Yiyao Luo*
School of teacher education, Nanjing University of Information Science & Technology, Nanjing, China, 210044
*Corresponding author:15335147142@163.com
Abstract: The growing intricacy and unpredictability within financial markets have given rise to more and more convoluted risk – transmission mechanisms and conventional analytical approaches find it difficult to depict nonlinear dynamic correlations and high – dimensional heterogeneous data so they are unable to satisfy the requirements for precise identification and early warning, thus this article put forward an analytical framework that combines big data with graph neural networks (GNNs) which builds a dynamic financial correlation graph on the basis of multi – source heterogeneous data and by integrating the evolutionary graph attention mechanism with LSTM, this model seizes the temporal evolution features of the network making it possible to identify and predict risk – transmission paths and a real – world dataset was used to validate the model with findings indicating that the proposed model attained an AUC of 0.862 which was 0.081 higher than that of GAT and it precisely portrayed the timing and intensity of risk – transmission and effectively identified key nodes and paths and research showed that this fusion method could enhance the accuracy and timeliness of risk identification, offer technical support for the development of dynamic early – warning and prevention strategies and have considerable value for financial regulation.
Keywords: Financial Risk Transmission; Big Data Analysis; Graph Neural Networks; Risk Modeling; Risk Early Warning; Network Structure; Risk Prevention
1 Overview of Financial Risk Transmission Mechanisms
1.1 Basic Concepts of Financial Risk Transmission
Financial risk spread pertains to the procedure where a local risk incident within the financial system proliferates and is conveyed to the whole system via the interconnections among financial institutions, financial markets, and financial products, which includes the operational perils of a single institution as well as cross – market, cross – industry, and cross – border risk spread mechanisms across diverse markets, and it can be categorized into direct and indirect spread according to different spread means, with direct spread mainly taking place through channels like asset – liability relations, credit chains, and investment linkages between financial institutions while indirect spread depends on elements such as market sentiment, information diffusion, and policy alterations to scatter the risk, and since modern financial markets have constituted an extensive market system covering all regions and multiple business sectors, the speed and extent of risk spread based on this system have far surpassed former situations, so merely relying on the risk prevention steps of individual financial institutions was unlikely to prevent the emergence of systemic risk.
Early research regarding the modeling and analysis of transmission mechanisms mainly depended on traditional econometric approaches like linear regression, cointegration, and VAR, which centered on linear transmission routes and static network configurations and were unable to precisely capture the many nonlinear dynamic features and complex network topologies that exist in financial systems; in the past few years, the growth of big – data technology has made it possible to more elaborately describe the interactions among financial entities, providing a firm basis for constructing dynamic and heterogeneous financial networks, so GNNs, because of their capacity to readily deal with non – Euclidean data, have been more and more applied to the financial risk transmission mechanism and have had some success in modeling and forecasting risk transmission routes by using node characteristics and graph information [1], but since most of the early studies modeled graphs at fixed time instants, the resulting risk transmission routes were likely to have lagged transmission and often could not represent the true risk transmission route; with the growing attention to the time – series dimension, many techniques and accomplishments related to time – series graph neural networks have emerged in recent years, which are very important for understanding how to add time – series information to explain risk evolution, where the time axis can be directly included in the graph network as a special feature, combining it with the instantaneous state of the financial system, and then we enter a real – world situation to finish our study of the dynamic transmission patterns of financial risk [1].
1.2 Traditional Risk Transmission Analysis Methods
Research on financial risk spread commonly employs techniques like linear regression models, VAR models, PCA, and static graph models founded on complex networks and these techniques have been of great significance in exploring the paths and intensity of risk spread within financial systems while early financial risk identification and warning systems frequently make use of linear regression models, VAR models, PCA for factor extraction, and static graph models based on complex networks and in terms of the VAR model, the spread of risk shocks could mirror the dynamic response relationships among different financial variables and PCA for factor extraction was able to disclose the main driving forces behind systemic risk and static graph models of complex networks were utilized to build the risk network topology, figure out whether each node was abnormal, and identify the nearest neighbor among all abnormal nodes.
