Stock portfolio recommendation method based on complex network multi-source information fusion

Abstract

A stock portfolio recommendation method based on complex network multi-source information fusion, comprising: collecting transaction data and text data of N stocks; data preprocessing; constructing a relationship matrix between the N stocks; constructing a multi-source information fusion neural network model based on a gated recurrent unit and a gated graph convolution, which is composed of a tensor fusion module, a gated recurrent unit neural network layer, a gated graph convolution network layer and an output layer in series; model iterative training, the cross loss value of the model is optimized in the iterative training, the trainable parameters of the model are updated; the stock prediction model considers the momentum spillover effect. Improve the prediction accuracy of the target stock price change, improve the robustness and reliability of the recommendation system, and provide a reference direction for the selection of stock portfolio for investors.

Description

Technical Field

[0001] This invention relates to a stock portfolio recommendation method. In particular, it relates to a stock portfolio recommendation method based on the fusion of multi-source information from complex networks. Background Technology

[0002] Financial markets, comprised of stocks, bonds, commodities, derivatives, and currencies, are characterized by nonlinearity, volatility, and chaos; they are influenced by various events and can be considered complex systems. With the development of computer technology, stock market forecasting and portfolio recommendation have become interdisciplinary fields between finance and computer science. Benefiting from powerful data analysis capabilities, the application of deep learning in financial markets has increased significantly.

[0003] Stock portfolio recommendations rely on the analysis of stock trading data, which refers to the method of studying price trends through quantitative indicators. These primarily depend on quantitative stock indicators such as opening price, closing price, trading volume, highest and lowest prices. Stock trading data directly reflects historical stock price trends, and studying the patterns of change in stock trading data helps predict future stock movements, thereby enabling stock portfolio recommendations. With stocks, funds, foreign exchange, and other financial derivatives becoming popular financial products in modern society, people can easily find news reports and investor comments about listed companies or related stocks online. This textual information influences investors' views on the future performance of related stocks to some extent, such as whether they are bullish or bearish. With the emergence of these issues, implementation plans based on big data applications have been proposed.

[0004] The momentum spillover effect is a common phenomenon in the market. In the stock market, it refers to the ability to predict the earnings of related companies by analyzing the past earnings of the company whose stock is listed. Because of supply and demand relationships or competition within the same industry, the price movements of one stock are closely related to the movements of other related stocks. However, most existing stock price prediction systems do not consider the impact of other companies on the future price fluctuations of the target stock.

[0005] Constructing a relationship network among stocks is a crucial step in studying the impact of related stock movements on stock prediction and portfolio recommendations. Therefore, finding efficient and reliable methods for constructing such networks is essential. Currently, common methods for calculating the correlation between stocks include Euclidean distance and cosine similarity. However, unlike ordinary, regular data sequences, real-world financial price sequences are subject to sudden and unpredictable price fluctuations due to unexpected events (such as terrorist attacks), which can be considered noise points. Existing methods lack specific solutions to address this problem. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a stock portfolio recommendation method based on complex network multi-source information fusion that can provide investors with a reference direction for stock portfolio selection.

[0007] The technical solution adopted in this invention is: a stock portfolio recommendation method based on multi-source information fusion of complex networks, comprising the following steps:

[0008] 1) Collect transaction data and text data for N stocks;

[0009] 2) Data preprocessing, including:

[0010] (2.1) Perform outlier detection and missing value completion on transaction data and text data;

[0011] (2.2) Convert the trading data and text data of each stock into an array of trading data sequences n. i and text feature sequence m i ;

[0012] 3) Construct a relationship matrix among the N stocks, including:

[0013] (3.1) Using the opening price sequence of each stock, construct a spatiotemporal dynamic hierarchical complex network among N stocks within a sliding window of size T days;

[0014] (3.2) Calculate the degree distribution and clustering coefficient distribution of each layer of the single stock network to obtain the mutual information of degree distribution and mutual information of clustering coefficient distribution between layers. These are used to characterize the degree correlation between pairs of stocks and the correlation of clustering coefficient between pairs of stocks, thus obtaining a dynamic stock relationship network with stocks as nodes and the correlation between pairs of stocks as the connecting edges.

