Intelligent transaction decision-making method and system based on hierarchical multi-round confrontation debate

Abstract

The invention discloses an intelligent transaction decision-making method and system based on hierarchical multi-round confrontation debate, and is suitable for stock and other financial asset transaction scenes. According to the method, a hierarchical multi-round adversarial debate mechanism is constructed, so that efficient, low-delay and interpretable intelligent transaction decision is realized, and the limitation of the prior art on deep reasoning, delay control and decision transparency is solved. The core of the method is to construct a complete process of multi-agent confrontation debate, dynamic reputation evaluation and transparent decision making, divide multi-role agents into an argument generation group and a reputation group, and execute R rounds of deep debate. In the debate process, the agent reputation is updated by adopting Bayesian reputation, and a transaction signal is generated within millisecond-level delay in combination with a Soft-Borda dynamic voting mechanism. And meanwhile, through a transparent decision-making link of multiple rounds of debate logs and interpretation vectors, the requirements of financial supervision on model interpretability and real-time auditing are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and financial technology, in particular, to an intelligent transaction decision-making method and system based on hierarchical multi-round confrontation debate. BACKGROUND

[0002] In the field of financial transactions, with the continuous rise of market complexity and the growing demand for transaction speed, intelligent transaction decision-making technology has made significant progress in recent years, especially in algorithmic trading, quantitative analysis, etc., with the help of artificial intelligence and machine learning algorithms. However, the current technology still shows weakness in dealing with certain key challenges. In the high-frequency trading scenario, there is a strict millisecond-level delay requirement for transaction decision-making speed. However, models that implement deep reasoning involve deep neural network models that simulate complex financial market dynamics and multi-factor linkage analysis, which often require a lot of time for data processing and complex calculations, making it difficult to complete decision output in a very short time and unable to meet the demand for immediacy in high-frequency trading.

[0003] The one-time voting mechanism exposes obvious drawbacks when dealing with conflicts of opinion in transaction decision-making. When transaction team members or agents have different judgments about market trends, simple one-time voting only makes decisions based on majority opinions, failing to fully exploit and utilize the information value contained in opposing opinions.

[0004] With the increasing strictness of the regulatory environment, transparency and compliance of financial transaction decision-making have become key requirements. Regulatory authorities expect transaction institutions to provide clear explanations of the transaction decision-making process and keep complete log records for supervision and review. However, many existing intelligent transaction decision-making models, especially some black-box models based on deep learning, have internal decision-making logic that is difficult to understand and explain, making it difficult to visually present how input data is transformed into the final transaction decision and providing detailed decision-making process logs.

[0005] Through the search of patent documents, it is found that the invention patent with publication number CN109376922A discloses a short-term transaction optimization management system and method based on big data. Through big data analysis, it predicts the yield rate, selects the most suitable stock for trading on the day, optimizes the control of the short-term trading process, and improves the short-term trading yield rate. At the same time, it tracks market transaction changes and adjusts and optimizes the voting situation according to market conditions. This patent makes a one-time decision and cannot resolve conflicts of opinion, involves a deep debate mechanism among multiple agents, and is difficult to deal with complex and changing market conditions. Moreover, the decision-making process is not transparent, making it difficult to meet the regulatory requirements for log records and explainability.

[0006] In summary, in view of the problems of the above-mentioned existing technology, it is a key task to develop an intelligent transaction decision-making method and system based on hierarchical multi-round confrontation debate. SUMMARY

[0007] In view of the defects in the prior art, the purpose of the present application is to provide an intelligent transaction decision-making method and system based on hierarchical multi-round confrontation debate.

[0008] According to the intelligent transaction decision-making method and system based on hierarchical multi-round confrontation debate provided by the present application, the following steps are included: Step S1, acquiring market data and constructing an input feature vector, initializing an agent set and dividing it into a point generation group and a refutation group, and assigning an initial credibility score to each agent; Step S2, performing R rounds of deep debate and dynamically adjusting the credibility score; Step S3, based on the adjusted credibility score, performing Soft-Borda dynamic voting and outputting a transaction signal.

