WO2022105016A1

WO2022105016A1 - Stock price trend prediction method and system, terminal, and storage medium

Info

Publication number: WO2022105016A1
Application number: PCT/CN2020/139674
Authority: WO
Inventors: 叶洁瑕; 赵娟娟; 叶可江; 须成忠
Original assignee: 中国科学院深圳先进技术研究院
Priority date: 2020-11-19
Filing date: 2020-12-25
Publication date: 2022-05-27
Also published as: CN112365075A

Abstract

A stock price trend prediction method and system, a terminal, and a storage medium. The method comprises: obtaining an original price feature of a stock set in a predetermined number of historical trading days; constructing a stock right network relationship graph of the stock set, each vertex of the stock right network relationship graph corresponding to one stock; inputting a historical price feature of the stock set and the stock right network relationship graph into a deep learning model, the deep learning model being used for performing feature aggregation and nonlinear transformation on vertices of the stock right network relationship graph by means of a graph convolutional neural network, extracting an interaction between the vertices of the stock right network relationship graph, and constructing a new price feature of the stock set; combining the new price feature and the original price feature of the stock set and then inputting same to a gated recurrent neural network, so as to obtain a new feature comprising a time series feature; and predicting a stock price trend of the stock set according to the new feature. The method can improve the accuracy of stock price trend prediction.

Description

A stock price trend prediction method, system, terminal and storage medium

technical field

The present application belongs to the technical field of stock data analysis, and in particular relates to a stock price trend prediction method, system, terminal and storage medium.

Background technique

With the development of China's economy, the stock market has continued to grow, and more and more individual and institutional investors have entered the stock market. The price fluctuation of the stock market affects the income of investors, thus becoming the focus of many investors. Accurate stock price forecasts can bring huge trading volume and income. However, the stock market is extremely complex, and stock price changes are affected by many factors. At the same time, the stock market data is numerous and complicated. How to comprehensively and effectively mine the massive stock data contained in it? It has become a hot technical issue to accurately predict the trend of stock prices.

In the prior art, traditional stock data analysis methods mainly include:

(1) Fundamental analysis: Fundamental analysis is based on traditional economic theories, through detailed analysis of the company's intrinsic value, macroeconomic situation, industry development prospects, business status and other information, to calculate the long-term investment value and marginal safety of listed companies, and form investment recommendations. Fundamental analysis believes that stock price fluctuations are unpredictable, and stocks can only be bought with sufficient margins and held for a long time.

(2) Technical analysis: Technical analysis is based on traditional securities theory, taking stock price as the research object, and studying the historical chart of stock price changes. It believes that market behavior digests everything, and stock price fluctuations can be quantitatively analyzed and predicted. The main theories are Dow Theory, Wave Theory, Gann Theory, etc.

(3) Evolutionary analysis: Based on evolutionary securities theory, evolutionary analysis takes the life movement characteristics of stock market fluctuations as the research object, and analyzes the metabolism, profit-seeking, and plasticity of the stock market. It believes that stock price volatility cannot be accurately predicted, and turned to a new analytical framework for investors.

The above traditional analysis methods require analysts to have professional financial knowledge and rich stock market trading experience. The era of big data has spawned a series of new methods and technologies such as mathematical statistics, machine learning, and deep learning. These technologies have reduced the dependence on professional knowledge, and are driven by data to discover the hidden complex laws in stock price changes. more and more attention.

At present, common mathematical statistical models include Vector autoregression (Vector autoregression, VAR for short) model and Conditional autoregressive range (Conditional autoregressive range, CARR for short) model. But they all need to analyze the influencing factors of stocks and rely on expert knowledge to a certain extent. At the same time, they can only simply fit time series and cannot handle more complex data structures.

Traditional machine learning methods such as support vector machines and Markov chains can model and train a large amount of historical data, mine high-dimensional nonlinear features of stock price fluctuations hidden in the data, and predict stock prices. However, they cannot simulate long-term time series characteristics in stock data.

