Multi-source information plate rotation investment portfolio system based on deep reinforcement learning

Abstract

The invention discloses a multi-source information plate rotation investment portfolio system based on deep reinforcement learning, and relates to the field of finance. The system comprises a stock and news data acquisition module, a data preprocessing module, a plate trend data acquisition module, a multi-agent plate trend prediction layer, a news emotion analysis layer, a hypergraph convolution module and a plate rotation layer. Compared with other existing portfolio inventions using news data, the news sentiment analysis layer has the advantages that unsupervised news datamation is realized, the manual labeling time is saved, and the multi-source data fusion efficiency is improved. According to the invention, more data sources are fused, the income capability of portfolio management is improved, and the portfolio management at a more macroscopic level is realized.

Description

Technical Field

[0001] The present invention belongs to the field of finance, and in particular relates to a multi-source information sector rotation investment portfolio system based on deep reinforcement learning. Background Art

[0002] Quantitative trading refers to replacing human subjective judgment with advanced mathematical models, and using computer technology to find multiple high-probability times that bring excess returns in a large amount of historical data, and then formulate corresponding strategies. As an important branch of quantitative trading, stock historical data analysis has a very high status in quantitative trading, and stock analysis can be broken down into many tasks, such as stock price and trend forecasting, risk analysis, and return forecasting, etc.

[0003] Portfolio management has always attracted people's attention. Although there are endless solutions, due to the variability of stock data, this problem is still a challenging task. At present, the solutions to this problem can be mainly divided into machine learning solutions and deep learning solutions. Deep learning methods have the advantages of strong adaptability, high accuracy, strong feature capture ability and less human intervention. As a new research direction in the field of machine learning, deep learning makes machine learning closer to the original goal of achieving artificial intelligence. Its concept originates from artificial neural networks, which imitate the human brain mechanism to interpret data. It combines the underlying features to form high-level representation attribute categories or features. Nowadays, with the rise of deep learning, deep learning has also been widely used in various fields, such as search technology, data mining, natural language processing, image recognition, etc. As a time series data prediction problem, there are other solutions that use deep learning as a solution to the sector rotation portfolio problem. Although they can achieve good results, there is a problem that the data source used is relatively single. These solutions often only use the historical data of stocks as the input of deep learning and lack consideration of sector-level data and news data. Summary of the invention

[0004] In view of the shortcomings of the existing technology, the present invention proposes a multi-source information sector rotation investment portfolio system based on deep reinforcement learning to solve the problem that the current investment portfolio management method lacks investment rotation at the sector level, and introduces more data sources to improve investment portfolio efficiency.

[0005] The technical solution of the present invention is:

[0006] A multi-source information sector rotation investment portfolio system based on deep reinforcement learning, including a stock and news data acquisition module, a data preprocessing module, a sector trend data acquisition module, a multi-agent sector trend prediction module, a news sentiment analysis module, a hypergraph convolution module and a sector rotation module;

[0007] The stock and news data acquisition module is used to acquire the historical data of the selected stock as the input of the data preprocessing module; acquire the industry news data as the input of the news sentiment analysis module;

[0008] The data preprocessing module is used to preprocess the historical data of the selected stock, and obtain the preprocessed historical data of the selected stock as the input of the sector trend data acquisition module and the hypergraph convolution module;

[0009] The sector trend data acquisition module is used to divide the pre-processed historical data of the selected stocks according to sectors, and perform mean statistics according to sectors to obtain sector trend data as input of the multi-agent sector trend prediction module;

[0010] The multi-agent sector trend prediction module is a feedforward neural network, which is used to predict the future trend of each sector based on the sector trend data of each sector input, and output a sector information vector as the input of the sector rotation module; the sector information vector is the predicted increase or decrease of each sector, and the length of the sector information vector is equal to the number of sectors;

[0011] The news sentiment analysis module is used to analyze the impact of each industry news data on each sector, and output a news information vector as the input of the sector rotation module; the news information vector includes the impact of news on each sector on different trading days;