[0009] Preferably, in step S1, the point generation group is used to analyze the input feature vector and propose an initial transaction view, and the initial transaction view includes buying, selling and watching; the refutation group is used to question the view proposed by the point generation group and find potential risks and logical loopholes.

[0010] Preferably, in step S1, the initial credibility score of each agent is set to an unbiased initial value.

[0011] Preferably, step S2 includes the following sub-steps: Step S2.1, view generation and refutation interaction in round r; Step S2.2, updating the credibility score of the agent based on the Bayesian mechanism; Step S2.3, constructing a round debate log and a credibility trajectory.

[0012] Preferably, in step S2.1, in each round of debate, the point generation group generates a set of transaction views in parallel , and the refutation group receives the set of transaction views , constructs a set of refutation points based on its own analysis model, historical data and debate information of previous rounds, and uses it to challenge or disprove the views of the point generation group ; evaluates the effectiveness of the points or refutations proposed by agent i in the current round r, and quantifies it as two key indicators: point-target consistency and point-market consistency , where A is the set of agents.

[0013] Preferably, in step S2.2, after each round of debate, the Bayesian updating mechanism is adopted to dynamically adjust the reputation score of the agent according to the performance of the agent in the round, and the reputation score of the agent i after the rth round is The updating formula is as follows:

[0014] Wherein: is the reputation score of the agent i in the last round; is the initial reputation score of the agent i; is the effectiveness likelihood of the argument or rebuttal of the agent i in the rth round, which is calculated by weighting the performance indicators;

[0015] Wherein: represents the argument-target consistency; represents the argument-market consistency; 、 respectively represent the experience weights of target consistency and historical consistency.

[0016] Preferably, in step S2.3, in each round of debate, a structured log is generated and stored to realize complete transparency and traceability of the decision-making process, and the log includes the argument and rebuttal texts in natural language format generated by each agent, the key indicators on which each reputation update is based and , the reputation scores before and after the update , and the SHAP-ExConsistency explanation vector for explaining the basis of the model decision.

[0017] Preferably, in step S3, after completing all R rounds of debate, the comprehensive score of each trading signal in all rounds is calculated based on the debate results of each round, and the comprehensive score is based on the Borda scoring method, that is, the ranking of each agent on the signal is considered comprehensively; at the same time, the dynamic temperature coefficient is calculated, the comprehensive score and the dynamic temperature coefficient are fused through the Soft-Borda dynamic voting formula, and the trading signal with the highest score is output.

[0018] Preferably, step S3 includes the following sub-steps: Step S3.1, calculating the dynamic temperature coefficient , the formula is:

[0019] =0.3； = ； is the realized volatility of the recent preset time; Step S3.2, using the Soft-Borda dynamic voting formula to calculate the final score of each trading signal:

[0020] wherein is the input The original numerical value of the function.

[0021] Output the trading signal with the highest score, and output the "neutral" signal when the score difference between the highest score and the second highest score signal is less than a preset threshold.

[0022] The application also provides an intelligent trading decision system based on hierarchical multi-round confrontation debate, comprising: Module M1, acquiring market data and constructing input feature vectors, initializing the agent set and dividing it into argument generation groups and refutation groups, and assigning each agent an initial credibility score; Module M2, performing R rounds of deep debate and dynamically adjusting the credibility score; Module M3, based on the adjusted credibility score, performing Soft-Borda dynamic voting and outputting a trading signal.

[0023] Compared with the prior art, the application has the following beneficial effects: 1. The application uses hierarchical multi-round confrontation debate protocol (H-DAP) and Bayesian credibility updating technology, which can fully eliminate the conflict of viewpoints of agents compared with traditional single-round voting, avoid one-sided decision-making and overfitting, and improve the accuracy of trading signals.

[0024] 2. The application uses Soft-Borda dynamic voting and temperature self-adaptive mechanism to overcome the defect of insufficient sensitivity of fixed weight voting in high volatility market, effectively reducing risk exposure and maximum drawdown.