Among deep learning methods, Recurrent Neural Network (RNN) and its variants can model time series data and extract long-term dependencies in stock data. However, the network cannot handle other factors outside the time series, such as the mutual influence between stocks due to the connection of various economic businesses.

SUMMARY OF THE INVENTION

The present application provides a stock price trend prediction method, system, terminal and storage medium, aiming to solve one of the above technical problems in the prior art at least to a certain extent.

In order to solve the above problems, the application provides the following technical solutions:

A stock price trend prediction method, including:

Obtain the original price characteristics of the stock set on the historical trading days of the predetermined number of days; the stock set includes at least two stocks;

Construct an equity network relationship graph of the stock set based on prior knowledge, and each vertex in the equity network relationship graph corresponds to a stock respectively;

The historical price features of the stock set and the equity network relationship graph are input into the deep learning model, and the deep learning model performs feature aggregation and nonlinear transformation on the vertices on the equity network relationship graph through the graph convolutional neural network, and extracts all the vertices. The interaction between the vertices on the equity network relationship graph, and construct the new price feature of the stock set;

After splicing the new price feature of the stock set with the original price feature, input the gated recurrent neural network, and output the new feature including the time series feature through the gated recurrent neural network;

The stock price trend of the stock set is predicted according to the new feature including the time series feature.

The technical solution adopted in the embodiment of the present application further includes: the acquisition of the original price characteristics of the stock set in a predetermined number of historical trading days includes:

Perform Z-score normalization processing on the original price feature;

The original price characteristics include closing price, opening price, trading volume, and trading turnover.

The technical solutions adopted in the embodiments of the present application further include: the equity network relationship diagram includes an equity relationship diagram, an industry relationship diagram, and a topic relationship diagram, specifically:

G _S =(V,E _S ,A _S )

G _I =(V,E _I ,A _I )

G _T =(V,E _T ,A _T )

In the above formula, G _S is the equity relationship diagram, G _I is the industry relationship diagram, and _GT is the topic relationship diagram; V={v ₁ ,...,v _N } is the vertex set of the graph, and v _i is the ^ith vertex , there are N vertices in total, and the vertex sets of each graph are the same, which is the stock set S; E is the graph edge set, each edge represents the connection relationship between stocks, and the edge sets of different stock graphs are different; A=(a _ij ) _N×N is the adjacency matrix of the graph, and the element a _ij of the matrix is the weight of the edge, which represents the influence factor between stock i and stock j.

The technical solutions adopted in the embodiments of the present application further include: the graph convolutional neural network performs feature aggregation and nonlinear transformation on the vertices on the equity network relationship graph, specifically:

The interaction of the stock set in each relation graph is extracted according to the Laplacian matrix of the stock network relation graph. The formula of each layer of the graph convolutional neural network is:

In the above formula, the adjacency matrix corresponding to the Laplacian matrix {L _S , L _I , L _T } is {A _S , A _I , A _T },

I _N is the identity matrix, D is the adjacency matrix of A;

is a trainable parameter, K is the order of the Laplacian matrix, indicating the radius of the graph convolution, θ _k is the trainable coefficient of the k-th order Laplacian matrix; ρ is the activation function,

are the trainable parameters of the lth layer,

is the first layer of the graph convolutional neural network,

is the original price feature of the stock set S on day t, and each stock includes F original price features;

The original price feature

The new price feature output after the graph convolutional neural network is

Each stock includes C new price features.

The technical solutions adopted in the embodiments of the present application further include: the hidden layer passing through the gated recurrent neural network is:

In the above formula, the input time step t∈[d-P+1,...,d-1],

is the hidden layer state of the gated recurrent neural network at the input time step t-1, r _t is the reset gate, u _t is the update gate; σ∈[0,1] is the activation function, ⋅ is the matrix multiplication, ⊙ is Huffman product; {W _r ,W _u ,W _h } are trainable weight parameters, { _br , _bu ,b _h } are trainable bias terms; the output layer of the gated recurrent neural network Yes

in

Each stock includes G new features.