[0015] (3.3) Using stocks as nodes, a dynamic stock relationship network is constructed with the degree correlation and clustering coefficient correlation between each pair of stocks as edges. This yields the degree correlation network and clustering coefficient correlation network among N stocks in each sliding window. The mean of these networks is calculated to obtain the final spatiotemporal dynamic hierarchical complex network ε, denoted as: The network representing the relationships between stocks from trading day t-T+1 to trading day t is denoted as: Where E i,j This indicates the correlation between stock i and stock j;

[0016] 4) Construct a multi-source information fusion neural network model based on gated recurrent units and gated graph convolution. This model is composed of tensor fusion module, gated recurrent unit neural network layer, gated graph convolutional network layer and output layer connected in series.

[0017] 5) Iterative training of the model, predicting the label y for the future rise or fall of each stock at time t. i And the predicted value of future stock price increases and decreases Calculate the cross-entropy loss value to obtain the cross-loss value of N stocks. Average these cross-loss values ​​yields the cross-loss value of a multi-source information fusion neural network model based on gated recurrent units and gated graph convolutions. The trainable parameters of this model are updated by optimizing the cross-loss value during iterative training. Specifically, when the opening price of the i-th stock on the (t+1)-th trading day is greater than the opening price on the t-th trading day, the future price change prediction label y for the i-th stock at time t is... i Set to 1 otherwise set to 0.

[0018] This invention presents a stock portfolio recommendation method based on complex network multi-source information fusion. By collecting various types of historical information and exploring historical patterns in trading data, it incorporates the influence of news and investor sentiment on stocks, enabling it to accurately grasp stock price fluctuations and recommend suitable stock portfolios. Inspired by the momentum spillover effect, this invention adds a stock relationship network, incorporating relevant stock information based on deep learning target stock data to improve prediction accuracy and derive suitable portfolios based on price movements. This invention innovatively introduces a gating mechanism into the graph convolutional network to filter out minor fluctuations in related stocks, increasing the stability and reliability of the system's predictions. It incorporates the momentum spillover effect into the stock prediction model, improving the accuracy of predicting target stock price changes and enhancing the robustness and reliability of the recommendation system. This invention utilizes a feedforward neural network to construct the model's output layer, enabling accurate prediction of the opening price fluctuations of each stock and deriving portfolios based on the prediction results. In summary, this invention can provide investors with a reference direction for stock portfolio selection. Attached Figure Description

[0019] Figure 1 This is a flowchart of the stock portfolio recommendation method based on complex network multi-source information fusion of the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of the multi-source information fusion neural network model of the present invention;

[0021] Figure 3 This is a schematic diagram of the construction of a single-layer network in the spatiotemporal dynamic hierarchical complex network of this invention. Detailed Implementation

[0022] The following detailed description of the stock portfolio recommendation method based on complex network multi-source information fusion of the present invention, with reference to embodiments and accompanying drawings, is provided in conjunction with the embodiments and accompanying drawings.

[0023] The present invention provides a stock portfolio recommendation method based on complex network multi-source information fusion. It collects transaction data and text data of target stocks and related stocks, combines them with a spatiotemporal dynamic hierarchical stock relationship network constructed on the stock transaction data, and performs multi-source information fusion and learning through the multi-source information fusion neural network model based on gated recurrent units and gated graph convolution proposed in this invention to predict the rise and fall of target stock prices.

[0024] like Figure 1 As shown, the stock portfolio recommendation method based on complex network multi-source information fusion of the present invention includes the following steps:

[0025] 1) Collect transaction data and text data for N stocks;

[0026] This embodiment takes the prediction of the opening price fluctuations of N stocks as an example, and uses stocks predicted to rise as the investment portfolio. It stipulates that if the opening price of stock i on day t is greater than the opening price on day t-1, it is considered a rise; if it is less than the opening price on day t-1, it is considered a fall. The system collects transaction data and text data for the N stocks.

[0027] The transaction data refers to the data information generated by each stock in the market, including five attributes: the stock's opening price, closing price, highest price, lowest price, and trading volume. The text data includes daily Tencent News articles about N stocks and investor discussions about N stocks on the Eastmoney Forum.