[0025] 3. The application uses millisecond-level parallel debate and log explanation chain technology to achieve the balance between deep reasoning and low latency in high-frequency trading and high interpretability of the model, and meets the regulatory transparency requirements.

[0026] 4. The application uses multi-modal data fusion technology to integrate multi-source information such as market and news, eliminate the information blind area of single-source data, and enhance the decision-making integrity and transparency. BRIEF DESCRIPTION OF DRAWINGS

[0027] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments thereof, when read in connection with the following accompanying drawings: Figure 1 A flow chart of an intelligent transaction decision-making method based on hierarchical multi-round confrontation debate according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] The present application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present application. These are within the scope of protection of the present application.

[0029] The present application discloses an intelligent transaction decision-making method and system based on hierarchical multi-round confrontation debate, which is suitable for stock and other financial asset trading scenarios. The method realizes efficient, low-latency and interpretable intelligent transaction decision-making by constructing a hierarchical multi-round confrontation debate mechanism, and solves the limitations of existing technologies in deep reasoning, delay control and decision transparency. The core of the method is to construct a complete process of multi-agent confrontation debate, dynamic reputation evaluation and transparent decision-making, which divides the multi-role agent into a thesis generation group and a refutation group, and executes R rounds of deep debate. In the debate process, the agent reputation is updated using the Bayesian reputation update, and the Soft-Borda dynamic voting mechanism is combined to generate a transaction signal within a millisecond delay. At the same time, through the transparent decision link of multi-round debate log and explanation vector, the requirements of financial supervision on model interpretability and real-time audit are met.

[0030] Embodiment 1 Figure 1 A flow chart of an intelligent transaction decision-making method based on hierarchical multi-round confrontation debate according to an embodiment of the present application.

[0031] As shown in Figure 1 , the present embodiment 1 provides an intelligent transaction decision-making method based on hierarchical multi-round confrontation debate, which includes the following steps: Step S1, acquire market data and construct input feature vectors, initialize agent set and divide into thesis generation group and refutation group, and assign initial reputation score to each agent.

[0032] Specifically, the market data includes market quotes, company announcements, and social public opinions. The argument generation group is configured to analyze the input feature vector and propose initial trading views, including buy, sell, and hold. The rebuttal group is configured to challenge the views proposed by the argument generation group and find potential risks and logical flaws. The initial credibility score of each agent is set to an unbiased initial value. In this embodiment, the initial credibility score of all agents is 0.5.

[0033] In step S2, R rounds of deep debate are performed, and the credibility score is dynamically adjusted.

[0034] In this embodiment, R is an integer greater than 3, and the single-round delay of R rounds of debate does not exceed 50 ms.

[0035] Specifically, step S2 includes the following sub-steps: In step S2.1, view generation and rebuttal interaction in round r is performed, where r is an integer from 1 to R.

[0036] Specifically, in each round of debate, the argument generation group generates a set of trading views in parallel The rebuttal group receives the set of trading views After receiving the set of trading views, the rebuttal group constructs a set of rebuttal arguments based on its own analysis model, historical data, and debate information from previous rounds, for challenging or falsifying the views of the argument generation group Evaluate the effectiveness of the arguments or rebuttals of agent i in the current round r, quantified as two key indicators: argument-target consistency and argument-market fit , where A is the set of agents.

[0037] Argument-target consistency is used to measure the consistency of the agent's view with the preset target function in direction, and the preset target function is to maximize expected return.

[0038] Argument-market fit is used to measure the degree of fitting or correlation between the agent's view and the historical real market trend.

[0039] In step S2.2, the credibility score of the agent is updated based on the Bayesian mechanism.

[0040] Specifically, after each round of debate, the Bayesian update mechanism is adopted to dynamically adjust the credibility score according to the performance of the agent in this round (i.e., argument effectiveness), and the update formula of the credibility score of agent i after the rth round is as follows:

[0041] Where: ​ is the reputation score of agent i in the last round; is the initial reputation score of agent i.

[0042] is the effectiveness likelihood of the argument or rebuttal of agent i in the rth round, calculated by weighting the performance indicators;

[0043] wherein: represents the argument-target consistency (value range).