[0012] The hypergraph convolution module is used to construct a hypergraph according to the preprocessed historical data of the selected stocks, and at the same time use the GRU network to extract features from the preprocessed historical data of the selected stocks to obtain feature-extracted data, and then input the feature-extracted data into the hypergraph convolution layer. The hypergraph convolution layer performs dimension transformation on the feature-extracted data according to the constructed hypergraph to obtain a stock information vector as the input of the sector rotation layer; the length of the stock information vector is equal to the number of sectors;

[0013] The sector rotation module is used to integrate the sector information vector, stock information vector and news information vector, use the GRU network to extract features from the integrated vector, and then input the extracted features into the fully connected layer and pass through the Softmax function to generate an investment portfolio result; the investment portfolio result includes the investment proportion of each sector;

[0014] Furthermore, the historical data of the stock is a time series, including the closing price, opening price, trading volume, highest price and lowest price of the stock at different times; the industry news data is news in different industry fields;

[0015] Further, the preprocessing is specifically as follows: firstly, filling the missing values ​​in the historical data of the selected stock, then dividing the historical data of the selected stock according to the year, then optimizing the historical data of the selected stock by peak elimination and valley filling, and finally normalizing the historical data of the selected stock after peak elimination and valley filling to obtain the preprocessed historical data of the selected stock;

[0016] Furthermore, the calculation formula of the plate trend data is as follows:

[0017]

[0018] Among them, e m is the plate trend data of plate m, x n is the historical data of stocks belonging to sector m, n is the number of stocks belonging to sector m, and N is the number of stocks in sector m;

[0019] Furthermore, the multi-agent plate trend prediction module is trained based on a Markov decision process of multiple agents, and the Markov decision process is defined as a four-tuple in is the state space, is the action space, is the state transition probability function, indicating that in state s t Through action a t Transition to state s t+1 The probability of a is the action in the action space, a t is the action on trading day t, s represents the state in the state space, and s t represents the status of trading day t, s t+1 represents the state of trading day t+1, r is the state s t Through action a t Transition to state s t+1 After the reward;

[0020] The states in the state space are defined as:

[0021]

[0022] in, represents the status of sector m on trading day t, S m is the state space of plate m;

[0023] The action space is defined as:

[0024]

[0025] in, is the action space of trading day t, which is a continuous set of real numbers, representing the prediction of the agent’s increase or decrease in the stocks belonging to the sector on trading day t;

[0026] Rewards are defined as:

[0027]

[0028] in, is the reward of sector m on trading day t, is the plate information vector P output by the multi-agent plate trend prediction module a The elements in , that is, the prediction of the rise and fall of stocks belonging to sector m by the agent on trading day t, is the actual rise or fall of sector m on trading day t;

[0029] Furthermore, the processing process of the news sentiment analysis module is specifically as follows: designing prompt words, using a generative large language model to analyze the impact of industry news data on each sector according to the prompt words, generating a sector impact label for each piece of industry news data, and counting the impact of news on each sector on each trading day according to timestamps, thereby obtaining a news information vector;

[0030] The sector impact label indicates the impact of the news on the sector, including 1, 0 and -1, representing positive, irrelevant and negative respectively;

[0031] The impact of news on each sector on each trading day is as follows:

[0032]

[0033] In the formula, is the situation in which sector m is affected by news on trading day t, that is, the elements in the news information vector output by the news sentiment analysis module after receiving industry news data. T is a time point less than t, indicating that each input is industry news data from trading day T to trading day t;

[0034] Furthermore, the hypergraph includes vertices v and hyperedges e, the vertices represent historical data of stocks, and the hyperedges represent sectors;

[0035] The hypergraph is represented as V and ε are the number of stocks and the number of sectors, respectively;

[0036] In a hypergraph, the relationship between vertices and hyperedges is:

[0037]

