Information extraction evaluation method and computer readable storage medium

Abstract

The invention discloses an information extraction and evaluation method, and aims to improve the precision and personalized service effect of financial information screening. Different from a traditional TF-IDF or simple machine learning model abstract generation method, the method combines a large language model and a personalized preference matching model to perform multi-dimensional evaluation on timeliness, propagation popularity and relevance of the information, so that efficient screening and sorting of the information are realized. According to the method, the risk preference and the attention demand of the user can be dynamically matched, and the user preference model is updated in real time along with the change of external information, so that the timeliness and the correlation of recommended contents are ensured. Meanwhile, based on dynamic adjustment of the user risk preference model, the system can adaptively optimize the recommended content, and it is ensured that the user continuously obtains high-value information. According to the method, influence classification is carried out in combination with a large language model, efficient and intelligent financial information services are provided for users, and the defects of timeliness, individuation and multi-dimensional information integration in the prior art are overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to an information extraction and evaluation method and a computer-readable storage medium. Background Art

[0002] In today's information society, especially in the field of financial services, users' demand for fast response and high-value information screening has become particularly urgent. However, traditional information screening and processing methods have many limitations when dealing with large-scale and diverse financial data, which are mainly reflected in the following aspects:

[0003] 1. Timeliness is difficult to guarantee

[0004] Traditional information screening methods usually rely on preset keywords or rule matching, and the processing flow is relatively simple, which makes it difficult to adapt to the rapid changes in information. For example, in the financial market, the products and market dynamics that users are concerned about may occur in a short period of time, and traditional screening methods, due to their manually set rules, are difficult to respond in time and filter out information that is highly relevant to users. This results in users being unable to obtain effective information in a timely manner, and may miss market opportunities.

[0005] 2. Personalized needs are difficult to meet

[0006] Existing methods lack in-depth understanding of users’ personalized needs in information screening, and usually rely on static keyword matching and generalized information push, ignoring the differences among users in terms of risk preferences, investment records, industry concerns, etc. Since dynamic preference matching of individual users is not achieved, the accuracy of information push is insufficient and cannot fully meet users’ personalized needs.

[0007] 3. Limited data processing capabilities

[0008] Existing machine learning models or text processing models based on statistical methods such as TF-IDF are difficult to guarantee high efficiency when facing massive data in the financial field. These methods usually lack multi-dimensional correlation analysis of data, such as the similarity, timeliness and popularity of related financial products, resulting in a lot of redundant information and insufficient identification of key information when screening information, and are unable to provide users with accurate and high-value information.

[0009] In the financial services sector, investors need to pay close attention to market dynamics, corporate information, and changes in their investment portfolios. However, in the context of information overload, it is often difficult for users to quickly extract information that has a significant impact on their investment decisions from the complex information. For example:

[0010] 1. Rapid market changes: In the stock or fund market, breaking news or corporate announcements often have a significant impact on the market. Users need to quickly determine the relevance of this information to the financial products they hold so that they can adjust their investment strategies in a timely manner. Traditional keyword matching is difficult to quickly respond to these high-frequency, short-term information updates.

[0011] 2. Diversity of user preferences: Different users have different investment styles and risk preferences. Some users prefer conservative investments, while others may prefer high-risk, high-return products. The existing push methods cannot effectively distinguish the individual needs of users and cannot accurately push information that meets their preferences to users.

[0012] 3. Multi-channel information processing: Financial information comes from a variety of channels, such as news platforms, social media, market announcements, etc. The information formats of these channels are diverse and updated frequently. The existing information screening methods are difficult to effectively integrate and analyze multi-source heterogeneous information, resulting in the failure of the pushed information to effectively meet the actual needs of users. Summary of the invention

[0013] The purpose of the present invention is to provide an information extraction and evaluation method and a computer-readable storage medium in view of the deficiencies in the prior art. The "information extraction and evaluation method" provided by the present invention effectively solves the deficiencies of the prior art in terms of timeliness, personalization, and data processing capabilities by combining a large language model and a personalized preference matching model. The method can not only accurately extract high-value information from large-scale, multi-source heterogeneous information, but also dynamically update the screening rules based on the user's risk preference, providing users with real-time, accurate, personalized financial information services.

