Intelligent financial engineering quantitative research and reproduction method and related assembly

Abstract

The invention discloses an intelligent financial engineering quantitative research and reproduction method and related components, and relates to the field of finance. The method comprises the following steps: firstly, determining a target workflow by combining a workflow template and meta-information extracted from a quantitative research document by an information extraction agent; when the target workflow is executed, the information extraction agent extracts quantitative calculation information from the document; the code agent generates a quantitative calculation code according to the quantitative calculation information and the data source information, and quantitative factor data is obtained after the quantitative calculation code is executed; performing quantitative back-testing on the quantitative factor data according to a back-testing tool corresponding to a back-testing method specified in the document to obtain a back-testing result; and the verification agent compares the back-test result with the quantitative research result in the document to obtain a quantitative research conclusion. The whole process of financial engineering quantitative research is automatically completed based on the large model agent, practitioners are enabled to concentrate on core work of quantitative design and strategy design, and the research efficiency of users is improved.

Description

Technical Field

[0001] The present invention relates to the field of finance, and particularly to an intelligent financial engineering quantitative research and replication method and related components. Background Art

[0002] Financial engineering quantitative research uses mathematics, statistics, and computer technology to analyze and process various data in the financial market, and formulates investment strategies and manages risks by constructing quantitative models. In the financial industry, financial engineering quantitative research is a field with relatively high professional thresholds. Practitioners not only need to be familiar with the professional knowledge of financial quantification but also need to understand the supporting facilities (such as professional market data services and computing services, etc.), which poses higher requirements for practitioners. While designing, implementing, and validating their own financial quantification research topics, practitioners also need to keep up with the market and research frontiers, read and replicate a large number of existing research topics and reports, with heavy workload and extremely high professional requirements.

[0003] Thus, it can be seen that even for industry insiders, the industry threshold for financial engineering quantitative research is relatively high. Therefore, there is an urgent need for a financial engineering quantitative research method that can improve work efficiency and reduce research difficulty. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent financial engineering quantitative research and replication method and related components, which can complete the whole process of financial engineering quantitative research based on a large model agent, enabling practitioners to focus on the core work of quantitative design and strategy design, and improving the research efficiency of users.

[0005] To solve the above technical problems, the present invention provides an intelligent financial engineering quantitative research and replication method, which includes:

[0006] Using an information extraction agent to extract meta-information related to the workflow structure from the quantitative research document provided by the user, and determining and executing a target workflow in combination with the meta-information and the workflow template;

[0007] When executing the target workflow, input the quantitative research document into the information extraction agent to obtain the quantitative calculation information extracted by the information extraction agent from the quantitative research document;

[0008] Input the quantitative calculation information into a code agent, so that the code agent generates quantitative calculation code based on the quantitative calculation information and the data source information obtained by calling a professional domain knowledge base, a data warehouse, and professional computing tools through an interface agent, and obtains professional data through the interface agent and executes the quantitative calculation code to generate quantitative factor data;

[0009] According to the information extraction agent, extract the specified backtesting method in the quantitative research document, and call the backtesting tool corresponding to the backtesting method through the interface agent. Use the backtesting tool to perform quantitative backtesting on the quantitative factor data to obtain the backtesting results;

[0010] Input the backtesting results and the quantitative research results extracted by the information extraction agent from the quantitative research document into the verification agent to obtain the quantitative research conclusion drawn by the verification agent by comparing the backtesting results with the quantitative research results.

[0011] Optionally, the information extraction agent extracts quantitative calculation information from the quantitative research document, including:

[0012] The information extraction agent sets each prompt word related to the quantitative calculation information from macro to detail in a step-by-step progressive manner, and extracts the quantitative calculation information from the quantitative research document according to each prompt word in the order from macro to detail;

[0013] And / or, the information extraction agent sets prompt words related to the quantitative calculation information from multiple preset dimensions, and extracts the quantitative calculation information from the quantitative research document according to each prompt word.