[0038] in, Represents the relationship between vertices and hyperedges. When vertex v iWhen the corresponding i-th stock belongs to the sector m corresponding to the hyperedge e, its value is 1, otherwise it is 0;

[0039] The degree of the hypergraph vertices: The degree matrix of the vertex

[0040] The degree of a hypergraph hyperedge is: Hyperedge degree matrix

[0041] The Laplacian of a hypergraph is defined as:

[0042]

[0043] Where W is the weight of the hyperedge, I is the unit matrix, and W=I is set here;

[0044] For the data after feature extraction Based on the Laplacian operator of the hypergraph H, the feature-extracted data X is processed through the hypergraph convolution layer (l) Perform dimension transformation to obtain the stock information vector. The specific formula is as follows:

[0045]

[0046] Where l is a portfolio round, f is the number of features in the data after feature extraction, HConv is the hypergraph convolution, is a learnable matrix, and ELU is the activation function;

[0047] Furthermore, the method for integrating the sector information vector, the stock information vector and the news information vector is as follows:

[0048]

[0049] Where M is the number of plates;

[0050] Furthermore, the plate rotation module is trained based on a Markov decision process of multiple agents, defined as a four-tuple in is the state space, is the action space, is the state transition probability function, indicating that in state s' t Through action a' t Transition to state s' t+1 The probability of a' is the action in the action space, a' t is the action on trading day t, s' represents the state in the state space, and s' t represents the status of trading day t, s' t+1represents the state of trading day t+1, r' is the state s' t Through action a' t Transition to state s' t+1 After the reward;

[0051] The state in the state space is defined as s' t ;

[0052] Actions are defined as:

[0053] a'=W t ∈R 1×M

[0054] Where W t The final output is the investment portfolio result;

[0055]

[0056] In the formula, FC(·) is the fully connected layer, GRU is the GRU network;

[0057] Rewards are defined as:

[0058]

[0059] In the formula, and are the closing price and opening price of trading day t respectively.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] 1. Compared with other existing investment portfolio inventions that use news data, the news sentiment analysis module in the present invention realizes unsupervised news digitization, saves time for manual labeling and improves the efficiency of multi-source data fusion.

[0062] 2. The present invention integrates more data sources and improves the profitability of investment portfolio management.

[0063] 3. The present invention generally achieves investment portfolio management at a more macro level. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flowchart of a multi-source information sector rotation investment portfolio system based on deep reinforcement learning in an embodiment of the present invention;

[0065] Figure 2 It is a schematic diagram of the process of the hypergraph convolution module and the plate rotation module in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0067] The present invention can be summarized as follows: first, stock data and news data are obtained through web crawlers and APIs provided by data institutions, and sector data is obtained after data preprocessing. Then, the sector data is used as the input of the multi-agent sector trend prediction module, the news data is used as the input of the news sentiment analysis module, and the stock data is used as the input of the hypergraph convolution layer through recurrent neural network feature extraction to obtain three information vectors. Finally, the sector rotation module obtains the final sector-level investment portfolio results, and the reinforcement learning method is used to train the model according to the results.

[0068] like Figure 1 As shown, a multi-source information sector rotation investment portfolio system based on deep reinforcement learning includes a stock and news data acquisition module, a data preprocessing module, a sector trend data acquisition module, a multi-agent sector trend prediction module, a news sentiment analysis module, a hypergraph convolution module and a sector rotation module;

[0069] The stock and news data acquisition module is used to acquire the historical data of the selected stock as the input of the data preprocessing module; acquire the industry news data as the input of the news sentiment analysis module; the historical data of the stock is a time series, including the closing price, opening price, trading volume, highest price and lowest price of the stock at different times; the industry news data is news in different industry fields;

[0070] The data preprocessing module is used to preprocess the historical data of the selected stock, and obtain the preprocessed historical data of the selected stock as the input of the sector trend data acquisition module and the hypergraph convolution module;

