Project research method and system based on agent reflection and debate

Abstract

The invention provides a throwing and research method and system based on agent reflection and debate, and the method comprises the following steps: S1, simulating an opposition field through an agent: generating viewpoints of'square 'and'reverse' through presetting an agent role, and forcing the positive and negative agents to sound; s2, searching for an expected difference: constructing a heterogeneous data fusion engine, integrating data sources, and discovering non-dominant association by applying a causal inference model; s3, decision weight decentralization: dynamically adjusting member viewpoint weights according to the historical win rate of each agent, wherein a leader viewpoint is only used as one of inputs; and S4, human traders participate in, wherein human researchers check the analysis logic chain of the intelligent agent, and wrong logic is eliminated. According to the method, a multi-agent debate mechanism is introduced, and multi-view debate is forcibly carried out, so that the problems in a traditional investment and research mode are solved, investors are helped to mine market opportunities, and risks are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of financial technology and generative AI technology, in particular to a research and investment method and system based on agent introspection and debate. BACKGROUND

[0002] In the traditional financial research and investment "morning meeting" mode, there are many drawbacks. According to a 2022 study in the Journal of Financial Economics, the decision-making error rate of traditional research and investment teams is significantly positively correlated with the degree of opinion convergence (r=0.67), and when the degree of opinion divergence in the team is less than 30%, the probability of missing major risks increases by 42%. The existing financial research and investment methods and systems mainly have the following defects:

[0003] (1) Opinion convergence: When most members of the team hold a bearish view, a few members who hold a bullish view and lack data support or are afraid of being questioned may choose to remain silent, resulting in a single view.

[0004] (2) Echo chamber effect: In the Internet era, investors tend to browse information that aligns with their own views, further strengthening their views and making it difficult to access different perspectives.

[0005] (3) Suppression of innovative thinking by authority: If the leader clearly expresses a bearish view, team members may abandon in-depth exploration of potential opportunities due to hierarchical pressure, limiting innovative thinking.

[0006] (4) Single perspective leads to risk exposure: When team opinions are too consistent, potential risks such as transaction congestion in the government bond market may be overlooked. SUMMARY

[0007] To address the shortcomings of existing technology, the present application proposes a research and investment method and system based on agent introspection and debate, which introduces a multi-agent debate mechanism to force multi-perspective debate, addressing the above-mentioned problems in traditional research and investment models and helping investors uncover market opportunities and reduce risks.

[0008] To achieve the above technical solution, the present application provides a research and investment method based on agent introspection and debate, comprising the following steps:

[0009] S1, simulate opposing positions through agents: generate "pro" and "con" views through pre-set agent roles, and force the pro and con agents to speak;

[0010] S2, find the expected difference: build a heterogeneous data fusion engine, integrate data sources, and apply a causal inference model to discover non-explicit correlations;

[0011] S3, Decentralized decision weight: dynamically adjust the member view weight according to the historical win rate of each agent, the leader view is only one of the inputs, design a double-layer Bayesian updating mechanism, the short-term weight is based on the prediction accuracy in the past 30 days, the long-term weight is based on the Sharpe ratio adjustment factor, and the initial weight of the leader view is set to 0.2 and floats ±0.15 with the overall win rate of the team;

[0012] S4, Human trader participation: human researchers check the analysis logic chain of the agent, and exclude the wrong logic.

[0013] Preferably, in step S1, the agent role adopts a generative adversarial network framework, sets a generator and a discriminator, dynamically optimizes the debate strategy through reinforcement learning, and the debate depth reaches the NLP semantic understanding level.

[0014] Preferably, in step S1, after starting the agent debate mechanism, preset the pro and anti agent roles, let them generate opposite views according to the collected data and preset algorithm, and debate, in the debate process, the agents constantly exchange information, dig data relationship and find expected difference.

