A systematic risk assessment method based on complex financial networks
By constructing a guarantee network model and combining attention neural networks and convolutional neural networks for multi-level feature learning, the accuracy and scale issues of risk assessment in large-scale guaranteed loan networks are solved, and efficient identification and monitoring of systemic risks are achieved.
Patent Information
- Application Number
- CN202210162187.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-22
- Publication Date
- 2025-09-26
- Estimated Expiration
- 2042-02-22
AI Technical Summary
Existing technologies have obvious limitations in analysis scale and accuracy when evaluating large-scale secured loan networks, making it difficult to accurately identify and monitor systemic risks. In particular, the financial status of small and medium-sized enterprises is difficult to measure through financial statements, resulting in complex risk transmission and affecting financial security.
A guarantee network model is constructed based on the breadth-first search algorithm, and multi-level feature learning is performed by combining intra-chain and inter-chain attention neural networks. The infection chain and risk rating are dynamically updated, and a convolutional neural network is used for systematic risk assessment.
It significantly improves the accuracy and effectiveness of predicting systemic risks in complex financial loan networks. It has high applicability and flexibility. The model prediction results are better than empirical judgment, can identify key enterprises and contagion paths, and improve the accuracy and interpretability of risk assessment.
Smart Images
Figure CN114549176B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of financial risk technology, and in particular to a systemic risk assessment method based on complex financial networks. Background Art
[0002] Small and medium-sized enterprise loans play a vital role in promoting technological innovation, boosting economic development, improving people's livelihoods, and increasing employment. Currently, due to the high credit assessment standards of commercial banks, small and medium-sized enterprises (SMEs) struggle to meet these requirements. To meet these standards, many SMEs resort to mutual guarantees to secure bank credit, forming a complex guarantee network. When a borrower defaults on their loan, the risk is transmitted through the guarantee network. This cascading risk creates potential systemic risks and poses significant challenges to national financial security and regulation. Compared to existing independent risk assessments of individual enterprises, analyzing risk from a contagion perspective can identify high-risk guarantee groups. This systemic perspective, identifying the impact of individual risks on the system, can provide early warning of potential correlated risks and provide a basis for monitoring these risks.
[0003] Since the 1950s, the banking industry has been developing risk quantification models. Subsequently, the Basel Committee issued a series of recommendations on banking supervision (the Basel Accords). Banks worldwide have widely adopted the Basel Accords, and some of the indicators proposed by the Basel Accords (such as probability of default, loss given default, and exposure to default) have been adopted by most banks as risk management metrics. However, these methods rely heavily on the financial statements of individual companies and often overlook the impact of defaults by other companies in the network. Furthermore, because the financial health of small and medium-sized enterprises is difficult to measure through publicly disclosed financial statements, monitoring the status of these loan networks becomes a complex undertaking.
[0004] To address the challenge of loan risk assessment, the industry has introduced numerous deep learning-based techniques. For example, recursive and graph neural networks have been used for enterprise default prediction and automatic feature learning. However, these methods are more suitable for assessing individuals and small businesses. When it comes to analyzing large-scale secured loan networks, these methods face significant limitations in scale and accuracy. Therefore, more accurate and interpretable models are needed to detect risks and prevent systemic crises. Summary of the Invention
[0005] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies in terms of obvious limitations in analysis scale and accuracy when facing large-scale secured loan networks, and to provide a systemic risk assessment method based on complex financial networks, which will help financial regulators and risk managers to better track and understand risk contagion and systemic risks in loan networks.
[0006] The purpose of the present invention can be achieved by the following technical solutions:
[0007] A systemic risk assessment method based on complex financial networks, comprising the following steps:
[0008] S1. Based on the breadth-first search algorithm, obtain the target enterprise's status data and guaranteed loan information, establish a guarantee network model, and construct all related transmission chains, and dynamically update the guarantee network model and transmission chains;
[0009] S2. Based on the constructed infection chain, an intra-chain attention neural network and an inter-chain attention neural network are established, and a multi-level feature learning layer is used to calculate the potential risk information of the infection chain and risk rating;
[0010] S3. Establish a convolutional neural network model to obtain the risk level assessment results of the target enterprise based on the potential risk information of the infection chain and risk rating.