However, because of the inborn complexity and interrelatedness within financial markets, traditional approaches struggle to model the intricate and dynamically varying structures of these financial networks, as on the one hand, linear models are unable to depict the non – linear relationships among financial entities, thus lacking an exact understanding of risk transmission routes and on the other hand, static graph models can’t seize the features of financial markets as they develop over time, so their predictive ability is greatly diminished, particularly when confronted with sudden occurrences or substantial changes in the market’s immediate structure, and traditional complex network risk simulations rely on fixed topological structures, overlooking the temporal differences between nodes during different time frames, consequently, in financial markets, the actual risk transmission often varies considerably from what our traditional simulations suggest.
More seriously, in view of the wide variety of financial products and the growing interconnectedness among financial market products and markets, risk transmission displays obvious multi – level, cross – market, and cross – asset features and traditional mathematical models find it hard to quickly model and calculate heterogeneous, high – dimensional data; in regional financial reforms such as in the Changjitu Development and Opening – up Pilot Zone, where capital flows are more frequent and market structures are more complex, the existing models are not easy to be used for analyzing regional financial risk transmission so only the modeling methods that are effective for both spatiotemporal evolution and high – dimensional systems could meet the current requirements of risk transmission analysis and accurately and timely capture risk transmission.
In brief, though the traditional analytical approaches to financial risk transmission lay a theoretical groundwork, they can’t depict the intricate dynamic features, non – linear aspects, and high – dimensional nature of financial systems, so they fall short of fulfilling the current development demands of financial markets and the risk analysis and early – warning requirements of regulatory bodies, thus future theoretical research on financial risk transmission mechanisms ought to make use of new technologies like big data and graph neural networks for expansion.
2 Application of Big Data in Financial Risk Analysis
2.1 Big Data Technology and Financial Risk Monitoring
The utilization of large – scale data technology in financial risk monitoring has radically transformed procedures like financial risk identification and early warning, as financial markets grew larger and more complex over time and their risk transmission routes showed complex, non – linear, and dynamic features, so conventional risk warning approaches relying on static financial indicators could no longer meet the requirements of modern financial regulation and risk prevention, while large – scale data, when combined with diverse data from multiple sources, such as transaction data, social media, corporate financial data, and macroeconomic data, was able to depict the systemic risk profile of the entire financial market more comprehensively and accurately, enabling real – time tracking and tracing of risk information as well as early warning, and it was a new method to combine deep learning with graph neural networks to model risk transmission routes, with the semantic reasoning and relationship extraction techniques based on graph neural networks effectively disclosing potential risk correlations concealed behind complex financial events [4].
When handling financial text information, context – aware semantic encoding was carried out by means of pre – trained language models (RoBERTa) to enhance the semantic comprehension of Chinese financial corpora, and based on this, a global graph structure made up of document nodes and entity nodes was constructed, which enhanced the capacity to express relations among entities, dealt with cross – sentence argument dispersion and long – distance dependencies, further strengthening the ability of the structured information extraction technology to automatically identify financial risk events and making it possible to detect risk signals early and perform structural modeling [3], and moreover, in dynamic financial scenarios, user behaviors and social relationships were constantly changing over time, thus forming dynamic graph structures with temporal characteristics, and dynamic graph neural networks could utilize the evolving graph attention mechanisms to capture the temporal evolution of graph structures, which played a certain auxiliary role in predicting user default behaviors, and by combining graph attention weights with the hidden state of a time – series LSTM, more effective dynamic model adaptability could be achieved, enabling the real – time modeling of financial risk transmission pathways [4].
To sum up, big data technology was able to augment the data dimension in financial risk monitoring and the combination of big data technology and graph neural networks could enhance the intelligence of financial risk warnings as well, this technological integration also furnished a theoretical foundation and a practical approach for constructing a financial risk warning system with dynamic perception capabilities, which had a certain significance for the advanced area of risk prevention and control in financial mathematics.
