Dynamic optimization method and device of model, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a dynamic optimization method and device of a model, equipment and a medium. According to the embodiment of the invention, the electronic equipment carries out the dynamic optimization of the weight of each analysis agent through an evolutionary learning algorithm based on the comparison between the transaction decision of the analysis agent agents on the target product and the current market trend of the target product, enables the analysis agent agents to evolve continuously along with the change of the market environment, and improves the efficiency. And a strategy structure is continuously optimized, and dynamic self-evolution and continuous adaptive capacity improvement of a system level are realized.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and medium for dynamic optimization of a model. Background Technology

[0002] In recent years, Large Language Models (LLMs) have been increasingly adopted in the field of financial investment decision-making due to their superior language understanding and knowledge reasoning capabilities. Furthermore, researchers have proposed a multi-agent architecture, assigning multiple LLM agents to different roles. Each LLM agent can analyze large-scale unstructured text and generate trading signals based on the analysis results. Then, through the collaboration of multiple LLM agents, a more stable and multi-faceted trading strategy is generated.

[0003] However, since the prompts for the LLM agents and the weights of the multiple LLM agents do not change after training or initial setup, they may be unable to adapt to market fluctuations, leading to inaccurate final trading decisions. Summary of the Invention

[0004] This application provides a method, apparatus, device, and medium for dynamic optimization of a model to adapt to market fluctuations and improve the accuracy of trading decisions.

[0005] In a first aspect, embodiments of this application provide a method for dynamic optimization of a model, the method comprising: If an instruction to perform transaction decision analysis on the target product is received, then each analysis agent will determine the transaction decision of the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent. The daily market trend of the target product is obtained, and an optimized agent is used to update the weight of each analytical agent based on whether the transaction decision corresponding to each analytical agent satisfies the daily market trend. The weight indicates the importance of each transaction decision generated by each analytical agent when the decision agent determines the final transaction decision based on the transaction decisions generated by each analytical agent.

[0006] Secondly, embodiments of this application also provide a dynamic optimization apparatus for a model, the apparatus comprising: The processing module is configured to, upon receiving an instruction to perform transaction decision analysis on a target product, use each analysis agent to determine the transaction decision of each analysis agent for the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent. An optimization module is used to obtain the daily market trend of the target product and, using an optimization agent, update the weight of each analysis agent based on whether the transaction decision corresponding to each analysis agent satisfies the daily market trend. The weight indicates the importance of each transaction decision generated by the analysis agent when the decision agent determines the final transaction decision based on the transaction decisions generated by each analysis agent.

[0007] Thirdly, embodiments of this application also provide an electronic device, the electronic device including a processor, the processor being configured to execute a computer program stored in a memory to implement the steps of the dynamic optimization method of any of the models described above.

[0008] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the dynamic optimization method for any of the models described above.

[0009] In this embodiment, the electronic device compares the transaction decisions of the analysis agent for the target product with the daily market trend of the target product, and uses an evolutionary learning algorithm to dynamically optimize the weight of each analysis agent, so that the analysis agent can continuously evolve with changes in the market environment, continuously optimize the strategy structure, and achieve dynamic self-evolution and continuous improvement of adaptability at the system level. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A schematic diagram illustrating the dynamic optimization process of a model provided in an embodiment of this application; Figure 2 A schematic diagram of the transaction logic provided in the embodiments of this application; Figure 3 A schematic diagram of the evolutionary process provided for embodiments of this application; Figure 4 A schematic diagram of a dynamic optimization device for a model provided in this application embodiment; Figure 5 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] Traditional quantitative trading primarily relies on rule-driven and statistical modeling methods, such as trend following, momentum strategies, mean reversion, and portfolio optimization, to generate trading signals. However, this traditional approach, which depends on rule-driven and statistical modeling, generally suffers from a limited perspective: models built through statistical modeling often focus only on quantifiable features such as prices and technical indicators, making it difficult to effectively integrate more complex and multi-dimensional financial information such as policy guidance, macroeconomic variables, and market sentiment. This limits the comprehensiveness and adaptability of trading decisions.

[0014] In recent years, LLM (Limited Language Modeling) has been increasingly adopted in the field of financial investment decision-making due to its superior language understanding and knowledge reasoning capabilities. LLM can process large-scale unstructured text and perform summarization, reasoning, and sentiment analysis, thus providing rich information support for trading decisions. Existing research has inputted news texts and macroeconomic information into LLM to generate trading signals by analyzing their sentiment tendencies. For example, some sentiment-driven strategies construct buy or sell decision logic based on the sentiment scores extracted from news by LLM, which significantly improves the problem of prediction noise interference.

[0015] Furthermore, a multi-agent architecture is proposed, assigning multiple LLM agents to different roles. Through collaboration or "debate" mechanisms between models, multi-dimensional information is integrated to generate more stable and multi-faceted trading strategies. For example, in the Trading Generative Pre-trained Transformer (TradingGPT) framework, different agents possess layered memory structures and distinct characteristics, and they can debate with each other to reach trading decisions, thereby improving the system's responsiveness to historical and real-time signals. The Trading Agents framework, mimicking the organizational structure of a real trading firm, includes fundamental analysts, sentiment analysts, technical analysts, researchers (Bull / Bear), traders, and a risk management team. Through structured communication and natural language debate, role collaboration and risk control are achieved.

[0016] Although considerable work has validated the potential of LLM multi-agent systems in financial transactions, the technology still faces several key limitations: 1. Lack of dynamic optimization mechanism: Many systems still use static prompts or agent roles with fixed weights. Their policy logic hardly changes after training or initial settings, and cannot adaptively optimize according to market fluctuations or real-time feedback.

[0017] 2. Difficulty in assessing contribution: In multi-agent systems, the decision-making contributions of each role (such as analyst agent, researcher agent, and trader agent) are difficult to quantify, making it difficult for the system to clearly identify high-value agents and inefficient agents, thus limiting the optimal allocation of resources (such as computation and API calls).

[0018] 3. Insufficient intelligence and professionalism: Currently, some LLM agents still lack the deep reasoning and interpretability of human experts in a financial context. Some systems struggle to understand complex financial statements, macroeconomic policies, and the professional logic of trading experts.

[0019] In summary, while existing technologies have demonstrated the enormous potential of LLM and multi-agent systems in financial transactions, significant gaps remain in areas such as dynamic optimization, agent contribution assessment, agent collaboration, and specialization.

[0020] Based on this, in order to adapt to market fluctuations and improve the accuracy of trading decisions, embodiments of this application provide a dynamic optimization method, apparatus, device, and medium for a model. The method includes: if an instruction to perform trading decision analysis on a target product is received, then each analysis agent is used to determine the trading decision of each analysis agent for the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent; the daily market trend of the target product is obtained, and an optimization agent is used to update the weight of each analysis agent based on whether the trading decision corresponding to each analysis agent satisfies the daily market trend; wherein, the weight indicates the importance of the trading decision generated by each analysis agent when the decision agent determines the final trading decision based on the trading decisions generated by each analysis agent.

[0021] The embodiments of this application can dynamically adjust the weights of multiple agents, continuously evolve the trading logic, and improve the profitability and risk control capabilities while ensuring robustness. This not only fills the gap in current technology, but also introduces a new level of adaptability, intelligence and interpretability into the financial AI trading system.

[0022] Example 1: Figure 1 This application provides a schematic diagram of a dynamic optimization process for a model, which includes: S101: If an instruction to perform transaction decision analysis on the target product is received, then each analysis agent determines the transaction decision of the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent.