[0011] The formula of the guarantee network model G is as follows:
[0012] G=(V,E)
[0013] Where V={v1,v2,…,v |V|} is a node set, each node represents the small and medium-sized enterprises in the guarantee network model, E = {e1, e2, ..., e |E|} is an edge set, each edge represents the guarantee relationship between nodes, |V| represents the number of nodes, and |E| represents the number of edges in the guarantee network model.
[0014] Furthermore, the status data of the target enterprise includes registered capital, number of employees and financial status, which are embedded in the attributes of the nodes of the guarantee network model; the guaranteed loan information of the target enterprise includes the guaranteed amount, loan amount and loan interest, which are stored in the attributes of the edges of the guarantee network model.
[0015] Furthermore, all relevant infection chains in step S1 form an infection chain set, and the specific formula is as follows:
[0016] C={c1,c2,…,c |C|}
[0017] c i ={(v,e)|v∈Diff(v i ,G)∪e∈Diff(v i ,G)}
[0018] Among them, C is the set of infection chains, and the infection chain c i is node v iThe maximum risk diffusion range, Diff(v i ,G) indicates that from v i Edges passing through the guarantee network model G.
[0019] Since the guarantee contract has a validity period, in step S1, the guarantee network model and the infection chain are dynamically updated during the training and prediction stages.
[0020] Step S1 also involves first connecting bank loan records by key value through a query operation to reconstruct an initial guarantee network model. During the feature learning process, since guarantee relationships are dynamic, if a guarantee is issued, a temporary edge e is added to the network G; when the guarantee expires, this edge is deleted. Finally, the attributes of the nodes and edges in the set are used as features of the infection chain.
[0021] The potential risk information of the infection chain and risk rating in step S2 includes node-level risk characteristics, chain-level risk characteristics, and network-level risk characteristics.
[0022] Furthermore, the node-level risk features include node features and edge features, which are calculated by the attribution graph learning layer to obtain the given node v i Individual characteristics The specific formula is as follows:
[0023]
[0024] in, is the node feature, is the edge feature, W I and a I is the trainable weight, NN I represents a fully connected neural layer in a single feature learning module, is node v i Neighborhood set of , || represents and operation.
[0025] Furthermore, the chain-level risk characteristics It is calculated through the in-chain attention neural network. The specific formula is as follows:
[0026]
[0027]
[0028]
[0029] Among them, node v j Indicates v sharing the same infection chain iNeighbor nodes, LeakyReLU(·) represents the nonlinear transformation of LeakyRectified Linear Unit, W c is the weighted matrix for learning by the in-chain attention neural network, is the attention weight of the in-chain attention neural network, is the node v on the same infection chain i The neighborhood set of .
[0030] Furthermore, the network-level risk feature is calculated through the inter-chain attention neural network and the chain-level risk feature. The specific formula is as follows:
[0031]
[0032] Among them, σ represents the sigmoid function, represents the network-level attention coefficient, Represents the node v in the network level i neighborhood, K is the number of multi-head attention on the last layer of the inter-chain neural network. i and c j share the same node, then c i and c j The nodes in are network-level neighborhoods.
[0033] Furthermore, in step S2, K independent attention mechanisms are used to perform linear combination of features, and the output features of each node are obtained by combining multi-level representations.
[0034] The loss function for calculating the risk level assessment result in step S3 is specifically as follows:
[0035]
[0036]
[0037] z i,r =MLP r (rep r ,i;θ r ),r∈{c,n}
[0038] Among them, y∈{1,2,…,l} is the risk rating category, rep r represents the infectious chain or network learned from the intra-chain attention neural network and the inter-chain attention neural network, where r∈{c,n}, where c represents the infectious chain, n represents the guarantee network model, and z i Indicates rep r As the predicted score of the i-th risk level corresponding to the input, NN r(rep,i;θ) is a shallow feedforward neural network that represents rep r Mapped into an actual score, θ is the learned weight, MLP r It consists of two layers of neurons whose activation function is ReLU function and one layer of neurons whose activation function is sigmoid function, σ i (z) represents the probability that the infection chain or guarantee network model is at risk level i, and the sigmoid function is used to map the probability to the value range [0,1.