[0015] Preferably, in step S1, the debate of the agent adopts a round design, which specifically includes:

[0016] (1) Point generation: T5 model generates initial views;

[0017] (2) Evidence retrieval: cross-modal retrieval based on ElasticSearch;

[0018] (3) Logic verification: formal argumentation framework ensures the validity of reasoning;

[0019] (4) Consensus reaching: Nash equilibrium solver balances the interests of multiple parties.

[0020] Preferably, in step S2, the pro and anti agents take the output of other research agents as input, use sufficient data sources, dig the data relationship missed in traditional investment research, support non-mainstream views, and find market alpha.

[0021] Preferably, in step S2, the agent analyzes a large amount of financial data, finds the potential relationship between different data, and provides a new perspective for investment decision-making.

[0022] The application also provides an investment research system based on agent reflection and debate, which is used to realize the above investment research method, and specifically includes:

[0023] Agent investment research module: including a plurality of agent modules, respectively responsible for different functions, setting corresponding algorithms and rules for each agent module, so that it can simulate different market roles and views;

[0024] Data collection and processing module: collect various financial data, and carry out cleaning and preprocessing, ensure the accuracy and availability of data;

[0025] Decision weight adjustment module: dynamically adjust the member opinion weight according to the historical winning rate of each agent, and the leading opinion is taken as one of the inputs to participate in the decision-making process with other agent opinions;

[0026] Heterogeneous data fusion module, integrating satellite remote sensing data, supply chain logistics data and social media text data;

[0027] Adversarial debate engine, using a generative adversarial network architecture, configured with a reinforcement learning driven dynamic strategy optimizer;

[0028] Interpretability guarantee module, realizing the graph structure visualization of decision logic chain and counterfactual reasoning detection.

[0029] Preferably, the decision weight adjustment module loads a Bayesian online learning algorithm with a forgetting factor, the forgetting factor λ=0.93, and the weight is updated at 23:00 every day.

[0030] Preferably, the agent investment research module specifically comprises:

[0031] Data layer: exchange trading data + fundamental data + financial data terminal API + alternative data crawler;

[0032] Computing layer: distributed graph database stores knowledge graph;

[0033] Display layer: interpretable AI visualization interface.

[0034] The investment research method and system based on agent introspection and debate provided by the application have the following advantages:

[0035] (1) Break the consensus of views: through the simulation of opposite positions by agents, forced multi-perspective debate is carried out, avoiding the situation that the majority opinion dominates in the traditional mode, so that various opinions can be fully expressed.

[0036] (2) Exploiting market opportunities: using agents to find expected differences and exploit the data relationships missed in traditional investment research can help discover potential investment opportunities in the market and obtain excess returns.

[0037] (3) Promote innovative thinking: decentralized decision-making weight reduces the suppression of authoritative thinking on innovation, encourages team members to think about problems from different angles, and stimulates innovative thinking.

[0038] (4) Reduce risk: multi-perspective analysis can help investors have a more comprehensive understanding of the market and discover potential risks, thereby reducing investment risk.

[0039] (5) Improve the scientificity of decision-making: The analysis logic of the agent is fully transparent, which facilitates human researchers to check and verify, making investment decisions more scientific and reasonable.

[0040] (6) Help human traders establish a more neutral and objective view: Human traders generally do not have personal emotions when judging the views of the agent, making them more receptive to views that contradict their own. Different views provided by the agent can broaden the human trader's perspective, helping them view the problem from multiple angles and make more neutral and objective investment decisions. For example, when a human trader originally holds an optimistic view of a market, the opposite view proposed by the agent can prompt them to reevaluate their judgment.

[0041] (7) Fully transparent process: The analysis logic chain of the agent is fully transparent, which facilitates human researchers to check and analyze the logic and eliminate incorrect logic. Human researchers can clearly understand the decision-making process and basis of the agent, and supervise and verify it. For example, after the agent gives a certain investment suggestion, the researcher can view the data and models used in the analysis process to judge its rationality. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of the present application.

