A financial risk dynamic monitoring and report generation method based on multi-source data

Abstract

The application provides a financial risk dynamic monitoring and report generation method based on multi-source data, comprising: acquiring multi-source data of credit default swaps, pre-processing and feature extraction on interest rate spread data of different maturities to obtain a maturity structure feature vector; obtaining a long-short term interest rate spread curve based on real-time acquired long-short term interest rate spread data, performing anomaly detection on the shape change of the long-short term interest rate spread curve, and triggering an evaluation process of credit risk contagion effect when detecting a distortion anomaly of the interest rate spread curve; in the risk evaluation and monitoring process, continuously collecting market evaluation, extracting information related to key credit default swap attributes, taking the information as a supplementary input of risk transmission mechanism analysis and monitoring strategy, and dynamically optimizing a risk management strategy.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for dynamic monitoring and reporting of financial risks based on multi-source data. Background Technology

[0002] Credit default swap (CDS) spreads are financial derivatives used to hedge credit risk, and their spreads reflect the market's assessment of credit risk. In the process of dynamic monitoring and reporting of financial risk based on multi-source data, the term structure characteristics of CDS spreads are an important analytical object. Accurately assessing the contagion effect of credit risk when the long-term and short-term interest rate spread curves become distorted, and adjusting monitoring strategies accordingly, is a complex technical problem. This problem involves several aspects: first, how to promptly detect signals of interest rate spread curve distortion from massive amounts of multi-source heterogeneous data; second, how to quantitatively assess the impact of curve distortion on credit risk and consider the transmission mechanism of risk between different entities and markets; and finally, how to transform complex quantitative analysis results into clear and easy-to-understand risk reports to provide effective support for decision-making. These interconnected problems are constrained by multiple factors such as data quality, model capabilities, and computing efficiency, posing significant challenges to technical implementation. Summary of the Invention

[0003] This invention provides a method for dynamic monitoring and reporting of financial risks based on multi-source data, mainly including:

[0004] This study acquires multi-source data on credit default swaps (CDS), preprocesses and extracts features from interest rate spreads across different maturities to obtain term structure feature vectors. Based on real-time acquired long-short interest rate spread data, it generates long-short interest rate spread curves and performs anomaly detection on the curves' shape changes. When anomalies in the spread curve distortion are detected, an assessment process for credit risk contagion effects is triggered. During this assessment, the standardized term structure feature vectors are used as node attributes of a graph, and anomalies in the spread curve distortion are used as edge attributes. A graph neural network is used to learn the transmission path and intensity of credit risk across different maturities and market participants to analyze the credit risk transmission mechanism based on the graph neural network. Monte Carlo simulation is employed to analyze the credit risk transmission mechanism learned by the graph neural network, generating risk contagion scenarios driven by abnormal changes in CDS interest rate spreads across different maturities, and assessing the cumulative losses of credit risk under different scenarios. The risk management system is structured to determine the thresholds and early warning levels for risk monitoring indicators centered on key credit default swap (CDS) attributes. Key CDS attributes are dynamically monitored, and when the CDS spread exceeds a preset threshold or other attributes show significant abnormal changes, a risk warning and emergency response process is triggered, with corresponding risk mitigation measures implemented based on the warning level. During risk assessment and monitoring, market evaluations are continuously collected, and information related to key CDS attributes is extracted. This information serves as supplementary input for risk transmission mechanism analysis and monitoring strategies, dynamically optimizing risk management strategies. Regular credit risk monitoring reports are generated, summarizing risk assessment results, early warning situations, and response measures, as well as the cumulative loss distribution under different risk contagion scenarios. Data visualization methods are used to visualize the credit risk monitoring content, including time series graphs of CDS attribute changes, credit risk transmission network diagrams, and probability density graphs of cumulative loss distributions.

[0005] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0006] This invention discloses a method for dynamic monitoring and reporting of financial risks based on multi-source data. The method acquires heterogeneous data from multiple sources, extracts standardized term structure features, monitors anomalies in the interest rate spread curve in real time, and triggers risk assessment. It utilizes graph neural networks to learn risk transmission mechanisms, combines Monte Carlo simulations to generate risk contagion scenarios, assesses the distribution of cumulative losses, and determines monitoring indicators. Key attributes are dynamically monitored, automatically triggering early warnings and emergency responses. Simultaneously, market evaluations are integrated to optimize risk management strategies. This method achieves comprehensive risk monitoring of the credit default swap market, enabling timely identification of anomalies, assessment of transmission paths, and prediction of potential losses, providing support for risk management decisions and effectively preventing systemic credit risk. Attached Figure Description

[0007] Figure 1This is a flowchart of a method for dynamic monitoring and reporting of financial risks based on multi-source data, according to the present invention.

[0008] Figure 2 This is a schematic diagram of a method for dynamic monitoring and reporting of financial risks based on multi-source data according to the present invention.

[0009] Figure 3 This is another schematic diagram of a method for dynamic monitoring and report generation of financial risks based on multi-source data according to the present invention. Detailed Implementation

[0010] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0011] like Figure 1-3 This embodiment of a method for dynamic monitoring and reporting of financial risks based on multi-source data may specifically include:

[0012] S101. Obtain multi-source data on credit default swaps, preprocess and extract features from interest rate spread data for different maturities, and obtain term structure feature vectors.

[0013] Interest rate spread data is retrieved from the database based on the time interval. An initial interest rate spread dataset is obtained by aligning dates and unifying decimal places. Linear interpolation is used to fill in the detected null points in the initial interest rate spread dataset, and duplicate data items are removed by using distribution characteristics to obtain a cleaned interest rate spread dataset. The cleaned interest rate spread dataset is then grouped by term and outliers are corrected using a moving average method. Data points exceeding the threshold range are removed using a percentile method to obtain smoothed interest rate spread data. The smoothed interest rate spread data is normalized using a min-max normalization method, and the term structure feature vector is calculated by fitting a support vector regression model.