[0014] To achieve the above object, in a first aspect, the present invention provides an information extraction and evaluation method, comprising:

[0015] Obtain external information;

[0016] Preprocessing the external information to extract the reference source, timestamp, and number of references, and outputting the structured external information to a database;

[0017] Collect and update the user's personal information, including risk preferences, current product holdings, historical holdings, and information about companies and industries of interest;

[0018] Based on the user's purchase record and product attributes, expert experience defines the user's risk preference rules and establishes a sample set. Experts give label definition rules for the various attributes of financial products on the market, establish risk preference labels, and use the Transformer model to build a label classification model, generate an association mapping with user preferences, and perform label prediction for all public financial products on the market.

[0019] Extract the financial product information associated with the user from the external information, calculate the correlation between the associated financial product information and the products held, concerned and risk-related by the user, and output the top N associated financial product information after sorting;

[0020] Constructing thought chain prompt words, inputting the current status of each associated financial product information output and corresponding external information into the large language model, so that the large language model outputs influence classification;

[0021] Map the impact classification to the weight, recalculate the value score of each piece of information, and output the top M related financial product information after re-ranking;

[0022] Combine the top M related financial product information with the thought chain prompt words, and call the large language model to output the analysis report.

[0023] Furthermore, it also includes:

[0024] Use tracking technology to evaluate user feedback on generated reports, combine users' actual behavior data, and regularly and dynamically update user preferences and corresponding content of interest.

[0025] Furthermore, the calculation method of the correlation is as follows:

[0026] Through keyword query, obtain all information related to the user's current products in external information and filter it according to time requirements;

[0027] The exponential decay function is used to quantitatively evaluate the timeliness of external information, as follows:

[0028]

[0029] Among them, P t is the quantitative evaluation result of the timeliness of external information; e is a natural constant, ω is the timeliness attenuation coefficient, t current The timestamp when the information is evaluated; t pubish The release timestamp of external information;

[0030] Use expert rules to define the weight of each citation source as follows:

[0031] P s =SourceScore(s)

[0032] Among them, P s is the weight of the citation source s, SourceScore is the score given by the expert to the citation source;

[0033] Calculate the relevance of external information to the enterprise and industry information that users are interested in based on topics or keywords, as follows:

[0034]

[0035] Among them, P c is the correlation between the calculated external information and the enterprise and industry information that the user is concerned about, E i is the embedding vector of external information, E u Embedding vectors for enterprise and industry information that users are interested in;

[0036] The propagation heat of the external information is calculated as follows:

[0037] P h =w r *Norm(Count r )+w f *Norm(Count f )+w c

[0038] *Norm(Count c )

[0039] Among them, P h is the calculated propagation heat of the external information, w r is the reading weight, Count r is the number of times external information is read, w f is the forwarding weight, Count f is the number of times external information is forwarded, w c is the counting weight of the information in the system, Count c is the number of counts of external information, Norm(.) is the regularization function;

[0040] The value of the external information is calculated as:

[0041] P I =w1*P t +w2*P s +w3*P c +w4*P h

[0042] Among them, P I is the calculated value of external information, w i is the weight of each dimension, i=1, 2, 3, 4.

[0043] Furthermore, the impact classification includes positive, negative and neutral.

[0044] Furthermore, the weight of the positive mapping is 1, the weight of the negative mapping is 1.5, and the weight of the neutral mapping is 0.5.

[0045] Furthermore, the risk preference labels include conservative, stable, balanced, active and aggressive.

[0046] Furthermore, the sources of the external information include news and market data.

[0047] In a second aspect, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0048] Beneficial effects: The present invention comprehensively utilizes the large language model prompt word engineering, Transformer classification model and relevance evaluation to achieve the sorting and identification of high-value information, and finally screens out a list of information that meets the user's risk preferences; and utilizes the summarizing ability of the large language model to output a risk report that is closely related to the user, providing users with services with high timeliness and accurate information. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flowchart of the information extraction and evaluation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. The embodiments are implemented based on the technical solution of the present invention. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0051] like Figure 1 As shown, an embodiment of the present invention provides an information extraction and evaluation method, comprising:

[0052] Obtain external information. The above external information sources include news, market data, etc., through which a more comprehensive data input system is built.