The technical solution adopted in the embodiment of the present application further includes: the prediction of the stock price trend of the stock set according to the new feature including the time series feature is specifically:

The new features output by the gated recurrent neural network are input into the fully connected layer with the sigmoid activation function, and the probability of rising and falling of the stock set is generated through the fully connected layer. The formula is as follows:

In the above formula,

is the stock rise and fall trend of the stock set S on the d day,

is the forecast of the rising and falling trend of stock i on the d day;

are trainable parameters.

The technical solutions adopted in the embodiments of the present application further include: the deep learning model uses cross-entropy as a loss function, specifically as follows:

In the above formula,

is the real price trend value of the stock s on the d day, and the real price trend value is the ratio of the gap between the opening price of the stock on the t day and the opening price of the previous day.

Another technical solution adopted by the embodiment of the present application is: a stock price trend prediction system, comprising:

Data acquisition module: used to acquire the original price characteristics of a stock set on historical trading days of a predetermined number of days; the stock set includes at least two stocks;

Relationship graph building module: used to construct an equity network relationship graph of the stock set based on prior knowledge, and each vertex in the equity network relationship graph corresponds to a stock respectively;

Graph convolution module: used to input the historical price features of the stock set and the equity network relationship graph into a deep learning model, and the deep learning model performs feature aggregation on the vertices on the equity network relationship graph through a graph convolutional neural network and nonlinear transformation, extract the interaction between the vertices on the equity network relationship graph, and construct the new price feature of the stock set;

Time series feature extraction module: used to input the gated recurrent neural network after splicing the new price features of the stock set with the original price features, and output new features including time series features through the gated recurrent neural network;

Result output module: used to predict the stock price trend of the stock set according to the new feature including time series features.

Another technical solution adopted by the embodiments of the present application is: a terminal, the terminal includes a processor and a memory coupled to the processor, wherein,

The memory stores program instructions for implementing the stock price trend prediction method;

The processor is configured to execute the program instructions stored in the memory to control stock price trend prediction.

Another technical solution adopted by the embodiments of the present application is: a storage medium storing program instructions executable by a processor, where the program instructions are used to execute the stock price trend prediction method.

Compared with the prior art, the beneficial effects of the embodiments of the present application are: the stock price trend prediction method, system, terminal and storage medium of the embodiments of the present application describe the relationship between stocks from multiple perspectives (equity, industry, topic). , and combined with the historical price features of stocks, using graph convolutional neural network to extract the cross influence between stock prices, capture the interaction between stock prices, and build new price features based on this. After splicing the new price features and historical price features, the time series features and cross-influence features of the stock are extracted through the gated recurrent neural network, and then the stock price trend prediction is carried out. The embodiment of the present invention fully considers the influence of the interaction between stock prices on the ups and downs of stocks, and extracts the time series features and cross-influence features in the price data, thereby greatly improving the accuracy of stock price trend prediction.

Description of drawings

FIG. 1 is a system schematic diagram of a deep learning model based on a diagram in an embodiment of the present application;

Fig. 2 is the flow chart of the stock price trend prediction method of the embodiment of the present application;

3 is a schematic structural diagram of a stock price trend prediction system according to an embodiment of the application;

FIG. 4 is a schematic structural diagram of a terminal according to an embodiment of the present application;

FIG. 5 is a schematic structural diagram of a storage medium according to an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

In view of the deficiencies of the prior art, the stock price trend prediction method of the embodiment of the present application builds a deep learning model based on a graph, and constructs a stock relationship diagram, an industry relationship diagram and a topic relationship diagram based on three relationship networks for the stocks that affect each other. . On each graph, combined with the historical price features of the stock, a graph convolutional neural network is used to simulate the cross influence between stock prices, capture the interaction between stock prices, and build new price features based on this. The new price features and historical price features are input into the gated recurrent neural network to extract the time series features and cross-influence features of the stock, and then predict the stock price trend. As shown in FIG. 1 , it is a system schematic diagram of a graph-based deep learning model according to an embodiment of the present application. The model formula is as follows:

In formula (1), d is the trading day for predicting the stock price trend, P is the number of historical trading days used for prediction,

is the prediction result of the stock price trend of the stock set S output by the model on the trading day d. G is the equity network relationship graph of the stock set S, and Θ is the trainable parameter of the model.