[0023] The dynamic optimization method for a model provided in this application is applied to an electronic device, which may be a PC, a trading system, a trading server, or a trading decision platform, etc.

[0024] In this embodiment, the electronic device receives an instruction to perform transaction decision analysis on a specific product. This instruction can be sent via a client, another device, or a scheduled task; no limitation is placed on this. The instruction carries at least an identifier for the product to be analyzed, which may be the product's name, number, serial number, etc. Upon receiving the instruction, the electronic device obtains the identifier and identifies the identified product as the target product for the transaction decision analysis. This target product can be stocks, contracts, funds, financial consumer products, etc., without limitation.

[0025] After the electronic device identifies the target product, it performs collaborative reasoning using analytical agents. These agents can be one or multiple. In this embodiment, to improve prediction accuracy, multiple analytical agents are used. Each agent predicts the future market trend of the target product from a different perspective, and each agent has corresponding prompts, which are different for each agent. Each agent utilizes a large language model, combined with its unique prompts, to perform analysis and prediction, generating a transaction decision for the target product. The large language model used by each agent can be the same or different.

[0026] For example, each analytics agent can share a single agent, or each analytics agent can use its own agent; there is no restriction on this.

[0027] In this embodiment, after receiving an instruction to perform transaction decision analysis on a target product, the electronic device, for each analysis agent, obtains the corresponding prompt word, inputs the prompt word and the identifier of the target product into the analysis agent, enabling the analysis agent to invoke the corresponding large language model to predict the future market trend of the target product from the corresponding business perspective, and generates a transaction decision for the target product. For example, the transaction decision can be buy, sell, or remain unchanged, and must also include the specific transaction quantity, which is not limited here.

[0028] In the embodiments of this application, the analysis agent can be at least one of an investment master agent, a news analysis agent, a fundamental analysis agent, and a social media analysis agent.

[0029] Figure 2 This is a schematic diagram of the transaction logic provided in the embodiments of this application, as shown below. Figure 2 As shown, the transaction process includes: 1. Multi-Agent Analysis: When the electronic device receives market information from the database and instructions to analyze transaction decisions for the target product, various analytical agents initiate analysis in parallel: the fundamental analysis agent (fundamental analysis agent) analyzes financial statements and calculates indicators such as net profit growth rate, ROE, and debt-to-equity ratio; the social media analysis agent (social media analysis agent) captures social media reactions and assesses market sentiment bias; the news analysis agent (news analysis agent) captures the latest news events and infers market trends; and the investment master analysis agent (investment master agent) integrates historical data and financial statements to determine the long-term value and competitive advantage of the target product. Each analytical agent outputs independent transaction decisions, namely O1, O2, O3, and O4.

[0030] 2. Signal Aggregation and Confidence Calibration: The electronic device employs a weighted voting mechanism to fuse transaction decisions from multiple analytical agents. Initial weights are evenly distributed among the analytical agents, namely W1, W2, W3, and W4.

[0031] 3. Portfolio Decision-Making and Trading: The trading agent (trading agent) aggregates the trading decisions of various analytical agents, allocates the importance of each analytical agent to the trading agent according to its weight, and generates the final decision trade (trading signal).

[0032] S102: Obtain the daily market trend (real market) of the target product, and use an optimized agent to update the weight of each analytical agent based on whether the transaction decision corresponding to each analytical agent satisfies the daily market trend; wherein, the weight indicates the importance of the transaction decision generated by each analytical agent when the decision agent determines the final transaction decision based on the transaction decision generated by each analytical agent.

[0033] In this embodiment, after each analytical agent generates a transaction decision for the target product, the electronic device can invoke the transaction agent to generate a final transaction decision based on the transaction decision generated by each analytical agent for the target product.

[0034] Specifically, in this embodiment, the electronic device is pre-configured with a weight corresponding to each analytical agent. This weight is used by the trading agent to indicate the importance of each analytical agent's trading decisions when determining the final trading decision based on the trading decisions generated by each analytical agent. The larger the weight of any analytical agent, the more important the analytical agent is, meaning the greater the impact of its trading decisions on the final trading decision; conversely, the smaller the weight of any analytical agent, the less important the analytical agent is, meaning the smaller the impact of its trading decisions on the final trading decision.

[0035] The transaction agent can determine the final transaction decision for the target product based on the transaction decisions generated by each analytical agent and the weight corresponding to each analytical agent.

[0036] To improve prediction capabilities, in this embodiment, the electronic device can dynamically adjust the influence of each analytical agent in the final transaction decision based on the transaction decision of each analytical agent on the target product and the market trend of the target product on that day. That is, it can adjust the weight of each analytical agent, thereby achieving a continuously evolving adaptive mechanism.

[0037] Specifically, electronic devices need to obtain the target product's market trend on the decision-making day. This market trend is a standardized label used to represent the price fluctuation of the target product, which can be upward, downward, or sideways. The market trend can be derived by comparing the closing price and the opening price, or by combining indicators such as trading volume and volatility for more refined calculations. For example, if the target product's closing price increases by 2.3% compared to the previous day, then the market trend for that day is determined to be upward.

[0038] In addition, in this embodiment of the application, the electronic device is also equipped with an optimized agent, which can be used to adjust the weights corresponding to each analysis agent.

[0039] Specifically, the electronic device inputs the target product's daily market trend, the trading decisions corresponding to each analytical agent, and the prompts for the optimization agent into the optimization agent. The optimization agent then compares each analytical agent's trading decision with the daily market trend. For example, if an analytical agent's trading decision is to buy the target product, and the target product's daily market trend is "rising," then the analytical agent's decision is considered correct; if an analytical agent's trading decision is to buy the target product, but the target product's daily market trend is "falling," then the analytical agent's decision is considered incorrect.

[0040] Based on this, the optimization agent adjusts the weight of each analytical agent according to the judgment information of whether the current decision is correct. In this embodiment, the initial weight of each analytical agent can be set to be equal. For example, if there are four analytical agents, the initial weight of each analytical agent is 25%. However, as the trading day progresses, the weight of each analytical agent will change dynamically. For example, if an analytical agent makes incorrect predictions for five consecutive days, the weight of that analytical agent will be gradually reduced; while if an analytical agent is consistently accurate in a volatile market, the weight of that analytical agent will be increased.

[0041] Furthermore, in this embodiment, the weights corresponding to each analytical agent are not only the basis for the transaction agent's decision-making but also the input for evolutionary learning. The optimized agent can periodically write data such as the weights, accuracy, and market tags corresponding to each analytical agent into a log for use by the monthly evolution module.

[0042] In addition, in this embodiment of the application, the electronic device can also update the weights of each analysis agent based on the gradient of the reinforcement learning (RL) policy.

[0043] In this embodiment, the electronic device compares the transaction decisions of the analysis agent for the target product with the daily market trend of the target product, and uses an evolutionary learning algorithm to dynamically optimize the weight of each analysis agent, so that the analysis agent can continuously evolve with changes in the market environment, continuously optimize the strategy structure, and achieve dynamic self-evolution and continuous improvement of adaptability at the system level.

[0044] Example 2: To adapt to market fluctuations and improve the accuracy of trading decisions, based on the above embodiments, the method in this application embodiment further includes: If the current date is the end date of this optimization cycle, after updating the weights corresponding to each analysis agent, obtain the updated weights of each analysis agent. If there exists a first analysis agent with an updated weight less than a preset weight threshold, then a second analysis agent is constructed and the second analysis agent is used to replace the first analysis agent. The updated weights corresponding to the first analysis agent are used as the weights corresponding to the second analysis agent in the next optimization cycle.