[0053] Preprocess external information. During the preprocessing process, extract key information such as the citation source, timestamp, and number of citations, and store the structured external information in the database. Preprocessing operations include data cleaning and deduplication, such as extracting key fields such as company name, stock name, bond name, fund name, and removing duplicate data. The specific method of data deduplication is to cluster information by content, select the earliest release time, summarize the content popularity, and evaluate the influence of data sources.

[0054] Collect and update users' personal information, including:

[0055] 1) Risk preference assessment: Understand the user's risk tolerance and investment preferences.

[0056] 2) Current product holdings: record the types and quantities of products currently held by the user.

[0057] 3) Historical holding records: Track users’ past investment behaviors and position changes.

[0058] 4) Companies and industries of interest: Understand the companies and industry fields that users are interested in.

[0059] Based on the user's purchase record and product attributes, the user's risk preference rules are defined by expert experience and a sample set is established. Experts give label definition rules for the various attributes of financial products on the market and establish risk preference labels. The Transformer model is used to establish a label classification model, generate an association map with user preferences, and perform label prediction for all public financial products on the market.

[0060] The definitions of the above risk appetite labels are as follows:

[0061] 1) Conservative: This type of user prefers investment products with low risks and stable returns, and is usually unwilling to take greater investment risks.

[0062] 2) Conservative type: This type of user is willing to take a certain degree of risk in the hope of obtaining higher returns than conservative investments.

[0063] 3) Balanced: Users seek a balance between risk and return and are willing to take moderate risks in order to obtain relatively high returns.

[0064] 4) Aggressive: This type of user is willing to take higher risks in the hope of obtaining higher returns on investment.

[0065] 5) Aggressive type: Users pursue high returns and are willing to take extremely high risks. Usually, this type of users has a deep understanding of the market and high investment skills.

[0066] The financial products with corresponding labels are as follows:

[0067] Conservative: Low-risk, stable return products.

[0068] Stable type: products with lower risks and stable returns.

[0069] Balanced type: products with balanced risks and returns.

[0070] Aggressive: Products with higher risks and expected higher returns.

[0071] Aggressive: high-risk, high-return products. The types of financial products that match the label include: low-volatility, large-cap stocks and bonds (conservative), medium-volatility and high-return stocks and funds (sound), etc., to meet the investment needs of different users.

[0072] Based on the Bert model, a classification mapping layer is constructed to implement a multi-classification algorithm for information to match the user's risk preference with financial products. The following steps and processes are required, including data preparation, model training, feature extraction, classification mapping layer construction and prediction.

[0073] Step 1: Data Preparation

[0074] Collect user information and relevant data of financial products, including user’s risk preference labels and characteristic data of financial products.

[0075] Based on the risk characteristics of financial products (such as volatility, yield, market value, P / E, P / B, etc.), label each financial product and determine its risk level (conservative, stable, balanced, active, aggressive).

[0076] Step 2: Data Preprocessing

[0077] Feature extraction is performed on the description of financial products. The BERT model can be used to convert the text description into a vector representation.

[0078] Numerical features were normalized to ensure that data were compared on the same scale.

[0079] Step 3: BERT model mapping classification algorithm

[0080] Text Embedding of BERT Model - Use the pre-trained BERT model to convert the description of financial products into embedding vectors. The BERT model converts the input text data into a fixed-length vector representation.

[0081] Build a classification mapping layer - Based on the BERT model, build a classification mapping layer that maps the BERT embedding vector to different risk categories. This is achieved using a fully connected layer and a softmax activation function.

[0082] Model Training - The model is trained using the cross entropy loss function and the Adam optimizer.

[0083] Step 4: Model Evaluation and Prediction

[0084] Model evaluation - Use the test data set to evaluate the performance of the model. Usually before training, the sample data is divided into three categories: training set (Train), validation set (Validate), and test set (Test). Among them:

[0085] Validate - The validation set is used for model tuning and selection, especially during model training, to evaluate the performance of the model on unseen data. It is mainly used to adjust the model's hyperparameters, such as learning rate, regularization parameters, etc. The validation set helps select the optimal model configuration and prevent overfitting. By evaluating the performance of the model on the validation set under different hyperparameter settings, the best performing model is selected.