[0014] Optionally, the information extraction agent extracts quantitative calculation information from the quantitative research document according to each prompt word, including:

[0015] The information extraction agent uses two or more different types of large models to extract quantitative calculation information from the quantitative research document respectively;

[0016] Deduplicate and integrate the quantitative calculation information extracted by each large model from the quantitative research document to obtain the final quantitative calculation information, and generate a prompt message when the difference between the quantitative calculation information extracted by each large model reaches a preset threshold to prompt the user to correct the quantitative calculation information.

[0017] Optionally, after the information extraction agent extracts the quantitative calculation information from the quantitative research document according to each prompt word, it further includes:

[0018] Judge whether the currently extracted quantitative calculation information meets the preset completeness verification condition, where the preset completeness verification condition is the condition corresponding to the quantitative calculation information meeting the information required for the code agent to generate the quantitative calculation code;

[0019] If so, enter the step of inputting the quantization calculation information into the code agent, so that the code agent generates quantization calculation code according to the quantization calculation information and the data source information obtained by calling the professional domain knowledge base, data warehouse, and professional calculation tools through the interface agent;

[0020] If not, enter the step of extracting the quantization calculation information from the quantization research document according to each of the prompt words.

[0021] Optionally, the code agent generates quantization calculation code according to the quantization calculation information and the data source information obtained by calling the professional domain knowledge base, data warehouse, and professional calculation tools through the interface agent, including:

[0022] The code agent generates quantization calculation code according to the quantization calculation information according to the preset prompt words, and when generating the quantization calculation code, calls the professional domain knowledge base through the interface agent to adjust the quantization calculation code in combination with the content in the professional domain knowledge base;

[0023] The code agent calls the data warehouse through the interface agent to obtain relevant data required for sample testing of the quantization calculation code, and calls the professional calculation tool through the interface agent to test the quantization calculation code using the test sample;

[0024] When the quantization calculation code passes the sample test, the quantization calculation code is stored in the library and saved as the final quantization calculation code;

[0025] When the quantization calculation code fails the sample test, the error and exception information that appears during the execution of the quantization calculation code is incorporated into the professional domain knowledge base, and re-enter the step of generating quantization calculation code according to the quantization calculation information according to the preset prompt words, and when generating the quantization calculation code, call the professional domain knowledge base through the interface agent to adjust the quantization calculation code in combination with the content in the professional domain knowledge base.

[0026] Optionally, after the code agent executes the quantization calculation code to generate quantization factor data, it further includes:

[0027] The information extraction agent extracts the specified backtesting method in the quantization research document, and calls the backtesting tool corresponding to the backtesting method through the interface agent, and uses the backtesting tool to perform quantization testing on the quantization factor data in combination with the current real market data to obtain the current test result, so as to determine the performance of the quantization factor data in the current market.

[0028] Optionally, calling the professional domain knowledge base, data warehouse, professional calculation tool, and backtesting tool through the interface agent includes:

[0029] Take each first alternative tool in the first layer of the target access object as the current alternative tool, and write each current alternative tool into the function call parameters of the interface agent, so that the interface agent selects the current target tool from each current alternative tool;

[0030] Wherein, the target access object includes the professional domain knowledge base, data warehouse, professional computing tool and backtesting tool, and the professional domain knowledge base, the data warehouse, the professional computing tool and the backtesting tool are all pre-reconstructed into a tree structure according to a preset layering standard;

[0031] Take each alternative tool in the next layer of the node where the current target tool is located as the new current alternative tool, and enter the step of writing each current alternative tool into the function call parameters of the interface agent, so that the interface agent selects the current target tool from each current alternative tool. When the current target tool is the actual tool, return the structured parameters required to call the actual tool to the interface agent, so that the interface agent calls the target access object to complete the corresponding function.