2.2 Data-Driven Risk Transmission Modeling
Beyond the task of modeling financial risk transmission, the existing statistical models and static network – based approaches find it hard to fully describe the transmission routes and mechanisms of financial risk in multi – layered financial systems considering the growing complexity and instability of the financial system, the rapid growth of big data technology has made it possible to identify risks and build models within the financial system, in such a situation, graph neural networks (GNNs), which have the capacity to represent complex networks, are an efficient tool for depicting financial risk transmission mechanisms, in comparison with previous models, GNNs driven by big data can not only extract multi – source heterogeneous financial data and get a network – related relationship map between interconnected institutions and market participants, but also represent complex network – structured data in a vector form.
In the process of modeling risk transmission, the initial step was to set up a financial correlation graph using data regarding bilateral transaction relations among banks, supply – chain relations among enterprises, equity structure relations, and public opinion to construct a semantically rich dynamic financial network, then Graph Attention Networks (GATs) or dynamic graph neural networks which employed an evolving graph attention mechanism were utilized to seize the temporal dependencies and the evolution of relations between nodes which helped in characterizing the risk transmission routes that developed over time among financial institutions [4], for instance, in offline lending situations, the structure of a user’s social graph showed significant temporal evolution and the influence of changes in user relations on default risk couldn’t be grasped by using conventional static graph models, so combining a long short – term memory network (LSTM) with an attention mechanism to form a dynamic GNN model could better describe the evolution of the social graph structure and further enhance the capacity to predict default risk [4].
Furthermore, in order to deal with the issue of extracting risk information within financial text data, the combination of pre – trained language models and global graph neural network approaches could not only resolve the problems of long – range dependencies and difficulties in cross – sentence reasoning but also transform unstructured texts like financial news, regulatory reports, and social communications into entity nodes and document nodes via semantic encoding and on this basis establish a heterogeneous graph structure, allowing the model to more comprehensively understand the semantic relationships among risk events and thus attain the aim of more precise identification of risk transmission paths.
In order to present the specific process of risk transmission modeling based on big data and graph neural networks more intuitively, a structured representation was specially produced.
Figure 1 Risk transmission modeling process based on big data and graph neural networks
3 Graph neural networks and financial network modeling
3.1 Basic principles of graph neural networks
Graph Neural Networks (GNNs), which were a significant advancement in deep – learning in recent years, were created to handle non – Euclidean data having complex topologies, like risk – transmission networks in financial systems and were especially suitable for depicting highly interconnected risk – transmission networks within financial systems, and the fundamental idea of GNNs was to incorporate information regarding nodes (financial institutions, trading entities, etc.) and edges (capital flows, business transactions, etc.) within a graph structure via representation learning and this information was then combined with the state information of neighboring nodes to keep updating the node’s own state representation, finally resulting in a representation of the graph data, and in contrast to traditional neural networks, GNNs could be utilized to find high – level interactions between any two points in unordered graph data and were also very effective in identifying systemically important financial institutions and the transmission paths of financial risks in financial markets, and in financial mathematics, GNNs had been employed in financial – crisis early – warning models based on information – spillover networks, and by building dynamic industry – correlation maps, they enhanced the model’s sensitivity to systemic risk [5].
The fundamental framework of a Graph Neural Network (GNN) generally encompasses crucial steps like the initialization of node features, the aggregation of neighborhood information, and the update of states and if the graph structure was G = (V, E), with V standing for the set of nodes and E for the set of edges and every node V i had an initial feature vector h i, during the propagation at the t – th level, the formula for updating the node state was.
Where (i) stands for the neighbor set of node i, is the activation function, W (t) and b (t) are parameters that can be learned, and AGGREGATE refers to the aggregation function which is usually carried out by means of mean, max pooling, or attention mechanisms, and the above – mentioned mechanism was applied to GNN models for the analysis of financial risk transmission, in this case the model was utilized to build cross – market risk spillovers among entities, and when a network model was constructed on the basis of Elastic – Net – VAR, GNNs could extract key nodes and edges with high weight coefficients.