[0045] In the dynamic environment of financial markets, static agent configurations struggle to maintain efficient decision-making capabilities over the long term. To address this issue, this application introduces a periodic agent update mechanism, enabling the identification, elimination, and replacement of consistently underperforming analytical agents at the end of a fixed time window, thereby ensuring the vitality and forward-looking nature of the overall decision-making system.

[0046] Specifically, once the current optimization cycle ends on the last day, a comprehensive review and elimination process for the analytical agent will be automatically triggered. The end date of this optimization cycle is typically set as the last trading day of the calendar month. On this end date, the electronic device will not only focus on the daily performance of the analytical agent but also on the overall accuracy of each agent's decisions throughout the entire optimization cycle. Since the weight of each analytical agent is dynamically adjusted daily based on its prediction accuracy, market feedback, and confidence level, this weight reflects the agent's overall contribution throughout the entire optimization cycle.

[0047] Based on this, at the end of the optimization cycle, after updating the weights corresponding to each analysis agent, the electronic device reads the updated weights of each analysis agent and compares them with a preset weight threshold. If the updated weight of an analysis agent is lower than the preset weight threshold, the analysis agent is marked as the first analysis agent, meaning the electronic device determines that the first analysis agent can no longer provide effective decision support in the current market environment and is considered an inefficient or even negatively impactful unit.

[0048] The electronic device will remove the first analytical agent from the overall architecture. In order to avoid a break in the transaction decision-making agent and thus cause decision fluctuations, the electronic device will build a brand-new second analytical agent to replace the first analytical agent. The second analytical agent will use some of the state information of the first analytical agent, ensuring that the electronic device maintains decision continuity while updating the structure and avoiding drastic policy fluctuations caused by agent switching.

[0049] Specifically, the initial weight of the second analytical agent is set to the weight of the first analytical agent at the end of this optimization cycle, so that the second analytical agent has stronger adaptability and reasoning ability while retaining its original functional positioning.

[0050] In this embodiment, the preset weight threshold is not set arbitrarily, but is obtained based on historical backtesting data and risk tolerance calibration.

[0051] For example, if the preset weight threshold is 0.05, and the weight of the first analytical agent continuously declines from 0.22 at the beginning of the month to 0.04, since 0.04 is less than the preset weight threshold of 0.05, the electronic device constructs a second analytical agent and replaces the first analytical agent with the second analytical agent. Furthermore, the electronic device sets the initial weight of the second analytical agent to 0.04.

[0052] To adapt to market fluctuations and improve the accuracy of trading decisions, based on the above embodiments, the method in this application embodiment further includes: A semantic adjustment model is used to perform semantic fusion on the second prompt words corresponding to any two third analysis agent intelligent bodies other than the first analysis agent intelligent body and the first prompt word corresponding to the first analysis agent intelligent body to obtain the third prompt word; The third prompt word will be used as the prompt word for the second analysis agent in the next optimization cycle.

[0053] In this embodiment, after the electronic device uses the constructed second analysis agent to identify and replace the first analysis agent with low accuracy, the electronic device will give the second analysis agent better decision logic.

[0054] Specifically, this application proposes a prompt word generation method based on semantic fusion. By simulating the crossover operation in biological genetics, the second prompt words of multiple high-performing analysis agent agents and the first prompt words of the first analysis agent agent are recombined at the semantic level, thereby creating a third prompt word that combines professional depth and cross-domain insight. This third prompt word is then used as the prompt word corresponding to the second analysis agent agent in the next optimization cycle.

[0055] The method for generating prompt words can be implemented by a specially trained semantic adjustment model, which can ensure that the generated third prompt words not only conform to language norms but also have a clear financial reasoning path.

[0056] Specifically, when an electronic device decides to build a second analytical agent to replace the first analytical agent, it excludes the first analytical agent from all currently applied analytical agents and then randomly selects two third analytical agents as the logical source. For example, it might choose a fundamental analysis agent and an investment guru agent, which respectively represent quantitative rigor and long-term value thinking.

[0057] The electronic device invokes the semantic adjustment model to perform deep semantic parsing and semantic fusion on the second prompt words of the two third analysis agent agents and the first prompt word of the first analysis agent agent, thereby generating and outputting the third prompt word.

[0058] The semantic adjustment model, pre-trained on large-scale financial text, identifies core functional modules in prompts, such as "ROE stability assessment," "safety margin calculation," and "public opinion heat threshold setting," mapping them to high-dimensional semantic vectors. In the vector space, the semantic adjustment model calculates the semantic similarity and complementarity between prompts. Second prompts from two third-party analytical agents are considered high-quality and retained; invalid or conflicting logic in the first prompt from the first analytical agent is removed. Subsequently, the semantic adjustment model performs semantic fusion, creatively combining prompts from different analytical agents to obtain initial third prompts. For example, combining "moat analysis" from the investment guru agent with "sentiment polarity judgment" from the social media agent creates a new rule: "If a target company possesses a significant moat, but the social media sentiment index remains in the extremely pessimistic range for three consecutive days, it is considered a short-term oversold opportunity." This cross-domain fusion not only preserves the original professional logic but also introduces new decision-making dimensions, giving the new agent a stronger environmental awareness.

[0059] The semantic adjustment model also performs natural language reconstruction on the initial third prompt words obtained through fusion, ensuring that the final generated third prompt words are grammatically correct, clearly defined, and unambiguous. For example, the third prompt word might be: "You are an investor who focuses on the intrinsic value of a company; please prioritize evaluating free cash flow and return on capital." However, before making a final decision, it is necessary to check whether there have been any major negative events in mainstream financial news over the past five days, and to combine this with social media sentiment indices to determine whether the market has overreacted.

[0060] In the embodiments of this application, the semantic adjustment model can be a large language model or a neural network model, and there is no limitation.

[0061] Example 3: To adapt to market fluctuations and improve the accuracy of transaction decisions, based on the above embodiments, in this embodiment, the step of using each analytical agent to determine the transaction decision for the target product based on the identifier of the target product and the corresponding prompt words of each analytical agent includes: For each analytical agent, the identifier of the target product and the corresponding prompt word are input into the analytical agent. After receiving the prompt word and the identifier, the analytical agent retrieves the information corresponding to the target product with the identifier from the pre-configured database corresponding to the analytical agent, and generates the transaction decision based on the information and the prompt word.

[0062] In this embodiment of the application, the process of analyzing the agent's generation of transaction decisions is essentially a highly structured professional reasoning activity. Its core lies in combining the unique identifier of the target product with the prompt words corresponding to the agent, obtaining domain-aligned information sources through a pre-configured data interface, and finally outputting an interpretable decision report.

[0063] In this embodiment, the input of each analytical agent consists of two parts: first, an identifier for the target product, used to anchor the object of analysis; and second, a unique prompt for the analytical agent, used to limit its analytical perspective and reasoning logic. For example, the prompt for the fundamental analysis agent explicitly requires: "Based on the latest financial report, calculate the net profit growth rate, gross profit margin, and debt-to-equity ratio, and assess its financial health." This prompt has undergone multiple rounds of iterative optimization to ensure that its language effectively guides the large language model to focus on key financial indicators, rather than discussing general industry trends.

[0064] The electronic device also configures a corresponding database for each analysis agent. The database for each analysis agent can be the same or different.

[0065] Specifically, for each analytical agent, after receiving input, the analytical agent does not directly call the general knowledge base, but instead calls the database corresponding to the analytical agent through a dedicated data interface pre-configured for the analytical agent, thereby pulling the real-time, structured professional data required for the analytical agent's decision analysis.