[0039] Furthermore, r={r1,r2,…,r l} represents the risk rating level of type l, each r i Represents a risk range, r1 is the lowest range, r l Is the highest range.
[0040] Furthermore, the convolutional neural network model concatenates node-level risk features, chain-level risk features, and network-level risk features in adjacent order. These features are first processed by convolutional and pooling layers, and then a feedforward layer is used to perform system risk assessment. The convolutional neural network model is trained through the following optimization process:
[0041]
[0042] Among them, θ s is the weight of the convolutional neural network model, represents the mean square error function, NN s It is a model composed of a single convolutional layer with a maximum pooling operation and a single-layer feed-forward network, and NN s (x i θ s ) is mapped to [0,1], and λ is the penalty parameter.
[0043] Compared with the prior art, the present invention has the following beneficial effects:
[0044] 1. The guarantee network model of the present invention adopts a graph learning model and uses a multi-level feature learning layer to calculate the potential risk information of the infection chain and risk rating, which improves the learning efficiency of risk features in the infection chain and significantly improves the accuracy of predicting systemic risk in complex financial loan networks.
[0045] 2. Since the guarantee contract has a validity period, the present invention dynamically updates the guarantee network model and the infection chain, which significantly improves the effectiveness of the systemic risk score in the network-level and chain-level risk assessment.
[0046] 3. The present invention applies graph neural networks and attention mechanisms to directly learn risk characteristics from networked loan behavior data, eliminating the need to rely on professional knowledge in the financial field and having high applicability and flexibility.
[0047] 4. In the simulation of risk mitigation and risk isolation, the present invention shows significant advantages in finding key small and medium-sized enterprises and infection paths in loan networks: the results predicted by the model are on average 100% to 220% better than the results based on human experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is a schematic diagram of the process of the present invention;
[0049] Figure 2 It is a logical schematic diagram of the present invention. DETAILED DESCRIPTION
[0050] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0051] Example
[0052] like Figure 1 As shown in FIG, a systemic risk assessment method based on a complex financial network includes the following steps:
[0053] S1. Based on the breadth-first search algorithm, obtain the target enterprise's status data and guaranteed loan information, establish a guarantee network model, and construct all related transmission chains, and dynamically update the guarantee network model and transmission chains;
[0054] S2. Based on the constructed infection chain, an intra-chain attention neural network and an inter-chain attention neural network are established, and a multi-level feature learning layer is used to calculate the potential risk information of the infection chain and risk rating;
[0055] S3. Establish a convolutional neural network model to obtain the risk level assessment results of the target enterprise based on the potential risk information of the infection chain and risk rating.
[0056] The formula of the guarantee network model G is as follows:
[0057] G=(V,E)
[0058] Where V={v1,v2,…,v |V|} is a node set, each node represents the small and medium-sized enterprises in the guarantee network model, E = {e1, e2, ..., e |E|} is an edge set, each edge represents the guarantee relationship between nodes, |V| represents the number of nodes, and |E| represents the number of edges in the guarantee network model.
[0059] The target enterprise's status data includes registered capital, number of employees and financial status, which are embedded in the attributes of the nodes of the guarantee network model; the target enterprise's guaranteed loan information includes the guarantee amount, loan amount and loan interest, which are stored in the attributes of the edges of the guarantee network model.
[0060] All relevant infection chains in step S1 form an infection chain set, and the specific formula is as follows:
[0061] C={c1,c2,…,c |C|}
[0062] c i ={(v,e)|v∈Diff(v i ,G)∪e∈Diff(v i ,G)}
[0063] Among them, C is the set of infection chains, and the infection chain c i is node v i The maximum risk diffusion range, Diff(v i ,G) indicates that from v i Edges passing through the guarantee network model G.
[0064] Since the guarantee contract has a validity period, in step S1, the guarantee network model and the infection chain are dynamically updated during the training and prediction stages.
[0065] Step S1 also involves first linking bank loan records by key value through a query operation to reconstruct the initial guarantee network model. During the feature learning process, since guarantee relationships are dynamic, if a guarantee is issued, a temporary edge e is added to the network G; when the guarantee expires, this edge is deleted. Finally, the attributes of the nodes and edges in the set are used as features of the contagion chain.