Besides the traditional graph convolutional networks (GCNNs), numerous variants of GNNs such as Graph Attention Networks (GATs), Graph Isomorphism Networks (GINs), and Gated Graph Neural Networks (GGNNs) were developed with the concept of GCNNs as their basis and all of these were utilized to model financial graphs whether they were heterogeneous graphs or time – series graphs, for instance in the task of extracting events from Chinese financial text, the RoBERTa encoder was used to extract document – level semantic representations and a global graph structure including both entity nodes and document nodes was used to carry out cross – sentence semantic reasoning via a GCNN which strengthened long – range dependencies and effectively solved the long – range dependency issue [3], additionally based on the directed – graph modeling of contact – relationship graphs in online lending, a new structural model that combined the improved GraphSAGE and GAT structures was fed into a GNN to model user overdue risk which made people realize that user creditworthiness could be evaluated not just by the aforementioned methods.
3.2 Methods for Constructing Financial Networks Based on GNN Methods
In the realm of financial risk transmission analysis, creating highly precise dynamic financial networks is crucial for the discovery and warning of systemic risks, but conventional financial network modeling approaches rely on linear correlation or causal inference at a single moment and are unable to depict the nonlinear and dynamic interplays among multiple entities within complex financial systems, while graph neural networks (GNNs) are a favored model for graph – structured data modeling so GNN – based modeling methods can effectively illustrate and represent the nonlinear interactions and dynamic impacts among multiple entities in complex financial systems, thus they can better mirror the elements influencing financial risk transmission, enhance the comprehension of the rules governing financial risk transmission, and reduce the occurrence of financial risks [5].
Taking the industry as an example, the use of the Elastic – Net – VAR model was able to extract multi – dimensional information spillovers such as the connection among risk spillover, volatility spillover, and yield spillover, thus building a financial risk transmission network with economic explanatory capacity [5], and based on this, a dynamic risk propagation path model was constructed by using gated graph neural networks (GGNNs) and graph isomorphism networks (GINs) which improved the prediction of systemic financial risk. For individual – level information like the credit rating of internet finance users, graph neural networks were applied to user credit ratings, and by creating a complex network of user lending relationships and including structural characteristics such as node centrality, graph neural networks combined information from adjacent circles and lending behavior data, greatly enhancing the predictive accuracy of credit scoring models, and meanwhile, as social relationship networks are more about graph data like address book relationship structures, the application of graph neural networks has also been proven in this case as it could unearth potential risk signals hidden beneath such unnoticeable connections and enhance the stability and generalization ability of overdue risk prediction.
In order to present the technical process of GNNs in building financial networks more directly, a GNN modeling flowchart founded on multi – source financial data was offered (Figure 2), precisely, commencing from the raw data, the process successively carried out the steps such as data pre – processing, graph structure construction, feature extraction, model training, and evaluation, and finally led to a final risk transmission network.
Figure 2 Schematic diagram of the GNN modeling process based on multi-source financial data
4 Model construction and analysis of financial risk transmission
4.1 Risk Model Combining Big Data and GNNs
Earlier attempts at financial risk transmission modeling mostly depended on conventional risk assessment techniques, but these traditional risk assessment approaches, mainly founded on structured data and linear models, had difficulty in revealing the intricate, nonlinear, non – Gaussian, and dynamically changing networks within complex financial systems. As big data developed and a large amount of graph – structured data emerged, graph neural networks (GNNs), which have attracted a great deal of attention in recent years, are now a crucial means for analyzing and resolving complex, interconnected problems in financial systems, and by employing the concept of GNN, financial entities like users, enterprises, and financial institutions are abstracted into nodes in the graph, and the relationships among them such as transactions, credit, and social relations are represented by the edges in the graph, so that the risk transmission path and the impact mechanism of risk transmission can be effectively realized. This article presents how to utilize big data and graph neural networks to build a financial risk transmission model through the development of actual cases, and moreover, dynamic graph evolution and the attention mechanism are incorporated into the model to achieve the ability to perceive temporal information and enhance the capacity for risk transmission and risk prediction.
