Enterprise valuation system and method based on knowledge graph and deep learning

Abstract

The invention relates to an enterprise valuation system and method based on a knowledge graph and deep learning, and the system comprises a data collection module, an enterprise similarity calculation module, a valuation model module, a macroeconomic scoring module, and an enterprise screening module. The data acquisition module acquires data from multiple dimensions such as enterprise basic information, financial statements, macroeconomic data, enterprise stock prices and analyst valuation; the enterprise similarity calculation module is used for calculating the similarity between enterprises by constructing a model based on a knowledge graph and a graph neural network; the valuation model module performs enterprise reasonable market value prediction by combining enterprise financial data, similarity weighted data and analyst dynamic weighted valuation data by using a deep learning technology; the macroeconomic scoring module predicts stock market trend based on the macroeconomic data and a neural network and dynamically adjusts an enterprise estimation value; and the enterprise screening module screens the enterprises with investment value by comparing the current stock prices of the enterprises with the reasonable stock prices. According to the method, multi-dimensional data can be comprehensively integrated, a deep learning technology and a knowledge graph are combined, the scientificity and efficiency of enterprise valuation are improved, and the method is suitable for enterprise value evaluation and investment analysis scenes.

Description

Technical Field

[0001] The present invention belongs to the fields of enterprise valuation and fintech, and particularly relates to an enterprise valuation system and method based on knowledge graph and deep learning, which are applicable to scenarios such as enterprise value assessment and investment analysis. Background Art

[0002] Enterprise valuation is an important task in the financial field, and its results directly affect investment decisions in the capital market. However, traditional enterprise valuation methods and deep learning-based enterprise valuation methods have the following problems: 1. Limited data dimension: Traditional methods mainly rely on enterprise financial data and ignore the impact of external factors such as news and macroeconomic environment on enterprise valuation, resulting in the lack of comprehensiveness of valuation results. 2. Improper use of model label data: Some researchers believe that due to the subjective bias of analysts, the given valuation data may be inappropriate, so stock price data is used as the label data for the valuation model. However, stock price data is easily affected by various factors such as market sentiment and speculative behavior. In addition, in some stocks with low market capitalization or poor liquidity, a small number of large transactions may cause large fluctuations in stock prices, seriously deviating from the true value of the enterprise. Therefore, the valuation data of analysts has been studied and analyzed by professionals, filtering out some of these short-term and irrational market volatility factors, providing relatively stable label data for the deep learning model. 3. Insufficient dynamic adjustment: Traditional methods fail to effectively combine the dynamic market environment and macroeconomic changes, resulting in valuation results lagging behind the actual market situation.

[0003] 4. Diverse analyst opinions are difficult to utilize: Different analysts may give different valuation results due to differences in personal experience, understanding of the industry, and valuation models used. Currently, the enterprise valuation method based on deep learning models fails to effectively integrate the opinions of authoritative analysts as label data to improve the accuracy and reliability of valuation.

[0004] In recent years, with the development of artificial intelligence technology, the application of knowledge graph and deep learning in the field of fintech has gradually emerged. The knowledge graph can intuitively present the relationship between an enterprise and other entities by constructing an association network between the enterprise and its attributes; deep learning technology can extract deep features from multi-dimensional data. However, there is currently a lack of a systematic enterprise valuation method that organically combines the knowledge graph and deep learning. Therefore, developing an enterprise valuation method and system based on the knowledge graph and deep learning can not only solve the deficiencies of the existing technology, but also more scientifically evaluate the enterprise value and provide strong support for investment decisions. Summary of the Invention

[0005] In view of the above, the present invention provides an enterprise valuation method and system based on knowledge graph and deep learning to solve the problems existing in traditional enterprise valuation methods, such as limited data dimensions, improper use of model label data, insufficient dynamic adjustment, and difficulty in utilizing scattered analyst opinions.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An enterprise valuation method based on knowledge graph and deep learning, comprising the following steps:

[0008] Step 1, system construction. The system includes a data collection module, an enterprise similarity calculation module, a valuation model module, a macroeconomic scoring module, and an enterprise screening module.

[0009] Step 2, data collection. The data collection module is subdivided into six parts. The first part is the enterprise basic information collection module, including a combined set of data such as the industry to which the enterprise belongs, the concepts to which it belongs, the region to which it belongs, the listing time, and the education background of senior executives. The second part is the financial statement data collection module, which is used to collect the core financial data of the enterprise, including a combined set of data such as asset-liability ratio, net assets per share, operating income, advance receipts, inventory, inventory turnover rate, gross profit margin, return on net assets, current ratio, cash flow from operating activities, cash flow from investing activities, and cash flow from financing activities. The third part is the real-time collection module for news data and e-commerce platform data, which is used to monitor news reports related to the enterprise in real time and data such as the sales volume, user ratings, and market share of the enterprise's products on the e-commerce platform. The fourth part is the macroeconomic data collection module, which is used to obtain key data on the current macroeconomic environment and capture the changing trend of the economic cycle, including a combined set of data such as GDP growth rate, unemployment rate, inflation rate, interest rate, and the rise and fall of stock market indexes of other major economies. The fifth part is the enterprise stock price data collection module, which is used to obtain the enterprise stock price and screen out stocks with the current enterprise stock price less than the reasonable stock price. The sixth part is the analyst valuation data collection module.