[0066] In this embodiment, a directed breadth-first search (BFS) algorithm is used to generate the infection chain of each node. The specific process is shown in Table 1:
[0067] Table 1 Directed breadth-first search algorithm generates infection chains
[0068]
[0069] The potential risk information of the infection chain and risk rating in step S2 includes node-level risk characteristics, chain-level risk characteristics, and network-level risk characteristics.
[0070] like Figure 2As shown in the figure, the node-level risk features include node features and edge features, which are calculated by the attribute graph learning layer to obtain the given node v i Individual characteristics The specific formula is as follows:
[0071]
[0072] in, is the node feature, is the edge feature, W I and a I is the trainable weight, NN I represents a fully connected neural layer in a single feature learning module, is node v i Neighborhood set of , || represents and operation.
[0073] Chain-level risk characteristics It is calculated through the in-chain attention neural network. The specific formula is as follows:
[0074]
[0075]
[0076]
[0077] Among them, node v j Indicates v sharing the same infection chain i Neighbor nodes, LeakyReLU(·) represents the nonlinear transformation of LeakyRectified Linear Unit, W c is the weighted matrix for learning by the in-chain attention neural network, is the attention weight of the in-chain attention neural network, is the node v on the same infection chain i The neighborhood set of .
[0078] The network-level risk characteristics are calculated through the inter-chain attention neural network and the chain-level risk characteristics. The specific formula is as follows:
[0079]
[0080] Among them, σ represents the sigmoid function, represents the network-level attention coefficient, Represents the node v in the network level i neighborhood, K is the number of multi-head attention on the last layer of the inter-chain neural network. i and c j share the same node, then c i and cj The nodes in are network-level neighborhoods.
[0081] In step S2, K independent attention mechanisms are used to perform linear combination of features and obtain the output features of each node through the combination of multi-level representations.
[0082] The loss function for calculating the risk level assessment result in step S3 is specifically as follows:
[0083]
[0084]
[0085] z i,r =MLP r (rep r ,i;θ r ),r∈{c,n}
[0086] Among them, y∈{1,2,…,l} is the risk rating category, rep r represents the infectious chain or network learned from the intra-chain attention neural network and the inter-chain attention neural network, where r∈{c,n}, where c represents the infectious chain, n represents the guarantee network model, and z i Indicates rep r As the predicted score of the i-th risk level corresponding to the input, NN r (rep,i;θ) is a shallow feedforward neural network that represents rep r Mapped into an actual score, θ is the learned weight, MLP r It consists of two layers of neurons whose activation function is ReLU function and one layer of neurons whose activation function is sigmoid function, σ i (z) represents the probability that the infection chain or guarantee network model is at risk level i, and the sigmoid function is used to map the probability to the value range [0,1].
[0087] r={r1,r2,…,r l} represents the risk rating level of type l, each r i Represents a risk range, r1 is the lowest range, r l is the highest range. In this embodiment, l=5.
[0088] The convolutional neural network model concatenates node-level risk features, chain-level risk features, and network-level risk features in adjacent order. These features are first processed by convolutional and pooling layers, and then a feedforward layer is used to perform system risk assessment. The convolutional neural network model is trained through the following optimization process:
[0089]
[0090] Among them, θ s is the weight of the convolutional neural network model, represents the mean square error function, NN s It is a model composed of a single convolutional layer with a maximum pooling operation and a single-layer feed-forward network, and NN s (x i θ s ) is mapped to [0,1], and λ is the penalty parameter.
[0091] In addition, it should be noted that the specific embodiments described in this specification may be named differently, and the above content described in this specification is merely an example of the structure of the present invention. Any equivalent changes or simple changes made based on the structure, features and principles of the present invention are included in the protection scope of the present invention. Those skilled in the art of the present invention may make various modifications or supplements to the specific examples described or adopt similar methods, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.
