An asset pricing method based on industry chain heterogeneous graph and market state self-adaption

Abstract

This invention provides an asset pricing method based on heterogeneous industry chain graphs and market state adaptation. By constructing a two-layer heterogeneous graph network of "industry-stock" and designing a heterogeneous graph neural network for message passing, it deeply integrates structured information such as supply and demand transmission and industry competition from upstream and downstream of the industry chain into the stock pricing model for the first time, solving the problem of fragmented industry chain information modeling in existing technologies. Simultaneously, it innovatively introduces a market state perception module and uses it as a weight parameter in the dynamic generation graph interaction network of gating signals, achieving end-to-end dynamic adaptation of feature interaction logic under different market environments. Furthermore, by designing a cross-level attention mechanism and a joint loss function with constraints, this invention achieves collaborative decision-making and deep coupling of three types of information: "macro-industry chain transmission," "micro-company feature interaction," and "dynamic adaptation of market state."

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an asset pricing method based on an adaptive approach to supply chain heterogeneity graphs and market conditions. Background Technology

[0002] Asset pricing is a core issue in the fields of financial investment and risk management, aiming to reveal the intrinsic relationship between expected asset returns and their influencing factors. As research deepens, the industry generally believes that asset returns are not only related to a company's own financial characteristics (such as momentum, value, and profitability), but also to the complex influence of the macroeconomic environment and the transmission of prosperity across the industrial chain. In recent years, artificial intelligence technologies, represented by deep learning, have demonstrated enormous potential in asset pricing research due to their powerful nonlinear fitting capabilities. For example, existing technologies include solutions such as "An Asset Pricing Method Based on Deep Interactive Learning" (CN119579323A), which achieves high-order interactive modeling between company characteristics by constructing a graph neural network (GNN), effectively capturing the nonlinear pricing information ignored by traditional linear factor models.

[0003] However, despite the significant progress made by the aforementioned methods, their technical solutions still have inherent limitations when dealing with the complex and ever-changing capital markets. On the one hand, existing technologies generally isolate the modeling object to the company characteristics of individual stocks, failing to effectively incorporate structured information at the industry chain level. For example, although the patent literature attempts to construct industry correlation graphs, it only uses a single-layer isomorphic graph to model linear relationships between industries, failing to distinguish between the two fundamentally different heterogeneous relationships of upstream and downstream supply and demand transmission and intra-industry competition. This results in a severe deficiency in capturing macro-level pricing information such as the transmission of industry chain prosperity and industry rotation. On the other hand, the interaction logic of existing models is static and cannot be dynamically adjusted according to the evolution of market conditions. For instance, although the patent literature proposes training multiple independent models based on market conditions, its multi-model separation approach is not only computationally expensive but also fails to achieve end-to-end adaptive weight updates. This leads to the inability to smoothly switch the model's pricing logic under different market conditions such as bull markets, bear markets, and volatile markets, resulting in a significant decrease in prediction accuracy and robustness under extreme market conditions.

[0004] Therefore, how to construct an adaptive asset pricing model that can deeply integrate macro-level transmission information of the industrial chain and dynamically adapt to different market conditions, in order to overcome the shortcomings of fragmented information modeling and lack of market condition adaptability in existing technologies, is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide an asset pricing method based on the heterogeneity of the industry chain diagram and market conditions to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following solution: This invention provides an asset pricing method based on an adaptive supply chain heterogeneity graph and market conditions, comprising the following steps: S1. Data Acquisition and Preprocessing: Acquire the original dataset of the target stock, which includes individual company characteristic data, macroeconomic indicator data, and upstream and downstream enterprise relationship data required to construct the industry chain map; S2. Construct a two-layer heterogeneous graph of the industrial chain: Construct a two-layer heterogeneous graph network G=(V,E) containing an industry layer and an individual stock layer, where the node set V contains industry nodes and individual stock nodes; the edge set E contains three types of heterogeneous edges: upstream and downstream supply and demand relationship edges between industries, intra-industry competition relationship edges, and the affiliation edge between individual stocks and their respective industries. S3. Market State Perception and Feature Encoding: The current market state is quantitatively encoded based on macroeconomic indicator data to obtain a market state vector; at the same time, initial representation learning of individual stock company features is performed using embedding technology to obtain an initial individual stock feature vector. S4. Supply chain-guided hierarchical interaction and adaptive state updates: S41. Message transmission at the industry chain level: Based on the constructed two-layer heterogeneous graph, message transmission between layers is carried out through the heterogeneous graph neural network, which aggregates pricing information from upstream and downstream industries and competing industries, updates the feature representations of industry nodes and individual stock nodes, and obtains individual stock features after industry chain enhancement. S42, State-Adaptive Feature Interaction: Using the market state vector as a gating signal, the weights of high-order interactions between company features are dynamically adjusted. A market state-adaptive graph interaction network is used to model the high-order interactions of individual stock features after the industry chain enhancement, resulting in high-order interaction features. S5. Cross-level feature aggregation and pricing: Through a cross-level attention mechanism, the initial stock features, the enhanced stock features of the industrial chain, and the high-order interaction features are dynamically fused to obtain the final fused feature vector, and the excess return prediction result of the stock is output through a multilayer perceptron.