[0032] To solve the above technical problems, the present invention also provides an intelligent financial engineering quantitative research and reproduction system, which includes:

[0033] A target workflow determination unit, configured to use an information extraction agent to extract meta-information related to the workflow structure from a quantitative research document provided by a user, and determine and execute a target workflow in combination with the meta-information and a workflow template;

[0034] A quantitative calculation information extraction unit, configured to input the quantitative research document into the information extraction agent when executing the target workflow, and obtain the quantitative calculation information extracted by the information extraction agent from the quantitative research document;

[0035] A quantitative factor data generation unit, configured to input the quantitative calculation information into a code agent, so that the code agent generates quantitative calculation code according to the quantitative calculation information and data source information obtained by calling a professional domain knowledge base, a data warehouse and a professional computing tool through an interface agent, and obtains professional data through the interface agent and executes the quantitative calculation code to generate quantitative factor data;

[0036] A backtesting unit, configured to extract the specified backtesting method in the quantitative research document by the information extraction agent, and call the backtesting tool corresponding to the backtesting method through the interface agent, and perform quantitative backtesting on the quantitative factor data by using the backtesting tool to obtain a backtesting result;

[0037] A verification unit for inputting the backtest result and the quantitative research result extracted by the information extraction agent from the quantitative research document into a verification agent, and obtaining the quantitative research conclusion drawn by the verification agent by comparing the backtest result with the quantitative research result.

[0038] To solve the above technical problems, the present invention also provides an intelligent financial engineering quantitative research and replication device, which includes:

[0039] A memory for storing computer programs;

[0040] A processor for implementing the steps of any of the above intelligent financial engineering quantitative research and replication methods when executing the computer program.

[0041] To solve the above technical problems, the present invention also provides a storage medium with a computer program stored thereon, and the computer program realizes the steps of any of the above intelligent financial engineering quantitative research and replication methods when executed by a processor.

[0042] The beneficial effect of the present invention lies in providing an intelligent financial engineering quantitative research and replication method and related components, systems, devices and storage media. First, determine the target workflow by combining the workflow template and the meta-information extracted by the information extraction agent from the quantitative research document. When executing the target workflow, the information extraction agent extracts quantitative calculation information from the document; the code agent generates quantitative calculation code based on the quantitative calculation information and data source information, and obtains quantitative factor data after executing the quantitative calculation code; perform a quantitative backtest on the quantitative factor data using the backtest tool corresponding to the specified backtest method in the document to obtain a backtest result; the verification agent compares the backtest result with the research result in the document to obtain a quantitative research conclusion. The entire process of financial engineering quantitative research is completed based on the large model agent, allowing practitioners to focus on the core work of quantitative design and strategy design, and improving the research efficiency of users. Description of the Drawings

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the prior art and the embodiments. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0044] Figure 1 The first flowchart of an intelligent financial engineering quantitative research and replication method provided by the present invention;

[0045] Figure 2 The second flowchart of an intelligent financial engineering quantitative research and replication method provided by the present invention;

[0046] Figure 3 The flowchart of an information extraction agent provided by the present invention;

[0047] Figure 4 The flowchart of a code agent provided by the present invention;

[0048] Figure 5 The flowchart of an interface agent provided by the present invention;

[0049] Figure 6 The structural schematic diagram of an intelligent financial engineering quantitative research and replication system provided by the present invention;

[0050] Figure 7 The structural schematic diagram of an intelligent financial engineering quantitative research and replication device provided by the present invention. Detailed implementation manners

[0051] The core of the present invention is to provide an intelligent financial engineering quantitative research and replication method and related components, which can complete the whole process of financial engineering quantitative research based on large model agents, enabling practitioners to focus on the core work of quantitative design and strategy design and improving the research efficiency of users.

[0052] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] Please refer to Figure 1 , Figure 1 The first flowchart of an intelligent financial engineering quantitative research and replication method provided by the present invention, and the method includes:

[0054] S101. Use an information extraction agent to extract meta-information related to the workflow structure from the quantitative research document provided by the user, and determine and execute the target workflow in combination with the meta-information and the workflow template.