In Figure 1, the deep learning model includes a multi-relation-driven graph convolutional neural network Multi-GCN, a gated recurrent neural network GRU, and a fully connected layer FC. The input of the model is the historical price data of the stock set S in the past P trading days and the equity network relationship diagram of the stock set S. After the graph convolutional neural network Multi-GCN aggregation, the new price features of the stock set S are output, and the stock set S is aggregated. After splicing the original price features and new price features of S, the total price features are obtained, and the total price features are input into the gated recurrent neural network GRU to extract the time series features of stocks, and the new features of the stock set S are output, and the new features are input with The fully connected layer FC of the sigmoid activation function obtains the prediction result of the stock price trend of the stock set S on the d day.

Please refer to FIG. 2 , which is a flowchart of a method for predicting a stock price trend according to an embodiment of the present application. The stock price trend prediction method in the embodiment of the present application includes the following steps:

S10: Obtain the original price characteristics of the stock set S in the past P historical trading days, and perform Z-score normalization processing on the original price characteristics;

In this step, the stock set S includes N stocks, each stock has F original price characteristics, and the original price characteristics include but are not limited to closing price, opening price, trading volume, and trading turnover rate.

S20: construct the influence relationship between the stocks in the stock set S based on the prior knowledge, and construct the equity network relationship diagram of the stock set S based on the influence relationship;

In this step, through the equity network relationship diagram constructed based on prior knowledge, including the equity relationship diagram, the industry relationship diagram, and the topic relationship diagram, the complex cross-influence among the stocks in the stock set S can be depicted. The equity relationship diagram, industry relationship diagram and topic relationship diagram are as follows:

G _S =(V,E _S ,A _S ) (2)

G _I =(V,E _I ,A _I ) (3)

G _T = (V, E _T , A _T ) (4)

In formulas (2), (3) and (4), G _S is the equity relationship diagram, G _I is the industry relationship diagram, and G _T is the topic relationship diagram. V={v ₁ ,...,v _N } is the vertex set of the graph, v _i is the ^ith vertex, each vertex corresponds to a stock, there are a total of N vertices (ie N stocks), the vertex of each graph The sets are all the same, which is the stock set S. E is a graph edge set, each edge represents the connection relationship between stocks, and different stock graph edge sets are different. A=(a _ij ) _N×N is the adjacency matrix of the graph, the element a _ij of the matrix is the weight of the edge, which represents the influence factor between stock i and stock j, that is, the strength of the mutual influence of prices, the adjacency of different stock graphs The matrix is also different.

Further, the equity relationship diagram is used to reflect the phenomenon that the prices of stocks affect each other due to the mutual shareholding of their companies. For example, if there is an equity relationship between two listed companies i and j, there is a directed edge between the stocks corresponding to i and j, and the weight of the edge is the shareholding ratio; conversely, if there is no equity between the two listed companies relationship, there is no edge between them. The value of the equity ratio is [0,1], and the value of the edge is also a _ij =[0,1].

The industry relationship graph is used to reflect the stock price interaction between stocks due to their industry and company size. There is a well-known lead-lag phenomenon in the stock market, that is, the impact of some leading stocks in the industry on the prices of other stocks ahead or lags behind. A theoretical explanation for the lead-lag phenomenon is that new information in the stock market is generally reflected in the transactions of leading stocks, and then transmitted to other stocks, and in the same industry, the prices of large leading stocks generally lead the smaller stocks. stock changes. Based on this theory, the embodiment of the present invention analyzes the stock price interaction between stocks due to their industry and company size according to the following rules: If two listed companies i and j are located in different industries, then their stock price interaction strength is a _ij =0, otherwise the influence strength is

Among them, M is the size of the company, which can be measured by the company's registered capital. At the same time, if the difference between M _i and M _j is higher, it means that the influence intensity a _ij of company i on company j is higher, and the influence intensity a _ji of company j on company i is weaker.