[0066] For example, a fundamental analysis agent retrieves financial statements from a financial statement knowledge base and calculates indicators such as net profit growth rate, ROE, and debt-to-equity ratio by analyzing the financial statements; a social media analysis agent retrieves social media reactions from a social media reaction knowledge base and assesses market sentiment bias by analyzing the social media reactions; a news analysis agent retrieves the latest news events from a news knowledge base and infers market trends by analyzing the latest news events; and an investment master agent retrieves historical data and financial statements from a historical knowledge base and judges the long-term value and competitive advantage of target products by analyzing the historical data and financial statements.

[0067] For example, the fundamental analysis agent connects to a financial statement knowledge base, enabling it to accurately obtain key fields such as year-on-year revenue growth, net profit margin, and net operating cash flow for a target product within a specific time period. The news analysis agent crawls recent news reports about the target product through news media APIs and uses an event extraction model to identify the subjects, actions, and potential impacts. The social media agent connects to social media APIs and uses a sentiment analysis model to quantify the optimism / pessimism ratio in user comments, discovering that recent discussions have increased in popularity but the sentiment leans towards caution.

[0068] Based on the information about the target product obtained by the analytical agent, the agent performs multi-step reasoning, including but not limited to: first extracting key facts, then making judgments based on the logical chain in the prompts, and finally organizing the information into a natural language report. For example, the natural language report generated by the fundamental analysis agent might be: "The target product has stable performance, ample cash flow, a healthy asset-liability structure, and its current valuation is at the historical median; we recommend holding." This natural language report not only includes the final decision but also provides supporting reasons and data to ensure that the decision-making process is transparent and auditable.

[0069] In this embodiment, each agent can maximize its effectiveness within its area of ​​expertise while avoiding information contamination. For example, sentiment data will not interfere with fundamental judgments, and policy interpretations will not obscure financial facts. Through role isolation and dedicated data supply, electronic devices achieve a unified "independence + collaboration" for multi-perspective analysis, providing a high-quality input source for subsequent signal aggregation.

[0070] Example 4: To adapt to market fluctuations and improve the accuracy of trading decisions, based on the above embodiments, in this embodiment, the use of optimized agent intelligence, and updating the weights of each analytical agent intelligence according to whether the trading decisions corresponding to each agent intelligence satisfy the market trend of the day, includes: For each analytical agent, the analytical accuracy of the agent in the current optimization cycle is updated based on whether the trading decision corresponding to the agent satisfies the market trend of the day. The optimized agent is used to update the weights of each analytical agent based on the transaction decisions corresponding to each analytical agent, the market trend of the day, and the analytical accuracy of each analytical agent in the current optimization cycle.

[0071] In this embodiment, the scientific nature of the weight update mechanism directly determines whether the electronic device can continuously optimize decision-making quality in a dynamic market. Based on this, in this embodiment, the electronic device updates the weights corresponding to each analytical agent based on the analytical accuracy of the analytical agent. This ensures that the weight adjustment of each analytical agent not only reflects its single-time performance but also embodies its long-term stability and trend consistency, thereby avoiding decision distortion caused by overfitting or excessive penalties.

[0072] Specifically, at the end of each trading day, the electronic device obtains the market trend of the target product for that day and calls upon the optimized agent to compare the predicted market trend of each analytical agent with the actual market trend for that day.

[0073] For each analytical agent, if the predicted market trend is consistent with the market trend on that day, the number of correct predictions made by the analytical agent is updated, i.e., the number of correct predictions is increased by one; if the predicted market trend is inconsistent with the market trend on that day, the number of incorrect predictions made by the analytical agent is updated, i.e., the number of incorrect predictions is increased by one; the analytical accuracy of the analytical agent is determined based on the number of correct and incorrect predictions made by the analytical agent.

[0074] The electronic device employs an optimized agent, updating the weights of each agent based on its trading decisions, daily market trends, and the accuracy of its analysis within the current optimization cycle.

[0075] Specifically, the electronic device inputs the trading decisions, daily market trends, analysis accuracy of each analysis agent in the current optimization cycle, and prompts for the optimization agent into the optimization agent. This allows the optimization agent to update the weights of each analysis agent based on the trading decisions, daily market trends, and analysis accuracy of each analysis agent in the current optimization cycle.

[0076] In order to adapt to market fluctuations and improve the accuracy of trading decisions, based on the above embodiments, in this application embodiment, the trading decision includes a forecast of the target product's market trend and a corresponding analysis report; The transaction decisions made by the analytical agent satisfy the market trend of the day, including: The predicted market trend carried in the transaction decisions of the analytical agent is consistent with the market trend of the day.

[0077] To ensure the objectivity and reproducibility of the contribution evaluation of each analytical agent, in the embodiments of this application, the transaction decisions generated in each analytical agent include the predicted market trend of the target product and the corresponding analytical report.

[0078] Among them, electronic devices can divide the daily market trend of the target product into the following categories: if the closing price of the target product increases by 1% or more compared with the previous trading day, the market trend for the day is "rising"; if it falls by 1% or more, the market trend for the day is "falling"; if the rise or fall is between -1% and +1%, the market trend for the day is "sideways". This classification standard has been verified by a large amount of historical data, which can effectively distinguish between trending market and volatile market, and avoid misjudgment caused by small fluctuations.

[0079] In this application, each analytical agent must include a predicted market trend in its trading decision. The classification of this predicted market trend must be consistent with the classification of the market trend for the day, and vague expressions such as "likely to rise" or "stronger" are prohibited. Only when the predicted market trend of the analytical agent is completely consistent with the market trend for the day will the electronic device determine that the analytical agent's trading decision conforms to the market trend for the day.

[0080] For example, if a trading decision generated by an analytical agent predicts an upward trend in the target product's market direction, but the actual market trend for that target product on that day is also upward, then the trading decision is determined to satisfy the market trend for that day. Conversely, if a trading decision generated by an analytical agent predicts an upward trend in the target product's market direction, but the actual market trend for that target product on that day is downward or sideways, then the trading decision is determined to not satisfy the market trend for that day.

[0081] In the logs of the electronic device, every matching result of each analysis agent is recorded for subsequent accuracy calculation and weight update.

[0082] Figure 3 A schematic diagram of the evolutionary process provided in the embodiments of this application, as shown below. Figure 3 As shown, the process includes: 1. Fitness Assessment: Electronic devices call upon optimized agent intelligence. The daily trading decisions O1, O2, O3, and O4 of each analysis agent intelligence are compared with the market trend (real market) of the day to calculate the contribution of each analysis agent intelligence (analysis agent contribution calculation). The predictive ability of each analysis agent intelligence is measured, and the weights of analysis agent intelligence intelligence with weak predictive ability are updated and assigned lower weights.

[0083] 2. Evolutionary Operation Process: Each month, electronic devices use an optimized agent (optimized agent) to summarize the prediction performance of each analytical agent, eliminate the first analytical agent whose weight is below the threshold, and use an evolutionary algorithm to generate new prompt words.

[0084] Enabling the evolutionary algorithm includes the following steps: Selection: Retain the second cue word of the analysis agent with the highest fitness ranking as the parent cue word; Crossover: Randomly select two parent prompt words and merge them with the first prompt word of the eliminated analysis agent to generate a new initial third prompt word; Variation: A semantic-level adjustment is randomly introduced into the initial third prompt word to obtain a new third prompt word. This semantic-level adjustment can include adding a "policy analysis" dimension to predict the market, etc.