[0044] represents the argument-market consistency (value range).

[0045] 、 respectively represent the experience weights of target consistency and historical consistency, which are configurable parameters for balancing forward-looking and historical experience. In this embodiment, = 0.6, = 0.4.

[0046] Step S2.3, constructing round-by-round debate logs and reputation trajectories.

[0047] Specifically, in each round of debate, structured logs are generated and stored to achieve complete transparency and traceability of the decision-making process, including the argument and rebuttal texts in natural language format generated by each agent, the key indicators on which each reputation update is based and the reputation scores before and after the update and the SHAP-ExConsistency explanation vector for explaining the basis of the model decision, ensuring that the requirements of financial supervision on model explainability are met.

[0048] Step S3, based on the adjusted reputation scores, performing Soft-Borda dynamic voting and outputting trading signals.

[0049] Specifically, after completing all R rounds of debate, the comprehensive scores of each trading signal (buy, sell, and hold) in all rounds are calculated based on the debate results of each round. The comprehensive score is based on the Borda scoring method , that is, the ranking of each agent on the signal is considered comprehensively; at the same time, the dynamic temperature coefficient is calculated, and the comprehensive score and the dynamic temperature coefficient are fused through the Soft-Borda dynamic voting formula to output the trading signal with the highest score.

[0050] In this embodiment, step S3 includes the following sub-steps: Step S3.1, calculating the dynamic temperature coefficient , the formula is:

[0051] =0.3； = ； is the realized volatility in the last 60 minutes, the greater the market volatility, the smaller, so that the voting results are more sensitive to high-score signals.

[0052] Step S3.2, using the Soft-Borda dynamic voting formula to calculate the final score of each trading signal:

[0053] wherein is the original numerical value of the input function.

[0054] The trading signal with the highest output score is output. To control risk, when the score difference between the highest score and the second highest score signal is less than a preset threshold, it indicates that the market consensus is not strong, and an "observation" signal is output.

[0055] In this embodiment, the preset threshold is 0.05.

[0056] Embodiment 2: The application also provides an intelligent trading decision system based on hierarchical multi-round confrontation debate. The intelligent trading decision system based on hierarchical multi-round confrontation debate can be realized by executing the process steps of the intelligent trading decision method based on hierarchical multi-round confrontation debate, that is, those skilled in the art can understand that the intelligent trading decision method based on hierarchical multi-round confrontation debate is the preferred implementation manner of the intelligent trading decision system based on hierarchical multi-round confrontation debate.

[0057] The intelligent trading decision system based on hierarchical multi-round confrontation debate comprises: Module M1, acquiring market data and constructing an input feature vector, initializing an agent set and dividing it into a point generation group and a refutation group, and assigning an initial credibility score to each agent; Module M2, performing R rounds of deep debate and dynamically adjusting the credibility score; Module M3, based on the adjusted credibility score, performing Soft-Borda dynamic voting and outputting a trading signal.

[0058] Specifically, module M2 comprises the following sub-modules: Module M2.1, generating points and refuting interactive for round r, wherein the value of round r is 1, 2, 3...R.

[0059] Specifically, in module M2.1, in each round of debate, the argument generation group generates a set of transaction views , the refutation group receives the set of transaction views , and based on its own analysis model, historical data, and debate information from previous rounds, constructs a set of refutation arguments for challenging or falsifying the views of the argument generation group ; evaluates the effectiveness of the arguments or refutations proposed by agent i in the current round r, quantified as argument-target consistency and argument-market fit , two key indicators, where A is the set of agents.

[0060] Argument-target consistency, which measures the consistency of the agent's view with the preset target function (such as "maximize expected return") in direction.

[0061] Argument-market fit, which measures the degree of fitting or correlation between the agent's view and the historical real market trend.

[0062] Module M2.2, based on the Bayesian mechanism to update the reputation score of the agent.

[0063] Specifically, in module M2.2, in each round of debate, the Bayesian updating mechanism is adopted to dynamically adjust the reputation score according to the performance of the agent in this round (i.e. argument effectiveness), and the update formula of the reputation score of agent i after the rth round is as follows:

[0064] Where: is the reputation score of agent i in the last round; is the initial reputation score of agent i.