[0014] Specifically, credit default swap spread data of different maturities are obtained from the stock exchange database within a specified time interval. Interbank market quoted spread data are collected through a financial data platform, and rating change information is obtained from international rating agencies. The collected spread data is then processed for date alignment and decimal place unification to obtain an initial spread dataset. Data quality verification is performed on the initial spread dataset. Missing value detection methods are used to identify data null points, which are then filled using linear interpolation. Duplicate data items are removed based on data distribution characteristics to obtain a cleaned spread dataset. The cleaned spread dataset is then grouped by maturity. A five-day moving average method is used to correct outliers. The upper and lower quartile thresholds for each maturity group are calculated using the percentile method, and data points exceeding these thresholds are removed to obtain smoothed grouped spread data. Based on the smoothed grouped spread data, autocorrelation analysis is used to calculate the correlation coefficient matrix between each maturity group. Highly correlated maturity groups are merged based on the correlation coefficient thresholds to obtain dimensionality-reduced term structure data. The reduced term structure data is normalized using the min-max standardization method. This is achieved by subtracting the minimum value from each data point and dividing by the range of values ​​in that range, resulting in standardized feature data. A support vector regression model is then constructed for this standardized feature data, using a Gaussian kernel. Cross-validation is employed to determine the penalty and kernel parameters, and the model is used to fit the data for each term group, yielding fitted interest rate spreads. The root mean square error method is used to calculate the residual vector between the fitted interest rate spreads and the standardized feature data. Based on this residual vector, principal component analysis is used to extract features, resulting in the term structure feature vector. Credit default swap interest rate spreads typically originate from multiple market participants, and the data formats differ between these sources. For example, stock exchange interest rate spreads are accurate to four decimal places, while interbank market quotes are accurate to two decimal places, necessitating standardized processing. Taking a corporate bond as an example, its 5-year credit default swap (CDS) spread on the exchange is 235.4568 basis points, while the interbank market quote is 235.46 basis points. After unified processing, the value is taken as 235.46 basis points. During the data quality verification phase, missing values ​​are filled using linear interpolation. For example, if the 5-year CDS spread for a company on February 1st and February 3rd is 245.36 and 255.36 basis points respectively, and data for February 2nd is missing, then the spread for February 2nd is calculated to be 250.36 basis points using linear interpolation. Furthermore, if duplicate data is found for the same day, such as a company recording the same 5-year spread as both 240.35 and 240.36 basis points on the same day, then the data at 240.36 basis points is retained. In outlier handling, a 5-day moving average method is used to smooth data points with large fluctuations.For example, if a company's 5-year yield spread over five consecutive trading days is 260.35, 262.45, 295.36, 265.42, and 268.35 basis points respectively, and the data on the third day deviates significantly, it is corrected by calculating the 5-day moving average of 271.59 basis points. Simultaneously, the upper and lower quartile intervals are calculated. If the upper quartile value is 280.36 basis points and the lower quartile value is 250.36 basis points, data points outside these intervals are removed. For yield spread data of different maturities, autocorrelation analysis is used to identify highly correlated maturity groups. Assuming the correlation coefficient between the 3-year and 5-year yield spreads is 0.95, exceeding the preset threshold of 0.9, the data from these two maturity groups are merged. During standardization, if the maximum value of a maturity group is 320.36 basis points and the minimum value is 220.36 basis points, then for a data point within that group with a value of 270.36 basis points, its standardized value is 0.5. When constructing the support vector regression model, a Gaussian kernel function was chosen as the kernel function. Cross-validation was used to determine the kernel parameter γ=0.01 and the penalty parameter C=100. For the fitting results, the root mean square error RMSE=√(Σ(y_pred-y_true)² / n) was calculated to evaluate the model performance, where y_pred is the predicted value, y_true is the actual value, and n is the number of samples. Finally, principal component analysis was performed based on the residual vectors, and principal components with an eigenvalue contribution rate exceeding 85% were extracted as the term structure feature vectors.

[0015] S102. Based on the real-time acquired long-short interest rate spread data, the long-short interest rate spread curve is obtained. Anomaly detection is performed on the shape changes of the long-short interest rate spread curve. When an abnormal distortion of the interest rate spread curve is detected, the assessment process of the credit risk contagion effect is triggered.

[0016] Long-term and short-term interest rate spread data are obtained from the market database according to a preset sampling frequency. A basic interest rate spread dataset is obtained through data standardization and timestamp alignment. For the basic interest rate spread dataset, the spline function coefficients are determined using cubic spline interpolation and least squares methods to obtain continuous and smooth interest rate spread curve data. The rate of change of interest rate spread between adjacent maturity points is calculated for the interest rate spread curve data, and the curve anomaly data is obtained using the sliding window method and autoregressive moving average method. If the curve anomaly data exceeds the warning threshold, the risk transmission probability is calculated, and a risk transmission matrix is ​​established using the risk transmission probability to obtain the term structure risk transmission coefficient.