[0010] Step 3, enterprise similarity calculation. The enterprise similarity calculation module uses an enterprise similarity calculation method based on knowledge graph and GNN model, and the specific steps are as follows:

[0011] The first step: knowledge graph construction. The construction of the knowledge graph is based on the enterprise basic information data in the first part of the data collection module. The enterprise and its related attributes are represented as a graph structure G=(V, E), where V is the set of nodes (representing the enterprise and its attribute nodes), and E is the set of edges (representing the relationship between the enterprise and the attributes). Each enterprise v i ∈V is used as a node, and the basic information of the enterprise (industry to which it belongs, concepts to which it belongs, region to which it belongs, listing time, education background of senior executives) is used as the attribute node v j∈V, the edges between enterprises and attributes represent the association relationships between enterprises and their attributes. Enterprises are connected through shared attributes, and the formed edges have edge weights, which can reflect the association strength between enterprises. The edge weights can be calculated through attribute similarity. For example, for enterprises A and B, the edge weight W is calculated based on the shared attributes A,B , and the formula is as follows:

[0012]

[0013] Finally, through the above steps, a knowledge graph G is constructed, which has a node set V and an edge set E. The formula is expressed as follows:

[0014] G = (V, E), V = {v1, v2, …, v n}, E = {e ij |v i , v j ∈V}

[0015] Step 2: Feature embedding based on the GNN model. Through the structural information and node attribute information of the knowledge graph G, use the graph neural network (GNN) to generate high-dimensional feature vectors for each enterprise node. Based on the GNN model, aggregate the features of neighbor nodes to update the features of each node. For the enterprise node v i , at the l-th layer, its feature representation is updated according to the following formula:

[0016]

[0017] where W (l) is the learnable weight matrix at the l-th layer, AGGREGATE is the aggregation function, is the feature vector of node v j at the l-1-th layer, N(i) is the set of neighbor nodes of node v j , and σ is the activation function.

[0018] The initial feature of the enterprise node is the core financial data in the second part of the data acquisition module and the real-time data of news data and e-commerce platform data in the third part. For news data, use natural language processing (NLP) technology to perform sentiment analysis on news texts and score each article (such as positive, negative, or neutral). Finally, the text data is represented by One-Hot encoding, and the numerical data is directly normalized as the initial feature.

[0019] Step 3: Similarity calculation. After L layers of propagation, the final feature vector i of each enterprise node v is obtained based on the embedded high-dimensional feature vector The cosine similarity formula is used to calculate the similarity between enterprises, and the formula is;

[0020]

[0021] where and are the final feature vectors of enterprise i and enterprise j respectively. represents the norm of the vector. Step 4, enterprise valuation. The valuation model module adopts a valuation method based on deep learning, and the specific steps are as follows:

[0022] The first step is to construct the model training features and labels. The feature vector of each enterprise i consists of a combined set of data including asset-liability ratio, net assets per share, operating income, advance receipts, inventory, inventory turnover rate, gross profit margin, return on net assets, current ratio, cash flow from operating activities, cash flow from investing activities, and cash flow from financing activities. This combined data is normalized or standardized and converted into a numerical vector Based on the enterprise similarity data, for enterprise i, a similarity-weighted feature vector can be generated The formula is as follows:

[0023]

[0024] Combine and to form the final feature vector X of enterprise i i , and the formula is as follows:

[0025]

[0026] The target variable (label) is the enterprise valuation given by the analyst which is based on the analyst valuation data in the sixth part of the data collection module and processed by "a dynamic analyst weighted valuation calculation method".

[0027] The second step is to build a deep learning model. The model consists of an input layer, a hidden layer, a fully connected layer, and an output layer. The output of the L-th layer of the hidden layer is:

[0028]

[0029] where W (l) is the weight matrix of the l-th layer, b (l) is the bias vector of the l-th layer, and the initial input The fully connected layer extracts the non-linear relationship of the input features through a multi-layer fully connected neural network. The output layer is a fully connected layer that outputs the predicted reasonable market value of enterprise i

[0030]

[0031] The model training uses the mean squared error (MSE) as the loss function. For a sample with a training quantity of N, calculate the predicted value of the model and the label value The error between them is calculated by the formula:

[0032]

[0033] Step 5, Macroeconomic scoring. The macroeconomic scoring module adopts a macroeconomic scoring method based on neural network. It uses the combined data of GDP growth rate, unemployment rate, inflation rate, interest rate, and the rise and fall of stock market indexes of other major economies as features, and uses the stock market core index data corresponding to the next reporting date as the label to predict the impact of the macroeconomy on the stock market performance, and dynamically adjusts the enterprise valuation based on this. According to the predicted rise and fall of the stock market index C t obtained from the neural network, map it to the macroeconomic score M t :

[0034] M t = 100 + k·C t

[0035] where k is a coefficient used to adjust the impact of the rise and fall on the score. The benchmark score is 100. The higher the predicted increase, the higher the score, and vice versa.

[0036] Step 6, Calculate the enterprise valuation under the current macroeconomic situation according to the macroeconomic score M t and the reasonable market value given by the valuation model module

[0037]

[0038] Step 7, Calculate the reasonable stock price of the enterprise currently according to the current enterprise valuation and the total number of shares of the enterprise:

[0039]

[0040] Step 8, The enterprise screening module screens out enterprises with investment value by comparing the current stock price of the enterprise and the reasonable stock price of the enterprise currently, providing candidate enterprises for further analysis from other aspects such as the technical side in the next step. In the said Step 4, "a dynamic analyst weighted valuation calculation method" considers from three aspects of historical valuation error, valuation stability, and analyst industry experience, and dynamically adjusts the weight of the analyst over time to reflect the latest market data and analyst performance. The specific steps of this method are as follows:

[0041] ​Step 1: Calculate the historical valuation error. The historical valuation error is used to evaluate the valuation accuracy of analysts, and the formula is as follows:

[0042]

[0043] where, E i is the historical valuation error of analyst i; is the valuation of the enterprise given by analyst i at time t, and V t is the actual market value of the enterprise at time t, and N is the total number of times analysts participate in valuation. The smaller E i , the more accurate the analyst's valuation.