[0065] is the effectiveness likelihood of the argument or refutation proposed by agent i in the rth round, calculated by weighting the performance indicators;

[0066] Where: argument-target consistency (value range).

[0067] argument-market fit (value range).

[0068] 、 respectively represent the experience weight of target consistency and historical consistency, which are configurable parameters, in the embodiment = 0.6, = 0.4, used to balance the forward-looking and historical experience.

[0069] Module M2.3, constructing round debate log and reputation trajectory.

[0070] Specifically, in module M2.3, in each round of debate, a structured log is generated and stored to realize complete transparency and traceability of the decision-making process, including the argument and rebuttal text in natural language format generated by each agent, the key indicators on which each reputation update is based and the reputation scores before and after the update and the SHAP-ExConsistency explanation vector for explaining the basis of the model decision, ensuring that the requirements of financial supervision on model explainability are met.

[0071] Specifically, module M3 includes the following sub-modules: Module M3.1, calculating dynamic temperature coefficient to adapt to market volatility , the formula is:

[0072] =0.3； = ； is the realized volatility in the last preset time (in the embodiment, the preset time is 60 minutes), the greater the market volatility, the smaller, so that the voting result is more sensitive to high-score signals.

[0073] Module M3.2, using Soft-Borda dynamic voting formula to calculate the final score of each trading signal:

[0074] The output trading signal with the highest score is used to control risk, and when the score difference between the highest score and the second highest score is less than a preset threshold (in the embodiment, the threshold is 0.05), it indicates that the market consensus is not strong, and the "wait and see" signal is output.

[0075] To verify the actual effect of the application, a control experiment is constructed, and the specific content is as follows: The experimental object selects a traditional single round majority voting transaction model (Baseline) as a comparison object, and is compared with the method of the application. Market data selects 5-minute K-line data of Shanghai-Shenzhen 300 component stocks, company announcements and financial news titles as data sources, and the time interval covers January 2022 to December 2024. A rolling back test method with a 60-day sliding window is adopted, and the test period is set to January 2023 to December 2024. In order to ensure the fairness of the experiment, the two models are set to perform at most one transaction per day, the holding period is not more than T+3 days, and the commission is uniformly set to 0.2%.

[0076] Comparison of key indicators

[0077] The experimental data shows that the decision delay of the method of the application is 43 milliseconds, which is within the acceptable range of millisecond-level transactions. The annual yield rate reaches 16.8%, the annual sharp ratio is 2.03, and the maximum drawdown is-6.1%. 100% of the explainable log coverage rate is achieved, and the structured debate log and explanation vector are recorded, which meets the requirements of financial supervision on model traceability and explainability.

[0078] The experimental results show that, compared with the existing single round voting mechanism, the application can more effectively identify noise signals and enhance market opinion cross-validation by introducing a multi-round debate and reputation updating mechanism. While controlling the maximum drawdown, the sharp ratio is significantly improved, achieving a better balance between yield and risk.

[0079] Those skilled in the art know that, in addition to implementing the system provided by the application and each device, module, unit thereof in pure computer readable program code, the system provided by the application and each device, module, unit thereof can also be realized in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps to achieve the same function. Therefore, the system provided by the application and each device, module, unit thereof can be considered as a hardware component, and the devices, modules and units included therein for realizing various functions can also be considered as structures within the hardware component; the devices, modules and units for realizing various functions can also be considered as both software modules for realizing methods and structures within hardware components.

[0080] The specific embodiments of the application are described above. It should be understood that the application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the application. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An intelligent trading decision-making method based on hierarchical multi-round adversarial debate, characterized by: The steps include: Step S1: Obtain market data and construct input feature vectors, initialize the agent set and divide it into argument generation group and rebuttal group, and assign an initial credibility score to each agent; Step S2: perform R rounds of in-depth debate and dynamically adjust the reputation score; Step S3: Based on the adjusted reputation score, Soft-Borda dynamic voting is performed and a transaction signal is output.