[0017] Specifically, long-term and short-term interest rate spread data are retrieved from the trading market database at a fixed frequency. Through data format standardization, data update timestamp alignment, and data validity verification, data points exceeding twice the standard deviation of the previous five-day average are removed to generate basic long- and short-term interest rate spread data. For this basic data, a continuous and smooth interest rate spread curve is constructed using cubic spline interpolation. Interpolation nodes are selected at key points for each maturity, and the spline function coefficients are determined using the least squares method to obtain the continuous and smooth interest rate spread curve data. The rate of change of interest rate spreads between adjacent maturity points is calculated from the interest rate spread curve data. The historical fluctuation range for each maturity point is calculated using the sliding window method, with the upper and lower limits of the fluctuation range set at plus or minus twice the volatility of the previous twenty days, yielding the volatility characteristic data of the interest rate spread curve. For this volatility characteristic data, an autoregressive moving average method is used to construct a predictive benchmark curve. The autoregressive order is set to five, and the moving average order to three. The root mean square error between the actual curve and the predicted curve is calculated to obtain the anomaly level data of the interest rate spread curve. Based on the anomaly data of the interest rate spread curve, a warning threshold was set at twice the root mean square error of the previous 30 days. Curve intervals exceeding the warning threshold were marked to obtain risk-concern interval data. For this risk-concern interval data, a random forest algorithm based on entropy was used to calculate the risk transmission probability between different maturity points. Feature variables included the rate of change in interest rate spreads, volatility, and the rate of change in trading volume, establishing a risk transmission matrix. The risk spillover coefficient and risk absorption coefficient for each maturity node were calculated using the risk transmission matrix. The risk spillover coefficient represents the ability of a maturity to transmit risk to other maturities, and the risk absorption coefficient represents the degree to which a maturity absorbs risk from other maturities, yielding the term structure risk transmission assessment results. In the credit default swap market, abnormal fluctuations in long-term and short-term interest rate spreads often reflect the market's level of concern about credit risk. Taking a large enterprise's credit default swap as an example, its 1-year spread is 150 basis points, and its 5-year spread is 280 basis points. When data anomalies occur, such as the 1-year spread suddenly rising to 220 basis points, exceeding twice the standard deviation of the previous 5-day average of 145 basis points (30 basis points), this data point will be marked as an anomaly and removed. When constructing a continuous and smooth spread curve using cubic spline interpolation, 1-year, 2-year, 3-year, 5-year, 7-year, and 10-year periods are selected as key maturity points. The spread values ​​for each maturity point are set to 150, 180, 220, 280, 320, and 350 basis points, respectively. The spline function coefficients are solved using the least squares method to obtain the smooth spread curve equation. Where x represents the term, and the coefficients a, b, c, and d are obtained by minimizing the sum of squared errors. When calculating the rate of change of interest rate spread, the rate of change R between adjacent term points is defined as... Y2 and Y1 represent the interest rate spreads for maturities T2 and T1, respectively. For example, the rate of change of the interest rate spread between 2-year and 1-year maturities is (180-150) / (2-1) = 30 basis points per year. A 20-day window is used to calculate the historical volatility of each maturity level. If the average daily volatility of a certain maturity level is 15 basis points, then the upper limit of the volatility range is set to 30 basis points, and the lower limit is -30 basis points. For the autoregressive moving average model, the fifth-order autoregressive term represents the prediction of the current value using data from the previous 5 trading days. For example, the formula for calculating the predicted value of the current 5-year interest rate spread is Y_t = 0.3Y_t-1 + 0.25Y_t-2 + 0.2Y_t-3 + 0.15Y_t-4 + 0.1Y_t-5, where Y_t represents the interest rate spread on day t, Y_t-1 represents the interest rate spread on day t-1, and so on. The coefficients are obtained by fitting historical data. After incorporating a third-order moving average term, the impact of random disturbances is considered. In the risk transmission analysis, the entropy method is used to calculate the feature weights, assuming the weights for the rate of change in interest rate spreads, volatility, and trading volume are 0.4, 0.35, and 0.25, respectively. In the constructed risk transmission matrix, the matrix element P_ij represents the probability of risk transmission from maturity i to maturity j, calculated using the random forest algorithm. The risk transmission path in the term structure often exhibits a characteristic of short-term to long-term spread. For example, the transmission probability from 2-year to 5-year is 0.65, while the transmission probability from 5-year to 2-year is only 0.25, indicating that abnormal fluctuations in short-term interest rate spreads are more likely to trigger adjustments in long-term interest rate spreads.

[0018] S103. In the assessment process of triggering the contagion effect of credit risk, the standardized term structure feature vector is used as the node attribute of the graph, and the abnormal distortion of the interest rate spread curve is used as the edge attribute. The transmission path and intensity of credit risk between different terms and market participants are learned through graph neural network in order to analyze the credit risk transmission mechanism based on graph neural network.

[0019] A node attribute matrix is ​​constructed based on the standardized term structure feature vector, and an edge attribute matrix is ​​constructed using the interest rate spread curve distortion data. The initial transmission graph data is obtained using the minimum-maximum normalization method. For the initial transmission graph data, a weighted Euclidean distance method is used to calculate the similarity score between nodes. If the similarity score is lower than the median similarity score between node pairs, node connections are pruned to obtain optimized transmission graph data. A graph convolution operator is designed using Chebyshev polynomial expansion to construct a feature extractor for the nodes in the optimized transmission graph data, calculating the weighted features of adjacent nodes to obtain local structural features. The dot product attention calculation method is used to obtain the attention weight matrix between local structural features, and the node features are weighted and aggregated based on the attention weights to obtain global transmission features.