This model framework mainly includes an evolving graph attention mechanism and in Graph Attention Network (GAT), the weight parameters and attention coefficients are presented in the hidden states of the Long – Short – Term Memory (LSTM), which enables the model to update the parameters of the LSTM in real – time according to the changing time series, and the LSTM takes the adjacency matrix at each moment as input, and the hidden states within it dynamically update the GAT weight parameters to show the characteristic changes in the graph structure over time, and it also includes graph convolution and information propagation mechanisms to imitate the process of risk spreading between nodes, and moreover, during the model training stage, nonlinear filters are used for parameter estimation, and adaptive methods are added to the estimation of the noise covariance matrix to enhance the model’s robustness and convergence speed.
In order to examine the model’s efficacy in gauging financial risk transmission, we carried out simulations using an experimental data set sourced from a genuine offline consumer finance platform which encompassed user information application forms, past repayment behaviors, and de – identified correspondences from user contact lists and we mapped the user call records onto the edges of a graph and modeled the temporal development of these relationships within the time – series dimension to get the default probability for each node in the next instant and in comparison with the benchmark models, namely XGBoost, GraphSAGE, and GAT, the AUC was enhanced by 0.12, and there were different extents of improvement in the F1 – score and KS values and Figure 3 presents the overall model structure.
Figure 3 Framework of financial risk transmission model integrating big data and graph neural networks
During the experiment, the model’s parameter configuration and performance are presented in the following table.
Table 1 Parameter configuration and performance of the model
| Model type | AUC | F1-score | KS value | Training time (minutes) | Number of parameters (10,000) |
| XGBoost | 0.742 | 0.681 | 0.39 | 15 | 0.2 |
| GraphSAGE | 0.763 | 0.702 | 0.41 | 28 | 1.5 |
| GAT | 0.781 | 0.724 | 0.44 | 35 | 3.2 |
| Model in this paper | 0.862 | 0.803 | 0.53 | 42 | 4.1 |
Furthermore, the crucial formula of the developing graph attention mechanism within the model is presented as follows.
Among them, represents the attention coefficient of node i with regard to its neighboring node j at time t, is the weight matrix updated through the LSTM, and are the embedding vectors for nodes i and j respectively, a is a learnable attention parameter vector, and | stands for a vector concatenation operation, and this formula showed the model’s capacity to carry out weighted aggregation of the relationships between nodes in a dynamic graph.
In brief, the risk model put forward, which combines large – scale data with graph neural networks, in theory broadens the approach of dynamic graph modeling and shows excellent performance in practical tasks of predicting financial risk transmission and presents a new way of modeling for the domain of financial mathematics and gives valuable perspectives to other areas.
4.2 Simulation Analysis of Risk Transmission Paths
A model that can genuinely mirror the dynamic development of intricate financial systems is of great significance for the simulation and analysis of financial risk transmission routes. In the past, such models mainly centered on risk transmission routes constructed via traditional static graph structures but as they overlooked their evolutionary characteristics, they had difficulty in precisely simulating the risk transmission process which lessened the foresight and accuracy of the model. Dynamic Graph Neural Networks (DGNNs), which have been developed in recent years, can effectively integrate temporal evolution patterns into the modeling process thus better uncovering the risk transmission within the financial system. In offline consumer finance default prediction tasks, the dynamic graph formed by users’ social relationships intrinsically contains nonlinear risk transmission routes ranging from individual social capital to the partially shared spillovers of the entire system and if an evolving graph attention mechanism were incorporated on this basis, the model could extract these nonlinear risk transmission routes more effectively and depict the mechanisms and effects of risk transmission more accurately.
To enhance the precision of risk transmission simulation further, the parameter estimation process in Graph Neural Networks (GNN) was improved by means of nonlinear filtering as traditional neural network training algorithms had shortcomings like low estimation accuracy and slow convergence in nonlinear and non – Gaussian noise settings, so an enhanced particle filter algorithm, namely APF – IEKF (Auxiliary Particle Filter – Iterative Extended Kalman Filter), was introduced which combined an auxiliary particle filter with an iterative extended Kalman filter and when formulating the importance density function, it utilized the current observation information to define the filtering and prediction variances, ensuring that the particle distribution was close to the true posterior distribution, thus making the model more sensitive to the changes in the dynamic graph structure.