[0071] The preprocessing is specifically as follows: firstly, filling the missing values ​​in the historical data of the selected stock, then dividing the historical data of the selected stock according to the year, then optimizing the historical data of the selected stock by peak elimination and valley filling, and finally normalizing the historical data of the selected stock after peak elimination and valley filling to obtain the historical data of the selected stock after preprocessing;

[0072] The normalized formula is:

[0073]

[0074] In the formula, x i is the normalized historical data of the ith stock, is the historical data of the i-th stock after peak-filling optimization, i is the stock number, ∈ is the minimum value set to prevent the denominator from being zero;

[0075] The sector trend data acquisition module is used to divide the pre-processed historical data of the selected stocks according to sectors, and perform mean statistics according to sectors to obtain sector trend data as input of the multi-agent sector trend prediction module;

[0076] For the sector trend data of sector m, the average value of the historical data of the selected stocks belonging to the sector is calculated. The calculation formula is as follows:

[0077]

[0078] Among them, e m is the plate trend data of plate m, x n is the historical data of stocks belonging to sector m, n is the number of stocks belonging to sector m, and N is the number of stocks in sector m;

[0079] In this implementation method, all stocks are divided into 19 sectors according to the industry sectors of Eastmoney.com, including: agriculture, animal husbandry and fishery, mining, manufacturing, water, electricity, gas, construction, wholesale and retail, transportation and warehousing, accommodation and catering, information technology, finance, real estate, business services, scientific research services, public environmental protection, resident services, education, health, cultural communication and comprehensive.

[0080] The multi-agent plate trend prediction module is a feedforward neural network, which is used to predict the future trend of each plate according to the input plate trend data of each plate, and output the plate information vector P a As the input of the sector rotation module; the sector information vector is the predicted rise and fall of each sector, and the length of the sector information vector is equal to the number of sectors;

[0081] The multi-agent plate trend prediction module is trained based on a Markov decision process of multiple agents. The Markov decision process is referred to as the MDP process and is defined as a four-tuple in is the state space, is the action space, is the state transition probability function, indicating that in state s t Through action a t Transition to state s t+1 The probability of a is the action in the action space, a t is the action on trading day t, s represents the state in the state space, and s t represents the status of trading day t, s t+1 represents the state of trading day t+1, r is the state s t Through action a t Transition to state s t+1 After the reward;

[0082] The states in the state space are defined as:

[0083]

[0084] in, represents the status of sector m on trading day t, S m is the state space of plate m;

[0085] In this embodiment, the input received by the multi-agent plate trend prediction module is plate trend data Set the time window of the sector trend data to the previous 30 days, e is the number of features (closing price, opening price, trading volume, highest price and lowest price) in the sector trend data;

[0086] The action space is defined as:

[0087]

[0088] in, is the action space of trading day t, which is a continuous set of real numbers, representing the prediction of the agent’s increase or decrease in the stocks belonging to the sector on trading day t;

[0089] Rewards are defined as:

[0090]

[0091] in, is the reward of sector m on trading day t, is the plate information vector P output by the multi-agent plate trend prediction module a The elements in , that is, the prediction of the rise and fall of stocks belonging to sector m by the agent on trading day t, is the actual rise or fall of sector m on trading day t;

[0092] Through this MDP process, the multi-agent plate trend prediction module receives the plate trend data e m , output plate information vector P a , and adjust the parameters of the multi-agent sector trend prediction module based on the rewards;

[0093] The news sentiment analysis module is used to analyze the impact of each industry news data on each sector and output the news information vector p g As the input of the sector rotation module; the news information vector includes the impact of news on each sector on different trading days;

[0094] Specifically, we design prompt words, use OpenAI API to analyze the impact of industry news data on each sector based on the prompt words, generate sector impact labels for each piece of industry news data, and count the impact of news on each sector on each trading day according to timestamps, thereby obtaining news information vectors.