[0007] Preferably, the construction of the two-layer heterogeneous graph of the industrial chain in step S2 specifically includes: S21. Define the industry node set V ind ={ v i ind | i =1,…, N ind},in N ind The number of industries, and the initial feature h for each industry node. i ind(0) Initialized by the average characteristics of all stocks in its industry; S22. Define the individual stock node set V stk ={ v j stk | j =1,…, N stk},in N stk For each stock, the initial feature h is the number of individual stocks. j stk (0) represents the initial individual stock feature vector e obtained through step S3. j ; S23. Define three types of heterogeneous edges: upstream and downstream supply and demand edges (E) sup edge weight w i,k sup Supply share; industry competition edge E com edge weight w i,k com Market share overlap; Individual stock-industry affiliation edge E bel Connecting individual stocks j With its industry i edge weight w j,i bel The value is 1.

[0008] Preferably, the market state perception and feature encoding in step S3 specifically includes: S31. Obtain a set of macroeconomic indicators M={ m 1, m 2,…, m T}; S32. Time-series encoding of macroeconomic indicator sequences is performed using a gated recurrent unit network to obtain the market state vector s. t : ; ; in, It is the hidden state of macroeconomic indicators. and It consists of learnable weight matrices and bias terms; S33. Embedding learning is performed on the features of individual stocks to obtain the initial individual stock feature vector e. j .

[0009] Preferably, the message passing at the industry chain layer in step S41 specifically includes: S411, Industry Node Update: Industry Node v i ind exist l Layer updates aggregate data from upstream and downstream industries. N sup ( i and competing industries N com ( i Information: ; ; in, and It is an adjacency matrix representing upstream and downstream businesses and competitors in the same industry. , , It is a learnable transformation matrix. It is an activation function; S412, Individual Stock Node Update: Individual Stock Node v j stk exist l Layer updates aggregate data from its respective industry. i Information: ; ; in, Individual stock-industry adjacency matrix , It is a learnable matrix; after passing through L layers, the individual stock characteristics after industry chain enhancement are obtained. (L), denoted as c j .

[0010] Preferably, the state adaptive feature interaction in step S42 specifically includes: S421. Construct a market state adaptive graph interaction network, where the nodes are company characteristics of multiple categories, and each node... n p The initial state is from c j The corresponding feature c extracted j p ; S422. In each interaction, the node... n p From neighboring nodes n q The weights of the aggregated information are determined by the market state vector s. t Dynamic modulation: ; in, , and It is the initial vector of the node. It is a connection operation; S423, the aggregation and update process is represented as follows: ; ; in, , After K interactions, higher-order interaction features are obtained. .

[0011] Preferably, the cross-level feature aggregation and pricing in step S5 specifically includes: S51. Constructing a multi-level feature set ,in It is the initial set of individual stock characteristics. It is a collection of individual stock characteristics resulting from the enhancement of the industrial chain. It is a set of high-order interaction features; S52, Dynamic fusion through cross-level attention mechanisms: ; ; in, These are the eigenvectors of set O. It is a learnable query vector; S53, Combine the features The final asset pricing result is obtained by inputting the multilayer perceptron: .