[0085] 3. Iterative update and feedback loop: A second analytical agent is constructed and replaced with the first analytical agent. The generated third prompt word is used as the prompt word configuration for the second analytical agent in the next optimization cycle. The updated weights of the first analytical agent are used as the weights for the second analytical agent in the next optimization cycle, and the process proceeds to the next optimization cycle. Through multiple rounds of "evaluation-evolution-re-evaluation" iterations, the electronic device achieves self-evolution of the analytical agent's prompt words and performance improvement, gradually forming a more stable and market-adaptable multi-agent decision-making system.

[0086] Example 5: In order to adapt to market fluctuations and improve the accuracy of trading decisions, based on the above embodiments, in this application embodiment, each analysis agent includes at least one of an investment master agent, a news analysis agent, a fundamental analysis agent, and a social media analysis agent.

[0087] The multi-agent architecture of this application is based on a high degree of specialization and role differentiation. The electronic devices include, but are not limited to, investment master analysis agent agents, news analysis agent agents, fundamental analysis agent agents, and social media analysis agent agents. Each analysis agent agent simulates the cognitive patterns and decision-making logic of a specific financial role in the real world, together forming a three-dimensional analysis network covering macro, meso, and micro dimensions.

[0088] Among them, the Investment Master Agent, guided by classic value investing theory, incorporates core concepts such as "moat," "discounted free cash flow," and "margin of safety" into its prompts, enabling it to assess a company's intrinsic value from a long-term perspective. The News Analysis Agent focuses on parsing unstructured texts such as policy documents, regulatory announcements, and industry news, using event-driven reasoning models to determine their potential impact on target products. The Fundamental Analysis Agent delves into the details of financial statements, excelling at extracting key indicators from income statements, balance sheets, and cash flow statements to assess a company's profitability and financial risk. The Social Media Analysis Agent captures the pulse of public sentiment, quantifying retail investor optimism or fear through natural language processing technology to warn of irrational market fluctuations.

[0089] In this embodiment, users can flexibly enable all or part of the proxies according to their own investment style. For example, long-term value investors can turn off social media proxies to avoid short-term noise interference; while short-term traders can strengthen the weight of news and sentiment proxies to improve their response speed to sudden events. The electronic device automatically adapts to the signal aggregation and weight update mechanism, and can stably output high-quality decisions regardless of changes in the proxies combination.

[0090] This application's embodiments revolve around multi-agent collaborative decision-making and dynamic evolutionary optimization to construct a complete intelligent investment research and trading system, forming an adaptive loop system of analytical agent construction and initialization, closed-loop execution of investment decisions, and evolutionary learning optimization feedback. The electronic device not only enables collaborative analysis by multi-role analytical agents but also achieves dynamic optimization of the analytical agent's capabilities through an evolutionary learning mechanism.

[0091] The analysis agent intelligent agent construction and initialization process aims to generate an intelligent agent group with multi-dimensional financial analysis capabilities, and construct the analysis foundation layer.

[0092] Among these, electronic devices can categorize analytical agents into two main types based on the professional specialization and decision-making logic differences in the financial investment field: investment master agents, used to simulate typical investment ideas and the trading logic of real-world investment masters to make trading decisions; professional analyst agents, used to focus on specific analytical dimensions, including but not limited to news analysis agents, which focus on analyzing news related to the stock market to capture future market trends; fundamental analysis agents, which focus on analyzing the fundamentals of various companies, such as financial reports, to infer market trends from this information; and social media analysis agents, which focus on analyzing public sentiment and opinions about the market on social media platforms, and to infer market trends from this information.

[0093] Electronic devices can leverage large language models to inject specialized features into each analytical agent through prompt word engineering. For example, the prompt words for an investment guru agent might include elements such as "assessing a company's competitive advantage, ROE stability, and margin of safety"; while the fundamental analysis agent might focus on a company's financial statements. Furthermore, different agents are configured with dedicated data interfaces: the fundamental analysis agent connects to a financial database, the social media analysis agent interfaces with social media APIs, and the news analysis agent interfaces with news APIs, thus ensuring professional consistency between data sources and analytical perspectives.

[0094] The dynamic optimization method for the model provided in this application will be described below with reference to a specific embodiment: Assuming the electronic device receives an instruction to perform trading decision analysis on target product A for the first time, on the first day of this optimization cycle, the electronic device assigns equal weights to the four analytical agents (Investment Master Agent, News Analysis Agent, Fundamental Analysis Agent, and Social Media Analysis Agent), meaning each agent receives 25% importance. The Investment Master Agent, after considering target product A's long-term financial trend, industry position, and management stability, outputs a trading decision predicting an upward market trend and providing the following reasons: 1. Continuously rising EPS and steady dividend growth over the past three years; 2. Strong ecosystem barriers and improved profit margins due to increased service revenue; 3. Stable management and strong capital repurchase efforts. The News Analysis Agent, after considering the latest news headlines, industry events, and regulatory changes, outputs a trading decision predicting a downward market trend and providing the following reasons: Regulatory agencies have announced a new round of review of target product A's privacy protection. Regulatory scrutiny carries potential negative consequences, particularly privacy issues that could lead to fines or business adjustments. The fundamental analysis agent, after reviewing the latest financial report (revenue, gross margin, net margin) and key financial ratios of target product A, outputs a trading decision predicting an upward market trend and providing the corresponding reasons: the financial report exceeded expectations, revenue growth recovered, and the market's concerns about a cyclical downturn in target product A did not materialize. The social media analysis agent, after considering the level of discussion on social media, outputs a trading decision predicting an upward market trend and providing the corresponding reasons: target product A has a high level of discussion on social media platform X, with more positive than negative reviews.

[0095] The trading agent first focuses on the weights of various analytical agents. Since the weights of all analytical agents are currently equal, it then focuses on the market trend predictions and reasoning output by each agent. After analysis, the trading agent finds that more agents predict an upward market trend, and the reasons given by these agents are more reliable. Therefore, it chooses to buy target product A, with a purchase amount of A. At the end of the first day of trading, the market trend for target product A is indeed upward. However, because the news analysis agent predicted a downward market trend, it was wrong and thus "dragged down" the trading performance. If the news analysis agent had predicted an upward market trend, it might have purchased a larger amount of target product A, thus reducing its weight.

[0096] The weighting process is achieved through optimizing the agent intelligence. For example, if the agent intelligence is told that the market trend for target product A is upward on a given day, and the investment master agent intelligence, news analysis agent intelligence, fundamental analysis agent intelligence, and social media analysis agent intelligence provide their respective market trend predictions and rationales, the agent intelligence will, based on this information, decide to reduce the weight of the news analysis agent intelligence to 23%, and also change the weight of the investment master analysis agent intelligence to 26%, and the weights of the fundamental analysis agent intelligence and social media analysis agent intelligence to 25.5%. The logic remains the same for the following days.