[0028] 2) Data preprocessing, including:

[0029] (2.1) Perform outlier detection and missing value completion on transaction data and text data;

[0030] (2.2) Convert the trading data and text data of each stock into an array of trading data sequences n. i and text feature sequence m i ;include:

[0031] (2.2.1) Construct the trading feature vector of the i-th stock on the t-th trading day using the L'-dimensional attribute data of that stock. The transaction data sequence of the i-th stock is represented as n i ;

[0032] (2.2.2) Sentiment extraction is performed on each text data of each stock by using the Chinese Financial Sentiment Dictionary (CFSD) to obtain the sentiment characteristics of the news related to each stock on the stock's market performance, including: positive, neutral and negative.

[0033] (2.2.3) Extract the sentiment features from the text information of the i-th stock on trading day t, and use an L-dimensional text feature vector. If we express this as m, then the text feature sequence of the i-th stock is represented as m. i .

[0034] 3) Construct a relationship matrix among the N stocks, including:

[0035] (3.1) Using the opening price sequence of each stock, construct a spatiotemporal dynamic hierarchical complex network among N stocks within a sliding window of size T days; specifically including:

[0036] (3.1.1) Set the finite crossing line of sight to e;

[0037] (3.1.2) Treat each trading day as a single stock network node;

[0038] (3.1.3) At two nodes and Horizontal lines are constructed between the histograms of the two nodes, with the height of the line being the height of the lower histogram of the two nodes. Among them, setting nodes This indicates that the i-th stock is in the t-th phase. a The opening price for the day was node This indicates that the i-th stock is in the t-th phase. b The opening price for the day was t a ,t b ∈[t-T+1,t] represents the range from the t-th trading day to the t-th trading day. a Tianhe t b Heaven, like Figure 3 As shown;

[0039] The two nodes are specified. and If a horizontal line intersects only e or fewer intermediate nodes, then these two nodes... and If there is an edge between them, then the nodes are considered to be in a single-stock network with a finite traversal distance of e. and nodes There are no connecting edges between them; therefore, a single-layer single-stock network G is constructed for each stock. i That is, to obtain an N-layer single-stock network for N stocks.

[0040] (3.2) Calculate the degree distribution and clustering coefficient distribution of each layer of the individual stock network to obtain the mutual information of degree distribution and clustering coefficient distribution between layers. These are used to characterize the pairwise degree correlation and pairwise clustering coefficient correlation of stocks, respectively, thus obtaining a dynamic stock relationship network with stocks as nodes and pairwise correlations of stocks as edges; where,

[0041] The obtained degree distribution mutual information between layers is used to characterize the degree correlation between any two stocks, specifically including:

[0042] The degree of a single stock network node in layer i is h. i The single-stock network of layer i at node i The degree value is The degree distribution of the nodes in this stock network is represented as p(h i ); by calculating the degree sequence of individual stock network nodes for the i-th stock. The degree sequence of a single stock network node for the j-th stock. The mutual information is used to describe the degree of correlation between the i-th stock and the j-th stock, and to determine the magnitude of the correlation between each pair of channels, and the degree of correlation between the i-th stock and the j-th stock. for:

[0043]

[0044] Among them, p(h i Let p(h) be the degree distribution of the network node of the i-th stock. i ,h j ) represents the node joint degree distribution of the i-th stock network and the j-th stock network.

[0045] The obtained mutual information of clustering coefficient distributions between layers is used to characterize the correlation of clustering coefficients between pairs of stocks, specifically including:

[0046] The clustering coefficient of the single-stock network in layer i is c. i The single-stock network of layer i at node i The clustering coefficient is in Represents a node The number of its neighboring nodes. Represents a node The actual number of edges between the node and its neighbors. Represents a node The maximum number of edges between the i-th stock and its neighboring nodes is used to determine the clustering coefficient distribution of the i-th stock as p(c i ); by calculating the clustering coefficient sequence of individual stock network nodes for the i-th stock. Clustering coefficient sequence of individual stock network nodes of the j-th stock The mutual information is used to represent the correlation of the pairwise clustering coefficients between the i-th stock and the j-th stock. for:

[0047]

[0048] Wherein, p(c i Let p(c) be the clustering coefficient distribution of the i-th stock. i ,c j ) is the joint clustering coefficient distribution of the i-th stock and the j-th stock.