[0043] Figure 2 is an argumentation process state machine diagram in the present application.

[0044] Figure 3 is a system architecture diagram of the present application.

[0045] Figure 4 is a weight dynamic adjustment algorithm pseudocode diagram in the present application.

[0046] Figure 5 is a visual interface example diagram in the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. All other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0048] Embodiment 1: A research and investment method based on agent reflection and argumentation.

[0049] Referring to Figure 1 and Figure 2 , a research and investment method based on agent reflection and argumentation specifically includes the following steps:

[0050] S1, agent simulation of opposite standpoints: by presetting agent roles, generate the viewpoints of "pro" and "anti", and force the pro and anti agents to speak. Adopt the framework of generative adversarial network (GAN), set the generator (pro) and discriminator (anti), dynamically optimize the debate strategy (PPO algorithm) through reinforcement learning, and debate to the depth of NLP semantic understanding level (BERT + graph neural network). In the research "morning meeting" scene, the agent can simulate different market viewpoints and break the traditional pattern of converging viewpoints. For example, when discussing the investment value of a stock, the pro agent can list the advantages and upside potential of the stock, while the anti agent can point out the risks it faces and the factors that may cause it to fall.

[0051] Among them, in the research "morning meeting", the agent debate mechanism is started. The pro and anti agent roles are preset, allowing them to generate opposing viewpoints based on collected data and preset algorithms, and debate. During the debate, agents constantly exchange information, dig into data relationships, and find expected differences. The debate of the agent adopts a round design (4 rounds of iteration, as shown in Figure 2 The specific contents include:

[0052] (1) Argument generation: T5 model generates initial viewpoints;

[0053] (2) Evidence retrieval: cross-modal retrieval (text + numerical + graph) based on ElasticSearch;

[0054] (3) Logic verification: formal argumentation framework (ASPIC+) ensures the validity of reasoning;

[0055] (4) Consensus reached: Nash equilibrium solver balances the interests of multiple parties.

[0056] S2, find the expected difference: build a heterogeneous data fusion engine, integrate alternative data sources (satellite images, supply chain logistics data, social media sentiment index), and apply causal inference models (DoWhy framework) to discover non-obvious correlations. The pro and anti agents take the output of other research agents as input, use sufficient data sources, and dig into the data relationships overlooked in traditional research to support non-mainstream viewpoints and find market alpha (excess returns adjusted by the Fama-French three-factor model). Agents can analyze vast amounts of financial data, discover potential relationships between different data, and provide new perspectives for investment decisions. For example, by analyzing macroeconomic data, industry data, and company financial data, etc., they can dig out investment opportunities overlooked by the market.

[0057] S3, Decentralized decision weight: The leader's view is only one of the inputs, and the agent investment research system dynamically adjusts the member's view weight based on the historical win rate of each agent. A double-layer Bayesian update mechanism is designed, the short-term weight is based on the past 30-day prediction accuracy (MAPE <15%), the long-term weight includes the Sharpe ratio adjustment factor, and the initial weight of the leader's view is set to 0.2, which will float ±0.15 with the overall team win rate. This can avoid the excessive influence of authority on decision-making, and make the decision more scientific and reasonable. The system will record the historical performance of each agent, and adjust its weight in decision-making according to its prediction accuracy and success rate. For example, an agent with a high historical win rate (rolling window prediction accuracy, 60-day window, Holt-Winters smoothing) will have a relatively high weight in decision-making.

[0058] S4, Human trader participation: Human researchers check the analysis logic chain of agents, and exclude incorrect logic. Human traders generally do not have personal emotions when judging the views of agents, which makes them more accepting of views that contradict their own, helping them to establish more neutral and objective views. Different views provided by agents can broaden the horizons of human traders, helping them to view problems from multiple angles and make more neutral and objective investment decisions. For example, when a human trader originally holds an optimistic attitude towards a certain market, the opposite view proposed by an agent can prompt them to reevaluate their judgment.