[0055] First of all, it should be noted that the quantitative research document in the present invention can be a quantitative research description document for financial engineering or a financial engineering research report on the market. The quantitative research description document for financial engineering usually includes detailed descriptions of the quantitative research background, objectives, methodology, data sources, model construction process, strategy implementation methods, and their applications. The financial engineering research report usually includes market views, strategy suggestions, investment suggestions, etc. Although the quantitative research description document for financial engineering and the financial engineering research report have different focuses, they both include the data, methods, and conclusions required for quantitative research in financial engineering. The intelligent financial engineering quantitative research and replication method provided by the present invention can automatically complete the quantitative research in financial engineering based on the above content. In addition, it is recommended that users provide the quantitative research document in the PDF file format.

[0056] Considering that there are many types of quantitative research in financial engineering, such as stock selection research, timing research, derivative research, and fundamental research, etc. There are differences in key information such as the source data schema and testing methods corresponding to different types of quantitative research. Therefore, there are also differences in the details of the workflows used to complete different types of quantitative research. Therefore, it is necessary to first determine the target workflow for this quantitative research based on the quantitative research document provided by the user.

[0057] Specifically, use the information extraction agent to extract the meta-information related to the workflow structure from the quantitative research document. The meta-information related to the workflow structure in the present invention includes, but is not limited to, the overall judgment of quantitative research, research type, source data information, and testing methods, etc. The dynamic prompt of the information extraction agent here needs to meet the above meta-information requirements and conform to the JSON structured output format.

[0058] The information extraction agent is an intelligent system that automatically identifies and extracts specified information from the quantitative research document based on artificial intelligence technology, and can interact with the user in natural language, including in the form of dialogue and document. In this step, after inputting the quantitative research document provided by the user to the information extraction agent, a prompt word with the theme of "Please extract the meta-information related to the workflow structure in this quantitative research document" can be input to the information extraction agent, and the information extraction agent can automatically extract the above meta-information from the quantitative research document. There are various information extraction agents that can achieve information extraction, and the present invention does not limit its specific type, and it can be selected according to actual needs.

[0059] Please refer to Figure 2 , Figure 2This is the second flowchart of an intelligent financial engineering quantitative research and replication method provided by the present invention. After inputting the research document or report provided by the user into the information extraction agent, the information extraction agent extracts the meta-information in the document (i.e., the quantitative research document provided by the user, hereinafter uniformly referred to as the document). The present invention further combines a pre-set workflow template to clarify a workflow whose structure conforms to the meta-information related to the workflow structure extracted this time, and determines this workflow as the target workflow for conducting the current financial engineering quantitative research.

[0060] The workflow stipulates the steps required to automatically complete the financial engineering quantitative research. Its main structure includes using the information extraction agent to extract quantitative calculation information, using the code agent to generate quantitative calculation code and quantitative factor data, and using professional tools and verification agents to complete quantitative testing and proofreading. The process of completing the quantitative research according to the target workflow using different large model agents will be described below.

[0061] S102. When executing the target workflow, input the quantitative research document into the information extraction agent to obtain the quantitative calculation information extracted by the information extraction agent from the quantitative research document.

[0062] First, it is necessary for the information extraction agent to extract the quantitative calculation information from the quantitative research document so as to generate the quantitative calculation code based on the quantitative calculation information subsequently. The quantitative calculation information includes but is not limited to the description of the quantitative factor calculation logic, calculation formula, step-by-step calculation independent planning, data source information, and factor structured meta-information (information such as frequency).

[0063] The process of the information extraction agent extracting the quantitative calculation information from the document is similar to the process of the information extraction agent extracting the meta-information from the document, except that the prompt words are different. The present invention does not limit the specific implementation manner of the information extraction agent extracting the quantitative calculation information from the document based on the prompt words.