[0049] (3.3) Using stocks as nodes, a dynamic stock relationship network is constructed with the degree correlation and clustering coefficient correlation between each pair of stocks as edges. This yields the degree correlation network and clustering coefficient correlation network among N stocks in each sliding window. The mean of these networks is calculated to obtain the final spatiotemporal dynamic hierarchical complex network ε, denoted as: The network representing the relationships between stocks from trading day t-T+1 to trading day t is denoted as: Where E i,j This indicates the correlation between stock i and stock j;

[0050] 4) Construct a multi-source information fusion neural network model based on gated recurrent units and gated graph convolution, such as... Figure 2 As shown, the model consists of a tensor fusion module, a gated recurrent neural network layer, a gated graph convolutional network layer, and an output layer connected in series; among them,

[0051] The tensor fusion module includes:

[0052] Used to capture transaction data sequence m i and text feature sequence n i The intrinsic connection of the K-dimensional bidirectional linear tensor product term m i Γ [1:K] n i The expression is as follows:

[0053]

[0054] in, Let l represent the l-th feature in the trading data sequence of the i-th stock on a given trading day. This represents the l'-th feature in the text feature sequence of the i-th stock on that trading day. It is a third-order tensor and Γ [1:K] =[Γ 1 ,Γ 2 ,...,Γk ,...Γ K ],Γ [1:K] One of the elements is

[0055] The trading sequence and text feature sequence of the i-th stock on a given trading day are weighted by a weight matrix. The feature fusion vector is obtained by concatenation and linear transformation. The expression is as follows:

[0056] x i =tanh(m i Γ [1:K] n i +W i [m i ||n i +b i ]),

[0057] Where || denotes matrix concatenation, It's the bias, tanh is the activation function, and the parameters are... Obtained through model learning.

[0058] The gated recurrent unit neural network layer is used to perform deep feature extraction on the feature fusion vector from day tT to day t, to obtain a deep fusion feature vector.

[0059] v i =GRU i (χ [t-T:t] ),

[0060] Where D is the hidden layer dimension of the gated recurrent unit neural network layer, and also the deep fusion feature vector v of the i-th stock. i Feature dimensions, The feature fusion vector x represents the sum of the features from trading day t-T+1 to trading day t. i t The set, using To represent the deep fusion feature vector v of each stock i A set of.

[0061] The construction of the gated graph convolutional network layer involves learning the relationship features s between the i-th stock and the j-th stock from the input stock relationship network. i,j The expression is as follows:

[0062]

[0063]

[0064] Among them, E i,j∈ε represents the degree of correlation between the standardized i-th stock and the j-th stock, while E i,j ∈[0,1], the closer to 1, the stronger the correlation between the i-th stock and the j-th stock; the closer to 0, the weaker the correlation between the i-th stock and the j-th stock. It is a weight matrix common to all stocks, which is obtained by training a multi-source information fusion neural network model based on gated recurrent units and gated graph convolution; v represents the deep fusion feature vector of each stock. i The set of values, i.e., the output of the gated recurrent unit neural network layer; σ is the sigmoid function; Let c(·) represent the feature vector of the relationship between the i-th stock and the other N-1 related stocks; It is the weight matrix in the gating mechanism. It is the bias, where D represents the deep fusion feature vector v of each stock. i The feature dimension, D', represents the feature vector s representing the relationship between the i-th stock and the other N-1 related stocks. i The feature dimensions.

[0065] The output layer consists of two layers of feedforward neural networks with a softmax function. Its output is a prediction of the future price movements of N stocks, where the expression for the prediction of the future price movement of each stock is:

[0066]

[0067] in, This is the weight matrix, obtained by training a multi-source information fusion neural network model based on gated recurrent units and gated graph convolution; C is the number of categories to which the stock's future price movement belongs; b i It is a bias, which is a multi-source information fusion neural network model based on gated recurrent units and gated graph convolution; It is the predicted future rise or fall value of the i-th stock. Based on the predicted future rise or fall value of each stock, all stocks predicted to rise are selected to obtain a stock portfolio. It is also the final output of a multi-source information fusion neural network model based on gated recurrent units and gated graph convolution.