[0059] Through the above investment research method, the multi-agent debate mechanism is introduced, and multi-perspective debate is forced to solve the above-mentioned problems in the traditional investment research mode, helping investors to uncover market opportunities and reduce risks. Moreover, the process is fully transparent, and the analysis logic chain of the agent is fully transparent, making it easy for human researchers to check and analyze the logic and exclude incorrect logic. Human researchers can clearly understand the decision-making process and basis of the agent, and supervise and verify it. For example, after an agent gives a certain investment suggestion, the researcher can view the data and models used in the analysis process to judge its rationality.

[0060] Through the above investment research method, the following progress has been made:

[0061] (1) Break the view convergence: By simulating opposing positions through agents, forced multi-perspective debate avoids the situation where the majority view dominates in the traditional mode, allowing various views to be fully expressed.

[0062] (2) Uncover market opportunities: Use agents to find expected differences and uncover data relationships missed in traditional investment research, which helps to discover potential investment opportunities in the market and achieve excess returns.

[0063] (3) Promote innovative thinking: Decentralized decision-making weight reduces the suppression of authoritative innovation thinking, encourages team members to think from different perspectives, and stimulates innovative thinking.

[0064] (4) Reduce risk: Multi-perspective analysis can help investors understand the market more comprehensively and discover potential risks, thereby reducing investment risk.

[0065] (5) Improve decision-making science: The analysis logic of the agent is fully transparent, making it easy for human researchers to check and verify, making investment decisions more scientific and reasonable.

[0066] Embodiment 2: A research and investment system based on agent reflection and debate.

[0067] Referring to Figure 3 and Figure 5 , a research and investment system based on agent reflection and debate, specifically includes:

[0068] (1) Agent research and investment module: including multiple agent modules, respectively responsible for different functions such as opinion generation, data analysis, decision weight calculation, etc. Set appropriate algorithms and rules for each agent module to simulate different market roles and opinions.

[0069] The agent research and investment module specifically includes:

[0070] Data layer: exchange trading data + fundamental data + financial data terminal API + alternative data crawler;

[0071] Calculation layer: distributed graph database (Neo4j) stores knowledge graph;

[0072] Display layer: interpretable AI visualization interface (LIME / SHAP value heat map).

[0073] (2) Data collection and processing module: collect various financial data, including macroeconomic data, industry data, company financial data, etc., and perform cleaning and preprocessing to ensure data accuracy and usability.

[0074] (3) Decision weight adjustment module: the agent research and investment system dynamically adjusts the member opinion weight according to the historical win rate of each agent. The leader's opinion is one of the inputs, participating in the decision-making process with other agent opinions. The decision weight adjustment module loads a Bayesian online learning algorithm with a forgetting factor (such as Figure 4 shown), the forgetting factor λ = 0.93, and the weight is updated at 23:00 every day.

[0075] (4) Heterogeneous data fusion module, integrating satellite remote sensing data, supply chain logistics data, and social media text data;

[0076] (5) The adversarial debate engine, configured with a generative adversarial network architecture, is equipped with a reinforcement learning-driven dynamic policy optimizer;

[0077] (6) The interpretability guarantee module realizes the graph structure visualization of the decision logic chain and the counterfactual reasoning detection (refer to Figure 5 ).

[0078] In the Q3 2024 semiconductor industry judgment experiment, the pro-party intelligent agent recommends to increase holdings based on equipment import customs declaration data, and the anti-party warns of technical blockade risks through EDA software authorization agreement text analysis (NLP sentiment score -0.47). Finally, the system generates a 'neutral' rating, avoiding a subsequent 12.7% decline in the industry index. Backtesting (A-share data from 2018 to 2024) shows that the system reduces portfolio maximum drawdown by 37.2%, annualized volatility by 28.4%, and captures 19 excess return events (α>3%, p<0.05).