[0097] Assuming the tenth day is the end date of this optimization cycle, the weights of each analytical agent are as follows: Investment Master Analysis Agent: 30%; News Analysis Agent: 18%; Fundamental Analysis Agent: 28%; Social Media Analysis Agent: 24%. Specifically, the trading decisions for each analytical agent on the tenth day are: Investment Master Analysis Agent predicts a market decline; News Analysis Agent predicts a market rise; Fundamental Analysis Agent predicts a market decline; Social Media Analysis Agent predicts a market rise; and the market trend for target product A on the tenth day is a decline. Based on this, the electronic device will provide the optimized agent with the following predictions: The market trend for target product A on that day is downward; the predictions from the investment master analysis agent are also downward, the news analysis agent is upward, the fundamental analysis agent is downward, and the social media analysis agent is upward, along with the reasons for each agent's prediction. Furthermore, the news analysis agent has made incorrect predictions 5 times in the past 10 days, and the social media analysis agent has made incorrect predictions 3 times in the past 10 days. In this case, the optimized agent will further reduce the weight of the news analysis agent and also decrease the weight of the social media analysis agent. However, because the news analysis agent has made more incorrect predictions, its weight will decrease more significantly. Conversely, the optimized agent will also increase the weights of the investment master analysis agent and the social media analysis agent.

[0098] If the updated weight of the news analysis agent falls below a preset weight threshold, the electronic device will construct another analysis agent and replace the current news analysis agent with that other analysis agent. Specifically, the electronic device uses a semantic adjustment model to semantically fuse any two corresponding second prompt words from the investment master agent, fundamental analysis agent, and social media analysis agent with the first prompt word corresponding to the news analysis agent, resulting in a third prompt word. The electronic device uses this third prompt word as the prompt word for the other analysis agent in the next optimization cycle, and also uses the updated weight of the news analysis agent as the weight for that other analysis agent in the next optimization cycle.

[0099] Compared with the prior art, the embodiments of this application have the following improvements: 1. Multi-role Professional Agent Collaborative Architecture: This application embodiment constructs a multi-dimensional collaborative system composed of investment master analysis agent agents and professional analysis agent agents to achieve cross-domain knowledge integration and professional financial decision-making. At the system design level, based on the specific needs of financial investment scenarios, the agents are first divided into multiple professional roles, such as value investment analysis agent agents, technical analysis agent agents, market sentiment analysis agent agents, and investment master analysis agent agents. Each role agent agent optimizes prompt words in a large language model based on corresponding domain knowledge to match different investment styles and analytical logics. For example, the investment master agent agent focuses on assessing company fundamentals and long-term intrinsic value, price patterns, volume-price relationships, and technical indicators.

[0100] The electronic device's operation process constructs a complete closed loop encompassing multi-agent analysis, signal aggregation, portfolio decision-making, risk assessment, and trade execution. Each analytical agent outputs its professional analysis results in parallel. The signal aggregation module uses a weighted voting mechanism combined with confidence level calibration to generate the final trading decision, with the weights of each analytical agent dynamically adjusted based on its historical contribution. The prediction results of each analytical agent are recorded, providing data support for subsequent contribution evaluation and model optimization.

[0101] Through the above-mentioned architectural design, this application breaks through the limitations of a single model in terms of analytical dimensions and cognitive breadth, and realizes the deep integration of cross-professional knowledge and multi-perspective collaborative decision-making, thereby significantly improving the system's comprehensive analytical capabilities, robustness and decision reliability.

[0102] 2. An agent dynamic optimization mechanism based on evolutionary learning: Addressing the existing problems of static decision-making logic and insufficient adaptability in intelligent agents, this application constructs a prompt word dynamic optimization framework based on evolutionary learning, enabling the agent to continuously self-evolve in a dynamic market environment. This framework includes three key stages: performance evaluation, population iteration, and agent mutation.

[0103] The process involves several phases: **Performance Evaluation Phase:** The electronic device compares the trading decisions of each analytical agent with the daily market trend, calculating quantitative indicators such as decision accuracy for each agent. This is combined with contribution analysis (e.g., weighting in signal aggregation, signal confidence) to generate a fitness score for each agent. **Population Iteration Phase:** On a daily timescale, the electronic device downgrades the weights of analytical agents with low prediction accuracy and allocates more weight to those with stronger prediction capabilities. On a monthly timescale, the electronic device eliminates analytical agents with prediction accuracy below a preset weight threshold and evolves new agents. A tournament selection method is used to retain the prompts of high-predictive-capability agents as "parent prompts." "Child prompts" are generated through cross-operations (merging logical modules from different types of agent prompts, such as value analysis logic and market sentiment assessment logic) and mutation operations (adjusting risk assessment thresholds, adding or deleting industry analysis dimensions, etc.).

[0104] Through the aforementioned evolutionary learning mechanism, electronic devices can achieve adaptive optimization and generational evolution of the decision-making logic of intelligent agents, continuously improve the accuracy of analysis and the stability of strategies in dynamic market environments, and achieve a fundamental breakthrough from static decision-making to dynamic evolutionary decision-making.

[0105] To address the shortcomings of existing technologies, this application aims to provide a multi-agent dynamic optimization financial trading system based on an evolutionary mechanism. This system achieves adaptive model evolution, accurate contribution assessment, and professional decision simulation, thereby significantly improving the system's decision-making performance and intelligence level in complex financial market environments. I. Construct a dynamic optimization mechanism to enhance the system's self-adaptation and evolution capabilities: Existing technologies generally lack dynamic optimization mechanisms for models or intelligent agents. Once a model is set, its decision-making logic remains fixed and it is difficult to continuously evolve according to market changes and its own performance, resulting in insufficient system adaptability.

[0106] This application introduces an evolutionary learning optimization module. Based on a comparison of the predictive capabilities of each agent with the real market, for analytical agents with low predictive capabilities, the module dynamically optimizes the prompt word parameters and policy weights of the analytical agent using an evolutionary learning algorithm. Through operations such as selection, crossover, and mutation, this module adaptively adjusts the analytical agent's analytical thinking and decision-making path, enabling the analytical agent to continuously evolve with changes in the market environment, continuously optimize its policy structure, and achieve dynamic self-evolution and continuous improvement of its adaptability at the system level.

[0107] II. Accurately assess agent contribution and optimize resource allocation and system performance: In existing multi-agent systems, the actual contribution of each analytical agent is difficult to quantify. The system cannot clearly identify high-value and inefficient analytical agents, resulting in unreasonable resource allocation and limited system performance optimization.

[0108] This application proposes a contribution analysis mechanism based on comparing the prediction results of each analytical agent with the real market: by extracting the trading decision-making process of each analytical agent under the same market conditions and comparing its decision results with the real market trend, the trading accuracy, risk-reward ratio, and relative contribution of the analytical agent are calculated. Based on this, the system generates quantitative performance indicators for each analytical agent, clarifying its role in the overall decision-making process.

[0109] Based on the evaluation results, the system can dynamically adjust the agent weights, prioritizing the enhancement of the capabilities of high-performing analytical agents while improving or eliminating poorly performing analytical agents, thereby achieving efficient resource allocation and continuous optimization of system performance.

[0110] Third, simulate professional investment thinking to improve the professionalism of decision-making: Existing technologies have significant shortcomings in handling complex financial knowledge and simulating the decision-making thinking of human investment experts, making it difficult to achieve truly professional and intelligent investment.

[0111] This application's embodiments introduce multiple types of professional role agents, such as investment master agent agents and professional analysis agent agents, and utilize large language models to perform knowledge modeling and thinking simulation for different investment styles and analysis systems.

[0112] For example, investment master agents can embody value investing principles, making judgments based on long-term returns and fundamental analysis of companies. The collaborative decision-making mechanism of multi-type analytical agents enables the system to integrate judgments across multiple levels, including macroeconomic policy analysis, fundamental research, and technical forecasting. This more closely resembles the comprehensive thinking logic of human experts, significantly improving the system's professional decision-making level and explanatory power.