[0012] Preferably, step S4 further includes a joint loss function for model training, which is used to calculate the excess return fitting loss. Based on this, add supply chain linkage constraints. Regularization terms adapted to market conditions The overall loss function is ,in , This is a hyperparameter.

[0013] Preferably, the supply chain association constraint item The formula used to constrain the consistency of feature representations of upstream and downstream industry nodes in the embedding space is as follows: .

[0014] Preferably, the market state adaptation regularization term The formula used to constrain the differences in model parameters under different market conditions is as follows: ; in, These are model parameters under specific market conditions. This is the average parameter across all market conditions.

[0015] Preferably, the multilayer perceptron (MLP) obtains the final asset pricing result using the following formula: ; ; in, It refers to the number of layers in an MLP. , They are the first l The learnable weight matrix and bias terms in the layer. and These are the learnable weight matrix and bias term in the output layer, respectively.

[0016] The present invention achieves the following beneficial technical effects compared to the prior art: This invention provides an asset pricing method based on heterogeneous industry chain graphs and market state adaptation, which has outstanding substantive features and significant progress. By constructing a two-layer heterogeneous graph network of "industry-stock" and designing a heterogeneous graph neural network for message passing, it deeply integrates structured information such as supply and demand transmission and industry competition from upstream and downstream of the industry chain into the stock pricing model for the first time. This solves the problem of fragmented industry chain information modeling in existing technologies and significantly enhances the ability to capture industry rotation and prosperity transmission. Simultaneously, it innovatively introduces a market state perception module and uses it as a weight parameter in the dynamic generation graph interaction network of gating signals, achieving end-to-end dynamic adaptation of feature interaction logic under different market environments. This fundamentally overcomes the defect of traditional static models where prediction accuracy drops sharply under extreme market conditions. Furthermore, by designing a cross-level attention mechanism and a joint loss function with constraints, this invention achieves collaborative decision-making and deep coupling of three types of information: "macro-industry chain transmission," "micro-company feature interaction," and "dynamic adaptation of market state." This results in significantly superior technical effects compared to existing technologies in terms of out-of-sample prediction accuracy, robustness in extreme scenarios, and long-term pricing stability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The flowchart of the asset pricing method based on the heterogeneity of the industrial chain diagram and market conditions provided by the present invention is shown. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The purpose of this invention is to provide an asset pricing method based on the heterogeneity of the industry chain diagram and market conditions to solve the problems existing in the prior art.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1: Please see Figure 1 This invention provides an asset pricing method based on a heterogeneous industry chain graph and market state adaptation. This method aims to address the core technical problems of fragmented industry chain information modeling, lack of market state adaptability, and single information fusion dimension in existing asset pricing technologies. By constructing a two-layer heterogeneous industry chain graph and a market state adaptation mechanism, it achieves cross-level pricing information fusion and dynamic adaptation to the entire market scenario.

[0023] In step S1, data acquisition and preprocessing are performed first. Specifically, the original dataset of the target stock is acquired, which includes three core types of data: The first type is individual company characteristic data, covering multiple dimensions of company attribute indicators such as trading friction, momentum, value, growth, profitability, and financial liquidity, such as turnover rate, book-to-market ratio, and asset growth rate; the second type is macroeconomic indicator data, including but not limited to GDP growth rate, consumer price index, market turnover rate, market volatility, and other macroeconomic variables that can reflect the overall economic environment and market sentiment; the third type is upstream and downstream enterprise relationship data required to construct the industry chain map. This data is generated based on input-output tables, industry classification standards, and enterprise supply chain information, and is used to characterize the supply and demand transmission relationship and the competitive relationship between industries.

[0024] In step S2, a two-layer heterogeneous graph of the industry chain is constructed. This invention constructs a two-layer heterogeneous graph network G = (V, E) containing an industry layer and an individual stock layer. The node set V contains industry nodes and individual stock nodes. Specifically, the industry node set V is defined as follows:ind ={ v i ind | i =1,…, N ind},in N ind The number of industries, and the initial feature h for each industry node. i ind (0) Initialize with the average characteristics of all stocks within its industry; define the stock node set V. stk ={ v j stk | j =1,…, N stk},in N stk For each stock, the initial feature h is the number of individual stocks. j stk (0) represents the initial individual stock feature vector e obtained through step S3. j Define three types of heterogeneous edges: upstream and downstream supply and demand edges E sup edge weight w i,k sup Supply share; industry competition edge E com edge weight w i,k com Market share overlap; Individual stock-industry affiliation edge E bel Connecting individual stocks j With its industry i edge weight w j,i bel A value of 1 is used to correlate individual stocks with the macroeconomic conditions of their respective industries.