Claims
1. A systemic risk assessment method based on complex financial networks, characterized by: The specific steps include: S1. Based on the breadth-first search algorithm, obtain the target enterprise's status data and guaranteed loan information, establish a guarantee network model, and construct all related transmission chains, and dynamically update the guarantee network model and transmission chains; S2. Based on the constructed infection chain, an intra-chain attention neural network and an inter-chain attention neural network are established, and a multi-level feature learning layer is used to calculate the potential risk information of the infection chain and risk rating; S3. Build a convolutional neural network model to obtain the risk level assessment results of the target enterprise based on the potential risk information of the infection chain and risk rating; The potential risk information of the infection chain and risk rating in step S2 includes node-level risk characteristics, chain-level risk characteristics, and network-level risk characteristics; The node-level risk features include node features and edge features, and the given node v is calculated by the attribution graph learning layer. i Individual characteristics The specific formula is as follows: in, is the node feature, is the edge feature, W I and a I is the training weight, NN I represents a fully connected neural layer in a single feature learning module, is node v i Neighborhood set of , || represents and operation.
2. A method for assessing systemic risk based on complex financial networks according to claim 1, characterized in that: The formula of the guarantee network model G is as follows: G=(V,E) Where V={v1,v2,…,v |V| } is a node set, each node represents the small and medium-sized enterprises in the guarantee network model, E = {e1, e2, ..., e |E| } is an edge set, each edge represents the guarantee relationship between nodes, |V| represents the number of nodes, and |E| represents the number of edges in the guarantee network model.
3. A method for assessing systemic risk based on complex financial networks according to claim 2, characterized in that: The status data of the target enterprise includes registered capital, number of employees and financial status, which are embedded in the attributes of the nodes of the guarantee network model; the guaranteed loan information of the target enterprise includes the guaranteed amount, loan amount and loan interest, which are stored in the attributes of the edges of the guarantee network model.
4. A method for assessing systemic risk based on complex financial networks according to claim 2, characterized in that: All relevant infection chains in step S1 form an infection chain set, and the specific formula is as follows: C={c1,c2,…,c |C| } c i ={(v,e)|v∈Diff(v i ,G)∪e∈Diff(v i ,G)} Among them, C is the set of infection chains, and the infection chain c i is node v i The maximum risk diffusion range, Diff(v i ,G) indicates that from v i Edges passing through the guarantee network model G.
5. A method for assessing systemic risk based on complex financial networks according to claim 1, characterized in that: The chain-level risk characteristics It is calculated through the in-chain attention neural network. The specific formula is as follows: Among them, node v j Indicates v sharing the same infection chain i Neighbor nodes, LeakyReLU(·) represents the nonlinear transformation of LeakyRectified Linear Unit, W c is the weighted matrix for learning by the in-chain attention neural network, is the attention weight of the in-chain attention neural network, is the node v on the same infection chain i The neighborhood set of .
6. A method for assessing systemic risk based on complex financial networks according to claim 5, characterized in that: The network-level risk characteristics are calculated through the inter-chain attention neural network and the chain-level risk characteristics. The specific formula is as follows: Among them, σ represents the sigmoid function, represents the network-level attention coefficient, Represents the node v in the network level i neighborhood, K is the number of multi-head attention on the last layer of the inter-chain neural network.
7. A method for systemic risk assessment based on complex financial networks according to claim 1, characterized in that: The loss function for calculating the risk level assessment result in step S3 is specifically as follows: With i,r =MLP r (rep r ,i;θ r ),r∈{c,n} Among them, y∈{1,2,…,l} is the risk rating category, rep r represents the infectious chain or network learned from the intra-chain attention neural network and the inter-chain attention neural network, where r∈{c,n}, where c represents the infectious chain, n represents the guarantee network model, and z i Indicates rep r As the predicted score of the i-th risk level corresponding to the input, NN r (rep,i;θ) is a shallow feedforward neural network that represents rep r Mapped into an actual score, θ is the learned weight, MLP r It consists of two layers of neurons whose activation function is ReLU function and one layer of neurons whose activation function is sigmoid function, σ i (z) represents the probability that the infection chain or guarantee network model is at risk level i.
8. A method for assessing systemic risk based on complex financial networks according to claim 7, characterized in that: The convolutional neural network model concatenates node-level risk features, chain-level risk features, and network-level risk features in adjacent order. These features are first processed by convolutional and pooling layers, and then a feedforward layer is used to perform system risk assessment. The convolutional neural network model is trained through the following optimization process: Among them, θ s is the weight of the convolutional neural network model, represents the mean square error function, NN s It is a model consisting of a single convolutional layer with a maximum pooling operation plus a single-layer feed-forward network, where λ is the penalty parameter.