[0095] The sector impact label indicates the impact of the news on the sector, including 1, 0 and -1, representing positive, irrelevant and negative respectively;

[0096]

[0097] In the formula, g j represents the jth piece of industry news data. ChatGPT divides the attitude towards this piece of industry news data into three categories: positive, irrelevant, and negative, and the corresponding labels are 1, 0, and -1 respectively;

[0098] The impact of news on each sector on each trading day is as follows:

[0099]

[0100] In the formula, is the situation in which sector m is affected by news on trading day t, that is, the elements in the news information vector output by the news sentiment analysis module after receiving industry news data. T is a time point less than t, indicating that each input is industry news data from trading day T to trading day t.

[0101] In this embodiment, the following prompt words are designed:

[0102] 'Forget all your previous instructions. Pretend you are a financial expert. You are a financial expert with stock recommendation experience. Answer only "1 to which sector" if good news, only "-1 to which sector" if bad news, or only "0" if uncertainty ro no sectors belongs to. The alternative sectors are as shown in the list: For example, the headline is "Shanghai promotes the resumption of business in stages", the answer is "1 financial industry 1 accommodation and catering". And the headline is "my country's summer grain is growing well and there is a basis for a good harvest", the answer is "1 agriculture, animal husbandry and fishery". Is this headline good or bad for the stock price of which sector? '

[0103] In this implementation, the OpenAI API version 'gpt-4o-turbo' is used;

[0104] The hypergraph convolution module is used to construct a hypergraph according to the preprocessed historical data of the selected stocks, and at the same time use the GRU network to extract features from the preprocessed historical data of the selected stocks to obtain feature-extracted data, and then input the feature-extracted data into the hypergraph convolution layer. The hypergraph convolution layer performs dimension transformation on the feature-extracted data according to the constructed hypergraph to obtain a stock information vector as the input of the sector rotation module; the length of the stock information vector is equal to the number of sectors;

[0105] The hypergraph It includes vertices v and hyperedges e. The vertices represent the historical data of stocks, the hyperedges represent the sectors, and V and ε are the number of stocks and the number of sectors respectively.

[0106] The hypergraph is constructed as follows:

[0107]

[0108] in, Represents the relationship between vertices and hyperedges. When vertex v iWhen the corresponding i-th stock belongs to the sector m corresponding to the hyperedge e, its value is 1, otherwise it is 0;

[0109] The degree of the hypergraph vertices: The degree matrix of the vertex

[0110] The degree of a hypergraph hyperedge is: Hyperedge degree matrix

[0111] The Laplacian of a hypergraph is defined as:

[0112]

[0113] Where W is the weight of the hyperedge, I is the unit matrix, and W=I is set here;

[0114] For the data after feature extraction Based on the Laplacian operator of the hypergraph H, the feature-extracted data X is processed through the hypergraph convolution layer (l) Perform dimension transformation to obtain the stock information vector. The specific formula is as follows:

[0115]

[0116] Where l is a portfolio round, f is the number of features in the data after feature extraction, HConv is the hypergraph convolution, is a learnable matrix, and ELU is the activation function;

[0117] like Figure 2 As shown, the sector rotation module is used to integrate the sector information vector, the stock information vector and the news information vector, and use the GRU network to extract features from the integrated vector. The extracted features are then input into the fully connected layer and passed through the Softmax function to generate an investment portfolio result. The investment portfolio result includes the investment proportion of each sector.

[0118] The method for integrating the sector information vector, the stock information vector and the news information vector is as follows:

[0119]

[0120] Where M is the number of plates;

[0121] The plate rotation module is trained based on a Markov decision process of multiple agents, defined as a four-tuple in is the state space, is the action space, is the state transition probability function, indicating that in state s't Through action a' t Transition to state s' t+1 The probability of a' is the action in the action space, a' t is the action on trading day t, s' represents the state in the state space, and s' t represents the status of trading day t, s' t+1 represents the state of trading day t+1, r' is the state s' t Through action a' t Transition to state s' t+1 After the reward;

[0122] The state in the state space is defined as s' t ;

[0123] Actions are defined as:

[0124] a'=W t ∈R 1×M

[0125] Where W t The final output is the investment portfolio result;

[0126] Rewards are defined as:

[0127]

[0128] In the formula, and are the closing price and opening price of trading day t respectively;

[0129]

[0130] In the formula, FC(·) is the fully connected layer;

[0131] In this implementation, the optimizer is set to Adam, the learning rate is 0.0001, and the model is trained according to the rewards and learning rates of the two MDP processes, and then it can be put into use.