[0025] In step S3, market state perception and feature encoding are performed. First, a set of macroeconomic indicator sequences M={ m 1, m 2,…, m T These indicators are arranged in chronological order, reflecting the dynamic changes in the market environment. Then, a gated recurrent unit (GRU) network is used to perform time-series encoding on the macroeconomic indicator series to capture the time dependence and evolutionary patterns of the macroeconomic state. The encoding process is as follows: ;in, It is the hidden state vector of macroeconomic indicators at time step t. This is the macroeconomic indicator input vector at time step t. Subsequently, the hidden state is nonlinearly transformed through a fully connected layer to obtain the final market state vector: ; in, and The weight matrix and bias term are learnable, and tanh, as the activation function, limits the output value to the range [-1, 1], giving the market state vector a definite numerical range. Simultaneously, embedding learning techniques are used to learn the initial representation of individual stock company features. This process can refer to step S2 in the original patent (CN119579323A), where the original company features are converted into real-valued vectors using the embedding matrix to obtain the initial individual stock feature vector e. j .

[0026] In step S4, hierarchical interaction and state adaptive update guided by the supply chain are performed. This step is one of the core innovations of this invention, and it is specifically divided into two sub-steps: supply chain layer message passing and state adaptive feature interaction.

[0027] In step S41, message passing at the industry chain level is performed. Based on the constructed two-layer heterogeneous graph, message passing between layers is achieved through a heterogeneous graph neural network (HGNN). First, the industry nodes are updated. (Industry nodes...) v i ind exist l Updating a layer requires aggregating data from upstream and downstream industry neighbors. N sup ( i ) and neighboring companies in the same industry N com ( i The information is as follows: ; ; in, and It is an adjacency matrix representing upstream and downstream businesses and competitors in the same industry. , , It is a learnable transformation matrix. It is the activation function. This update mechanism allows the feature representations of industry nodes to incorporate transmission information from upstream and downstream of the industry chain, as well as competitive information from peers. Then, the individual stock nodes are updated. Individual stock nodes v j stk exist l Layer updates aggregate information from its industry node i: ; ; in, Individual stock-industry adjacency matrix , It is a learnable matrix; after passing through L layers, the individual stock characteristics after industry chain enhancement are obtained. (L), denoted as c j Thus, the characteristics of individual stocks have incorporated the prosperity of their respective industries and the transmission effect of upstream and downstream industries in the industrial chain.

[0028] In step S42, state-adaptive feature interaction is performed. This invention constructs a market state-adaptive graph interaction network (SA-GNN), whose nodes are six categories of firm features (such as transaction friction, momentum, value, etc.). Each node... n p The initial state is from c j The corresponding feature c extracted j p The core innovation of this network lies in the fact that the weights for higher-order interactions between feature nodes are no longer fixed, but are determined by the current market state vector s. t Dynamic modulation. In each interaction (the kth time), the node... n p From neighboring nodes n q The formula for calculating the weight of aggregated information is as follows: ; in, , and It is the initial vector of the node. This is a connection operation. In this way, the intensity of interaction under different market conditions can be adaptively adjusted. The aggregation and update process can be represented as: ; ; in, , After K interactions, higher-order interaction features are obtained. This set not only includes non-linear interaction information between company characteristics, but also incorporates the dynamic impact of market conditions on the interaction logic.

[0029] In step S5, cross-level feature aggregation and pricing are performed. First, a multi-level feature set is constructed. Where I is the initial set of individual stock features, and C is the set of individual stock features after industry chain enhancement. This is a set of high-order interaction features. These three sets represent micro-level individual information, macro-level transmission information across the industry chain, and dynamic interaction information, respectively, forming a multi-level feature pyramid. Then, these features are dynamically fused through a cross-level attention mechanism to adaptively assess the importance of information at different levels to the current pricing task. ; ; in, These are the eigenvectors of set O. It is a learnable query vector. and They are learnable weight matrices and vectors. It is a learnable bias term. This mechanism allows the model to dynamically assign different weights to features at different levels based on the characteristics of the input data, achieving optimal information fusion. Finally, the fused feature Z is input into a multilayer perceptron (MLP) to obtain the final asset pricing result: ; ; in, It refers to the number of layers in an MLP. , They are the first l The learnable weight matrix and bias terms in the layer. and These are the learnable weight matrix and bias term in the output layer, respectively.