[0020] Specifically, a node attribute matrix of the graph is constructed based on standardized term structure feature vectors, and an edge attribute matrix is ​​constructed using anomaly data from interest rate spread curve distortion. The node and edge attribute matrices are standardized using the min-max normalization method to obtain initial transmission graph data. Node metrics are calculated for the initial transmission graph data. The similarity score between nodes of different terms is calculated using the weighted Euclidean distance method. A minimum similarity score threshold is set as the median of the similarity scores for all nodes. Connections to nodes below the threshold are pruned to obtain optimized transmission graph data. For each node in the optimized transmission graph data, a three-layer graph convolutional feature extractor with a receptive field is constructed. A graph convolution operator is designed using Chebyshev multinomial expansion to calculate the weighted features of adjacent nodes and obtain the local structural features of the nodes. For the local structural features of the nodes, the attention weight matrix between nodes is obtained using the dot product attention calculation method. Based on the attention weights, the node features are weighted and aggregated to obtain global transmission features. A four-layer graph neural network structure is constructed, with each layer containing a graph convolutional layer and an attention layer. Features are transferred between adjacent layers using skip connections. The network parameters are trained through a node classification task to obtain the risk transmission map. Based on the risk transmission graph, the shortest paths and path weights between nodes are extracted, and the cumulative transmission probability and transmission delay on the paths are calculated to quantitatively characterize the credit risk transmission path and transmission strength, thus obtaining the risk transmission mechanism matrix. In the credit market, the term structure eigenvector reflects the risk characteristics of debts with different maturities. Taking a corporate bond as an example, its standardized eigenvalues ​​for 1-year, 3-year, and 5-year maturities are 0.3, 0.5, and 0.7, respectively. These eigenvalues ​​are mapped to the 0-1 interval through min-max normalization and used as attribute values ​​for nodes in the graph. Simultaneously, the distortion of the spread curve, such as a sudden increase in the 1-3 year spread from 50 basis points to 120 basis points, is standardized and used as the attribute value of the edge connecting these two maturity nodes. When constructing the graph structure, the similarity between nodes is calculated using weighted Euclidean distance, with weights set according to the proximity of the maturities: adjacent maturities have a weight of 0.6, a one-maturity interval has a weight of 0.3, and a two-maturity interval has a weight of 0.1. Assuming the feature vectors of two nodes are [0.3, 0.4, 0.5] and [0.4, 0.5, 0.6] respectively, the calculated weighted distance is 0.15. If the median distance between all node pairs is 0.2, the connection between these two nodes is preserved. Graph convolution feature extraction uses a third-order Chebyshev polynomial expansion, with polynomial coefficients obtained through training [0.8, -0.3, 0.1]. For feature updates of the center node, the mean features of the first-order neighbors are first calculated. For example, if the feature values ​​of neighboring nodes are [0.4, 0.5], then the first-order feature is 0.45, the second-order feature is 0.42, and the third-order feature is 0.40. Finally, these features are weighted and combined to obtain the new features of the node.In the attention mechanism, attention scores are calculated using dot products. For example, the dot product of the feature vectors of nodes i and j is 0.85. After softmax normalization, the attention weight is 0.3, representing the strength of node j's influence on node i during information transmission. In the four-layer graph neural network, the output dimensions of each layer are set to 64, 32, 16, and 8, respectively. Features are transmitted between adjacent layers through skip connections to avoid information loss in deep networks. The risk transmission path is extracted using Dijkstra's algorithm, with the negative logarithm of the transmission probability as the path weight. Assuming the shortest path from 1 year to 5 years takes 3 years, and the transmission probabilities of each segment are 0.8 and 0.7, the cumulative transmission probability is 0.56, and the transmission delay is 2 time units.

[0021] S104. The Monte Carlo simulation method is used to analyze the credit risk transmission mechanism learned by the graph neural network, generate risk contagion scenarios driven by abnormal changes in credit default swap interest rates with different maturities, assess the cumulative loss distribution of credit risk under different scenarios, and determine the threshold and early warning level of risk monitoring indicators with key credit default swap attributes as the core.

[0022] Based on historical credit spread data, drift rate and volatility parameters are calculated. A set of random fluctuation paths for credit spreads is generated using a geometric Brownian motion equation, which includes a drift term and a disturbance term. For this set of random fluctuation paths, an exponentially weighted moving average method is used to calculate dynamic fluctuation interval values. The ratio of these dynamic fluctuation interval values ​​to the standard deviation is used to identify abrupt changes in credit status, resulting in a multidimensional risk feature sequence. For this multidimensional risk feature sequence, a Gaussian kernel density estimation method is used to fit the risk loss distribution function. The bandwidth parameter of the Gaussian kernel density estimation method is obtained through cross-validation. For the risk loss distribution function, a generalized Pareto distribution is used to fit the tail risk characteristics, obtaining the risk loss quantile values ​​under the confidence interval. A hierarchical clustering method is then used to determine the risk warning threshold.

[0023] Specifically, based on the risk transmission mechanism extracted by graph neural networks, a Monte Carlo simulator for abnormal interest rate changes is constructed, and the geometric Brownian motion equation is used. A stochastic fluctuation path for the credit spread of a reference entity is generated, where S represents the credit spread, dS represents the change in the credit spread over the time interval dt, the drift rate μ and volatility σ are estimated based on historical data, and dW is the Brownian motion increment, a random variable used to introduce randomness. A set of spread paths is obtained by adjusting the spreads of different maturities using a Markov transition probability matrix. For the spread path set, an exponentially weighted moving average method is used to calculate the dynamic fluctuation range, with upper and lower limits set at two standard deviations to identify abrupt changes in credit conditions. A term structure imbalance index is calculated by combining the ratio of short-term to long-term spreads, and liquidity volatility is characterized by the daily trading volume change rate, resulting in a multidimensional risk feature sequence. A hierarchical level is constructed for the multidimensional risk feature sequence, classifying risk levels based on credit rating, term structure, and liquidity level. Within each level, a Gaussian kernel density estimation method is used to fit the risk loss distribution function, and the kernel function bandwidth is selected using cross-validation to obtain a family of risk distribution functions. Based on a family of risk distribution functions, a generalized Pareto distribution is used to fit the tail risk, and risk loss quantiles at 99%, 97.5%, and 95% confidence intervals are calculated to generate loss threshold sequences at different confidence levels. A hierarchical clustering algorithm is applied to the loss threshold sequences, and the number of clusters is determined by maximizing inter-cluster distance. Risk level intervals are then divided based on cluster centers to obtain hierarchical risk warning thresholds. Based on the risk warning thresholds, the contribution of each dimension of risk indicators to the warning results is calculated using the conditional entropy method. Indicators with the highest contribution are selected to construct a monitoring indicator set, and red, orange, and yellow warning level thresholds are set to obtain a risk warning indicator system. The random fluctuation process of credit default swap spreads is simulated using the geometric Brownian motion equation. Taking a company's 5-year credit default swap as an example, based on historical data, the drift rate μ=0.02 and volatility σ=0.15 are estimated. The initial spread is set at 200 basis points, and the time step is 1 day. 1000 random paths are generated through Monte Carlo simulation. In the linkage adjustment of adjacent maturity spreads, the correlation between maturities is described using a Markov transition matrix. For example, the probability of a transition from 3-year to 5-year spreads is 0.7, indicating that changes in the 3-year spread will propagate to the 5-year spread with a 70% probability. In the risk characteristic identification process, an exponentially weighted moving average method is used to calculate the dynamic fluctuation range, assigning weights [0.4, 0.25, 0.15, 0.12, 0.08] to the data from the most recent five trading days. A credit mutation point is marked when the spread exceeds two standard deviations on a given day. The term structure imbalance index is calculated using the ratio of short-term to long-term spreads. For example, if the 2-year spread is 180 basis points and the 5-year spread is 240 basis points, the imbalance index is 0.75, lower than the normal level of 0.85. Liquidity characteristics are characterized by the daily trading volume change rate. For example, a 30% decrease in daily trading volume compared to the previous day indicates a significant deterioration in liquidity. In the risk loss distribution estimation, a Gaussian kernel density function is used, and the optimal bandwidth h=0.08 is selected through cross-validation.For tail risk, a generalized Pareto distribution is used for fitting, and its distribution function is given by [formula missing]. With location parameter ξ=0.3 and scaling parameter β=50, the risk loss quantile at a 99% confidence level is calculated to be 350 basis points. In risk level classification, inter-class distance is calculated using a hierarchical clustering algorithm, and Euclidean distance is used to measure the difference between different risk levels. Assuming the loss threshold sequence is [150, 200, 250, 300, 350] basis points, it is divided into three level intervals: [150, 200], [200, 300], and [300, 350], using the maximum inter-class distance criterion. The contribution of risk indicators is calculated based on the conditional entropy method; a higher entropy value indicates a stronger discriminative ability of the indicator. For example, the conditional entropy of the credit spread change rate is 0.85, the conditional entropy of the term structure imbalance index is 0.72, and the conditional entropy of the liquidity indicator is 0.65. Based on this, the top three indicators in terms of contribution are selected as core monitoring indicators. The final thresholds were set at 300 base points for red alerts, 200 base points for orange alerts, and 150 base points for yellow alerts, forming a complete risk warning system.