[0086] Test set (Test) - The test set is used for the final evaluation of the model. After training and validation are completed, the test set is used to evaluate the performance of the model on completely unseen data. The test set provides an estimate of the model's performance in the real world and is used to report the final model performance indicators (such as accuracy, F1 score, mean squared error, etc.).

[0087] Through the above model, we can achieve a five-level classification of information related to stocks, funds, and bonds in the market, and link it with the risk preferences of users, so as to achieve the grouping of risk preference "stratospheres" and map them to different risk preference groups. In addition, as time goes by, the labels of each stock, fund, and bond will change with the changes in prices, so the classification model also needs to be updated regularly.

[0088] Extract financial product information associated with the user from external information, calculate the correlation between the associated financial product information and the products that the user holds, pays attention to, and has risk preferences, and outputs the top N associated financial product information after sorting. Specifically include:

[0089] Timeliness evaluation: quantify the timeliness of external information through an exponential decay function;

[0090] Citation source weight: score and weight citation sources according to expert rules;

[0091] Embedding similarity: Use vector embedding technology to calculate the relevance of external information and the content that users are interested in;

[0092] Communication heat: The communication heat of information is calculated by comprehensively considering the number of readings, forwardings and displays.

[0093] The details are as follows:

[0094] Through keyword query, obtain all information related to the user's current products in external information and filter it according to time requirements;

[0095] The exponential decay function is used to quantitatively evaluate the timeliness of external information, as follows:

[0096]

[0097] Among them, Pt It is the quantitative evaluation result of the timeliness of external information, that is, a score used to indicate the freshness of information during the information screening process. e is a natural constant, approximately equal to 2.718, which is used as the basis of the exponential function. ω is the timeliness attenuation coefficient, which is an adjustable parameter used to control the speed at which information decays over time. Its value needs to be debugged according to the actual application scenario. Generally, the larger the value, the faster the decay. It is usually between 0 and 1 and can be adjusted appropriately according to needs. t current is the current time, i.e. the timestamp when the information is evaluated. pubish It is the release timestamp of external information, which is used to calculate the time difference of information to evaluate its timeliness.

[0098] Use expert rules to define the weight of each citation source as follows:

[0099] P s =SourceScore(s)

[0100] Among them, P s is the weight of the citation source s, and SourceScore is the score given by experts to the citation source. The weight of each data source can also be defined in a normalized manner after weighted processing based on the frequency of reading and forwarding of the information source.

[0101] Use natural language processing (NLP) technology to evaluate the relevance and quality of content, mainly using vector embedding technology (Embedding), converting text into arrays, and then using distance formulas (such as cosine distance formula, etc.) to calculate relevance. For example, calculate the relevance of external information and corporate and industry information that users are concerned about based on topics or keywords, as follows:

[0102]

[0103] Among them, P c is the correlation between the calculated external information and the enterprise and industry information that the user is concerned about, E i is the embedding vector of external information, E u Embedding vectors for enterprise and industry information that users are interested in;

[0104] The number of times the information is read, forwarded, and repeated can be used as a reference to evaluate the value of information dissemination and to normalize the data. Specifically, the dissemination heat of the external information is calculated as follows:

[0105] P h =w r *Norm(Count r )+w f *Norm(Couut f )+wc

[0106] *Norm(Count c )

[0107] Among them, P h Score the spread of external information, indicating the popularity or attention of the information within the system. r Reading weight, which represents the weight coefficient of the number of times the information is read in the spread heat, is generally determined through experiments or user behavior data analysis, and the value range is usually between 0 and 1. r is the number of times external information is read. f Forwarding weight, which indicates the influence coefficient of the frequency of information forwarding in the transmission heat, is set between 0 and 1 and is determined by experimental data. f is the number of times the external information is forwarded. c The number of times the information is displayed or counted in the system, that is, the exposure of the information in the system. c is the number of counts of external information. Norm(.) is a regular normalization function, which is used to normalize each count item (number of reads, number of reposts, number of impressions) to ensure that the difference in magnitude of different items does not affect the calculation results. Maximum normalization or mean normalization is usually used.