[0113] Based on this, the dynamic optimization method for the model provided in this application has the following advantages: 1. Significantly improved system adaptability and self-evolution capability: This application's embodiment introduces an evolutionary learning-based agent dynamic optimization mechanism, enabling each analytical agent to continuously adjust its decision-making logic based on its historical performance, market feedback, and risk preferences. Compared to traditional static multi-model integration systems (such as fixed-weight voting mechanisms), this application's embodiment can maintain stable decision-making performance in dynamic environments such as market fluctuations, policy adjustments, and industry rotation.

[0114] 2. Multi-agent collaborative decision-making enhances the depth and robustness of information fusion: The heterogeneous collaborative architecture of the investment master agent and the professional analysis agent designed in this application enables the system to simultaneously cover multi-dimensional information such as macro policies, industry trends, technical indicators, and market sentiment. Through the weighted voting and confidence calibration mechanism of the signal aggregation module, the system effectively reduces the risk of misjudgment caused by the bias of a single model.

[0115] 3. The agent contribution evaluation mechanism enhances system transparency and controllability: Through a parallel transaction evaluation mechanism, the embodiments of this application can accurately quantify the contribution of each analytical agent in the decision-making process, supporting dynamic adjustment of model weights. This mechanism enables the system to automatically identify low-performing analytical agents and reduce their weights, while retaining high-performing analytical agents.

[0116] Example 5: Based on the same inventive concept, embodiments of this application provide a dynamic optimization device for a model. Figure 4 A schematic diagram of a dynamic optimization device for a model provided in this application embodiment is shown below. Figure 4 The device includes: The processing module 401 is configured to, if it receives an instruction to perform transaction decision analysis on the target product, use each analysis agent to determine the transaction decision of each analysis agent for the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent; The optimization module 402 is used to obtain the daily market trend of the target product and use an optimization agent to update the weight of each analysis agent based on whether the transaction decision corresponding to each analysis agent satisfies the daily market trend; wherein, the weight indicates the importance of the transaction decision generated by each analysis agent when the decision agent determines the final transaction decision based on the transaction decision generated by each analysis agent.

[0117] In one possible implementation, the optimization module 402 is further configured to: If the current date is the end date of this optimization cycle, after updating the weights corresponding to each analysis agent, obtain the updated weights of each analysis agent. If there exists a first analysis agent with an updated weight less than a preset weight threshold, then a second analysis agent is constructed and the second analysis agent is used to replace the first analysis agent. The updated weights corresponding to the first analysis agent are used as the weights corresponding to the second analysis agent in the next optimization cycle.

[0118] In one possible implementation, the optimization module 402 is further configured to: A semantic adjustment model is used to perform semantic fusion on the second prompt words corresponding to any two third analysis agent intelligent bodies other than the first analysis agent intelligent body and the first prompt word corresponding to the first analysis agent intelligent body to obtain the third prompt word; The third prompt word will be used as the prompt word for the second analysis agent in the next optimization cycle.

[0119] In one possible implementation, the processing module 401 is specifically used for: For each analytical agent, the identifier of the target product and the corresponding prompt word are input into the analytical agent. After receiving the prompt word and the identifier, the analytical agent retrieves the information corresponding to the target product with the identifier from the pre-configured database corresponding to the analytical agent, and generates the transaction decision based on the information and the prompt word.

[0120] In one possible implementation, the optimization module 402 is specifically used for: For each analytical agent, the analytical accuracy of the agent in the current optimization cycle is updated based on whether the trading decision corresponding to the agent satisfies the market trend of the day. The optimized agent is used to update the weights of each analytical agent based on the transaction decisions corresponding to each analytical agent, the market trend of the day, and the analytical accuracy of each analytical agent in the current optimization cycle.

[0121] In one possible implementation, the transaction decision includes a forecast of the target product's market trend and a corresponding analysis report; The transaction decisions made by the analytical agent satisfy the market trend of the day, including: The predicted market trend carried in the transaction decisions of the analytical agent is consistent with the market trend of the day.

[0122] In one possible implementation, each analytical agent includes at least one of an investment guru agent, a news analysis agent, a fundamental analysis agent, and a social media analysis agent.

[0123] Example 6: Based on the same inventive concept, embodiments of this application provide an electronic device that can implement the steps of the dynamic optimization method for the model described above. Figure 5 This application provides a schematic diagram of an electronic device structure, such as... Figure 5 As shown, it includes: processor 501, communication interface 502, memory 503 and communication bus 504, wherein processor 501, communication interface 502 and memory 503 communicate with each other through communication bus 504. The memory 503 stores a computer program. When the program is executed by the processor 501, the processor 501 performs the following steps: If an instruction to perform transaction decision analysis on the target product is received, then each analysis agent will determine the transaction decision of the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent. The daily market trend of the target product is obtained, and an optimized agent is used to update the weight of each analytical agent based on whether the transaction decision corresponding to each analytical agent satisfies the daily market trend. The weight indicates the importance of each transaction decision generated by each analytical agent when the decision agent determines the final transaction decision based on the transaction decisions generated by each analytical agent.

[0124] In one possible implementation, the method further includes: If the current date is the end date of this optimization cycle, after updating the weights corresponding to each analysis agent, obtain the updated weights of each analysis agent. If there exists a first analysis agent with an updated weight less than a preset weight threshold, then a second analysis agent is constructed and the second analysis agent is used to replace the first analysis agent. The updated weights corresponding to the first analysis agent are used as the weights corresponding to the second analysis agent in the next optimization cycle.

[0125] In one possible implementation, the method further includes: A semantic adjustment model is used to perform semantic fusion on the second prompt words corresponding to any two third analysis agent intelligent bodies other than the first analysis agent intelligent body and the first prompt word corresponding to the first analysis agent intelligent body to obtain the third prompt word; The third prompt word will be used as the prompt word for the second analysis agent in the next optimization cycle.

[0126] In one possible implementation, the step of using each analytical agent to determine the transaction decision of each analytical agent for the target product based on the identifier of the target product and the corresponding prompt words of each analytical agent includes: For each analytical agent, the identifier of the target product and the corresponding prompt word are input into the analytical agent. After receiving the prompt word and the identifier, the analytical agent retrieves the information corresponding to the target product with the identifier from the pre-configured database corresponding to the analytical agent, and generates the transaction decision based on the information and the prompt word.

[0127] In one possible implementation, the step of employing optimized agent intelligence, and updating the weights of each analytical agent intelligence based on whether the trading decisions corresponding to each agent intelligence conform to the market trend of the day, includes: For each analytical agent, the analytical accuracy of the agent in the current optimization cycle is updated based on whether the trading decision corresponding to the agent satisfies the market trend of the day. The optimized agent is used to update the weights of each analytical agent based on the transaction decisions corresponding to each analytical agent, the market trend of the day, and the analytical accuracy of each analytical agent in the current optimization cycle.

[0128] In one possible implementation, the transaction decision includes a forecast of the target product's market trend and a corresponding analysis report; The transaction decisions made by the analytical agent satisfy the market trend of the day, including: The predicted market trend carried in the transaction decisions of the analytical agent is consistent with the market trend of the day.

[0129] In one possible implementation, each analytical agent includes at least one of an investment guru agent, a news analysis agent, a fundamental analysis agent, and a social media analysis agent.

[0130] Since the principle of the above-mentioned electronic device in solving the problem is similar to the dynamic optimization method of the model, the implementation of the above-mentioned electronic device can be found in the embodiments of the method, and repeated parts will not be described again.