[0030] Furthermore, to further enhance the model's generalization ability and robustness, this invention introduces a joint loss function during model training. This loss function, based on the excess return fitting loss (typically mean squared error), adds a supply chain correlation constraint term and a market state fit regularization term. The overall loss function is as follows: ,in , This is a hyperparameter. Specifically, the industry chain association constraint term... The formula used to constrain the consistency of feature representations of upstream and downstream industry nodes in the embedding space is as follows: .

[0031] The market state adaptation regularization term The formula used to constrain the differences in model parameters under different market conditions is as follows: ; in, These are model parameters under specific market conditions. The parameter is the average parameter across all market conditions. This regularization term prevents excessive parameter shifts in the model under different market conditions, thereby enhancing the model's stability throughout the entire market cycle.

[0032] Through the above steps, this invention realizes a complete asset pricing process from embedding industry chain information, perceiving market status, hierarchical interaction modeling to cross-level fusion output, effectively solving the core defects in the existing technology and significantly improving the prediction accuracy and generalization ability of the asset pricing model.

[0033] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0034] It should be noted that the components mentioned in the above embodiments are all general standard parts or components known to those skilled in the art. Their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0035] This invention has illustrated its principles and implementation methods using specific examples. The descriptions of these embodiments are merely illustrative of the method and its core ideas; furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and application scope based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the invention.

Claims

1. An asset pricing method based on an adaptive approach of supply chain heterogeneity graph and market conditions, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Acquire the original dataset of the target stock, which includes individual company characteristic data, macroeconomic indicator data, and upstream and downstream enterprise relationship data required to construct the industry chain map; S2. Construct a two-layer heterogeneous graph of the industrial chain: Construct a two-layer heterogeneous graph network G=(V,E) containing an industry layer and an individual stock layer, where the node set V contains industry nodes and individual stock nodes; the edge set E contains three types of heterogeneous edges: upstream and downstream supply and demand relationship edges between industries, intra-industry competition relationship edges, and the affiliation edge between individual stocks and their respective industries. S3. Market State Perception and Feature Encoding: The current market state is quantitatively encoded based on macroeconomic indicator data to obtain a market state vector; at the same time, initial representation learning of individual stock company features is performed using embedding technology to obtain an initial individual stock feature vector. S4. Supply chain-guided hierarchical interaction and adaptive state updates: S41. Message transmission at the industry chain level: Based on the constructed two-layer heterogeneous graph, message transmission between layers is carried out through the heterogeneous graph neural network, which aggregates pricing information from upstream and downstream industries and competing industries, updates the feature representations of industry nodes and individual stock nodes, and obtains individual stock features after industry chain enhancement. S42, State-Adaptive Feature Interaction: Using the market state vector as a gating signal, the weights of high-order interactions between company features are dynamically adjusted. A market state-adaptive graph interaction network is used to model the high-order interactions of individual stock features after the industry chain enhancement, resulting in high-order interaction features. S5. Cross-level feature aggregation and pricing: Through a cross-level attention mechanism, the initial stock features, the enhanced stock features of the industrial chain, and the high-order interaction features are dynamically fused to obtain the final fused feature vector, and the excess return prediction result of the stock is output through a multilayer perceptron.

2. The asset pricing method based on supply chain heterogeneity graph and market state adaptation as described in claim 1, characterized in that, Step S2, which involves constructing a two-layer heterogeneous graph of the industrial chain, specifically includes: S21. Define the industry node set V ind ={ v i ind | i =1,…, N ind },in N ind The number of industries, and the initial feature h for each industry node. i ind (0) Initialized by the average characteristics of all stocks in its industry; S22. Define the individual stock node set V stk ={ v j stk | j =1,…, N stk },in N stk For each stock, the initial feature h is the number of individual stocks. j stk (0) represents the initial individual stock feature vector e obtained through step S3. j ; S23. Define three types of heterogeneous edges: upstream and downstream supply and demand edges (E) sup edge weight w i,k sup Supply share; industry competition edge E com edge weight w i,k com Market share overlap; Individual stock-industry affiliation edge E bel Connecting individual stocks j With its industry i edge weight w j,i bel The value is 1.