Drawing on the developing graph attention mechanism, this research put forward a means of depicting the route of risk spread in networks via dynamic graph neural networks by fashioning the graph attention mechanism weight parameters and attention parameters into the states of the hidden layers of the LSTM network so that the model was able to adaptively seize the temporal alterations in the graph structure, in contrast to the initial method, this method was more effective for molecular structures and attained fairly good outcomes without the need for extra manual interference and experimental confirmation showed that this model outperformed traditional methods considerably in emulating risk transmission paths, especially in portraying the temporal sequence of risk transmission from one node to another and the strength of risk transmission, and the following table makes a comparison of the performance of this model across a number of financial risk transmission tasks.
Table 2 Performance Comparison of Models in Multiple Financial Risk Transmission Tasks
| Model Name | Accuracy (%) | AUC Value | F1 Score | Recall (%) | Specificity (%) | Average Loss |
| GCN | 82.1 | 0.832 | 0.798 | 78.5 | 85.3 | 0.412 |
| GAT | 84.6 | 0.857 | 0.819 | 80.2 | 87.1 | 0.389 |
| DCRNN | 86.4 | 0.875 | 0.836 | 82.4 | 89.0 | 0.362 |
| Our Model (EGAT-LSTM) | 89.3 | 0.901 | 0.867 | 85.6 | 91.8 | 0.321 |
Table 2 shows that compared to the GCN, GAT, and DCRNN models, the proposed EGAT-LSTM model not only achieves higher accuracy but also higher AUC and F1 scores. Furthermore, when compared to other models with the same or similar evaluation metrics, our proposed method exhibits a significant advantage in recall. Compared to other models, with the exception of MCN, our model achieves the highest specificity, a high accuracy at low loss, and the lowest average loss.
This paper employed a relationship graph to show how financial risk was transmitted to diverse nodes within a dynamic graph, every node in the graph stood for a financial entity (such as a bank, a company, or an individual) and edges were created from capital flows or investment relations among entities, thicker edges signified greater risk or transmission risk, moreover, graph neural network embeddings could be utilized to figure out which nodes were crucial risk transmission points, or the most harmful or fragile nodes, along with their risk transmission routes.
Figure 4 Financial risk transmission pathways in a dynamic graph
At the model – theory level, the risk – transmission simulation framework put forward in this paper combined a dynamic graph attention mechanism with nonlinear filtering techniques and its mathematical formulation was as follows:
Among them, represents the concealed state of the node at a particular time , represents the set of its neighboring nodes at that time, and is a parameter that can be learned within the graph attention mechanism, and this formula allowed for the dynamic modeling and real – time prediction of financial risk transmission routes, providing a theoretical foundation and technical support for financial supervision and risk prevention.
5 Research on Financial Risk Prevention and Control Strategies
5.1 Model-Based Risk Early Warning System
In the design and implementation of financial risk early – warning mechanisms, model – driven methods have drawn attention as they can effectively depict the dynamic features of complex financial systems, but as financial markets show more and more prominent nonlinearity, non – Gaussian nature, and high – frequency fluctuations, traditional risk early – warning models based on linear assumptions and static graph structures are no longer sufficient, so models that can reflect time – varying and nonlinear development were required to better identify sudden or potential risk factors in risk early – warning.
Dynamic Graph Neural Networks (DGNNs), which deal with the issue of time – varying graph structures, are able to cut down on noise in nodes and edges that keep changing all the time and thus get effective and valuable information, are highly fitting for financial situations where the relationships among entities are in a constant state of change over time and place; in the consumer finance field, the connection between users can be figured out according to their social networks and even if an order is regular, alterations in the borrowing intentions of other users or abnormal conditions can result in risk spreading, so this project takes the consumer finance field as an example; a dynamic graph neural network model based on the evolving graph attention mechanism makes the graph attention weights and parameters the hidden states of a long – short – term memory (LSTM) network to catch the temporal features of the graph structure and enhance the capacity to spot default risk [4], which allows for the setting up of a risk warning system that goes along with the development of underlying things.
