is in a unimodal, bimodal, or skewed state. This visualization can not only be used to verify the clustering effect but also provide regulatory and risk management personnel with an intuitive basis for risk perception.
[0133] In summary, this step effectively transforms the original continuous financial risk indicators into label variables that can be used for classification and prediction, and uses the GMM model to achieve automatic determination of risk status and structure identification, significantly improving the accuracy and interpretability of systemic risk prediction modeling.
[0134] S3. Causal Network Construction and Granger Relationship Identification Mechanism
[0135] In identifying risk transmission relationships between financial institutions, this invention introduces the Granger causality test mechanism. This method, based on time series analysis principles, determines whether the historical information of one variable can effectively predict the current changes of another variable. If the historical returns of one institution statistically have explanatory power for the returns of another institution, a directional risk influence relationship can be considered to exist between the two, thus constructing directed edges in the financial network.
[0136] Step S3.1: Variable Definition and Causality Testing Framework
[0137] In identifying risk transmission relationships between financial institutions, this invention introduces the Granger causality test mechanism. This method, based on time series analysis principles, determines whether the historical information of one variable can effectively predict the current changes of another variable. If the historical returns of one institution statistically have explanatory power for the returns of another institution, a directional risk influence relationship can be considered to exist between the two, thus constructing directed edges in the financial network.
[0138] Specifically, let the first Family of companies in time The logarithmic return is , No. The corresponding sequence of the companies is If the variable Historical information is useful for prediction If it has significant statistical explanatory power, it is defined as "company". Granger caused the company , is represented by the following symbol:
[0139]
[0140] To test this causal relationship, two regression models need to be constructed and hypothesis testing performed. The null hypothesis is (…). ) is: Company All Lagged variables on the company The current rate of return has no significant impact, i.e., all lagged coefficients. ,in The two sets of models established are as follows:
[0141] (1) Constraint Model (considering only companies) Its own historical gains):
[0142]
[0143] (2) Unrestricted model (adding a company to the original model) (lagged variables)
[0144]
[0145] Next, the two models are compared using the F-statistic to examine whether the addition of the lag term significantly improves the model's explanatory power. The formula for calculating the F-statistic is as follows:
[0146]
[0147] in, This represents the sum of squared residuals of the constrained model. This represents the sum of squared residuals of the unrestricted model. The sample size is given. If the significance level corresponding to the F-value is lower than the preset threshold (e.g., 0.05), the null hypothesis is rejected, and the company is considered to be... For the company Granger causality exists.
[0148] Through this testing mechanism, the present invention can identify potential dynamic risk impact paths among enterprises, providing basic data support for the subsequent construction of causal network graphs.
[0149] The identification of Granger causality depends heavily on the chosen lag order. This refers to the extent to which historical information should be included in the explanatory variables. If this parameter is set too small, it may lead to the omission of true causal effects; conversely, if it is set too large, it may introduce noise and reduce the robustness of the model. Therefore, to ensure the accuracy and scientific nature of the causal test, this invention introduces two mainstream lag-order optimization criteria: AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion).
[0150] The common goal of these two criteria is to achieve an optimal structural trade-off by suppressing excessive complexity while maintaining the model's explanatory power. They are defined as follows:
[0151]
[0152]
[0153] in, This indicates that the lag order is The residual variance estimated in the model, The number of model parameters, The total number of samples. Both are based on residual variance, but BIC penalizes the number of parameters more strictly, making it suitable for model comparison scenarios with large sample sizes.
[0154] In the implementation of this invention, multiple candidate lag orders are model-fitted, and their AIC and BIC values are calculated respectively. Finally, the lag order that performs best or second best in both criteria is selected as the final causal test order, thereby improving the robustness of causal relationship identification.
[0155] When encountering situations with small sample sizes, discontinuous data from some enterprises, or abnormal sequence fluctuations, if neither AIC nor BIC provides effective convergence results, this invention sets a default lag order of [missing information]. As a safety net strategy, it balances recognition ability and stability.
[0156] This mechanism ensures the adaptability of parameter settings during causal network modeling, effectively improving the scientific rigor and credibility of the entire financial risk propagation network map.