[0108] Through the weighted processing of the quantitative data of the above four dimensions, the value of external input information can be quantified. The specific calculation value of external information is:

[0109] P I =w1*Pt+w2*P s +w3*P c +w4*P h

[0110] Among them, P I is the comprehensive value score of external information, which is used to measure the overall value of the information to the user. i is the weight parameter of each score, corresponding to the timeliness score, reference source weight, relevance score and communication heat score. Each weight can be adjusted according to the needs of the application scenario to reflect the user's attention to different dimensions, i = 1, 2, 3, 4. s Score the citation source to indicate the authority and credibility of the information source. The source score can be determined based on expert experience or the historical performance of the source. c Score the relevance by calculating the relevance between the information content and the content that the user is interested in. hScore the spread of external information, indicating the popularity or attention of the information within the system. It comprehensively evaluates the spread of information by measuring the frequency of information being read, forwarded, and displayed.

[0111] Construct thought chain prompts, input the current status of each related financial product information and corresponding external information into the large language model, and use the text comprehension and reasoning capabilities of the large language model to construct output impact classification. The above impact classification includes positive, negative and neutral.

[0112] Map the impact classification to weights, recalculate the value score of each piece of information, and output the top M related financial product information after re-sorting. For example, the weight of the positive mapping is 1, the weight of the negative mapping is 1.5, and the weight of the neutral mapping is 0.5. The information of risk factors is sorted in advance, and the neutral information is pushed back, so as to meet the demand of sorting valuable information in the front. Such weights can be adjusted dynamically according to different user preferences.

[0113] Combine the top M related financial product information with the thought chain prompt words, and call the large language model to output the analysis report. It can provide detailed information analysis and future impact prediction to help users make wise decisions.

[0114] You can also use tracking technology to evaluate user feedback on generated reports, combine it with users' actual behavior data, and regularly and dynamically update user preferences and corresponding content of interest, so as to provide more accurate services in the future.

[0115] The large language model calling process of the embodiment of the present invention is as follows:

[0116] In the information extraction and evaluation process of the present invention, the calling process of the large language model is crucial to ensure the high timeliness and stability of information processing. The following are the calling conditions, priority processing content and specific processing logic of the large language model:

[0117] Call trigger conditions: The call of the large language model is triggered based on specific conditions, ensuring that the system is called when in-depth analysis or complex classification is required to save computing resources and improve response efficiency. Trigger conditions include:

[0118] 1) New information input: When the system detects the latest financial information from news, market announcements or social platforms, it triggers large language model analysis.

[0119] 2) User behavior feedback: The system records user feedback on recommended information (such as likes, click-through rate, etc.) through tracking technology. When the feedback data changes significantly, the system calls the large language model to re-evaluate and generate updated recommended content.

[0120] 3) Periodic batch processing: To adapt to the periodic changes in user preferences and market environment, the system regularly calls the large language model to update the user preference model and market information to ensure the continued relevance of the recommended content.

[0121] Prioritize content types:

[0122] 1) Information on user holdings and products of interest: Dynamic updates and related analysis of financial products currently held or followed by users to ensure that this information can be reflected in the analysis report in a timely manner.

[0123] 2) High-impact information: Information with high dissemination popularity or reading frequency (such as market trends and major events) is given higher priority, and the model prioritizes its impact classification and relevance assessment.

[0124] 3) Personalized demand-related information: Prioritize information that meets the user’s risk preferences to ensure that the model’s analysis results can directly improve the accuracy of personalized recommendations.

[0125] Measures to ensure timeliness and stability:

[0126] 1) Batch processing and asynchronous analysis: The system processes information in batches according to time and priority. High-priority content triggers immediate analysis, and low-priority content is processed in batches and asynchronously called to save resources, ensuring a quick response when the amount of information is large.

[0127] 2) Cache and short-term memory: Use a cache mechanism for similar data that appears frequently in a short period of time, store the processed results in the cache, and avoid repeated calls to the model. For similar analysis requests, the system directly returns the cached results to improve response speed.

[0128] 3) Load balancing and automatic expansion: When the number of user visits increases or the frequency of information input increases, the system dynamically allocates computing resources and automatically expands the number of model instances to ensure processing stability and smoothness.