Besides setting aside funds for the operation of other systems, this project also looks into the applicability and robustness of nonlinear filtering techniques in financial time – series modelling and state estimation, and taking full account of the non – Gaussian noise, time – varying volatility, and jump features of financial data, it made use of the benefits of particle filtering or the APF – IEKF algorithm to deal with the state estimation problem of nonlinear and non – Gaussian systems and applied it to parameter estimation and volatility modelling of financial market microstructure models [8], and based on this research, this paper put forward a method that combined nonlinear filtering with dynamic graph neural networks to model the hidden state of the financial system and the risk transmission paths within the financial system, thus enhancing the dynamic response ability and the accuracy of risk warnings, and the combination of nonlinear filtering and dynamic graph neural networks (DGNNs) in financial systems could be used to model the underlying state of the financial system and the risk transmission paths, strengthening the dynamic response capabilities and the predictive accuracy of risk early – warning systems.
Dynamic graph neural networks are capable of depicting the dynamically varying interconnected structures among financial institutions and market participants and by employing enhanced nonlinear filtering techniques, real – time estimation and updating of market state variables (such as volatility, jump intensity, and leverage effect) can be accomplished, so the combination of these two approaches could simultaneously capture local risk transmission routes and the overall trend of system – risk evolution and in markets with heavy – tailed distributions and asymmetric fluctuations, integrating MCMC parameter – estimation methods into market microstructure models based on the Student’s t – distribution would more precisely reflect asset – price dynamics and strengthen the robustness of early warnings.
In sum, the model – driven risk early – warning mechanism that combines dynamic graph neural networks and nonlinear filtering approaches has great theoretical and practical value and could be utilized not only to deal with the intricate dynamic features of financial systems but also to trace and forecast risk transmission routes in real – time within uncertain settings, offering crucial references for the scientific execution of financial risk prevention and control.
5.2 Policy and Risk Prevention
Concerning the design of financial risk prevention and control policies, it is crucial to concentrate on transforming and applying the most recent research outcomes in large – scale data and graph – based neural networks to enhance the precision and foresight of risk identification and transmission routes, previous research showed that dynamic graph – based neural networks (DGNNs) were more responsive to the alterations in the relationships among financial entities as time passed and had substantial merits in foretelling risks like default or fraud, demonstrating remarkable performance in default prediction and fraud detection, thus regulators and financial institutions should be spurred to intensify their attention on dynamic graph – based modeling and set up financial relationship networks that were capable of continuous updating and learning, allowing them to have a better comprehension of the risk transmission routes within the system.
Furthermore, research demonstrated that the utilization of time – series graph neural networks was able to incorporate the temporal structure of financial events into risk prediction models, enhancing their sensitivity to sudden risk events within complex financial systems [1], thus a real – time risk monitoring system based on graph neural networks might be introduced into financial market regulation by continuously constructing dynamic graph structures, like transaction behaviors and credit relationships among financial entities across different time dimensions to identify high – risk transmission nodes and paths, allowing for timely early warning and intervention.
Furthermore, since financial risk transmission frequently goes beyond market and institutional limits, a cross – market risk information sharing mechanism ought to have been established at the policy level and by means of graph neural network technology and graph databases, a unified risk map across institutions and markets could be constructed which was highly effective for pinpointing the sources of systemic risk, providing precise early warnings prior to the occurrence of risk events, and guiding regulation; additionally, in situations like offline consumer finance, users’ social relationships were constantly changing and if relevant institutions had been guided to notice and detect the influence of changes in user relationships on potential defaults, and thus strengthen the construction and maintenance of user relationship graphs, risk assessment models would have been helped to better adapt to the changing risk requirements [4].
At the institutional level, enhance the strengthening of the financial data governance framework to guarantee that the multi – source heterogeneous data demanded by graph neural networks are of high – quality, standardized, and up to the requirements and when dealing with relevant user social relationship data, pay attention to the desensitization and access control of such data to ensure the security and ethical compliance of the data during model training and utilization and it is necessary to intensify the promotion of interdisciplinary research like financial mathematics and artificial intelligence, strengthen the innovation in the basic theory and engineering of graph – neural – network – related algorithms, and lay a solid foundation for financial risk control.