3. The asset pricing method based on the adaptive relationship between supply chain heterogeneity graph and market conditions as described in claim 1, characterized in that, Step S3, market state perception and feature encoding, specifically includes: S31. Obtain a set of macroeconomic indicators M={ m 1, m 2,…, m T }; S32. Time-series encoding of macroeconomic indicator sequences is performed using a gated recurrent unit network to obtain the market state vector s. t : ； ； in, It is the hidden state of macroeconomic indicators. and It consists of learnable weight matrices and bias terms; S33. Embedding learning is performed on the features of individual stocks to obtain the initial individual stock feature vector e. j .

4. The asset pricing method based on supply chain heterogeneity graph and market state adaptation as described in claim 1, characterized in that, The message passing at the supply chain layer in step S41 specifically includes: S411, Industry Node Update: Industry Node v i ind exist l Layer updates aggregate data from upstream and downstream industries. N sup ( i and competing industries N com ( i Information: ； ； in, and It is an adjacency matrix representing upstream and downstream businesses and competitors in the same industry. , , It is a learnable transformation matrix. It is an activation function; S412, Individual Stock Node Update: Individual Stock Node v j stk exist l Layer updates aggregate data from its respective industry. i Information: ； ； in, Individual stock-industry adjacency matrix , It is a learnable matrix; after passing through L layers, the individual stock characteristics after industry chain enhancement are obtained. (L), denoted as c j .

5. The asset pricing method based on supply chain heterogeneity graph and market state adaptation as described in claim 1, characterized in that, The state-adaptive feature interaction in step S42 specifically includes: S421. Construct a market state adaptive graph interaction network, where the nodes are company characteristics of multiple categories, and each node... n p The initial state is from c j The corresponding feature c extracted j p ; S422. In each interaction, the node n p From neighboring nodes n q The weights of the aggregated information are determined by the market state vector s. t Dynamic modulation: ； in, , and It is the initial vector of the node. It is a connection operation; S423, the aggregation and update process is represented as follows: ； ； in, , After K interactions, higher-order interaction features are obtained. .

6. The asset pricing method based on supply chain heterogeneity graph and market state adaptation as described in claim 1, characterized in that, Step S5, cross-level feature aggregation and pricing, specifically includes: S51. Constructing a multi-level feature set ,in It is the initial set of individual stock characteristics. It is a collection of individual stock characteristics resulting from the enhancement of the industrial chain. It is a set of high-order interaction features; S52, Dynamic fusion through cross-level attention mechanisms: ； ； in, These are the eigenvectors of set O. It is a learnable query vector; S53, Combine the features The final asset pricing result is obtained by inputting the multilayer perceptron: 。 7. The asset pricing method based on supply chain heterogeneity graph and market state adaptation as described in claim 1, characterized in that, Step S4 also includes the joint loss function for model training, which is the excess return fitting loss. Based on this, add supply chain linkage constraints. Regularization terms adapted to market conditions The overall loss function is ,in , This is a hyperparameter.

8. The asset pricing method based on supply chain heterogeneity graph and market state adaptation as described in claim 7, characterized in that, The industry chain related constraints The formula used to constrain the consistency of feature representations of upstream and downstream industry nodes in the embedding space is as follows: 。 9. The asset pricing method based on supply chain heterogeneity graph and market state adaptation as described in claim 7, characterized in that, The market state adaptation regularization term The formula used to constrain the differences in model parameters under different market conditions is as follows: ； in, These are model parameters under specific market conditions. This is the average parameter across all market conditions.

10. The asset pricing method based on the adaptive relationship between supply chain heterogeneity graph and market conditions as described in claim 1, characterized in that, The multilayer perceptron (MLP) obtains the final asset pricing result using the following formula: ； ； in, It refers to the number of layers in an MLP. , They are the first l The learnable weight matrix and bias terms in the layer. and These are the learnable weight matrix and bias term in the output layer, respectively.

Images