[0079] The above is a preferred embodiment of the present application, but the present application should not be limited to the disclosed content of the embodiment and the drawings, so any equivalent or modification made without departing from the disclosed spirit of the present application falls within the scope of protection of the present application.

Claims

1. An investment research method based on agent introspection and debate, characterized in that Comprise the following steps: S1, simulate the opposite stand by agent: generate the views of "pro" and "anti" through preset agent roles, and force the pro and anti agents to speak; S2, find the expected difference: build a heterogeneous data fusion engine, integrate data sources, and apply a causal inference model to discover non-explicit correlations; S3, decentralized decision weight: dynamically adjust the member view weight according to the historical win rate of each agent, the leader's view is only one of the inputs, design a double-layer Bayesian update mechanism, the short-term weight is based on the prediction accuracy in the past 30 days, the long-term weight is adjusted by the Sharpe ratio, and the initial weight of the leader's view is set to 0.2, which will fluctuate ±0.15 with the overall team win rate; S4, human traders participate: human researchers check the analysis logic chain of the agent, and exclude the wrong logic. 2.The investment research method based on agent introspection and debate according to claim 1, characterized in that, In step S1, the agent role adopts a generative adversarial network framework, sets a generator and a discriminator, and dynamically optimizes the debate strategy through reinforcement learning, with a debate depth reaching the NLP semantic understanding level. 3.The investment research method based on agent introspection and debate according to claim 1, characterized in that, In step S1, after starting the agent debate mechanism, preset the pro and anti agent roles, let them generate opposite views according to the collected data and preset algorithm, and debate, in the process of debate, the agents constantly exchange information, dig data relationship, and find the expected difference.

4. The agent-based reflection and debate based investment research method according to claim 3, characterized in that, In step S1, the debate of the agent adopts a round design, which specifically includes: (1) argument generation: T5 model generates initial views; (2) evidence retrieval: cross-modal retrieval based on ElasticSearch; (3) logic verification: formal argument framework ensures the validity of reasoning; (4) consensus reached: Nash equilibrium solver balances the interests of multiple parties. 5.The investment research method based on agent introspection and debate according to claim 1, characterized in that, In step S2, the pro and anti agents take the output of other research agents as input, use sufficient data sources, dig the data relationship missed in traditional investment research, support non-mainstream views, and find market alpha. 6.The investment research method based on agent introspection and debate according to claim 1, characterized in that, In step S2, the agent analyzes a large amount of financial data, discovers the potential relationship between different data, and provides a new perspective for investment decision-making.

7. An investment research system based on agent introspection and debate, characterized in that, The investment research method for realizing any one of claims 1-6, specifically comprises: Agent investment research module: including multiple agent modules, respectively responsible for different functions, setting corresponding algorithms and rules for each agent module, so that it can simulate different market roles and views; Data collection and processing module: collect various financial data and perform cleaning and preprocessing to ensure data accuracy and usability; Decision weight adjustment module: dynamically adjust the member view weight according to the historical win rate of each agent, and the leader's view as one of the inputs participates in the decision-making process with other agent views; Heterogeneous data fusion module, integrating satellite remote sensing data, supply chain logistics data and social media text data; Adversarial debate engine, using a generative adversarial network architecture, equipped with a reinforcement learning driven dynamic strategy optimizer; Interpretability guarantee module, realizing the graph structure visualization and counterfactual reasoning detection of decision logic chain. 8.The investment research system based on agent introspection and debate according to claim 7, characterized in that, The decision weight adjustment module is loaded with a Bayesian online learning algorithm with a forgetting factor, and the forgetting factor λ=0.93, and the weight is updated at 23:00 every day. 9.The investment research system based on agent reflection and debate according to claim 7, characterized in that, The intelligent agent investment research module specifically comprises: Data layer: exchange trading data + fundamental data + financial data terminal API + alternative data crawler; Computing layer: distributed graph database stores knowledge graph; Display layer: interpretable AI visualization interface.

Images