[0064] However, it should be emphasized that the present invention sets different prompt words from macro to detail, as well as a series of prompt words related to the quantitative calculation information from different angles, enabling the information extraction agent to extract the quantitative calculation information from the quantitative research document in a gradually progressive and multi-angle redundant extraction manner, ensuring the completeness of the quantitative calculation information, and further ensuring that a rigorous and professional quantitative calculation code can be generated based on the quantitative calculation information. The specific process will be described in the subsequent embodiments and will not be elaborated here for the time being.

[0065] S103. Input the quantization calculation information into the code intelligent agent, so that the code intelligent agent generates quantization calculation code according to the quantization calculation information and the data source information obtained by calling the professional domain knowledge base, data warehouse and professional calculation tools through the interface intelligent agent, obtains professional data through the interface intelligent agent, and executes the quantization calculation code to generate quantization factor data.

[0066] Generating quantization calculation code based on quantization calculation information is a core part of quantitative research in financial engineering. In the past, manually writing quantization calculation code had problems such as high thresholds, time-consuming, and laborious. The present invention uses a code intelligent agent to generate quantization calculation code after obtaining the quantization calculation information. In order to ensure that the code intelligent agent can smoothly and accurately generate quantization calculation code that meets the requirements, during the process of the code intelligent agent generating quantization calculation code according to the quantization calculation information, the interface intelligent agent is used to call the professional domain knowledge base, data warehouse, and professional calculation tools.

[0067] Specifically, the present invention pre-collects general materials in the field of financial quantization based on RAG (Retrieval-Augmented Generation) and stores them in the professional domain knowledge base. When the code intelligent agent generates quantization calculation code, it will use the interface intelligent agent to query relevant professional information from the professional domain knowledge base first, so as to generate quantization calculation code that meets industry requirements. At the same time, a number of cases where code errors are likely to occur in the scenarios of financial data processing and quantization calculation are also collected and stored in the professional domain knowledge base. These cases are used as supplementary information to assist the code intelligent agent in generating correct and effectively executable quantization calculation code.

[0068] In addition, a self-feedback iteration mechanism can be adopted to enable the code intelligent agent to self-diagnose and repair the current problem code. After multiple rounds of iteration, the problems of the quantization calculation code generated in the previous round can be gradually repaired until the quantization calculation code no longer reports errors and the results meet the expectations.

[0069] The present invention further uses the interface intelligent agent to call the data in the data warehouse and professional calculation tools, and uses test samples to test the generated quantization calculation code. When the quantization calculation code passes the sample test, it is considered that the quantization calculation code passes the inspection and the quantization calculation code is saved; when the quantization calculation code fails the sample test, the quantization calculation code is adjusted again according to the problems that occur, and the adjusted quantization calculation code is retested with samples until the quantization calculation code passes the sample test. The quantization calculation code that passes the sample test is saved as the final quantization calculation code.

[0070] In summary, the code agent uses the interface agent to call the professional domain knowledge base, data warehouse, and professional computing tools to ensure that the quantitative calculation code generated based on the quantitative calculation information meets the conditions such as the description of the quantitative factor calculation logic, the alignment of data and variable field names, and the timely call of industry professional data and professional computing tools. The data source schema information determined by the previous dynamic workflow template will dynamically update the prompt information here, so that the generated quantitative calculation code can reflect core information such as the correct data dictionary information. Combining with the professional domain knowledge base, supplement professional descriptions for algorithms of concepts and indicators that are prone to ambiguity in the financial field, so that the generated quantitative calculation code can accurately express the calculation logic. The professional domain knowledge base also includes a knowledge base of error-prone quantitative calculation codes, and supplements the correct writing methods of codes in common financial quantitative scenarios in the professional domain knowledge base to generate accurate and error-free codes. It can also combine with the enterprise system interface technical document knowledge base to generate necessary code lines for calling relevant enterprise internal interfaces during the code generation process. The code agent will dynamically integrate the above information and the application of engineering methods such as iterative self-correction to generate robust and usable quantitative calculation codes for subsequent execution and quantitative testing.