[0157] Step S3.3: Causal Relationship Graph Construction and Weight Extraction
[0158] After completing the Granger causality test between each pair of companies, this invention further constructs a causal graph between financial institutions to visualize the risk propagation structure within the financial system. This graph is a directed weighted graph, where each edge represents a set of statistically significant causal relationships, the direction of the edge reflects the path of risk impact, and the weight of the edge quantifies the intensity of the impact.
[0159] Specifically, if a certain company's Perform the Granger test and find that its corresponding significance level satisfies That is, by using the commonly used 5% confidence threshold, the company is considered... For the company There is a statistically significant causal relationship. Based on this, a causal relationship is established within the graph structure, starting from the nodes... Pointing to node Directed edge .
[0160] At the same time, the system records the following information to characterize the edge's properties:
[0161] (1) Lag order : Indicates from the company Arrive at the company The optimal historical lag period;
[0162] (2) Significance level : Used to measure the statistical significance of Granger relationships;
[0163] (3) Edge weight Used to measure the strength of an impact; common forms include the negative logarithmic form. A larger value indicates a more significant causal relationship; alternatively, the corresponding F-statistic value can be used directly. The final constructed graph is denoted as: Where: the node set is This indicates all within the system. Family of enterprises; the set of directed edges is:
[0164]
[0165] Threshold represents a statistically significant causal path. Generally, it is taken as 0.05; for each edge The corresponding weight can be attached. This is used for subsequent propagation modeling or centrality analysis.
[0166] Through this step, the present invention integrates the original bilateral Granger test results, which were originally based on companies, into a systematic structural network, realizing the structural modeling of risk propagation paths within the system. This provides a solid foundation for subsequent identification of core risk nodes and construction of causal feature graph inputs.
[0167] Step S3.4: Graph Structure Analysis and Key Node Identification
[0168] After constructing the causal network map of systemic risk, this invention further conducts structural analysis to extract key nodes and propagation path characteristics with systemic importance. This analysis not only reveals the propagation patterns of risk in the financial system but also provides network structure input variables for subsequent deep learning models, thereby enhancing the explanatory power and structural sensitivity of systemic risk prediction.
[0169] First, by counting the number of incoming and outgoing edges at each node in the causal network, its role in risk transmission can be assessed:
[0170] Out-degree measures the extent to which a company influences other companies, i.e., the strength of a company as a "source of risk propagation".
[0171] In-degree measures the amount of influence a company receives from other firms, reflecting its sensitivity as a "receptor".
[0172] The mathematical expression for the out-degree is:
[0173]
[0174] For the indicator function, if the edge If it exists, the value is 1; otherwise, it is 0.
[0175] Secondly, by identifying the subgraph structure and connected components in the causal graph, local risk concentration areas can be further extracted. For example, certain nodes forming tightly connected subnetworks may correspond to systemic risk "clusters" within the same industry chain, region, or regulatory framework.
[0176] In measuring network influence, this invention introduces a typical graph theory centrality metric to quantify the importance of a firm within the system structure. One of the most representative metrics is the PageRank score, which is based on a random walk model to evaluate the structural weight of each node. The formula for calculating the PageRank value is as follows:
[0177]
[0178] in: Indicates the company The centrality score; This is the damping factor, which is usually set to 0.85;
[0179] This represents the total number of nodes in the graph; For all pointers The set of nodes; For nodes The out-degree, that is, the number of other companies it points to.
[0180] Furthermore, this invention can also introduce various structural indicators such as betweenness centrality and eigenvector centrality as inputs to corporate-level risk characteristics in subsequent machine learning models, thereby improving the sensitivity of predictions to structural risks.
[0181] The causal network constructed in this step not only depicts the risk transmission path between enterprises, but also provides structural features for subsequent deep learning models, helping to enhance the explanatory power of systemic risk prediction.
[0182] S4. Multi-model fusion prediction and LSTM / XGBoost / PSO collaborative structure
[0183] This step aims to intelligently predict the risk labels constructed above, using a fusion model structure: on the one hand, LSTM (Long Short-Term Memory) network is used to process the company's multi-period feature sequences to extract time dynamic information; on the other hand, XGBoost is used to extract nonlinear interaction information between variables; finally, particle swarm optimization (PSO) algorithm is used to globally search and optimize the model parameters and structure to achieve intelligent identification of systemic risks.