Topic graphs are used to reflect the significant impact of topics such as news on stock prices. Often, a stock will react to similar news that belongs to the same topic. At the same time, news on the same topic can affect a group of related stocks, triggering similar volatility in the stocks. For example, the 2019 coronavirus news affected pharmaceutical stocks, entertainment stocks, and restaurant stocks. In the embodiment of the present invention, the topic relationship is represented by the number of common concepts owned by stocks, so as to describe the similar fluctuation trends in prices between stocks due to the influence of the same topic. Specifically: if listed company i has M _i concepts and listed company j has M _j concepts, and the number of their common concepts is T _ij , then the impact factor of company i on company j

The influence factor of company j on company i is

If there is no number of concepts in common between two companies, the mutual influence factor between them is 0.

S30: Input the historical price features of the stock set S in the past P trading days and the equity network relationship graph of the stock set S into the deep learning model, and the deep learning model performs feature aggregation and non-discrimination on the vertices on each graph through the graph convolutional neural network. Linear transformation, extracting the interaction between the vertices on each graph, and constructing a new price feature including the interaction between stocks in each historical trading day t;

In this step, the graph convolutional neural network extracts the interaction of stocks in various relational graphs according to the Laplacian matrix of the stock network relation graph. The formula of each layer of the graph convolutional neural network is as follows:

In formula (5), the adjacency matrix corresponding to the Laplacian matrix {L _S , L _I , L _T } is {A _S , A _I , A _T },

I _N is the identity matrix, and D is the adjacency matrix of A.

is the trainable parameter, K is the order of the Laplacian matrix, which represents the radius of the graph convolution, and θ _k is the trainable coefficient of the k-th order Laplacian matrix. ρ is the activation function, such as ReLU, Sigmoid.

are the trainable parameters of the lth layer.

is the first layer of the graph convolutional neural network,

is the original price feature of the stock set S on day t, and each stock includes F original price features. In the embodiment of the present invention, the graph convolutional neural network includes three layers, and the Laplace order is K=3.

Raw price characteristics on the t-th trading day

The new price feature output after the graph convolutional neural network is

Each stock has C new price features that incorporate interactions between stocks.

S40: After splicing the original price feature and the new price feature of the stock set S, input the gated recurrent neural network to extract the time series feature of the stock set S, and output the new feature including the time series feature;

In this step, use the gated recurrent neural network GRU to analyze the original price features X _t and new price features of the past P historical trading days

Extract time series features. The hidden layer of GRU is as follows:

In Equation (6), the input time step t∈[d-P+1,…,d-1].

is the hidden layer state of the GRU at input time step t-1, r _t is the reset gate, and u _t is the update gate. σ∈[0,1] is the activation function, · is the matrix multiplication, and ⊙ is the Huffman product. {W _r ,W _u ,W _h } are trainable weight parameters, and { _{br ,b u} _, b _h } are trainable bias terms. The output layer of the GRU is

in

Each stock includes G new features.

S50: Input the new features output by the gated recurrent neural network into the fully connected layer with the Sigmoid activation function to obtain the stock price trend of the stock set S on the trading day d;

In this step, the probability of rising and falling of the stock set S is generated by using a fully connected layer FC, and the formula is as follows:

In formula (7),

is the stock rise and fall trend of the stock set S output by the model on the d day,

is the forecast of the up and down trend of stock i on the d day.

are trainable parameters. In the embodiment of the present invention, the cross entropy is used as the loss function, and the details are as follows:

In formula (8),

is the real price trend value of the stock s on the d day. The present invention calculates the trend value by calculating the ratio of the gap between the opening price of the stock on the t day and the opening price of the previous day, that is,

Based on the above, the stock price trend prediction method of the embodiment of the present application depicts the relationship between stocks through multiple perspectives (equity, industry, topic), and combines the historical price characteristics of stocks to use graph convolutional neural network to extract the relationship between stock prices. to capture the interactions between stock prices and build new price features based on them. After splicing the new price features and historical price features, the time series features and cross-influence features of the stock are extracted through the gated recurrent neural network, and then the stock price trend prediction is carried out. Compared with the prior art, the embodiment of the present invention fully considers the influence of the interaction between stock prices on the ups and downs of stocks, and extracts the time series features and cross-influence features in the price data, thereby greatly improving the accuracy of stock price trend prediction. .