[0129] 4) Dynamic priority adjustment based on weight: monitor the timeliness of the content during the call process, reduce the priority of outdated content through the exponential decay function, and give priority to submitting the latest information to the large language model for processing to ensure that the output analysis report is timely. Through the above trigger conditions and processing logic, the method of the present invention ensures that the large language model can provide timely, stable and accurate analysis results under the premise of efficient use of resources, and finally outputs personalized financial information analysis reports that meet user needs.

[0130] Based on the above embodiments, those skilled in the art can easily understand that the present invention also provides a computer-readable storage medium storing a computer program, which implements the above method when executed by a processor.

[0131] The above is only a preferred embodiment of the present invention. It should be noted that for ordinary technicians in this technical field, other parts not specifically described belong to the prior art or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention.

Claims

1. An information extraction evaluation method, characterized in that: include: Obtain external information; Preprocessing the external information to extract the reference source, timestamp, and number of references, and outputting the structured external information to a database; Collect and update the user's personal information, including risk preferences, current product holdings, historical holding records, and information about companies and industries of interest; Based on the user's purchase record and product attributes, expert experience defines the user's risk preference rules and establishes a sample set. Experts give label definition rules for the various attributes of financial products on the market, establish risk preference labels, and use the Transformer model to build a label classification model, generate an association mapping with user preferences, and perform label prediction for all public financial products on the market. Extract the financial product information associated with the user from the external information, calculate the correlation between the associated financial product information and the products held, concerned and risk-related by the user, and output the top N associated financial product information after sorting; Constructing thought chain prompt words, inputting the current status of each associated financial product information output and corresponding external information into the large language model, so that the large language model outputs influence classification; Map the impact classification to the weight, recalculate the value score of each piece of information, and output the top M related financial product information after re-ranking; Combine the top M related financial product information with the thought chain prompt words, and call the large language model to output the analysis report.

2. The information extraction and evaluation method according to claim 1, characterized in that: Also includes: Use tracking technology to evaluate user feedback on generated reports, combine users' actual behavior data, and regularly and dynamically update user preferences and corresponding content of interest.

3. The information extraction and evaluation method according to claim 1, characterized in that: The calculation method of the correlation is as follows: Through keyword query, obtain all information related to the user's current products in external information and filter it according to time requirements; The exponential decay function is used to quantitatively evaluate the timeliness of external information, as follows: Among them, P t is the quantitative evaluation result of the timeliness of external information; e is a natural constant, ω is the timeliness attenuation coefficient, t current is the timestamp when the information is evaluated, t pubish The release timestamp of external information; Use expert rules to define the weight of each citation source as follows: P s =SourceScore(s) Among them, P s is the weight of the citation source s, SourceScore is the score given by the expert to the citation source; Calculate the relevance of external information to the enterprise and industry information that users are interested in based on topics or keywords, as follows: Among them, P c is the correlation between the calculated external information and the enterprise and industry information that the user is concerned about, E i is the embedding vector of external information, E u Embedding vectors for enterprise and industry information that users are interested in; The propagation heat of the external information is calculated as follows: P h =w r *Norm(Count r )+w f *Norm(Count f )+w c *Norm(Count c ) Among them, P h is the calculated propagation heat of the external information, w r is the reading weight, Count r is the number of times external information is read, w f is the forwarding weight, Count f is the number of times external information is forwarded, w c is the counting weight of the information in the system, Count c is the number of counts of external information, Norm(.) is the regularization function; The value of the external information is calculated as: P I =w1*P t +w2*P s +w3*P c +w4*P h Among them, P I is the calculated value of external information, w i is the weight of each dimension, i=1,2,3,4.

4. The information extraction and evaluation method according to claim 1, characterized in that: The impact categories include positive, negative, and neutral.

5. The information extraction and evaluation method according to claim 4, characterized in that: The weight to which the positive mapping is applied is 1, the weight to which the negative mapping is applied is 1.5, and the weight to which the neutral mapping is applied is 0.

5.

6. The information extraction and evaluation method according to claim 1, characterized in that: The risk preference labels include conservative, sound, balanced, aggressive and aggressive.

7. The information extraction and evaluation method according to claim 1, characterized in that: The sources of external information include news and market data.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images