[0184] Step S4.1: Input sequence construction
[0185] To effectively capture the dynamic evolution of corporate risk status over time, this invention employs a sliding window mechanism to organize company-level data into a structured time-series input. By constructing a sequence of observations spanning multiple consecutive quarters, the model can learn intertemporal correlation patterns in risk status.
[0186] Specifically, let the time step be... That is, each sequence contains consecutive Feature information for the first quarter. For the first quarter A company, assuming it is in the first The feature vector for each quarter is ,in Let be the feature dimension. Then, taking this enterprise as the unit, the input sequence can be constructed in the following form:
[0187]
[0188] This input matrix represents the sequence from the first... quarter to the The continuous characteristic trajectory of the quarter, including the company's past... All information at each point in time is used to predict its risk status in the current quarter.
[0189] Accordingly, the first Quarterly risk status label The input sequence serves as the target variable. The labels here are determined by the high / low risk status identified by the aforementioned CoVaR-GMM module, where a value of 1 indicates that the company is in a systemically high-risk state during the quarter, and 0 indicates that its risk level is low.
[0190] When constructing the sample set, this invention extracts the input-output pairs corresponding to each time step sequentially using a sliding window method for the complete time series of each enterprise, forming a structurally complete supervised learning dataset. This approach not only enhances the temporal continuity of the samples but also provides a training foundation for deep learning models to capture trend changes and temporal dependencies.
[0191] Step S4.2: LSTM Deep Neural Network Modeling
[0192] This invention employs a Long Short-Term Memory Network (LSTM) to model the multi-quarter time series characteristics of a company during the risk prediction modeling stage. LSTM networks are a type of recurrent neural network architecture optimized for time series data, capable of capturing long-term dependencies and dynamic time series patterns, and are particularly suitable for handling lag effects and trend signals in the evolution of corporate risk states.
[0193] Suppose the input is a multi-quarter time series. ,in For time step, For the feature dimension, the state update mechanism of the LSTM unit includes the following steps:
[0194] Forget Gate: Controls the degree to which a network retains its previous state, defined as:
[0195]
[0196] Input Gate: Determines the strength of the current input information, defined as:
[0197]
[0198] The candidate state is generated as follows:
[0199]
[0200] Unit state update: Update memory state by combining the results of forgetting and input gates.
[0201]
[0202] Output Gate: Controls the generation of output results, defined as:
[0203]
[0204] The final hidden state is:
[0205]
[0206] After completing the state updates for all time steps, extract the hidden state of the last time point. This is then fed into a fully connected layer with a Sigmoid activation function to generate the systemic risk prediction probability for the current quarter:
[0207]
[0208] To optimize this model, this invention uses the standard binary classification cross-entropy as the loss function, which is defined as follows:
[0209]
[0210] This loss function achieves a balanced fit between samples with risk status labels of 0 and 1, and is suitable for class imbalance problems in financial risk identification.
[0211] Through the above modeling structure, the LSTM module of this invention can effectively extract intertemporal risk features, providing powerful sequence modeling capabilities for the intelligent identification of multi-quarter systemic risk states.
[0212] Step S4.3: XGBoost Model Construction and Training
[0213] To enhance the modeling capability of risk identification systems in terms of structural features, this invention introduces the eXtremeGradient Boosting (XGBoost) model as an additional risk prediction channel constructed in parallel with the LSTM model. XGBoost is an ensemble learning method based on gradient boosting mechanism, characterized by high efficiency, scalability, and robustness, making it particularly suitable for classification tasks involving structured data.
[0214] In this invention, the XGBoost model is specifically designed to handle non-time-series structural variables, such as static financial indicators, industry attributes, company size, etc. These features cannot be fully captured by time-series models such as LSTM, but they play a significant role in the judgment of risk status.
[0215] The basic form of this model is an ensemble of multiple regression trees, each tree... All come from the function space Its output is the sample Predicted probability:
[0216]
[0217] in, For the regression tree function space, the loss function includes a penalty term for the structural complexity of each tree; The total number of trees, Indicates the first The input variable vector for each sample.