[0024] S105. Dynamically monitor key credit default swap attributes. When the credit default swap spread exceeds the preset threshold or other attributes show significant abnormal changes, trigger the risk warning and emergency response process, and take corresponding risk mitigation measures according to the warning level.

[0025] A monitoring indicator matrix is ​​constructed based on credit rating change data, term spread data, and market trading volume data. Principal component analysis is used to reduce the dimensionality of this matrix, and Mahalanobis distance is used to calculate anomaly scores. For these anomaly scores, an exponentially weighted moving average control chart is used for anomaly detection. Multi-level early warning thresholds are set based on the weighted moving average and weighted standard deviation to obtain tiered early warning data. For these tiered early warning data, a cumulative sum control chart is used to calculate trend characteristic values. These trend characteristic values ​​determine the direction of indicator changes; if the indicators change in the same direction, linked early warning data is obtained. For these linked early warning data, a Gaussian kernel function support vector machine is used for classification. A risk knowledge graph is constructed based on the classification results, and the risk response rule base is queried through this risk knowledge graph to obtain risk mitigation solutions.

[0026] Specifically, a multi-dimensional monitoring indicator matrix is ​​constructed based on changes in credit ratings, fluctuations in term spreads, and changes in market trading volume. Principal component analysis is used to reduce the dimensionality of the multi-dimensional indicators, and Mahalanobis distance is used to calculate the anomaly scores of the indicators. The Mahalanobis distance uses the inverse of the covariance matrix as the metric matrix to obtain real-time anomaly monitoring data. For the real-time anomaly monitoring data, an exponentially weighted moving average control chart is used for anomaly detection, with the weighting coefficient set to λ=0.2. The weighted mean and weighted standard deviation are calculated, and three standard deviations are set as the upper limit for red warnings, two standard deviations as the upper limit for orange warnings, and one standard deviation as the upper limit for yellow warnings, resulting in tiered warning data. The trend characteristics of the tiered warning data are calculated using a cumulative sum control chart, and thresholds for consecutive rising points and cumulative distances are set to identify persistent abnormal trends. Combined with the unidirectional changes of multiple indicators, linked warning data is obtained. The linked warning data is classified using a Gaussian kernel function support vector machine. The kernel function parameter σ is determined through cross-validation. The classifier is trained based on historical risk event samples to obtain warning level discrimination data. A risk knowledge graph is constructed based on the early warning level data. Nodes represent risk event types, and edges represent risk transmission relationships. Risk response rules are stored and queried through a graph database to obtain risk management data. For the risk management data, a pre-set risk mitigation measure library is used for matching. Appropriate mitigation schemes are selected according to the degree of indicator anomaly, early warning level, and risk type, and execution priorities are set to obtain risk response plans. Taking a large enterprise's credit default swap contract as an example, the multi-dimensional monitoring indicators include three dimensions: frequency of credit rating changes, magnitude of term spread changes, and market trading volume change rate. Principal component analysis is used to reduce the dimensionality of these indicators. The first principal component explains 75% of the variance, representing the overall credit risk level, while the second principal component explains 15% of the variance, reflecting the term structure change characteristics. When calculating Mahalanobis distance, the inverse of the sample covariance matrix is ​​used as the metric matrix. An anomaly is identified when the Mahalanobis distance between the observed value and the mean exceeds 10.0. In the anomaly detection phase, the weighting coefficient λ=0.2 in the exponentially weighted moving average control chart, indicating that the weight of the most recent data is 0.2, the previous period is 0.16, the period before that is 0.128, and so on, decreasing sequentially. Assuming a weighted mean of 250 basis points and a weighted standard deviation of 30 basis points at a certain point in time, the upper limit for a red alert is 340 basis points, for an orange alert it is 310 basis points, and for a yellow alert it is 280 basis points. In the cumulative control chart, a trend warning is triggered when five consecutive points rise and the cumulative distance exceeds 50 basis points. Analysis of the correlation between multiple indicators shows that when a credit rating is downgraded by one level, the 5-year yield spread increases by 80 basis points, and daily trading volume decreases by 40%, this combination of deteriorating trends often foreshadows accelerated exposure to credit risk. A Gaussian kernel function is used in the support vector machine classifier. Where x and y are sample vectors from the dataset, which can be feature vectors of any dimension. The distance between x and y is the Euclidean distance, exp is an exponential function, and the width parameter σ=0.5 of the Gaussian kernel function is determined through five-fold cross-validation, achieving a classification accuracy of 85%. The risk knowledge graph is constructed based on historical risk event data. Nodes include risk types such as credit rating downgrades, term structure distortions, and liquidity shortages, with edge weights representing the probability of risk transmission. When a credit rating downgrade risk is identified, the graph database is used to query relevant risk transmission paths and calculate the cumulative risk exposure. At the risk management level, for red alerts, such as a 5-year yield spread exceeding 340 basis points, the highest level of response is initiated, including limiting trading volume and increasing margin requirements. For orange alerts, such as a five-day consecutive rise in yield spreads exceeding 50 basis points cumulatively, a medium-level response is adopted, such as increasing monitoring frequency and adjusting limits. For yellow alerts, such as a continued decline in trading volume, general preventative measures are taken, such as strengthening market observation and preparing emergency funds.