[0044] Step 2: Calculate the valuation stability. The valuation stability is used to evaluate whether there are large fluctuations in the analysts' valuations, and the formula is as follows:

[0045]

[0046] where, S i represents the valuation volatility of analyst i; represents the enterprise valuation of analyst i at time t. The smaller S i , the more stable the analyst's valuation.

[0047] Step 3: Calculate the analyst experience weight. Considering the number of years of industry experience or the number of cases of analysts, it is considered whether analysts have more experience in a specific industry or field. The more experienced, the higher the weight. The formula is as follows:

[0048]

[0049] where, W i represents the experience weight of analyst i; M is the total number of all analysts. The larger W i , the more experienced the analyst in this field.

[0050] Step 4: Calculate the comprehensive score. Considering the historical performance, valuation stability and industry experience weight of analysts comprehensively, the comprehensive score Q i of analysts can be calculated. The formula is as follows:

[0051]

[0052] Q i The larger it is, the better the comprehensive performance of the analyst.

[0053] Step 5: Integrate the valuations of multiple analysts. For the valuation of an enterprise, assuming that M analysts have given valuations , the valuations can be weighted and averaged according to the comprehensive score Q i of analysts to obtain

[0054]

[0055] This formula can effectively integrate the valuations of different analysts, and at the same time assign higher weights to excellent analysts.

[0056] Step 6: Dynamically adjust the analyst weights. Considering that the enterprise valuation may change over time, the weight Q of the analyst i can be dynamically updated. Every once in a while, recalculate the E of the analyst i , S i and W i , and adjust its comprehensive score Q i . The dynamic adjustment formula is as follows:

[0057]

[0058] Among them, α is a regulation parameter (such as 0.7), indicating the retention degree of the historical weight.

[0059] Compared with the prior art, the advantages of the present invention are as follows:

[0060] 1. Advantages of combining multi-source data: Integrating the basic information data of each stock based on the knowledge graph can comprehensively present the enterprise-related knowledge system. Using the stock relationship data and fundamental data as input features can consider from multiple dimensions of enterprise association and its own operating conditions. Choosing analyst data as the label is because analyst data comprehensively considers factors such as market expectations and industry competition patterns. These information play an important supplementary role in the model's learning of the relationship between enterprise value and various factors. Using stock price data as the evaluation criterion, although the stock price has the problem of being interfered, after all, it is the comprehensive reflection of the market on the enterprise value, including the collective judgment of market participants on various aspects of the enterprise. Using it as the evaluation criterion can, to a certain extent, test the fit between the model valuation and the actual market situation and objectively evaluate the performance of the model in the actual market environment. Generally speaking, this model explores the connection between data features and value from multiple dimensions, improving the accuracy and interpretability of the model.

[0061] 2. Dynamically adapt to the macro environment: After obtaining the enterprise valuation based on the valuation model, dynamically adjust it in combination with macroeconomic factors. In this way, the model can adapt to different market environment changes, and the final valuation result is closer to the actual situation, improving the adaptability of the model. Brief Description of the Drawings

[0062] Figure 1 is the flowchart of the embodiment of the present invention,

[0063] Figure 2 is the partial flowchart of the dynamic analyst weighted valuation calculation method of the embodiment of the present invention. Specific implementation mode

[0064] The following combines the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention.

[0065] Combined with Figure 1 , the present invention provides an enterprise valuation method based on a knowledge graph and deep learning, including the following steps:

[0066] Step 1, system construction. The system includes a data collection module, an enterprise similarity calculation module, a valuation model module, a macroeconomic scoring module, and an enterprise screening module.

[0067] Step 2, data collection. The data collection module is subdivided into six parts. The first part is the enterprise basic information collection module, including a combined set of data such as the industry to which it belongs, the concepts to which it belongs, the region to which it belongs, the listing time, and the education level of senior executives; the second part is the financial statement data collection module, which is used to collect the core financial data of the enterprise, including a combined set of data such as the asset-liability ratio, net assets per share, operating income, advance receipts, inventory, inventory turnover rate, gross profit margin, return on net assets, current ratio, cash flow from operating activities, cash flow from investing activities, and cash flow from financing activities; the third part is the real-time collection module for news data and e-commerce platform data, which is used to monitor real-time news reports related to the enterprise and data such as the sales volume, user ratings, and market share of the enterprise's products on the e-commerce platform; the fourth part is the macroeconomic data collection module, which is used to obtain key data on the current macroeconomic environment and capture the changing trend of the economic cycle, including a combined set of data such as the GDP growth rate, unemployment rate, inflation rate, interest rate, and the rise and fall of stock market indexes of other major economies; the fifth part is the enterprise stock price data collection module, which is used to obtain the enterprise stock price and screen out stocks whose current enterprise stock price is less than the reasonable stock price; the sixth part is the analyst valuation data collection module.