[0218] The XGBoost objective function consists of two parts: a loss term for the sample prediction error and a structural penalty term for model complexity. The overall objective function is as follows:
[0219]
[0220] in, To determine the loss function between the predicted label and the true label, logarithmic loss is typically used; structure regularization term Then it is defined as:
[0221]
[0222] In the formula, This indicates the number of leaf nodes in the current tree. The leaf node score vector. and Hyperparameters are used to control model complexity and balance model fitting ability and generalization ability.
[0223] In actual training, this invention employs a target function based on binary cross-entropy, combined with sample imbalance adjustment, feature importance screening, and early stopping mechanisms. This enables the XGBoost model to not only achieve excellent prediction accuracy but also possess good interpretability and stability. After training, the model outputs the risk prediction probability for each company in each quarter, serving as a crucial input for subsequent fusion predictions.
[0224] Step S4.4: Automatic Adjustment Mechanism for Particle Swarm Optimization (PSO) Parameters
[0225] In the construction of deep learning and ensemble learning models, model performance is often highly sensitive to the setting of hyperparameters, such as the number of hidden units and the dropout ratio in LSTM networks, or the tree depth (max_depth) and learning rate (learning_rate) in XGBoost models. Manually tuning hyperparameters or using grid search is not only inefficient, but also makes it difficult to guarantee global optimization.
[0226] To achieve automated and intelligent joint parameter tuning, this invention introduces Particle Swarm Optimization (PSO) to perform unified search and collaborative optimization of multiple model structural parameters and training hyperparameters. PSO is a global optimization algorithm based on swarm intelligence that finds the optimal solution by simulating the movement of particles in the solution space.
[0227] Let the first In the nth iteration, the 1st Each particle represents a set of hyperparameter configurations, and its current position is... The speed is The best position reached by this particle in history is The known optimal position in the entire group is The particle state update formula is as follows:
[0228] Speed update formula:
[0229]
[0230] Position update formula:
[0231]
[0232] in, The inertia factor is used to control the degree to which the current velocity of a particle affects the next generation. , These are learning factors that regulate the degree to which particles depend on their own experience and the experience of the group, respectively. , It is a random number between 0 and 1, used to increase the randomness and diversity of the search.
[0233] In this invention, the optimization objective of PSO is to jointly minimize the prediction loss function of the two sub-models, that is:
[0234]
[0235] in, This represents the set of all parameters to be adjusted. and These are the loss functions (such as cross-entropy) for the LSTM and XGBoost models, respectively. To adjust the fusion coefficient of the two weights, it is usually set to 0.5 to maintain balance.
[0236] During algorithm iteration, the system continuously generates new parameter combinations to train the LSTM and XGBoost models, and updates particle positions and velocities based on prediction performance feedback until the maximum number of iterations is reached or the loss function converges. This mechanism significantly improves model parameter tuning efficiency and prediction performance while avoiding getting trapped in local optima. The introduction of PSO in this invention significantly enhances the overall optimization capability of the systemic risk identification model, achieving intelligent joint tuning throughout the entire process from feature selection and structural design to parameter configuration.
[0237] The LSTM–XGBoost–PSO fusion structure described above achieves integrated modeling of structure learnability, parameter adaptation, and multi-dimensional risk information, significantly improving the predictive performance and generalization ability of systemic risk.
[0238] S5. Risk Prediction Output and System Deployment Logic
[0239] This phase integrates the prediction results from the aforementioned multiple models, outputs standardized systemic risk scores and state prediction results, and constructs model interfaces, decision support graphical modules, and system deployment frameworks to serve the construction of risk management and early warning mechanisms for financial institutions.
[0240] Step S5.1: Multi-model prediction fusion and final risk probability output
[0241] To fully leverage the strengths of different models in predicting systemic financial risks, this invention introduces a multi-model prediction fusion mechanism to improve overall identification accuracy and stability. This mechanism combines the long-range dependency capture capabilities of the LSTM model in time series modeling with the nonlinear interaction mining advantages of the XGBoost model in static structured data, constructing a unified risk probability output.
[0242] Specifically, for each company in each quarter, two types of models output the corresponding risk prediction probabilities:
[0243] The first LSTM model gives Risk probability of a sample (company-quarterly portfolio);
[0244] : Risk probability output by the XGBoost model.