Claims

1. A multi-source information sector rotation investment portfolio system based on deep reinforcement learning, characterized in that: It includes stock and news data acquisition module, data preprocessing module, sector trend data acquisition module, multi-agent sector trend prediction module, news sentiment analysis module, hypergraph convolution module and sector rotation module; The stock and news data acquisition module is used to acquire the historical data of the selected stock as the input of the data preprocessing module; acquire the industry news data as the input of the news sentiment analysis module; The data preprocessing module is used to preprocess the historical data of the selected stock, and obtain the preprocessed historical data of the selected stock as the input of the sector trend data acquisition module and the hypergraph convolution module; The sector trend data acquisition module is used to divide the pre-processed historical data of the selected stocks according to sectors, and perform mean statistics according to sectors to obtain sector trend data as input of the multi-agent sector trend prediction module; The multi-agent sector trend prediction module is a feedforward neural network, which is used to predict the future trend of each sector based on the sector trend data of each sector input, and output a sector information vector as the input of the sector rotation module; the sector information vector is the predicted increase or decrease of each sector, and the length of the sector information vector is equal to the number of sectors; The news sentiment analysis module is used to analyze the impact of each industry news data on each sector, and output a news information vector as the input of the sector rotation module; the news information vector includes the impact of news on each sector on different trading days; The hypergraph convolution module is used to construct a hypergraph according to the preprocessed historical data of the selected stocks, and at the same time use the GRU network to extract features from the preprocessed historical data of the selected stocks to obtain feature-extracted data, and then input the feature-extracted data into the hypergraph convolution layer. The hypergraph convolution layer performs dimension transformation on the feature-extracted data according to the constructed hypergraph to obtain a stock information vector as the input of the sector rotation module; the length of the stock information vector is equal to the number of sectors; The sector rotation module is used to integrate the sector information vector, stock information vector and news information vector, use the GRU network to extract features from the integrated vector, and then input the extracted features into the fully connected layer and pass through the Softmax function to generate an investment portfolio result; the investment portfolio result includes the investment proportion of each sector.

2. According to the multi-source information sector rotation investment portfolio system based on deep reinforcement learning according to claim 1, it is characterized in that: The historical data of the stock is a time series, including the closing price, opening price, trading volume, highest price and lowest price of the stock at different times; the industry news data is news in different industry fields.

3. According to the multi-source information sector rotation investment portfolio system based on deep reinforcement learning in claim 1, it is characterized in that: The preprocessing is specifically as follows: first, the missing values ​​in the historical data of the selected stock are filled, then the historical data of the selected stock is divided according to the year, then the historical data of the selected stock is optimized by eliminating peaks and filling valleys, and finally the historical data of the selected stock after eliminating peaks and filling valleys is normalized to obtain the historical data of the selected stock after preprocessing.

4. According to the multi-source information sector rotation investment portfolio system based on deep reinforcement learning according to claim 1, it is characterized in that: The calculation formula of the plate trend data is as follows: Among them, e m is the plate trend data of plate m, x n is the historical data of stocks belonging to sector m, n is the number of stocks belonging to sector m, and N is the number of stocks in sector m.