Please refer to FIG. 3 , which is a schematic structural diagram of a stock price trend prediction system according to an embodiment of the present application. The stock price trend prediction system 40 in the embodiment of the present application includes:

Data acquisition module 41: used to acquire the original price characteristics of a stock set in a predetermined number of historical trading days; the stock set includes at least two stocks;

Relationship graph construction module 42: used to construct an equity network relationship graph of the stock set based on prior knowledge, and each vertex in the equity network relationship graph corresponds to a stock respectively;

Graph convolution module 43: used to input the historical price characteristics of the stock set and the equity network relationship graph into a deep learning model, and the deep learning model uses a graph convolutional neural network to characterize the vertices on the equity network relationship graph Aggregate and nonlinear transformation, extract the interaction between the vertices on the equity network relationship graph, and construct the new price feature of the stock set;

Time series feature extraction module 44: used to input the gated recurrent neural network after splicing the new price features of the stock set with the original price features, and output new features including time series features through the gated recurrent neural network;

Result output module 45: used to predict the stock price trend of the stock set according to the new feature including the time series feature.

Please refer to FIG. 4 , which is a schematic structural diagram of a terminal according to an embodiment of the present application. The terminal 50 includes a processor 51 and a memory 52 coupled to the processor 51 .

The memory 52 stores program instructions for implementing the above-mentioned stock price trend prediction method.

The processor 51 is configured to execute the program instructions stored in the memory 52 to control the stock price trend prediction.

The processor 51 may also be referred to as a CPU (Central Processing Unit, central processing unit). The processor 51 may be an integrated circuit chip with signal processing capability. The processor 51 may also be a general purpose processor, digital signal processor (DSP), application specific integrated circuit (ASIC), off-the-shelf programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component . A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

Please refer to FIG. 5 , which is a schematic structural diagram of a storage medium according to an embodiment of the present application. The storage medium of this embodiment of the present application stores a program file 61 capable of implementing all the above methods, wherein the program file 61 may be stored in the above-mentioned storage medium in the form of a software product, and includes several instructions to make a computer device (which may It is a personal computer, a server, or a network device, etc.) or a processor that executes all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes , or terminal devices such as computers, servers, mobile phones, and tablets.

The above description of the disclosed embodiments enables any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined in this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

A stock price trend forecasting method, characterized in that it includes:

Obtain the original price characteristics of the stock set on the historical trading days of the predetermined number of days; the stock set includes at least two stocks;

Construct an equity network relationship graph of the stock set based on prior knowledge, and each vertex in the equity network relationship graph corresponds to a stock respectively;

The historical price features of the stock set and the equity network relationship graph are input into the deep learning model, and the deep learning model performs feature aggregation and nonlinear transformation on the vertices on the equity network relationship graph through the graph convolutional neural network, and extracts all the vertices. The interaction between the vertices on the equity network relationship graph, and construct the new price feature of the stock set;

After splicing the new price feature of the stock set with the original price feature, input the gated recurrent neural network, and output the new feature including the time series feature through the gated recurrent neural network;

The stock price trend of the stock set is predicted according to the new feature including the time series feature.
The method for predicting a stock price trend according to claim 1, wherein the acquiring the original price characteristics of the stock set in a predetermined number of historical trading days includes:

Perform Z-score normalization processing on the original price feature;

The original price characteristics include closing price, opening price, trading volume, and trading turnover.
The stock price trend prediction method according to claim 1, wherein the equity network relationship diagram includes an equity relationship diagram, an industry relationship diagram and a topic relationship diagram, specifically:

G S =(V,E S ,A S )

G I =(V,E I ,A I )

G T =(V,E T ,A T )