[0071] After generating the quantitative calculation code, the code agent uses the interface agent to obtain professional data and execute the quantitative calculation code to generate quantitative factor data.

[0072] S104. The information extraction agent extracts the specified backtesting method in the quantitative research document and calls the corresponding backtesting tool of the backtesting method through the interface agent, and uses the backtesting tool to perform quantitative backtesting on the quantitative factor data to obtain the backtesting result.

[0073] Quantitative testing is a key step in quantitative research in financial engineering. Using quantitative factor testing and quantitative backtesting as basic methods to verify the correctness and feasibility of the quantitative research ideas and methods provided in the quantitative research document. When performing quantitative backtesting, it is necessary to use the information extraction agent to analyze the quantitative research document again to extract the backtesting method for quantitative backtesting. Then, call the corresponding backtesting tool of the backtesting method through the interface agent, and use the backtesting tool to perform quantitative backtesting on the quantitative factor data to obtain the backtesting result. For example, if the information related to research and verification in the quantitative research document belongs to the general backtesting strategy, use the interface agent to call the corresponding backtesting tool of the above backtesting method to perform quantitative backtesting on the quantitative factor data based on real market data to obtain the backtesting result. For example, call the RiceQuant backtester to execute the backtesting and generate backtesting indicators (such as annualized return rate and Sharpe ratio, etc.). If the information related to research and verification in the quantitative research document belongs to the user-defined backtesting strategy, use the code agent to generate the backtesting code corresponding to the backtesting strategy, and then use the backtesting code to perform quantitative backtesting on the quantitative factor data to obtain the backtesting result.

[0074] S105. Input the backtest results and the quantitative research results extracted by the information extraction agent from the quantitative research document into the verification agent, and obtain the quantitative research conclusion drawn by the verification agent by comparing the backtest results with the quantitative research results.

[0075] Finally, qualitatively and quantitatively proofread the quantitative research results extracted by the information extraction agent from the quantitative research document and the real backtest results generated through the previous steps to obtain the quantitative research conclusion, that is, to judge whether the quantitative research ideas and conclusions proposed in the quantitative research document are correct.

[0076] In addition, the evaluation and optimization agent can be used to provide optimization suggestions for this quantitative research, such as the design direction of quantitative factors and the parameter optimization direction, provide research optimization and ideas from the perspective of AI for users, expand the ideas of users' research, and help users improve and optimize the theoretical research. If the user accepts, the AI - suggested ideas can be integrated with the current information data as a new quantitative research document, and the overall quantitative research process can be re - executed to obtain new conclusions and suggestions. This can continuously iterate and optimize the quantitative theory research of financial engineering until the user approves and is satisfied.

[0077] In summary, the present invention provides an intelligent quantitative research and replication method for financial engineering based on large - model agents, aiming to provide an efficient, professional and intelligent financial quantitative virtual research assistant for enterprises and practitioners (here, the virtual research assistant is an integration of the above - mentioned information extraction agent, interface agent, code agent and verification agent, which organically connects the above agents through a workflow and automatically executes each step of the quantitative research). It can interact with users in natural language, receive research tasks issued by users in the form of quantitative research documents, intelligently select a suitable target workflow, and gradually complete each step required for quantitative research of financial engineering according to the target workflow, and finally achieve the task from task description to research conclusions at the data and code levels. Please refer to Figure 2 , Figure 2 which is the second flowchart of the intelligent quantitative research and replication method for financial engineering provided by the present invention. The main processes for quantitative research of financial engineering include: clarifying the target workflow by combining the workflow template and the meta - information extracted by the information extraction agent from the quantitative research document; when executing the target workflow, the information extraction agent extracts quantitative calculation information from the document, including quantitative research logic, etc.; the code agent generates quantitative calculation code according to the quantitative calculation information and data source information, and obtains quantitative factor data after executing the quantitative calculation code; perform quantitative backtesting on the quantitative factor data using the backtesting tool corresponding to the backtesting method specified in the document to obtain the backtesting results; the verification agent compares the backtesting results with the conclusion information extracted by the information extraction agent from the document to obtain the quantitative research conclusion. During this process, the interface agent provides support for professional data and calculation tools.