5. According to the multi-source information sector rotation investment portfolio system based on deep reinforcement learning in claim 1, it is characterized in that: The multi-agent plate trend prediction module is trained based on a Markov decision process of multiple agents. The Markov decision process is defined as a four-tuple in is the state space, is the action space, is the state transition probability function, indicating that in state s t Through action a t Transition to state s t+1 The probability of a is the action in the action space, a t is the action on trading day t, s represents the state in the state space, and s t represents the status of trading day t, s t+1 represents the state of trading day t+1, r is the state s t Through action a t Transition to state s t+1 After the reward; The states in the state space are defined as: in, represents the status of sector m on trading day t, S m is the state space of plate m; The action space is defined as: in, is the action space of trading day t, which is a continuous set of real numbers, representing the prediction of the agent’s increase or decrease in the stocks belonging to the sector on trading day t; Rewards are defined as: in, is the reward of sector m on trading day t, is the plate information vector P output by the multi-agent plate trend prediction module a The elements in , that is, the prediction of the rise and fall of stocks belonging to sector m by the agent on trading day t, It is the actual rise or fall of sector m on trading day t.

6. According to the multi-source information sector rotation investment portfolio system based on deep reinforcement learning in claim 1, it is characterized in that: The processing process of the news sentiment analysis module is specifically as follows: design prompt words, use the generative large language model to analyze the impact of industry news data on each sector based on the prompt words, generate a sector impact label for each piece of industry news data, and count the impact of news on each sector on each trading day according to the timestamp, thereby obtaining a news information vector.

7. A multi-source information sector rotation investment portfolio system based on deep reinforcement learning according to claim 6, characterized in that: The sector impact label indicates the impact of the news on the sector, including 1, 0 and -1, representing positive, irrelevant and negative respectively; The impact of news on each sector on each trading day is as follows: In the formula, is the situation in which sector m is affected by news on trading day t, that is, the elements in the news information vector output by the news sentiment analysis module after receiving industry news data. T is a time point less than t, indicating that each input is industry news data from trading day T to trading day t.

8. According to the multi-source information sector rotation investment portfolio system based on deep reinforcement learning in claim 1, it is characterized in that: The hypergraph includes vertices v and hyperedges e, where vertices represent historical data of stocks and hyperedges represent sectors. The hypergraph is represented as V and ε are the number of stocks and the number of sectors, respectively; In a hypergraph, the relationship between vertices and hyperedges is: in, Represents the relationship between vertices and hyperedges. When vertex v i When the corresponding i-th stock belongs to the sector m corresponding to the hyperedge e, its value is 1, otherwise it is 0; The degree of the hypergraph vertices: The degree matrix of the vertex The degree of a hypergraph hyperedge is: Hyperedge degree matrix The Laplacian of a hypergraph is defined as: Where W is the weight of the hyperedge, I is the unit matrix, and W=I is set here; For the data after feature extraction Based on the Laplacian operator of the hypergraph H, the feature-extracted data X is processed through the hypergraph convolution layer (l) Perform dimension transformation to obtain the stock information vector. The specific formula is as follows: Where l is a portfolio round, f is the number of features in the data after feature extraction, HConv is the hypergraph convolution, is a learnable matrix and ELU is the activation function.

9. The multi-source information sector rotation investment portfolio system based on deep reinforcement learning according to claim 1 is characterized in that: The method for integrating the sector information vector, the stock information vector and the news information vector is as follows: Where M is the number of plates.

10. The multi-source information sector rotation investment portfolio system based on deep reinforcement learning according to claim 1, characterized in that: The plate rotation module is trained based on a Markov decision process of multiple agents, defined as a four-tuple in is the state space, is the action space, is the state transition probability function, indicating that in state s' t Through action a' t Transition to state s' t+1 The probability of a' is the action in the action space, a' t is the action on trading day t, s' represents the state in the state space, and s' t represents the status of trading day t, s' t+1 represents the state of trading day t+1, r' is the state s' t Through action a' t Transition to state s' t+1 After the reward; The state in the state space is defined as s' t ; Actions are defined as: a'=W t ∈R 1×M Where W t The final output is the investment portfolio result; In the formula, FC(·) is the fully connected layer, GRU is the GRU network; Rewards are defined as: In the formula, and are the closing price and opening price of trading day t respectively.

Images