[0068] Step 3, enterprise similarity calculation. The enterprise similarity calculation module uses an enterprise similarity calculation method based on a knowledge graph and a GNN model. The specific steps are as follows:

[0069] The first step: Knowledge graph construction. The construction of the knowledge graph is based on the enterprise basic information data in the first part of the data collection module. The enterprise and its related attributes are represented as a graph structure G=(V, E), where V is the node set (representing the enterprise and its attribute nodes), and E is the edge set (representing the relationship between the enterprise and the attribute). Each enterprise v i ∈V is used as a node, and the basic information of the enterprise (industry to which it belongs, concepts to which it belongs, region to which it belongs, listing time, education level of senior executives) is used as the attribute node v j∈V, the edge between an enterprise and an attribute represents the association relationship between the enterprise and its attribute. Enterprises are connected through shared attributes, and the formed edges have edge weights, which can reflect the association strength between enterprises. The edge weights can be calculated through attribute similarity. For example, for enterprises A and B, the edge weight W is calculated based on the shared attributes A,B , and the formula is as follows:

[0070]

[0071] Finally, through the above steps, a knowledge graph G is constructed, which has a node set V and an edge set E. The formula is expressed as follows:

[0072] G=(V, E), V={v1, v2, …, v n}}, E={e ij |v i , v j ∈V}

[0073] Step 2: Feature embedding based on the GNN model. Through the structural information and node attribute information of the knowledge graph G, use the graph neural network (GNN) to generate high-dimensional feature vectors for each enterprise node. Based on the GNN model, aggregate the features of neighbor nodes to update the features of each node. For the enterprise node v i , at the l-th layer, its feature representation is updated according to the following formula:

[0074]

[0075] where W (l) is the learnable weight matrix at the l-th layer, AGGREGATE is the aggregation function, is the feature vector of node v j at the (l - 1)-th layer, N(i) is the set of neighbor nodes of node v j , and σ is the activation function.

[0076] The initial feature of the enterprise node is the core financial data in the second part of the data collection module and the real-time data of news data and e-commerce platform data in the third part. For news data, use natural language processing (NLP) technology to perform sentiment analysis on news texts and score each article (such as positive, negative, or neutral). Finally, the text data is represented by One-Hot encoding, and the numerical data is directly normalized as the initial feature.

[0077] Step 3: Similarity calculation. After L layers of propagation, the final feature vector i of each enterprise node v is obtained based on the embedded high-dimensional feature vector The cosine similarity formula is used to calculate the similarity between enterprises, and the formula is;

[0078]

[0079] where, and are the final feature vectors of enterprise i and enterprise j respectively. represents the norm of the vector. Step 4, enterprise valuation. The valuation model module adopts a valuation method based on deep learning, and the specific steps are as follows:

[0080] The first step is to construct the model training features and labels. The feature vector of each enterprise i consists of a combined set of data including asset-liability ratio, net assets per share, operating income, advance receipts, inventory, inventory turnover rate, gross profit margin, return on net assets, current ratio, cash flow from operating activities, cash flow from investing activities, and cash flow from financing activities. This combined data is processed through normalization or standardization to be converted into a numerical vector Based on the enterprise similarity data, for enterprise i, a similarity-weighted feature vector can be generated The formula is as follows:

[0081]

[0082] Combine and to form the final feature vector X of enterprise i i , and the formula is as follows:

[0083]

[0084] The target variable (label) is the enterprise valuation given by the analyst which is based on the analyst valuation data in the sixth part of the data collection module and processed by "A Dynamic Analyst Weighted Valuation Calculation Method".

[0085] The second step is to build a deep learning model. The model consists of an input layer, a hidden layer, a fully connected layer, and an output layer. The output of the L-th layer of the hidden layer is:

[0086]

[0087] where W (l) is the weight matrix of the l-th layer, b (l) is the bias vector of the l-th layer, and the initial input The fully connected layer extracts the non-linear relationship of the input features through a multi-layer fully connected neural network. The output layer is a fully connected layer that outputs the predicted reasonable market value of enterprise i

[0088]

[0089] The model training uses the mean squared error (MSE) as the loss function. For a sample with a training quantity of N, calculate the predicted value of the model and the label value The error between them is calculated by the formula:

[0090]

[0091] Step 5, macroeconomic scoring. The macroeconomic scoring module adopts a macroeconomic scoring method based on a neural network. It uses a combination of data such as GDP growth rate, unemployment rate, inflation rate, interest rate, and the rise and fall of stock market indexes of other major economies as features, and uses the stock market core index data corresponding to the next reporting date as the label to predict the impact of the macroeconomy on the stock market performance, and based on this, dynamically adjusts the enterprise valuation. According to the predicted rise and fall of the stock market index C t predicted by the neural network, map it to the macroeconomic score M t :

[0092] M t = 100 + k·C t

[0093] where k is a coefficient used to adjust the impact of the rise and fall on the score. The benchmark score is 100. The higher the predicted increase, the higher the score, and vice versa.

[0094] Step 6, calculate the enterprise valuation under the current macroeconomic situation according to the macroeconomic score M t and the reasonable market value given by the valuation model module

[0095]

[0096] Step 7, calculate the reasonable stock price of the enterprise according to the current enterprise valuation and the total number of shares of the enterprise:

[0097]

[0098] Step 8, the enterprise screening module screens out enterprises with investment value by comparing the current stock price of the enterprise and the reasonable stock price of the current enterprise, providing candidate enterprises for further analysis from other aspects such as the technical side in the next step. Combining Figure 2 The specific steps of a dynamic analyst weighted valuation calculation method are as follows:

[0099] The first step is to calculate the historical valuation error. The historical valuation error is used to evaluate the valuation accuracy of analysts. The formula is as follows:

[0100] ​

[0101] Among them, E i is the historical valuation error of analyst i; is the valuation of the enterprise given by analyst i at time t, and V t is the actual market value of the enterprise at time t. N is the total number of times analysts participate in valuation. The smaller E i is, the more accurate the analyst's valuation is.