In the above formula, G S is the equity relationship diagram, G I is the industry relationship diagram, and GT is the topic relationship diagram; V={v 1 ,...,v N } is the vertex set of the graph, and v i is the ith vertex , there are N vertices in total, and the vertex sets of each graph are the same, which is the stock set S; E is the graph edge set, each edge represents the connection relationship between stocks, and the edge sets of different stock graphs are different; A=(a ij ) N×N is the adjacency matrix of the graph, and the element a ij of the matrix is the weight of the edge, which represents the influence factor between stock i and stock j.
The stock price trend prediction method according to claim 3, wherein the graph convolutional neural network performs feature aggregation and nonlinear transformation on the vertices on the equity network relationship graph, specifically:

The interaction of the stock set in each relation graph is extracted according to the Laplacian matrix of the stock network relation graph. The formula of each layer of the graph convolutional neural network is:

In the above formula, the adjacency matrix corresponding to the Laplacian matrix {L S , L I , L T } is {A S , A I , A T },
I N is the identity matrix, D is the adjacency matrix of A;

is a trainable parameter, K is the order of the Laplacian matrix, indicating the radius of the graph convolution, θ k is the trainable coefficient of the k-th order Laplacian matrix; ρ is the activation function,
are the trainable parameters of the lth layer,
is the first layer of the graph convolutional neural network,
is the original price feature of the stock set S on day t, and each stock includes F original price features;

The original price feature
The new price feature output after the graph convolutional neural network is
Each stock includes C new price features.
The stock price trend prediction method according to claim 4, wherein the hidden layer passing through the gated recurrent neural network is:

In the above formula, the input time step t∈[d-P+1,...,d-1],
is the hidden layer state of the gated recurrent neural network at the input time step t-1, r t is the reset gate, u t is the update gate; σ∈[0,1] is the activation function, ⋅ is the matrix multiplication, ⊙ is Huffman product; {W r ,W u ,W h } are trainable weight parameters, { br , bu ,b h } are trainable bias terms; the output layer of the gated recurrent neural network Yes
in
Each stock includes G new features.
The stock price trend prediction method according to claim 5, wherein the predicting the stock price trend of the stock set according to the new feature including the time series feature is specifically:

The new features output by the gated recurrent neural network are input into the fully connected layer with the sigmoid activation function, and the probability of rising and falling of the stock set is generated through the fully connected layer. The formula is as follows:

In the above formula,
is the stock rise and fall trend of the stock set S on the d day,
is the forecast of the rising and falling trend of stock i on the d day;
are trainable parameters.
The stock price trend prediction method according to any one of claims 1 to 6, wherein the deep learning model uses cross-entropy as a loss function, which is as follows:

In the above formula,
is the real price trend value of the stock s on the d day, and the real price trend value is the ratio of the gap between the opening price of the stock on the t day and the opening price of the previous day.
A stock price trend prediction system, characterized in that it includes:

Data acquisition module: used to acquire the original price characteristics of a stock set on historical trading days of a predetermined number of days; the stock set includes at least two stocks;

Relationship graph building module: used to construct an equity network relationship graph of the stock set based on prior knowledge, and each vertex in the equity network relationship graph corresponds to a stock respectively;

Graph convolution module: used to input the historical price features of the stock set and the equity network relationship graph into a deep learning model, and the deep learning model performs feature aggregation on the vertices on the equity network relationship graph through a graph convolutional neural network and nonlinear transformation, extract the interaction between the vertices on the equity network relationship graph, and construct the new price feature of the stock set;

Time series feature extraction module: used to input the gated recurrent neural network after splicing the new price features of the stock set with the original price features, and output new features including time series features through the gated recurrent neural network;

Result output module: used to predict the stock price trend of the stock set according to the new feature including time series features.
A terminal, characterized in that the terminal includes a processor and a memory coupled to the processor, wherein,

The memory stores program instructions for implementing the stock price trend prediction method according to any one of claims 1-7;

The processor is configured to execute the program instructions stored in the memory to control stock price trend prediction.
A storage medium, characterized in that it stores program instructions executable by a processor, and the program instructions are used to execute the stock price trend prediction method according to any one of claims 1 to 7.