2. The intelligent trading decision-making method based on hierarchical multi-round confrontational debate according to claim 1 is characterized in that: In step S1, the argument generation group is used to analyze the input feature vector and propose initial trading opinions, which include buy, sell and wait and see; the rebuttal group is used to question the opinions proposed by the argument generation group and find potential risks and logical loopholes.

3. The intelligent trading decision-making method based on hierarchical multi-round confrontational debate according to claim 1 is characterized in that: In step S1, the initial reputation score of each agent is set to an unbiased initial value.

4. The intelligent trading decision-making method based on hierarchical multi-round confrontational debate according to claim 1 is characterized in that: The step S2 includes the following sub-steps: Step S2.1, conduct round r of opinion generation and refutation interaction; Step S2.2, updating the agent's reputation score based on the Bayesian mechanism; Step S2.3: Construct the round debate log and reputation track.

5. The intelligent trading decision-making method based on hierarchical multi-round confrontational debate according to claim 4 is characterized in that: In step S2.1, in each round of debate, the argument generation group Generate transaction view sets in parallel , rebuttal group Receive the transaction view set Then, based on its own analysis model, historical data and the debate information of the previous rounds, it constructs a set of rebuttal arguments. , used to challenge or falsify the argument generation group point of view; evaluate the effectiveness of the argument or rebuttal proposed by agent i in the current round r, quantified as argument-goal consistency and Argument-Market Fit Two key indicators, where A is the set of intelligent agents.

6. The intelligent trading decision-making method based on hierarchical multi-round confrontational debate according to claim 5 is characterized in that: In step S2.2, after each round of debate, the Bayesian update mechanism is used to dynamically adjust the reputation score of the agent according to its performance in the current round. The reputation score of agent i after the rth round is The update formula is as follows: in: is the reputation score of agent i in the previous round; is the initial reputation score of agent i; is the likelihood of the validity of the argument or rebuttal proposed by agent i in round r, calculated by weighting the performance indicators; in: Indicates argument-goal consistency; Expressing argument-market conformity; 、 They represent the experience weights of goal consistency and historical compliance respectively.

7. The intelligent trading decision-making method based on hierarchical multi-round confrontational debate according to claim 6 is characterized in that: In step S2.3, in each round of debate, a structured log is generated and stored to achieve full transparency and traceability of the decision-making process. The log includes the arguments and rebuttal texts generated by each agent in natural language format, and the key indicators based on which each reputation update is based. and , reputation score before and after update and the SHAP-ExConsistency explanation vector used to explain the basis for the model's decisions.

8. The intelligent trading decision-making method based on hierarchical multi-round confrontational debate according to claim 1 is characterized in that: In step S3, after all R rounds of debate are completed, the comprehensive score of each transaction signal in all rounds is calculated based on the debate results of each round. The comprehensive score is based on the Borda scoring method, that is, the ranking of the signals by each intelligent agent is comprehensively considered; at the same time, the dynamic temperature coefficient is calculated, and the comprehensive score and the dynamic temperature coefficient are integrated through the Soft-Borda dynamic voting formula to output the transaction signal with the highest score.

9. The intelligent trading decision-making method based on hierarchical multi-round confrontational debate according to claim 8 is characterized in that: The step S3 includes the following sub-steps: Step S3.1, calculate the dynamic temperature coefficient , the formula is: =0.3； = ； is the realized volatility at the most recent preset time; In step S3.2, the final score of each trading signal is calculated using the Soft-Borda dynamic voting formula: Output the trading signal with the highest score. When the score difference between the highest and second highest scoring signals is less than the preset threshold, output a "wait and see" signal.

10. An intelligent trading decision-making system based on hierarchical multi-round adversarial debate, characterized by: include: Module M1 obtains market data and constructs input feature vectors, initializes the agent set and divides it into argument generation group and rebuttal group, and assigns an initial credibility score to each agent; Module M2 performs R rounds of in-depth debate and dynamically adjusts the reputation score; Module M3, based on the adjusted reputation score, performs Soft-Borda dynamic voting and outputs transaction signals.

Images