Search

CN-122023012-A - Multi-agent financial analysis collaboration method based on LLM

CN122023012ACN 122023012 ACN122023012 ACN 122023012ACN-122023012-A

Abstract

The invention provides a multi-agent financial analysis collaborative method based on LLM, which relates to the technical field of artificial intelligence and comprises the steps of evaluating and fine-tuning LLM, registering agents such as a main pipe, a financial analyst, a data assistant, a data analyst, a summary and the like, submitting demands by natural language of a user, decomposing the demands into dependent subtasks by the main pipe through an LLM thinking chain, generating a dependency graph, constructing a first-stage team according to the vector similarity of task description and metadata of each analyst, decomposing the subtasks again by the first-stage team, constructing a second-stage team according to similarity, matching the data assistant with the analyst, calling tools by the second-stage team to complete tasks and writing in a shared message pool, summarizing an agent weighted fusion result output strategy, reporting and obtaining feedback by the main pipe to the user, and triggering an observation-inverse-iteration loop re-planning task if the deviation occurs. The invention improves the efficiency, quality and flexibility of financial decision and reduces the labor cost.

Inventors

  • WANG SHAODONG

Assignees

  • 浙商银行股份有限公司
  • 易企银(杭州)科技有限公司

Dates

Publication Date
20260512
Application Date
20251217

Claims (10)

  1. 1. A multi-agent financial analysis cooperative method based on LLM is characterized by comprising the following steps: evaluating the basic large language model, performing parameter fine adjustment according to the evaluation result, registering the intelligent agent based on the large language model and constructing an intelligent agent management pool; receiving financial analysis requirements submitted by a user through natural language; Under a two-stage layering collaboration mechanism, the main pipe intelligent body calls a large language model through a first preset prompting template to conduct thinking chain reasoning, and the financial analysis requirements are decomposed; Summarizing that the agent acquires all output results from the shared message pool, carrying out weighted fusion based on preset evaluation indexes to generate financial strategies and suggestions, reporting analysis results to the user by the agent, and acquiring user feedback.
  2. 2. The LLM-based multi-agent financial analysis collaborative method according to claim 1, wherein the agents include a manifold agent, a financial analyst, a data assistant, a data analyst, and a summary agent, each agent associated with a metadata profile containing capability descriptions, tools, and domain keywords; The evaluation index includes: the accuracy rate, recall rate and F1 value of text classification; a text generated BLEU value, ROUGE value, met eor value; MRR value and NDCG value called by interface; the smoothness, correlation, information amount and diversity dimension of the manual evaluation.
  3. 3. The LLM-based multi-agent financial analysis collaborative method according to claim 1, wherein decomposing financial analysis requirements refers to decomposing financial analysis subtasks having dependencies and generating task dependency graphs; based on the semantic similarity calculation of the description text of each financial analysis subtask and the metadata configuration file of the financial analyst agent, a first-level team comprising matched financial analysts is built; The financial analyst of the first-level team calls a large language model to carry out secondary decomposition on the subtasks through a second preset prompt template to obtain a data processing step and an analysis executing step, and builds a second-level team for executing specific tasks based on the similarity of the steps with metadata of the data assistant and the data analyst; The agent of the second team calls its own ability and external tools to execute tasks, and outputs the result to the shared message pool.
  4. 4. The LLM-based multi-agent financial analysis collaborative method according to claim 1, wherein the task dependency graph includes JSON format representation including a task ID field, a task description field, and a dependency list field; The semantic similarity calculation specifically includes: Converting the task description text D and the capability description text M in the agent metadata configuration file into vectors to be embedded respectively by adopting a sentence vector model based on a Transformer And 。
  5. 5. The LLM-based multi-agent financial analysis collaborative method according to claim 1, characterized in that when constructing a team, it comprises the steps of: If the semantic similarity calculation results are lower than the preset threshold, triggering a new agent registration process, including: Performing parameter efficient fine adjustment on the basic large language model; creating a metadata configuration file containing missing capability descriptions, tools and keywords; Registering the new agent to the agent management pool.
  6. 6. The LLM-based multi-agent financial analysis collaborative method according to claim 1, wherein in the agent invoking self-capability and external tool executing task, the specific steps are as follows: the data assistant agent calls an external data API or a database interface to perform data collection and preparation; The data analyst agent invokes a data visualization tool or a statistical analysis tool to perform data processing and chart generation.
  7. 7. The LLM-based multi-agent financial analysis collaborative method according to claim 1, wherein an observe-dislike-iterate loop comprises: observing, namely detecting deviation between a task result and an expected target and corresponding task nodes; the retum is that a large language model is called through a preset retum prompting template, the deviation reason is analyzed, and a correction scheme is generated; and iterating, namely creating a new subtask or adjusting task parameters according to the correction scheme, and reassigning to the step.
  8. 8. The LLM-based multi-agent financial analysis collaborative method according to claim 1, wherein inter-agent communication is in a structured data format, the message comprising: Message ID, sender ID, receiver ID, associated task ID, message type, and payload field; the message types include synchronous task allocation, asynchronous task allocation and result feedback.
  9. 9. The LLM-based multi-agent financial analysis collaborative method according to claim 1, wherein the manifold agent comprises reporting analysis results to a user and obtaining user feedback comprises the steps of: if the feedback indicates that the analysis has deviation, triggering an observation-jettance-iteration loop, namely detecting unsatisfactory task nodes, analyzing the deviation reasons and generating a correction scheme, re-planning tasks according to the correction scheme, and returning to the step.
  10. 10. The LLM-based multi-agent financial analysis collaborative method according to claim 1, wherein the metadata configuration file comprises JSON format comprising: an agent ID field, a role type field, a capability description list field, a tool list field, and a professional domain keyword list field.