[0245] This invention employs a simple weighted fusion strategy to integrate the outputs of two types of models, and the final prediction probability is defined as:
[0246]
[0247] in, The model fusion weights are used to adjust the relative influence of the two models on the final result. These weights can be set based on historical validation set performance or determined through automatic parameter tuning mechanisms (such as particle swarm optimization).
[0248] After generating the fusion probability, to achieve binary classification output, this invention adopts a fixed threshold classification strategy: if the final predicted probability is greater than or equal to 0.5, the sample is classified as a high-systematic-risk state; otherwise, it is classified as a low-risk state. The determination rules are as follows:
[0249]
[0250] This fusion mechanism offers several advantages: Firstly, LSTM can effectively capture the trends and abrupt changes in risk over time, enhancing its responsiveness to temporal dynamics. Secondly, XGBoost demonstrates excellent generalization and nonlinear interaction modeling capabilities when handling non-temporal, high-dimensional financial features. Through weighted fusion, the two complement each other in a single prediction result, thereby enhancing the model's robustness and accuracy.
[0251] Furthermore, this invention retains the continuous risk probabilities output by the model, which facilitates subsequent risk ranking, scoring, and flexible threshold strategy adjustments, providing more interpretable and adaptable output results for various practical application scenarios such as financial regulatory authorities and internal bank risk control systems.
[0252] Step S5.2: Risk Index Construction and Ranking Mechanism
[0253] To enhance the readability and ease of application of systemic risk prediction results, this invention further constructs a continuous "systemic risk index" based on model predictions to quantify the degree of systemic financial risk faced by enterprises in a given quarter. This index can be used for risk ranking among companies, industry / regional comparisons, and trend visualization.
[0254] Specifically, in obtaining the fusion prediction probability Then, it is mapped to the standardized interval [0, 100] through a normalization transformation to form a continuous risk scoring index. The definition is as follows:
[0255]
[0256] in, This means linearly scaling the original probability values so that the predicted value range for all companies in a given quarter is standardized to 0 to 100, where 0 represents the lowest risk and 100 represents the highest risk.
[0257] This invention uses this risk index to output a relative ranking of the systemic risk levels of various companies on a quarterly basis, assisting financial regulatory agencies in identifying high-risk entities and investigating the sources of systemic risk. Simultaneously, the risk index can be grouped by industry, region, or asset attribute to construct risk heat maps and evolution trend charts, presenting the risk accumulation and diffusion trends in different sub-markets.
[0258] In addition, this risk index can also serve as a standard input indicator for subsequent risk warning, credit decision-making, and asset allocation strategies. It has good intuitiveness and business compatibility, and is suitable for the construction of quantitative scoring standards in actual risk control systems.
[0259] Step S5.3: Predict Explainability Module (via SHAP)
[0260] To enhance the interpretability and regulatory transparency of systemic risk prediction results, this invention introduces the SHAP (SHapley Additive exPlanations) value interpretation mechanism on top of multi-model prediction, constructing a feature attribution analysis module for the model output. The SHAP method, based on the Shapley value principle in game theory, assigns a set of marginal contribution values to each input feature, explicitly identifying the main driving factors of the model's prediction results.
[0261] Let the first The input feature vector for each sample is:
[0262]
[0263] in If we represent the feature dimension, then the SHAP model output can be interpreted as follows:
[0264]
[0265] in: For the model to sample The predicted risk probability; The global baseline prediction of the model represents the average output when there is no input information. Indicates the first The marginal contribution of each feature to the predicted value, i.e., the change in output caused when the feature is "added" to the model. Through the above decomposition method, this invention can accurately explain the specific direction and intensity of the influence of each input feature on the prediction result, solving the inherent problem of the uninterpretability of black box model predictions.
[0266] The SHAP method supports a variety of visualization outputs, including but not limited to: a global feature importance ranking plot (SummaryPlot): showing the average contribution of each feature in all samples, making it easier to identify overall risk drivers; and a single-sample interpretation plot (Force Plot): visually showing the positive and negative contributions of each feature of a specific company in a certain quarter and their inference paths.