[0068] 5) Iterative training of the model, predicting the label y for the future rise or fall of each stock at time t. i And the predicted future rise and fall of stocks Calculate the cross-entropy loss value to obtain the cross-loss value of N stocks. Average these cross-loss values ​​yields the cross-loss value of a multi-source information fusion neural network model based on gated recurrent units and gated graph convolutions. The trainable parameters of this model are updated by optimizing the cross-loss value during iterative training. Specifically, when the opening price of the i-th stock on the (t+1)-th trading day is greater than the opening price on the t-th trading day, the future price change prediction label y for the i-th stock at time t is... i Set to 1 otherwise set to 0.

[0069] This invention presents a stock portfolio recommendation method based on complex network multi-source information fusion. It innovatively proposes a multi-information fusion neural network model (GG-MFNet) based on gated recurrent units and gated graph convolutions. GG-MFNet improves the model's stock portfolio recommendation capability by designing corresponding modules from two aspects: in-depth feature extraction and fusion from multiple sources, and leveraging momentum spillover effects to utilize more comprehensive market information.

[0070] The feature extraction and fusion module in the GG-MFNet model designed in this invention includes a tensor fusion module for fusing features from textual and transaction data, and a gated recurrent unit module for deep feature extraction to find time dependencies in the multi-source data fusion features of stocks. Compared to traditional methods based solely on transaction data, this invention further utilizes a deep learning model to fuse multi-source information: stock transaction data and massive amounts of textual information, extracting more effective features to achieve effective multi-source information fusion, improving the accuracy of stock prediction and the rationality of portfolio recommendations. Finding the time-varying trends of transaction and textual data is crucial for predicting stock price trends. To address this issue, this invention uses a gated recurrent unit (GRU) network in the proposed GG-MFNet model to further reduce the dimensionality and extract features from the fused features of stock net asset value. The gated recurrent unit network is an improvement on the LSTM neural network, which not only solves the gradient vanishing problem in RNNs but also further reduces the number of parameters on the basis of the LSTM neural network, reducing the risk of overfitting.

[0071] Secondly, the GG-MFNet designed in this invention innovatively utilizes the momentum spillover effect. By controlling the dynamic correlations between stocks, it uses more comprehensive market information for portfolio selection. Compared to existing methods that independently predict each stock and then select and package them into a portfolio, this invention innovatively proposes a method to find a more reliable stock relationship network by constructing a spatiotemporally dynamic hierarchical complex network among multiple stocks. This relationship network is then integrated into the model using a gated graph convolution module, innovatively utilizing the market momentum spillover effect and thus improving the model's ability to recommend stock portfolios.

[0072] This invention fully considers the impact of momentum spillover effect on stocks when constructing a stock recommendation system, combines the correlation of trading information of multiple stocks, predicts the future return trend of stocks, and makes a reasonable stock recommendation system based on the prediction results of each stock.

[0073] This invention innovatively proposes a method for constructing a spatiotemporal dynamic hierarchical complex network to calculate the correlation network between stocks. The proposed spatiotemporal dynamic hierarchical complex network can not only be used to analyze multivariate time series and construct correlation networks for multiple stocks, but also has better noise resistance, making it more suitable for the analysis of measured time signals. This invention constructs a relationship network for multiple related stocks using a spatiotemporal dynamic hierarchical complex network. This stock relationship network uses stocks as network nodes and the correlation between each pair of stocks as network edges. After obtaining the relationship network, how to organically combine the stock relationships with the fusion characteristics of each stock, and realize the use of relevant stock feature information to assist in predicting the price changes of the target stock, is also an important aspect that CG-MFNet in this invention needs to consider. This invention innovatively adds a proposed Gate-graph convolutional neural network (GGCN) to the CG-MFNet model. This GGCN calculates a weighted sum of the deep features of the target stock and other related stocks. Essentially, it obtains the correlation between the target stock and other related stocks through a stock relationship network. Based on this correlation, the deep features of the related stocks are fused into the features of the target stock, allowing for more comprehensive market information and improving prediction accuracy. However, in actual financial markets, not all changes in related stocks affect the price of the target stock. For example, if the price of a related stock falls but the trading volume is small, the change in the related stock will not affect the target stock. Therefore, compared to traditional graph convolutional neural networks, this invention innovatively introduces a gating mechanism into the graph convolutional network to filter out such small changes in related stocks, increasing the stability and reliability of the system's predictions. This process allows the stock prediction model to consider the momentum spillover effect, improving the prediction accuracy of target stock price changes and enhancing the robustness and reliability of the recommendation system. Finally, a feedforward neural network is used to construct the model output layer to accurately predict the opening price fluctuations of each stock, and an investment portfolio is generated based on the prediction results for each stock.