[0102] Second step, calculate the valuation stability. The valuation stability is used to evaluate whether there are large fluctuations in the analyst's valuation. The formula is as follows:

[0103]

[0104] Among them, S i represents the valuation volatility of analyst i; represents the enterprise valuation of analyst i at time t. The smaller S i is, the more stable the analyst's valuation is.

[0105] Third step, calculate the analyst experience weight. Based on the number of years of industry experience or the number of cases of analysts, consider whether analysts have more experience in a specific industry or field. The more experienced they are, the higher the weight. The formula is as follows:

[0106]

[0107] Among them, W i represents the experience weight of analyst i; M is the total number of all analysts. The larger W i is, the more experienced the analyst is in this field.

[0108] Fourth step, calculate the comprehensive score. Considering the analyst's historical performance, valuation stability and industry experience weight comprehensively, the comprehensive score Q i of the analyst can be calculated. The formula is as follows:

[0109]

[0110] Q i The larger it is, the better the comprehensive performance of the analyst is.

[0111] Fifth step, integrate the valuations of multiple analysts. For the valuation of an enterprise, assuming that M analysts give valuations the valuations can be weighted and averaged according to the comprehensive score Q i of the analysts to obtain

[0112]

[0113] This formula can effectively integrate the valuations of different analysts while assigning higher weights to outstanding analysts.

[0114] Step 6: Dynamically adjust the analyst weights. Considering that the enterprise valuation may change over time, the weight Q of the analyst i can be updated dynamically. Every once in a while, recalculate the E i , S i and W i of the analyst and adjust their comprehensive score Q i . The dynamic adjustment formula is as follows:

[0115]

[0116] Among them, α is a regulation parameter (such as 0.7), indicating the degree of retention of the historical weight.

[0117] Table 1 shows the comparative experimental data of the valuation given by the present invention and the analyst valuations. Based on the valuation method proposed by the present invention, enterprise valuations are given using the fundamental data from December 31, 2023, December 31, 2022, December 31, 2021, and their respective previous time nodes. Taking the enterprise stock price data on December 31, 2024 as the evaluation criterion, the accuracy of the valuation results in the previous three years is tested. At the same time, the valuation data given by 12 analysts at different time periods within 3 years are selected for comparison to evaluate the reliability of this valuation method from multiple dimensions.

[0118] Table 1: Comparative experimental data of the valuation given by the present invention and the analyst valuations

[0119] Based on the Shanghai and Shenzhen 300 becomes component stock valuation error experimental data

[0120]

[0121] Based on the data in the table, it can be seen that the average valuation errors given by the method of the present invention on December 31, 2023 and December 31, 2022 are higher than the average errors of the 12 analysts. However, after removing the influence of extreme values, the median error of this method and the average error after removing the largest 5% of the error data of this method are not much different from the average errors of the 12 analysts. In addition, the valuation data given by this method on December 31, 2021 has been tested over three years, and the average error is better than the average valuation data given by the 12 analysts.

[0122] "Probability that the actual stock price > valuation by this method" refers to the probability that the stock price data of the constituent stocks of the CSI 300 Index on December 31, 2024 is greater than the valuation data given by this method on different valuation dates. This value is increasing over time, indicating that the enterprise valuation given by this method is more verified by the market over time. At the same time, the data for "Probability that the actual stock price > valuation by this method" for three years are all greater than "Probability that the actual stock price > analyst valuation (average)", indicating that the enterprise valuation given by this method is generally more accurate than that of analysts. Therefore, from the experimental data in multiple aspects, this method has good accuracy.

[0123] It should be noted that the above embodiments are not used to limit the protection scope of the present invention. Equivalent transformations or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. An enterprise valuation system based on knowledge graph and deep learning, characterized in that The system includes a data collection module, an enterprise similarity calculation module, a valuation model module, a macroeconomic scoring module and an enterprise screening module. The data collection module collects data from multiple dimensions such as enterprise basic information, financial statements, news and e-commerce platform data, macroeconomic data, enterprise stock prices and analyst valuations; The enterprise similarity calculation module calculates the similarity between enterprises by building a model based on knowledge graph and graph neural network (GNN); the valuation model module uses deep learning technology to combine enterprise financial data, similarity weighted data and analyst dynamic weighted valuation data to predict the reasonable market value of enterprises; The macroeconomic scoring module predicts stock market trends based on macroeconomic data and neural networks and dynamically adjusts corporate valuations; the corporate screening module screens companies with investment value by comparing their current stock prices with reasonable stock prices.

2. The enterprise valuation system based on a knowledge graph and deep learning according to claim 1, wherein The data collection module is divided into six parts. The first part is the enterprise basic information collection module, including the industry, concept, region, listing time, and senior management education. The second part is the financial statement data collection module, which is used to collect the core financial data of the enterprise, including the debt-to-asset ratio, net assets per share, operating income, accounts receivable, inventory, inventory turnover, gross profit margin, return on net assets, current ratio, cash flow from operating activities, cash flow from investment activities, and cash flow from financing activities. The third part is the news data and e-commerce platform data real-time collection module, which is used to monitor the news reports related to the enterprise and the sales volume, user ratings, and market share data of the enterprise's products on the e-commerce platform in real time. The fourth part is the macroeconomic data collection module, which is used to obtain key data of the current macroeconomic environment and capture the changing trends of the economic cycle, including GDP growth rate, unemployment rate, inflation rate, interest rate, and the rise and fall of stock market indices of other major economies. The fifth part is the enterprise stock price data collection module, which is used to obtain the enterprise stock price and screen out stocks whose current stock price is less than the reasonable stock price. The sixth part is the analyst valuation data collection module.