[0131] The communication bus mentioned in the aforementioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used in the figure, but this does not indicate that there is only one bus or one type of bus. Communication interface 502 is used for communication between the aforementioned electronic device and other devices. The memory can include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0132] The processors mentioned above can be general-purpose processors, including central processing units, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0133] Example 7: Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a processor. When the program runs on the processor, it causes the processor to perform the following steps: If an instruction to perform transaction decision analysis on the target product is received, then each analysis agent will determine the transaction decision of the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent. The daily market trend of the target product is obtained, and an optimized agent is used to update the weight of each analytical agent based on whether the transaction decision corresponding to each analytical agent satisfies the daily market trend. The weight indicates the importance of each transaction decision generated by each analytical agent when the decision agent determines the final transaction decision based on the transaction decisions generated by each analytical agent.

[0134] In one possible implementation, the method further includes: If the current date is the end date of this optimization cycle, after updating the weights corresponding to each analysis agent, obtain the updated weights of each analysis agent. If there exists a first analysis agent with an updated weight less than a preset weight threshold, then a second analysis agent is constructed and the second analysis agent is used to replace the first analysis agent. The updated weights corresponding to the first analysis agent are used as the weights corresponding to the second analysis agent in the next optimization cycle.

[0135] In one possible implementation, the method further includes: A semantic adjustment model is used to perform semantic fusion on the second prompt words corresponding to any two third analysis agent intelligent bodies other than the first analysis agent intelligent body and the first prompt word corresponding to the first analysis agent intelligent body to obtain the third prompt word; The third prompt word will be used as the prompt word for the second analysis agent in the next optimization cycle.

[0136] In one possible implementation, the step of using each analytical agent to determine the transaction decision of each analytical agent for the target product based on the identifier of the target product and the corresponding prompt words of each analytical agent includes: For each analytical agent, the identifier of the target product and the corresponding prompt word are input into the analytical agent. After receiving the prompt word and the identifier, the analytical agent retrieves the information corresponding to the target product with the identifier from the pre-configured database corresponding to the analytical agent, and generates the transaction decision based on the information and the prompt word.

[0137] In one possible implementation, the step of employing optimized agent intelligence, and updating the weights of each analytical agent intelligence based on whether the trading decisions corresponding to each agent intelligence conform to the market trend of the day, includes: For each analytical agent, the analytical accuracy of the agent in the current optimization cycle is updated based on whether the trading decision corresponding to the agent satisfies the market trend of the day. The optimized agent is used to update the weights of each analytical agent based on the transaction decisions corresponding to each analytical agent, the market trend of the day, and the analytical accuracy of each analytical agent in the current optimization cycle.

[0138] In one possible implementation, the transaction decision includes a forecast of the target product's market trend and a corresponding analysis report; The transaction decisions made by the analytical agent satisfy the market trend of the day, including: The predicted market trend carried in the transaction decisions of the analytical agent is consistent with the market trend of the day.

[0139] In one possible implementation, each analytical agent includes at least one of an investment guru agent, a news analysis agent, a fundamental analysis agent, and a social media analysis agent.

[0140] Since the principle of solving the problem using the above-mentioned computer-readable storage medium is similar to the dynamic optimization method of the model, the implementation of the above-mentioned computer-readable storage medium can be found in the implementation of the method, and the repeated parts will not be described again.

[0141] Example 10: Based on the same inventive concept, this application also provides a computer program product, which includes computer program code. When the computer program code is run on a computer, it causes the computer to execute the dynamic optimization method of any of the models discussed above. Since the principle of the above computer program product in solving the problem is similar to the dynamic optimization method of the model, the implementation of the above computer program product can refer to the implementation of the method, and the repeated parts will not be described again.

[0142] In this embodiment, the electronic device compares the transaction decisions of the analysis agent for the target product with the daily market trend of the target product, and uses an evolutionary learning algorithm to dynamically optimize the weight of each analysis agent, so that the analysis agent can continuously evolve with changes in the market environment, continuously optimize the strategy structure, and achieve dynamic self-evolution and continuous improvement of adaptability at the system level.

[0143] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0144] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0145] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0147] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A dynamic optimization method for a model, characterized in that, The method includes: If an instruction to perform transaction decision analysis on the target product is received, then each analysis agent will determine the transaction decision of the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent. The daily market trend of the target product is obtained, and an optimized agent is used to update the weight of each analytical agent based on whether the transaction decision corresponding to each analytical agent satisfies the daily market trend. The weight indicates the importance of each transaction decision generated by each analytical agent when the decision agent determines the final transaction decision based on the transaction decisions generated by each analytical agent.

2. The method according to claim 1, characterized in that, The method further includes: If the current date is the end date of this optimization cycle, after updating the weights corresponding to each analysis agent, obtain the updated weights of each analysis agent. If there exists a first analysis agent with an updated weight less than a preset weight threshold, then a second analysis agent is constructed and the second analysis agent is used to replace the first analysis agent. The updated weights corresponding to the first analysis agent are used as the weights corresponding to the second analysis agent in the next optimization cycle.

3. The method according to claim 2, characterized in that, The method further includes: A semantic adjustment model is used to perform semantic fusion on the second prompt words corresponding to any two third analysis agent intelligent bodies other than the first analysis agent intelligent body and the first prompt word corresponding to the first analysis agent intelligent body to obtain the third prompt word; The third prompt word will be used as the prompt word for the second analysis agent in the next optimization cycle.

4. The method according to claim 1, characterized in that, The step of using each analytical agent to determine the transaction decision for the target product based on the identifier of the target product and the corresponding prompt words of each analytical agent includes: For each analytical agent, the identifier of the target product and the corresponding prompt word are input into the analytical agent. After receiving the prompt word and the identifier, the analytical agent retrieves the information corresponding to the target product with the identifier from the pre-configured database corresponding to the analytical agent, and generates the transaction decision based on the information and the prompt word.

5. The method according to claim 1, characterized in that, The step of employing optimized agent intelligence, and updating the weights of each analytical agent intelligence based on whether the trading decisions corresponding to each agent intelligence conform to the market trend of the day, includes: For each analytical agent, the analytical accuracy of the agent in the current optimization cycle is updated based on whether the trading decision corresponding to the agent satisfies the market trend of the day. The optimized agent is used to update the weights of each analytical agent based on the transaction decisions corresponding to each analytical agent, the market trend of the day, and the analytical accuracy of each analytical agent in the current optimization cycle.

6. The method according to claim 5, characterized in that, The transaction decision includes a forecast of market trends for the target product and corresponding analysis reports; The transaction decisions made by the analytical agent satisfy the market trend of the day, including: The predicted market trend carried in the transaction decisions of the analytical agent is consistent with the market trend of the day.

7. The method according to any one of claims 1-6, characterized in that, Each analytical agent includes at least one of an investment guru agent, a news analysis agent, a fundamental analysis agent, and a social media analysis agent.

8. A dynamic optimization device for a model, characterized in that, The device includes: The processing module is configured to, upon receiving an instruction to perform transaction decision analysis on a target product, use each analysis agent to determine the transaction decision of each analysis agent for the target product based on the identifier of the target product and the prompt words corresponding to each analysis agent. An optimization module is used to obtain the daily market trend of the target product and, using an optimization agent, update the weight of each analysis agent based on whether the transaction decision corresponding to each analysis agent satisfies the daily market trend. The weight indicates the importance of each transaction decision generated by the analysis agent when the decision agent determines the final transaction decision based on the transaction decisions generated by each analysis agent.

9. An electronic device, characterized in that, The electronic device includes a processor that executes a computer program stored in a memory to implement the steps of the dynamic optimization method for the model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the dynamic optimization method for any of the models described in claims 1-7.

Images