[0074] The above description of the present invention and its embodiments is not limited thereto. The description in the embodiments is only one of the implementation methods of the present invention. Without departing from the spirit of the present invention, any structure or embodiment similar to the technical solution designed without creativity shall be within the protection scope of the present invention.

Claims

1. A multi-source information fusion method based on gated recurrent units and gated graph convolution, characterized in that, Includes the following steps: 1) Collect transaction data and text data for N stocks; 2) Data preprocessing; 3) Construct a relationship matrix among the N stocks; 4) Construct a multi-source information fusion neural network model based on gated recurrent units and gated graph convolutions. This model consists of a tensor fusion module, a gated recurrent unit neural network layer, a gated graph convolutional network layer, and an output layer connected in series. The tensor fusion module includes: a module for capturing transaction data sequences. and text feature sequences The intrinsic connection of the K-dimensional bidirectional linear tensor product terms The expression is as follows: ； in, Let l represent the l-th feature in the trading data sequence of the i-th stock on a given trading day. This represents the l'-th feature in the text feature sequence of the i-th stock on that trading day. It is a third-order tensor and , One of the elements is ; The trading sequence and text feature sequence of the i-th stock on a given trading day are weighted by a weight matrix. The feature fusion vector is obtained by concatenation and linear transformation. The expression is as follows: ； in, Indicates matrix concatenation. It's the bias, tanh is the activation function, and the parameters are... Obtained through model learning; The gated recurrent unit neural network layer is used to perform deep feature extraction on the feature fusion vector from day tT to day t, to obtain the deep fusion feature vector. : ； Where D is the hidden layer dimension of the gated recurrent unit neural network layer, and also the deep fusion feature vector of the i-th stock. Feature dimensions, This represents the feature fusion vector from trading day t-T+1 to trading day t. The set, using To represent the deep fusion feature vector of each stock A set; The construction of the gated graph convolutional network layer involves learning the relationship features between the i-th stock and the j-th stock from the input stock relationship network. The expression is as follows: ； in, This represents the degree of correlation between the standardized i-th stock and the j-th stock, while... The closer the correlation is to 1, the stronger the correlation between the i-th stock and the j-th stock; the closer the correlation is to 0, the weaker the correlation between the i-th stock and the j-th stock. It is a weight matrix common to all stocks, which is obtained by training a multi-source information fusion neural network model based on gated recurrent units and gated graph convolution; Represents the deep fusion feature vector of each stock The set of results, namely, the output of the gated recurrent unit neural network layer; It is the sigmoid function; Let the feature vector representing the relationship between the i-th stock and the other N-1 related stocks be represented. Indicates a gating mechanism; It is the weight matrix in the gating mechanism. It is the bias, where D represents the deep fusion feature vector of each stock. The feature dimension, D', represents the feature vector of the relationship between the i-th stock and the other N-1 related stocks. The feature dimensions.

2. The multi-source information fusion method based on gated recurrent units and gated graph convolution according to claim 1, wherein the data preprocessing in step 2) includes: (2.1) Perform outlier detection and missing value completion on transaction data and text data; (2.2) Convert the trading data and text data of each stock into a sequence of trading data in array form. and text feature sequences ,include: (2.2.1) Construct the trading feature vector of the i-th stock on the t-th trading day using the L'-dimensional attribute data of that stock. Then the transaction data sequence of the i-th stock is represented as: ; (2.2.2) Sentiment extraction is performed on each text data of each stock. This is done by using a Chinese financial sentiment dictionary to obtain the sentiment characteristics of the news related to each stock on the stock's market performance, including: positive, neutral and negative. (2.2.3) Extract the sentiment features from the text information of the i-th stock on trading day t, and use an L-dimensional text feature vector. If we express this as an expression, then the text feature sequence of the i-th stock is represented as follows: .