3. The enterprise valuation system based on knowledge graph and deep learning according to claim 1, characterized in that, The enterprise similarity calculation module uses an enterprise similarity calculation method based on knowledge graph and GNN model. The specific steps are as follows: Step 1: Knowledge graph construction. The construction of the knowledge graph is based on the enterprise basic information data in the first part of the data collection module. The enterprise and its related attributes are represented as a graph structure G=(V, E), where V is the set of nodes, representing the enterprise and its attribute nodes, and E is the set of edges, representing the relationship between the enterprise and the attribute. Each enterprise v i ∈V is used as a node, and the basic information of the enterprise, namely the industry it belongs to, the concept it belongs to, the region it belongs to, the listing time, and the educational background of senior executives, are used as attribute nodes v j ∈V. The edge between the enterprise and the attribute represents the association relationship between the enterprise and its attribute. Enterprises are connected through shared attributes, and the formed edges have edge weights, reflecting the association strength between enterprises. The edge weights are calculated through attribute similarity. For enterprises A and B, the edge weight W A,B is calculated based on the shared attributes, and the formula is as follows: Finally, through the above steps, a knowledge graph G is constructed, which has a node set V and an edge set E. The formula is as follows: G = (V, E), V = {v1, v2, …, v n}, E = {e ij | v i , v j ∈ V} Step 2: Based on the feature embedding of the GNN model, using the structural information and node attribute information of the knowledge graph G, generate high-dimensional feature vectors for each enterprise node using a graph neural network (GNN). Aggregate the features of neighbor nodes based on the GNN model to update the features of each node. For enterprise node v i , at the l-th layer, its feature representation The update formula is as follows: Among them, W (l) is the learnable weight matrix of the l-th layer, AGGREGATE is the aggregation function, is the feature vector of node v j at the (l - 1)-th layer, N(i) is the set of neighbor nodes of node v j and σ is the activation function. Initial features of enterprise nodes They are the core financial data in the second part of the data collection module and the real-time data of news data and e-commerce platform data in the third part. For news data, natural language processing (NLP) technology is used to perform sentiment analysis on news texts and score each article. Finally, text data is represented by One-Hot encoding, and numerical data is directly normalized as initial features. Step 3: Similarity calculation. After L-layer propagation, the final feature vector of each enterprise node v i is obtained Based on the embedded high-dimensional feature vector The cosine similarity formula is used to calculate the similarity between enterprises, and the formula is; Among them, and are the final feature vectors of enterprise i and enterprise j respectively, represents the norm of the vector.

4. The enterprise valuation system based on knowledge graph and deep learning according to claim 1, characterized in that The valuation model module adopts a valuation method based on deep learning, which includes the following steps: The first step is to construct model training features and labels. The feature vector of each enterprise i consists of a combined data set of asset-liability ratio, net assets per share, operating revenue, advance receipts, inventory, inventory turnover ratio, gross profit margin, return on net assets, current ratio, cash flow from operating activities, cash flow from investing activities, and cash flow from financing activities. This combined data is converted into a numerical vector through normalization or standardization processing. Based on enterprise similarity data, for enterprise i, a similarity-weighted feature vector is generated. The formula is as follows: Concatenate and to form the final feature vector X of enterprise i i , and the formula is as follows: The target variable (label) is the enterprise valuation given by the analyst which is obtained based on the analyst valuation data in the sixth part of the data collection module and processed by the dynamic analyst weighted valuation calculation method The second step is to build a deep learning model. The model consists of an input layer, a hidden layer, a fully connected layer, and an output layer. The output of the Lth layer of the hidden layer is: Among which W (l) is the weight matrix of the l-th layer, b (l) is the bias vector of the l-th layer, and the initial input The fully connected layer extracts the non-linear relationship of the input features through a multi-layer fully connected neural network. The output layer is a fully connected layer that outputs the reasonable market value prediction value of enterprise i The model training uses the mean squared error (MSE) as the loss function. For a sample with a training quantity of N, calculate the predicted value of the model and the label value The error between them is calculated by the formula:

5. The enterprise valuation system based on knowledge graph and deep learning according to claim 1, wherein The macroeconomic scoring module uses a neural network-based macroeconomic scoring method, taking the combined data of GDP growth rate, unemployment rate, inflation rate, interest rate, and the rise and fall of stock market indices of other major economies as features, and the core stock market index data corresponding to the next reporting day as labels, to predict the impact of the macroeconomy on stock market performance, and dynamically adjust corporate valuations based on this. The increase or decrease rate C of the stock market index predicted by the neural network t , which is mapped to the macroeconomic score M t : M t = 100 + k·C t Among them, k is a coefficient used to adjust the impact of the increase or decrease on the score. The benchmark score is 100. The higher the predicted increase, the higher the score, and vice versa. According to the macroeconomic score M t and the reasonable market value given by the valuation model module calculate the enterprise valuation under the current macroeconomic situation Based on the current enterprise valuation and the total number of shares of the enterprise, calculate the reasonable stock price of the current enterprise: The enterprise screening module screens out enterprises with investment value by comparing the current stock price of an enterprise with its reasonable stock price at present, providing candidate enterprises for further analysis from other perspectives such as technical aspects in the next step.