[0267] This module not only enhances the transparency of model results, but also helps risk managers or investors understand the decision-making logic behind the model, effectively supporting key aspects such as compliance auditing, strategy formulation, and risk disclosure.
[0268] Step S5.4: System Architecture and Deployment Mode
[0269] To achieve efficient execution and scalable application of the method of this invention, a modular microservice architecture is recommended for system deployment. The overall architecture follows the principles of layered decoupling, standardized interfaces, and service independence, ensuring good maintainability and horizontal scalability under multiple data sources, heterogeneous platforms, and high-frequency task execution scenarios. The main functional modules include:
[0270] Data preprocessing module (Data Processor): Responsible for tasks such as standardizing financial data, imputing missing values, and generating yield and risk factors to ensure the consistency and completeness of subsequent modeling data;
[0271] CoVaR and GMM Risk Labeler Module: Performs conditional value at risk calculation and Gaussian mixture model clustering to generate enterprise-level high and low risk status labels;
[0272] Causal Network Engine: Performs Granger causality tests, lag order selection, and causal graph construction, and outputs risk propagation structural characteristics;
[0273] Fusion Predictor: Performs multi-model prediction based on LSTM and XGBoost, and integrates the prediction results through a weighted mechanism;
[0274] Risk scoring and explanation module (Scorer+Explainer): Outputs a standardized risk index and provides SHAP value-driven feature explanation function;
[0275] Visualization and Interface Display Module (Dashboard Frontend): Constructs a graphical interactive interface that supports the display of risk trends, network structure, and individual interpretation results.
[0276] The system's modules communicate via standard RESTful API interfaces and support JSON-formatted data exchange, facilitating integration into internal bank risk systems, regulatory portals, or web service platforms. The system supports a periodic batch scheduling mechanism, automatically updating risk prediction results on a quarterly, monthly, or custom-defined basis and writing model outputs in batches to the database, thus automating and normalizing the prediction service.
[0277] Step S5.5: System demonstration examples and typical output screenshots
[0278] To demonstrate the output of the system in practical applications, multiple visualization interfaces and chart display modules are constructed to create an interactive environment for interpreting risk prediction results. Typical output results include:
[0279] Risk probability trend chart: Displays the trajectory of predicted risk probability changes for a company or all samples in different quarters, which can be used to observe cyclical risk fluctuations and abnormal upward signals;
[0280] Enterprise Risk Heat Map: The color intensity encodes the systemic risk level of an enterprise in a given quarter, and supports filtering by industry and region, making it easy for regulatory agencies to quickly identify high-risk concentration areas;
[0281] Systemic Risk Network Diagram: Displays a risk propagation map based on Granger causality, marking core nodes and main paths, used to analyze the risk aggregation and diffusion structure;
[0282] Single-Enterprise Force Plot: Shows which characteristics dominate the risk forecast results for a particular enterprise in the current quarter, distinguishing between positive driving and negative inhibiting factors, and enhancing the transparency of model interpretation.
[0283] All the above images are generated by standard visualization tools and can be deployed to web-based dashboards, facilitating interactive risk queries, model result interpretation, and export of reporting materials for users.
[0284] This invention integrates time series deep models, ensemble learning models, and causal structure recognition mechanisms to construct a complete, highly interpretable, and deployable intelligent risk identification framework, which is particularly suitable for scenarios such as financial supervision, bank risk control, credit rating, and counterparty risk management.
[0285] Furthermore, the following embodiments are all based on the technical solutions of the present invention:
[0286] Example 1: Systemic Risk Identification Process Based on Quarterly Financial Data
[0287] include:
[0288] Construct a multi-quarter sliding window input sequence;
[0289] Calculate CoVaR and risk label;
[0290] Construct a Granger causal network and extract in-degree, out-degree, and centrality;
[0291] Prediction is performed using LSTM and XGBoost;
[0292] The final model is obtained through PSO parameter tuning;
[0293] Output risk probability, risk index, and SHAP explanation diagram.
[0294] Example 2: Identification of Risk Transmission Paths Among Financial Institutions
[0295] Path identification is based on Granger causality, and weights and lag orders are labeled in the network graph.