[0027] S106. During the risk assessment and monitoring process, continuously collect market evaluations, extract information related to key credit default swap attributes, and use the information as a supplementary input for risk transmission mechanism analysis and monitoring strategies to dynamically optimize risk management strategies.

[0028] Market evaluation texts are obtained from financial websites according to preset rules. Named entities are extracted from the text using a conditional random field method to obtain initial text data. For the initial text data, text vectors are constructed using term frequency and inverse document frequency. Text features are extracted using a neural network with two convolutional layers to obtain text semantic data. The text semantic data is used to calculate an information reliability score based on the rating agency level, release time interval, and historical accuracy. If the reliability score is higher than a preset threshold, a text knowledge graph is constructed by extracting entity relationships to obtain information knowledge data. For the information knowledge data, an attention mechanism is used to identify risk-related entities, and a long short-term memory network is used to predict the risk transmission path to obtain risk trend data. The risk transmission probability is dynamically updated using a Markov decision process.

[0029] Specifically, market evaluation texts are obtained from financial news websites, rating agency announcements, and research reports according to preset information collection rules. These rules include a list of issuer names, keywords related to rating changes, and keywords related to rating outlook. The text is preprocessed using a bidirectional maximum matching word segmentation algorithm, and named entities are extracted from the text using a conditional random field method to obtain initial text data. For this initial text data, text vectors based on word frequency and inverse document frequency are constructed. A word vector model with a sliding window size of 5 is used to calculate text semantic similarity. Text features are extracted using a neural network with two convolutional layers, and text sentiment is classified into three categories to obtain text semantic data. Information reliability scores are calculated for the text semantic data based on the rating agency level of the information source, the time interval since publication, and historical accuracy. A threshold of reliability scores higher than 0.75 is set for filtering. A text knowledge graph is constructed through entity relationship extraction to obtain information knowledge data. A risk information matrix is ​​constructed based on the information knowledge data. An attention mechanism is used to identify risk-related entities in the text. The risk transmission probability is calculated based on the co-occurrence frequency and semantic distance between entities to obtain risk-related data. Early warning scores are set for risk-related data, and a Long Short-Term Memory (LSTM) network is used to predict the evolution trend of risk transmission paths. The network input dimension is the same as the number of risk entities, and the number of hidden layer nodes is set to twice the input dimension to obtain risk trend data. Based on the risk trend data, a Markov decision process is used to construct risk response rules. The risk transmission probability is dynamically updated based on the latest market evaluation information, and a risk management plan is generated by combining the risk warning threshold. In financial market text analysis, the preset information collection rules include a specific keyword list, such as rating change related words including "downgrade," "upgrade," "maintain," and "watch list," and rating outlook related words including "positive," "negative," and "stable." Taking a credit default swap issued by a company as an example, when the text "The company's debt scale continues to expand, cash flow pressure increases, and the rating outlook is adjusted to negative" is detected, the conditional random field method is used to identify "company" as the issuer entity and "negative" as the rating outlook entity. In the process of text vector construction, the term frequency-inverse document frequency (TF-IDF) method is used to calculate the word weights. Assuming the phrase "rating downgrade" appears 3 times in a document with a total of 100 words, and 10 out of 1000 documents contain this phrase, its TF value is 0.03, its IDF value is log(1000 / 10) = 2, and its final weight is 0.06. Text features are extracted using a two-layer convolutional neural network. The first layer uses 64 3×3 convolutional kernels, and the second layer uses 32 3×3 convolutional kernels to classify the text into three categories: positive, neutral, and negative. In the information reliability assessment, different weights are assigned to information from different sources; for example, information released by rating agencies has a weight of 0.9, research institution reports have a weight of 0.7, and news reports have a weight of 0.5.Regarding timeliness, the weight of information on the day of release is 1.0, decreasing by 0.05 for each subsequent day. Historical accuracy is calculated based on the accuracy of previous predictions; for example, if an information source has a past prediction accuracy of 85%, it is assigned a weight of 0.85. The weighted average of the three dimensions yields the final information reliability score. In the risk information matrix, the strength of association between entities is calculated using co-occurrence frequency and semantic distance. For instance, if "real estate" and "liquidity risk" co-occur 50 times in the text, and each appears 100 and 80 times respectively, the co-occurrence frequency is 50 / √(100×80) = 0.56. A Long Short-Term Memory (LSTM) network is used to predict risk transmission trends, with an input dimension of 20 risk entities, 40 hidden nodes, and a prediction window of 5 trading days. In constructing risk response rules, a Markov decision process is used, with a state space containing four states: normal, attention, warning, and danger. Appropriate risk management measures are selected based on the current state and market evaluation information. If market evaluation shows a sustained negative trend and the reliability score of negative information exceeds 0.8, the status will be shifted from "attention" to "warning", and the risk management plan will be adjusted accordingly.

[0030] S107. Regularly generate credit risk monitoring reports, summarizing risk assessment results, early warning situations and response measures, as well as the cumulative loss distribution under different risk contagion scenarios. Use data visualization methods to visualize the credit risk monitoring content, including time series graphs of changes in credit default swap attributes, credit risk transmission network diagrams, and probability density graphs of cumulative loss distribution.