6. The enterprise valuation system based on knowledge graph and deep learning according to claim 4, characterized in that Dynamic Analyst Weighted Valuation Calculation Method Comprises the following steps: First step, calculate the historical valuation error, which is used to evaluate the valuation accuracy of analysts. The formula is as follows: Among them, E i The historical valuation error of analyst i; is the valuation of the enterprise given by analyst i at time t, V t is the actual market value of the enterprise at time t, N is the total number of times analysts participate in valuation, E i The smaller it is, the more accurate the analyst's valuation is. Second step, calculate the valuation stability, which is used to evaluate whether there are large fluctuations in the valuations of analysts. The formula is as follows: Among them, S i represents the valuation volatility of analyst i; represents the enterprise valuation of analyst i at time t, and the smaller S i is, the more stable the analyst's valuation is. Third step, consider whether an analyst has more experience in a specific industry or field based on the analyst's years of industry experience or the number of cases. The more experienced, the higher the weight. The formula is as follows: Among them, W i represents the experience weight of analyst i; M is the total number of all analysts, and the larger W i is, the more experienced the analyst is in this field. Step 4: Calculate the comprehensive score. Considering the analyst's historical performance, valuation stability, and industry experience weights, calculate the analyst's comprehensive score Q i , and the formula is as follows: Q i The larger it is, the more excellent the comprehensive performance of the analyst is. Step 5: For the valuation of an enterprise, assume that M analysts give valuations According to the comprehensive score Q of the analysts i Perform a weighted average on the valuations to obtain Step 6. Considering that the enterprise valuation may change over time, the weight Q of the analyst i is updated dynamically. Every once in a while, recalculate the E i , S i and W i of the analyst, and adjust their comprehensive score Q i . The dynamic adjustment formula is as follows: Where α is a regulatory parameter (such as 0.7), representing the retention degree of the historical weight.

7. An enterprise valuation method based on knowledge graph and deep learning, characterized in that, Using the system according to any one of claims 1-6, the method comprises the following steps: Step 1, system construction. The system includes a data collection module, an enterprise similarity calculation module, a valuation model module, a macroeconomic scoring module, and an enterprise screening module. Step 2, data collection. The data collection module is subdivided into six parts. The first part is the enterprise basic information collection module, including the combined data of the industry, concept, region, listing time, and executive education background to which the enterprise belongs; the second part is the financial statement data collection module, used to collect the core financial data of the enterprise, including the combined data of asset-liability ratio, net assets per share, operating income, advance receipts, inventory, inventory turnover rate, gross profit margin, return on net assets, current ratio, cash flow from operating activities, cash flow from investing activities, and cash flow from financing activities; the third part is the real-time collection module for news data and e-commerce platform data, used to monitor in real time the news reports related to the enterprise and the sales volume, user ratings, and market share data of the enterprise's products on the e-commerce platform; the fourth part is the macroeconomic data collection module, used to obtain the key data of the current macroeconomic environment and capture the changing trend of the economic cycle, including the combined data of GDP growth rate, unemployment rate, inflation rate, interest rate, and the gains and losses of stock market indexes of other major economies; the fifth part is the enterprise stock price data collection module, used to obtain the enterprise stock price and screen out the stocks whose current enterprise stock price is less than the reasonable stock price; the sixth part is the analyst valuation data collection module. Step 3, enterprise similarity calculation. The enterprise similarity calculation module uses an enterprise similarity calculation method based on a knowledge graph and a GNN model. The specific steps are as follows: Step 1: Knowledge graph construction. The construction of the knowledge graph is based on the enterprise basic information data in the first part of the data collection module. The enterprise and its related attributes are represented as a graph structure G = (V, E), where V is the set of nodes, representing the enterprise and its attribute nodes, and E is the set of edges, representing the relationship between the enterprise and the attributes. Each enterprise v i ∈V is used as a node, and the basic information of the enterprise, namely the industry it belongs to, the concepts it belongs to, the region it belongs to, the listing time, and the education background of senior executives, are used as attribute nodes v j ∈V. The edge between the enterprise and the attribute represents the association relationship between the enterprise and its attribute. Enterprises are connected through shared attributes, and the formed edges have edge weights, reflecting the association strength between enterprises. The edge weights are calculated through attribute similarity. For enterprises A and B, the edge weight W A,B is calculated based on the shared attributes, and the formula is as follows: Finally, through the above steps, a knowledge graph G is constructed, having a node set V and an edge set E, which is represented by the formula as follows: G = (V, E), V = {v1, v2, …, v n}, E = {e ij | v i , v j ∈ V} Step 2: Based on the feature embedding of the GNN model, using the structural information and node attribute information of the knowledge graph G, generate high-dimensional feature vectors for each enterprise node using a graph neural network (GNN). Aggregate the features of neighbor nodes based on the GNN model to update the features of each node. For enterprise node v i , at the l-th layer, its feature representation The update formula is as follows: Among them, W (l) is the learnable weight matrix of the l-th layer, AGGREGATE is the aggregation function, is the feature vector of node v j at the (l - 1)-th layer, N(i) is the set of neighbor nodes of node v j , σ is the activation function, and the initial feature of the enterprise node is the core financial data in the second part of the data acquisition module and the real-time data of news data and e-commerce platform data in the third part. For news data, natural language processing (NLP) technology is used to perform sentiment analysis on news texts and score each article. Finally, the text data is represented by One-Hot encoding, and the numerical data is directly normalized as the initial feature. Step 3: Similarity calculation. After L-layer propagation, the final feature vector of each enterprise node v i is obtained Based on the high-dimensional feature vector of the embedding The cosine similarity formula is used to calculate the similarity between enterprises. The formula is; Among them, and are the final feature vectors of enterprise i and enterprise j respectively, represents the norm of the vector. Step 4, enterprise valuation. The valuation model module adopts a valuation method based on deep learning, and the specific steps are as follows: Step 1: Model training feature and label construction. The feature vector of each enterprise i consists of a combined data set of asset-liability ratio, net assets per share, operating revenue, advance receipts, inventory, inventory turnover ratio, gross profit margin, return on net assets, current ratio, cash flow from operating activities, cash flow from investing activities, and cash flow from financing activities. This combined data set is processed through normalization or standardization to be converted into a numerical vector. Based on the enterprise similarity data, for enterprise i, generate a similarity-weighted feature vector. The formula is as follows: Combine and to form the final feature vector X of enterprise i i , and the formula is as follows: The target variable (label) is the enterprise valuation given by the analyst which is obtained based on the analyst valuation data in the sixth part of the data collection module and processed by the dynamic analyst weighted valuation calculation method Second step, deep learning model construction. The model consists of an input layer, a hidden layer, a fully connected layer, and an output layer. The output of the Lth layer of the hidden layer is: where W (l) is the weight matrix of the l-th layer, b (l) is the bias vector of the l-th layer, and the initial input The fully connected layer extracts the non-linear relationship of the input features through a multi-layer fully connected neural network. The output layer is a fully connected layer that outputs the reasonable market value prediction value of enterprise i The model training uses the mean squared error (MSE) as the loss function. For a sample with a training quantity of N, calculate the predicted value of the model and the label value The error between them is calculated by the formula: Step 5, Macroeconomic Scoring. The macroeconomic scoring module adopts a macroeconomic scoring method based on neural networks. It uses the combined data of GDP growth rate, unemployment rate, inflation rate, interest rate, and the rise and fall of stock market indexes of other major economies as features, and the stock market core index data corresponding to the next reporting date as labels to predict the impact of the macroeconomy on stock market performance. Based on this, it dynamically adjusts the enterprise valuation. The increase or decrease rate C of the stock market index predicted by the neural network t , which is mapped to the macroeconomic score M t : M t = 100 + k·C t Among them, k is a coefficient used to adjust the impact of the rise and fall on the score. The benchmark score is 100. The higher the predicted increase, the higher the score, and vice versa. Step 6, according to the macroeconomic score M t and the reasonable market value given by the valuation model module calculate the enterprise valuation under the current macroeconomic situation Step 7, based on the current enterprise valuation and the total number of shares of the enterprise, calculate the reasonable stock price of the current enterprise: Step 8, The enterprise screening module screens out enterprises with investment value by comparing the current stock price of the enterprise and the reasonable stock price of the current enterprise, providing candidate enterprises for further technical analysis in the next step.