Description

Multi-agent financial analysis collaboration method based on LLM Technical Field The invention relates to the technical field of artificial intelligence, in particular to a multi-agent financial analysis collaborative method based on LLM. Background In the financial field, with the increasing complexity and data volume of the market, conventional financial analysis methods face a number of challenges. Traditional financial analysis relies primarily on manual operations, which require an analyst to spend a significant amount of time collecting data, performing complex calculations and analysis, and composing reports. This method is not only inefficient, but also susceptible to artifacts, resulting in insufficient accuracy and reliability of the analysis results. In recent years, with the development of artificial intelligence technology, especially the application of large language model LLM, new opportunities for financial analysis are brought. The large language model can quickly understand and generate texts through the strong natural language processing capability, and provides a new technical means for financial analysis. However, most of the existing financial analysis methods based on large language models stay on the application of a single model, and lack cooperativity and flexibility. These methods generally deal with simple financial analysis tasks and often fail to provide a comprehensive and thorough solution to complex financial problems. In addition, the prior art often lacks effective task decomposition and team cooperation mechanisms when processing financial analysis tasks, which results in an inefficient analysis process and an inability to fully exploit the potential of large language models. In the process of realizing the embodiment of the invention, at least the following problems or defects exist in the prior art that the prior art cannot effectively integrate the capability of various intelligent agents, the accuracy and stability of the output of a single intelligent agent are required to be improved and a refined cooperative mechanism is lacked, so that the analysis efficiency is low, a flattened task allocation mechanism is usually adopted in the prior art when the multi-intelligent agents are coordinated, complex analysis tasks requiring multi-level and multi-role deep cooperation in the financial field are difficult to process, the prior art has the defects in task decomposition and team construction, proper intelligent agents cannot be flexibly matched according to task requirements, the prior art lacks an effective iterative optimization mechanism when the complex financial analysis tasks are processed, and the analysis strategy cannot be timely adjusted according to user feedback. Disclosure of Invention The invention provides a multi-agent financial analysis cooperative method based on LLM, which comprises the following steps: evaluating the basic large language model, performing parameter fine adjustment according to the evaluation result, registering the intelligent agent based on the large language model and constructing an intelligent agent management pool; receiving financial analysis requirements submitted by a user through natural language; Under a two-stage layering collaboration mechanism, the main pipe intelligent body calls a large language model through a first preset prompting template to conduct thinking chain reasoning, and the financial analysis requirements are decomposed; Summarizing that the agent acquires all output results from the shared message pool, carrying out weighted fusion based on preset evaluation indexes to generate financial strategies and suggestions, reporting analysis results to the user by the agent, and acquiring user feedback. Further, the agents include manifold agents, financial analysts, data assistants, data analysts, and summary agents, each agent associated with a metadata profile containing capability descriptions, tools, and domain keywords; The evaluation index includes: the accuracy rate, recall rate and F1 value of text classification; a text generated BLEU value, ROUGE value, met eor value; MRR value and NDCG value called by interface; the smoothness, correlation, information amount and diversity dimension of the manual evaluation. The analysis of the financial analysis requirements refers to the generation of task dependent graphs for financial analysis subtasks with dependency relations, and the construction of a first-level team comprising matching financial analysts based on the semantic similarity calculation of the description text of each financial analysis subtask and the metadata configuration file of the financial analyst agent; The financial analyst of the first-level team calls a large language model to carry out secondary decomposition on the subtasks through a second preset prompt template to obtain a data processing step and an analysis executing step, and builds a second-level team for executing specific task