[0031] Based on credit status data, term structure data, and liquidity data, a weighted average method is used to calculate the moving average of indicators, resulting in a risk indicator sequence. For this risk indicator sequence, a rolling time window is used for segmented statistical calculations, and hierarchical clustering is used to group the indicator features, resulting in a risk feature sequence. The risk feature sequence is then fitted using a Gaussian kernel function for kernel density estimation, and cumulative loss sample data is generated through Latin hypercube sampling to obtain risk probability data. A transmission network graph is constructed based on the risk probability data, and gravity coefficients between nodes are set. Iterative optimization is used to determine the node positions, resulting in network layout data.

[0032] Specifically, a monitoring indicator matrix is ​​created based on data on changes in credit status, term structure, and liquidity. A multidimensional array is used to store early warning records and response measures for different risk types. The moving average and rate of change of each indicator are calculated using a weighted average method to obtain a risk indicator sequence. The risk indicator sequence is then segmented using a rolling time window to calculate the mean, variance, and skewness of each indicator at different time scales. Hierarchical clustering is used to group the indicator features to obtain a risk feature sequence. For the risk feature sequence, a Gaussian kernel function is used for kernel density estimation to fit the loss distribution under different risk scenarios. Cumulative loss samples are generated using Latin hypercube sampling, and risk levels are divided into risk probabilities by setting 90%, 95%, and 99% confidence intervals. A transmission network diagram is constructed based on the risk probability data. A stress-based force-directed layout algorithm is used to set the attraction and repulsion coefficients between nodes, and the node positions are determined through iterative optimization to obtain network layout data. The network layout data is then visualized. A time-series line graph is drawn using spline interpolation, a kernel density curve is used to display the loss distribution, and a heatmap is used to display the risk transmission intensity, resulting in chart component data. A monitoring report template is constructed based on chart component data. A convolutional neural network is used to optimize the chart layout spatially, and a responsive grid system is used to adaptively layout different screen sizes to obtain the monitoring report content. In the risk monitoring report generation process, a monitoring indicator matrix is ​​first constructed. Taking a company's credit default swap as an example, its credit status indicators include credit rating, distance to default, and debt-to-equity ratio; term structure indicators include short-term spread, long-term spread, and term spread ratio; and liquidity indicators include daily trading volume, bid-ask spread, and trading frequency. A 30-day sliding window is used to calculate the mean and rate of change of each indicator. Weights are set according to the importance of the indicators, such as a credit rating weight of 0.4, a distance to default weight of 0.3, and a debt-to-equity ratio weight of 0.3. When extracting risk features, multiple time scales are used to statistically analyze each indicator, such as using 5-day, 10-day, and 20-day windows to calculate statistics. Taking the credit spread as an example, its 5-day mean is 280 basis points, the variance is 45, and the skewness is 0.8, indicating a right-skewed distribution in the short term. Hierarchical clustering is used to group similar index features together, with Euclidean distance used for clustering and average connection method employed. A Gaussian kernel function is used for loss distribution fitting. The bandwidth parameter σ is selected through cross-validation. 10,000 sample points are generated using Latin hypercube sampling, and the risk value is calculated under different confidence intervals; for example, the loss value is 350 basis points under a 95% confidence interval and 420 basis points under a 99% confidence interval. The visualization of the risk transmission network uses a force-oriented layout algorithm, with the attraction coefficient between nodes set to 0.3 and the repulsion coefficient set to 0.7. Node positions are optimized through 1000 iterations. Node size represents risk intensity, edge thickness represents transmission probability, and color intensity represents transmission speed. For the visualization of time-series data, cubic spline interpolation is used to ensure curve smoothness, with an interpolation node spacing of one trading day. In the monitoring report layout, an 8-layer convolutional neural network is used to optimize the chart layout. The input is the chart's position and size information, and the output is the optimized layout parameters. The responsive grid system uses a 12-column layout, automatically adjusting the number of columns under different screen sizes; for example, maintaining 12 columns on large screens, reducing to 8 columns on medium screens, and reducing to 4 columns on small screens, while maintaining the chart's display effect. All charts use a unified color scheme: red indicates high risk, yellow indicates medium risk, and green indicates low risk. The gradual change in color reflects the continuous change in risk.

[0033] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring and reporting of financial risks based on multi-source data, characterized in that, The method includes: S1. Obtain multi-source data on credit default swaps, preprocess and extract features from interest rate spread data for different maturities to obtain term structure feature vectors; S2, based on real-time acquired long-short interest rate spread data, derives the long-short interest rate spread curve, and performs anomaly detection on the shape changes of the curve. When an abnormal distortion of the spread curve is detected, it triggers the assessment process for the credit risk contagion effect, specifically including: S21. Obtain long-term and short-term interest rate spread data from the market database according to the preset sampling frequency, and obtain the basic interest rate spread dataset through data standardization and timestamp alignment. S22, For the aforementioned basic interest rate spread dataset, the spline function coefficients are determined using cubic spline interpolation and least squares methods to obtain continuous and smooth interest rate spread curve data; S23, calculate the rate of change of interest rate spread between adjacent maturity points for the interest rate spread curve data, and obtain the curve anomaly data by the sliding window method and the autoregressive moving average method; S24. If the abnormality data of the curve exceeds the warning threshold, calculate the risk transmission probability, establish the risk transmission matrix through the risk transmission probability, and obtain the term structure risk transmission coefficient. S3, in the assessment process of triggering the contagion effect of credit risk, uses the standardized term structure feature vector as the node attribute of the graph and the abnormal distortion of the interest rate spread curve as the edge attribute. The graph neural network is used to learn the transmission path and intensity of credit risk between different terms and market participants in order to analyze the credit risk transmission mechanism based on the graph neural network. S4. The Monte Carlo simulation method is used to analyze the credit risk transmission mechanism learned by the graph neural network, generate risk contagion scenarios driven by abnormal changes in credit default swap interest rates of different maturities, assess the cumulative loss distribution of credit risk under different scenarios, and determine the threshold and early warning level of risk monitoring indicators with key credit default swap attributes as the core. S5 dynamically monitors key credit default swap attributes. When the credit default swap spread exceeds a preset threshold, it triggers a risk warning and emergency response process and takes corresponding risk mitigation measures according to the warning level. S6. During the risk assessment and monitoring process, continuously collect market evaluations, extract information related to key credit default swap attributes, and use the information as a supplementary input for risk transmission mechanism analysis and monitoring strategies to dynamically optimize risk management strategies. S7 generates credit risk monitoring reports regularly, summarizing risk assessment results, early warnings and response measures, as well as the distribution of cumulative losses under different risk contagion scenarios. It uses data visualization methods to visualize the credit risk monitoring content, including time series graphs of changes in credit default swap attributes, credit risk transmission network diagrams, and probability density graphs of cumulative loss distributions.