3. The multi-source information fusion method based on gated recurrent units and gated graph convolution according to claim 1, wherein step 3) of constructing the relationship matrix among N stocks includes: (3.1) Using the opening price sequence of each stock, construct a spatiotemporal dynamic hierarchical complex network among N stocks within a sliding window of size T days; (3.2) Calculate the degree distribution and clustering coefficient distribution of each layer of the single stock network to obtain the mutual information of degree distribution and mutual information of clustering coefficient distribution between layers. These are used to characterize the degree correlation between pairs of stocks and the clustering coefficient correlation between pairs of stocks, thus obtaining a dynamic stock relationship network with stocks as nodes and the correlation between pairs of stocks as the connecting edges. (3.3) Using stocks as nodes, construct a dynamic stock relationship network with the degree correlation and clustering coefficient correlation between each pair of stocks as edges. This yields the degree correlation network and clustering coefficient correlation network among N stocks under each sliding window. Calculate the mean of the degree correlation network and clustering coefficient correlation network to obtain the final spatiotemporal dynamic hierarchical complex network. , represented as The network representing the relationships between stocks spanning from trading day t-T+1 to trading day t is denoted as: ,in This represents the correlation between stock i and stock j.

4. The multi-source information fusion method based on gated recurrent units and gated graph convolution as described in claim 1, characterized in that, Step 3) Step (3.1) specifically includes: (3.1.1) Set the finite crossing line of sight to e; (3.1.2) Treat each trading day as a single stock network node; (3.1.3) At two nodes and Horizontal lines are constructed between the histograms of the two nodes, with the height of the line being the height of the lower histogram of the two nodes. Among them, setting nodes This indicates that the i-th stock is in the t-th phase. a The opening price for the day was ,node This indicates that the i-th stock is in the t-th phase. b The opening price for the day was , It represents the range from trading day t-T+1 to trading day t. a Tianhe t b sky; The two nodes are specified. and If a horizontal line intersects only e or fewer intermediate nodes, then these two nodes... and If there is an edge between them, then the nodes are considered to be in a single-stock network with a finite traversal distance of e. and nodes There are no connecting edges between them; therefore, a single-stock network is constructed for each stock. That is, to obtain an N-layer single-stock network for N stocks. .

5. The multi-source information fusion method based on gated recurrent units and gated graph convolution according to claim 1, characterized in that, Step 3) Step (3.2) obtains the degree distribution mutual information between layers to characterize the degree correlation between pairs of stocks, specifically including: The degree value of a single stock network node in layer i is The single-stock network of layer i at node i The degree value is The degree distribution of the nodes in this stock network is represented as follows: ; By calculating the degree sequence of individual stock network nodes for the i-th stock. The degree sequence of a single stock network node for the j-th stock. The mutual information is used to describe the degree of correlation between the i-th stock and the j-th stock, and to determine the magnitude of the correlation between each pair of channels, and the degree of correlation between the i-th stock and the j-th stock. for: ； in, It represents the degree distribution of the network nodes of the i-th stock. It represents the node joint degree distribution of the i-th stock network and the j-th stock network; The mutual information of clustering coefficient distributions between layers is used to characterize the correlation of clustering coefficients between pairs of stocks, specifically including: The clustering coefficient of the single-stock network in layer i is The single-stock network of layer i at node i The clustering coefficient is ,in Represents a node The number of its neighboring nodes. Represents a node The actual number of edges between the node and its neighbors. Represents a node The maximum number of edges between the i-th stock and its neighboring nodes is used to determine the clustering coefficient distribution of the i-th stock. ; By calculating the clustering coefficient sequence of individual stock network nodes for the i-th stock. Clustering coefficient sequence of individual stock network nodes of the j-th stock The mutual information is used to represent the correlation of the pairwise clustering coefficients between the i-th stock and the j-th stock. for: ； in, It is the clustering coefficient distribution of the i-th stock. It is the distribution of the joint clustering coefficients of the i-th stock and the j-th stock.

Images