References
[1] Qi Jingjing. Research on the systemic risk transmission mechanism and modeling based on the multi-layer financial network between banks and enterprises [J]. Finance, 2019, 009(4): P.350-364
[2] Qi Lili, Sun Kena. Application of neural network in the modeling of commercial bank financial risk assessment system [J]. Journal of Tianjin Vocational and Technical Normal University, 2004, 14(4): 46-49
[3] Song Lingyun, Ma Zhuoyuan, Li Zhanhuai, Shang Xuequn. A review of temporal graph neural network for financial risk prediction [J]. Journal of Software, 2024, 35(8): 3897-3922
[4] Ren Yinghua, Jiang Jinfeng, Ni Qingshan. Research on financial crisis early warning based on graph neural network model – from the perspective of information spillover across industries [J]. Statistical Research, 2022, 39(8): 141-160
[5] Ning Pengxiang. Research on financial fraud detection method based on graph neural network [D]. Donghua University, 2023
[6] Wang Guoliang. Research on dynamic graph neural network and its application in offline consumer finance default prediction [D]. Shanghai University of Finance and Economics, 2022
[7] Zhou Yixuan. Research on overdue risk prediction of Internet financial users based on graph neural network [D]. Shanghai University of Finance and Economics, 2022
[8] Li Sidi, Guo Binghui, Yang Xiaobo. Research on financial credit investigation based on graph neural network [J]. Computer Science, 2021, 48(4): 85-90
[9].Reports Outline Networks Study Findings from Tianjin University(Text-rich Graph Neural Networks With Subjective-objective Semantic Modeling)[J].Network Daily News, 2024, (5):45-45
[10].Recent Studies from University of Virginia Add New Data to Networks(Multifideli ty Graph Neural Networks for Efficient and Accurate Mesh-based Partial Different ial Equations Surrogate Modeling)[J].Network Daily News, 2024, (23):56-57
[11]Zhang Z., Liao Z., Tian W.Yuan Z.Lu C..Graph neural network-based surrogate modelling for real-time hydraulic prediction of urban drainage networks[J].Water research:A journal of the international water association, 2024, 263(1):1.1-1.13
[12] Hu Jie, He Wei, Zeng Zhangfan. Document-level Chinese financial event extraction based on global graph neural network based on RoBERTa [J]. Journal of Chinese Information Science, 2023, 37(2): 107-118
[13]Danhao Zhu.PRGNN:Modeling high-order proximity with relational graph neural network for knowledge graph completion[J].Neurocomputing, 2024, 594(14):127857.1-127857.8
[14]Wang Ting-Yun, Chen Chiao-Ting, Huang Ju-ChunHuang Szu-Hao.Modeling Cross-session Information with Multi-interest Graph Neural Networks for the Next-item Recommendation[J].ACM transactions on knowledge discovery from data, 2023, 17(1):1.1-1.28
[15] Xi Yanhui. Nonlinear filtering algorithm and its application in neural network and financial market modeling [D]. Central South University, 2013
[16] Liu Yixin, Zhang Dingli. Research on the construction strategy of Changjitu regional separation type international financial market [J]. Contemporary Economic Research, 2014, 0(11): 73-77
[17] Li Junfeng, Li Huilin. Financial support strategy for the construction of Changjitu development and opening pilot zone [J]. Jilin Financial Research, 2010, 12(3): 60-62
[18]Roesch Moritz, Nolde Michael, Ullmann TobiasRiedlinger Torsten.Data-Driven Wildfire Spread Modeling of European Wildfires Using a Spatiotemporal Graph Neural Network[J].Fire, 2024, 7(6), 112-117.
[19]Gallos Ioannis K., Lehmberg Daniel, Dietrich FelixSiettos Constantinos.Data-driven modelling of brain activity using neural networks, diffusion maps, and the Koopman operator[J].Chaos, 2024, 34(1):223-248.
[20]Sondip Poul Singha, Md.Mamun Hossain, Md.Ashiqur RahmanNusrat Sharmin.Investigation of graph-based clustering approaches along with graph neural networks for modeling armed conflict in Bangladesh[J].International Journal of Data Science and Analytics, 2024, 18(2):187-203