2. The method according to claim 1, characterized in that, In the assessment process for triggering the contagion effect of credit risk, standardized term structure feature vectors are used as node attributes of the graph, and abnormal distortion of the interest rate spread curve is used as edge attributes. A graph neural network is used to learn the transmission path and intensity of credit risk across different maturities and market participants, in order to analyze the credit risk transmission mechanism based on the graph neural network, including: A node attribute matrix is ​​constructed based on the standardized term structure feature vector, and an edge attribute matrix is ​​constructed using the interest rate spread curve distortion data. The initial transmission graph data is obtained by using the minimum-maximum normalization method. The weighted Euclidean distance method is used to calculate the similarity score between nodes for the initial transmission graph data. If the similarity score is lower than the median similarity score of the node pairs, the node connections are pruned to obtain optimized transmission graph data. By designing a graph convolution operator using Chebyshev polynomial expansion, a feature extractor is constructed for nodes in the optimized transmission graph data to calculate the weighted features of adjacent nodes and obtain local structural features. The dot product attention method is used to obtain the attention weight matrix between local structural features. Based on the attention weights, the node features are weighted and aggregated to obtain the global transmission features.

3. The method according to claim 1, characterized in that, The method employs Monte Carlo simulation to analyze the credit risk transmission mechanism learned by the graph neural network, generating risk contagion scenarios driven by abnormal changes in credit default swap differentials of different maturities. It assesses the cumulative loss distribution of credit risk under different scenarios and determines the thresholds and early warning levels of risk monitoring indicators centered on key credit default swap attributes, including: The drift rate and volatility parameters are calculated based on historical credit spread data. A set of random fluctuation paths of credit spread is generated using the geometric Brownian motion equation, which includes a drift term and a disturbance term. For the set of random fluctuation paths, the exponentially weighted moving average method is used to calculate the dynamic fluctuation interval value. The ratio of the dynamic fluctuation interval value to the standard deviation is used to identify the credit status change point and obtain a multidimensional risk feature sequence. For the multidimensional risk feature sequence, the risk loss distribution function is fitted using the Gaussian kernel density estimation method, and the bandwidth parameter of the Gaussian kernel density estimation method is obtained through cross-validation. For the aforementioned risk loss distribution function, a generalized Pareto distribution is used to fit the tail risk characteristics, and the risk loss quantile values ​​under the confidence interval are obtained. The risk warning threshold is then determined by a hierarchical clustering method.

4. The method according to claim 1, characterized in that, The system dynamically monitors key credit default swap attributes. When the credit default swap spread exceeds a preset threshold, it triggers a risk warning and emergency response process, and takes corresponding risk mitigation measures based on the warning level, including: A monitoring indicator matrix is ​​constructed based on credit rating change data, term spread data, and market trading volume data. The matrix is ​​then dimensionality-reduced using principal component analysis, and anomaly score data is obtained by calculating Mahalanobis distance. For the abnormal score data, an exponentially weighted moving average control chart is used for anomaly detection. Multi-level early warning thresholds are set based on the weighted moving average and weighted standard deviation to obtain graded early warning data. For the aforementioned graded early warning data, a trend characteristic value is calculated using a cumulative sum control chart. The direction of indicator change is determined by the trend characteristic value. If the direction of indicator change is the same, then linked early warning data is obtained. For the aforementioned linked early warning data, a Gaussian kernel function support vector machine is used for classification. A risk knowledge graph is constructed based on the classification results. The risk response rule base is queried through the risk knowledge graph to obtain a risk mitigation plan.

5. The method according to claim 1, characterized in that, During the risk assessment and monitoring process, market evaluations are continuously collected, and information related to key credit default swap attributes is extracted. This information is used as supplementary input for risk transmission mechanism analysis and monitoring strategies to dynamically optimize risk management strategies, including: According to preset rules, market evaluation text is obtained from financial websites, and named entities in the text are extracted using the conditional random field method to obtain initial text data; For the initial text data, text vectors are constructed using term frequency and inverse document frequency, and text features are extracted through a neural network with two convolutional layers to obtain text semantic data; The text semantic data is used to calculate an information reliability score based on the rating agency level, release time interval, and historical accuracy. If the reliability score is higher than a preset threshold, a text knowledge graph is constructed by extracting entity relationships to obtain information knowledge data. For the aforementioned information and knowledge data, an attention mechanism is used to identify risk-related entities, a long short-term memory network is used to predict the risk transmission path to obtain risk trend data, and a Markov decision process is used to dynamically update the risk transmission probability.

6. The method according to claim 1, characterized in that, The system regularly generates credit risk monitoring reports, summarizing risk assessment results, early warning situations, and response measures, as well as the cumulative loss distribution under different risk contagion scenarios. It uses data visualization methods to visualize the credit risk monitoring content, including time series graphs of changes in credit default swap attributes, credit risk transmission network diagrams, and probability density diagrams of cumulative loss distributions. Based on credit status data, term structure data, and liquidity data, a weighted average method is used to calculate the moving average of the indicators, resulting in a risk indicator series. The risk indicator sequence is segmented and statistically calculated using a rolling time window, and the indicator features are grouped using a hierarchical clustering method to obtain the risk feature sequence. The risk feature sequence is fitted with a Gaussian kernel function for kernel density estimation, and cumulative loss sample data is generated by Latin hypercube sampling to obtain risk probability data. A transmission network diagram is constructed based on the risk probability data, the attraction coefficients between nodes are set, and the node positions are determined through iterative optimization to obtain network layout data.

Images