[0296] Example 3: System Deployment Example
[0297] It adopts a microservice architecture, including a data module, a modeling module, an interpretation module, and a RESTful API, and supports quarterly risk updates.
[0298] The above embodiments are merely typical illustrative methods of the present invention, and the scope of protection of the present invention is not limited thereto. All equivalent substitutions and improvements made under the concept of the present invention should fall within the scope of protection. It should be emphasized that any modifications or minor adjustments made by those skilled in the art without departing from the basic principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A method for intelligent identification of systemic financial risk based on causal networks and multi-model optimization, characterized in that, Includes the following steps: Step S1: Read and clean financial market and corporate financial data, and construct a quarterly logarithmic return feature series; Step S2: Calculate the company-level risk exposure value based on the CoVaR mechanism, and binarize the risk status label of the enterprise using Gaussian Mixture Model (GMM) clustering. Step S3: Construct a causal directed network among financial institutions using the Granger causality test method, and extract lag parameters, significance coefficients, and structural measures; Step S4: Construct a risk prediction model based on LSTM neural network and XGBoost tree model, and jointly optimize the model parameters and structure using particle swarm optimization algorithm; Step S5: Weight and fuse the output results of multiple models to generate a systematic risk score and state classification results, and deploy the interpretability and visualization service module to realize online application.
2. The method according to claim 1, characterized in that, In step S2, the CoVaR value is defined by the following formula: in, For the company The logarithmic return series, To set the confidence level, CoVaR was calculated based on historical simulation.
3. The method according to claim 1, characterized in that, In step S2, the Gaussian Mixture Model (GMM) is used for risk status identification. The GMM model is defined as follows: in, The number of components in a Gaussian mixture model is defined as the number of high-risk and low-risk clusters. and through the mean Determine the risk category; This is an index of the mixture components in a Gaussian mixture model, used to distinguish different clustering components. and The first The mean and variance of a Gaussian distribution.
4. The method according to claim 1, characterized in that, In step S3, the causal network is constructed based on the Granger causality test, with the null hypothesis being: Among them, the null hypothesis For: Company All Lagged variables on the company The current rate of return has no significant impact, i.e., all lagged coefficients. ,in ; The significance of the test was performed using the following F-statistic: in, , These are the sums of squared residuals for the restricted and unrestricted models, respectively. This represents the number of samples.
5. The method according to claim 1, characterized in that, In step S3, the lag order is determined by minimizing the AIC and BIC criteria. The AIC is defined as follows: in For the number of model parameters, For the sample size, To predict the residual variance.
6. The method according to claim 1, characterized in that, In step S4, the state update of the LSTM neural network model includes the following calculation process: Cell status update: Output status: in , , These are the activation vectors for the forget gate, input gate, and output gate, respectively.
7. The method according to claim 1, characterized in that, In step S4, the prediction function of the XGBoost tree model is defined as: in For the regression tree function space, the loss function includes a penalty term for the structural complexity of each tree; The total number of trees, Indicates the first The input variable vector for each sample.
8. The method according to claim 1, characterized in that, In step S4, the particle swarm optimization (PSO) algorithm is used to jointly tune the prediction model. The optimization process is as follows: Let the first step be... In the nth iteration, the 1st Each particle represents a set of hyperparameter configurations, and its current position is... The speed is The best position reached by this particle in history is The known optimal position in the entire group is ; Speed updates: Location update: in Inertia factor , The learning coefficient, .
9. The method according to claim 1, characterized in that, In step S5, the final output of the fused prediction probability is: in The fusion coefficient is adjustable. The first given by the LSTM model The risk probability of the sample The risk probability output by the XGBoost model; Constructing a risk index: Used to uniformly quantify and rank the risk levels of all companies in the same quarter; The model prediction results are interpreted using the SHAP method and satisfy the following form: in, For the model to sample The predicted risk probability; The global baseline prediction of the model represents the average output when there is no input information. Representation of features The marginal contribution of the prediction results to the explanation.
10. The method according to claim 1, characterized in that, Deployed in a modular microservice architecture, it includes at least: a data processing module, a label building module, a causal graph building module, a predictive modeling module, a parameter optimization module, an interpretation and display module, and a predictive service interface module, supporting RESTful API format request returns.