8. A method for enterprise valuation based on knowledge graph and deep learning according to claim 7, characterized in that: In step 4, the dynamic analyst weighted valuation calculation is considered from three perspectives: historical valuation error, valuation stability, and analyst industry experience, and the weight of the analyst is dynamically adjusted over time to reflect the latest market data and analyst performance. The specific steps are as follows: The first step is to calculate the historical valuation error, which is used to evaluate the valuation accuracy of the analyst. The formula is as follows: Among them, E i The historical valuation error of analyst i; is the valuation of the enterprise given by analyst i at time t, V t is the actual market value of the enterprise at time t, N is the total number of times analysts participate in valuation, E i The smaller it is, the more accurate the analyst's valuation is. The second step is to calculate the valuation stability, which is used to evaluate whether there are large fluctuations in the analyst's valuation. The formula is as follows: Among them, S i represents the valuation volatility of analyst i; represents the enterprise valuation of analyst i at time t, and the smaller S i is, the more stable the analyst's valuation is. The third step is to calculate the analyst experience weight. Based on the number of years of industry experience or the number of cases of the analyst, it is considered whether the analyst has more experience in a specific industry or field. The more experienced, the higher the weight. The formula is as follows: Among them, W i represents the experience weight of analyst i, M is the total number of all analysts, and the larger the W i is, the more experienced the analyst is in this field. Step 4: Calculate the comprehensive score. Considering the analyst's historical performance, valuation stability, and industry experience weights comprehensively, calculate the analyst's comprehensive score Q i , and the formula is as follows: Q i The larger it is, the better the comprehensive performance of the analyst is. Step 5: Integrate the valuations of multiple analysts. For the valuation of an enterprise, assume that M analysts have given valuations According to the comprehensive score Q of the analysts i Perform a weighted average on the valuations to obtain This formula effectively integrates the valuations of different analysts and assigns higher weights to excellent analysts. Step 6: Dynamically adjust the analyst weights. Considering that the enterprise valuation may change over time, the weights of the analysts are updated dynamically. Every once in a while, recalculate the E i , S i and W i , and adjust their comprehensive score Q i . The dynamic adjustment formula is as follows: Among them, represents the comprehensive score of analyst i at time t. α is a regulatory parameter (such as 0.7), representing the degree of retention of historical weights.

Images