[0078] The intelligent financial engineering quantitative research and reproduction method provided by the present invention no longer requires professionals to spend a lot of time and energy to prepare data and write code, allowing practitioners to focus on more core work such as quantitative involvement and strategy design, thereby expanding the work efficiency and ability boundaries of human employees.

[0079] In summary, the intelligent financial engineering quantitative research and reproduction method provided by the present invention uses the powerful semantic understanding, logical reasoning and code generation capabilities of the large model intelligent body, so that after the user provides the quantitative research document, the overall research process from data to quantitative calculation code to quantitative backtesting and conclusion can be automatically realized, which promotes the efficiency and feasibility of financial engineering quantitative research, lowers the industry threshold, and allows ordinary practitioners to handle complex quantitative research tasks and realize and verify their own quantitative research ideas. At the same time, it allows practitioners to release more energy to apply to other creative, guiding and strategic work, enhance their own value, and help enterprises optimize talent allocation.

[0080] Based on the above embodiments:

[0081] As an optional embodiment, the information extraction agent extracts quantitative calculation information from the quantitative research document, including:

[0082] The information extraction agent sets prompt words related to the quantitative calculation information in a step-by-step manner from macro to detail, and extracts the quantitative calculation information from the quantitative research document according to each prompt word in order from macro to detail; and / or, the information extraction agent sets prompt words related to the quantitative calculation information from multiple preset dimensions, and extracts the quantitative calculation information from the quantitative research document according to each prompt word.

[0083] The information extraction agent is the key to the successful completion of the overall quantitative research task. In this embodiment, the information extraction agent uses a variety of information extraction techniques for information extraction. On the one hand, prompt words related to quantitative calculation information are set from macro to detail, so that the information extraction agent extracts quantitative calculation information from the quantitative research document according to each prompt word in the order from macro to detail, that is, extracts quantitative calculation information from the quantitative research document in a gradually progressive extraction manner. For example, first extract the summary document theme and core elements from the quantitative research document at the macro level, and then extract the detailed information related to the core elements based on the core elements at the detail level. On the other hand, prompt words related to quantitative calculation information are also set from multiple preset dimensions, so as to make the information extracted by the information extraction agent from the quantitative research document more complete and sufficient by means of multi-angle extraction. For example, extract information from multiple angles related to the core elements in the quantitative research document to verify and confirm each other; extract the calculation logic description from the perspective of factor construction in the quantitative research document, and extract the factor calculation logic description from the perspective of the calculation paragraphs in the document.

[0084] Please refer to Figure 3 , Figure 3 which is a flowchart of the working process of an information extraction agent provided by the present invention. Taking Figure 3 as an example, first take the theme description extraction as a prompt word and extract the research theme and core idea from the document; then take the research object extraction as a prompt word and extract the main research object in the document from the content related to the research theme and core idea; then take the calculation logic detail extraction as a prompt word and extract the calculation logic detail from the content related to the main research object in the document.

[0085] In addition, it can be seen from the foregoing description that during the entire quantitative research process of financial engineering, there are multiple nodes where an information extraction agent needs to extract information from the quantitative research document, such as extracting quantitative calculation information, extracting backtesting methods, and quantitative research results. Each node can adopt the gradually progressive and multi-angle extraction methods in this embodiment to complete the extraction of the required information, and only need to adjust the prompt words accordingly.

[0086] On this basis, when the information extraction agent extracts quantitative calculation information from the quantitative research document according to each prompt word, it can simultaneously use two or more different types of large models to extract quantitative calculation information from the quantitative research document respectively. Then, the quantitative calculation information extracted by each large model from the quantitative research document is de-duplicated and integrated to obtain the final quantitative calculation information.

[0087] When using an information extraction agent to extract information from a quantitative research document, multiple large models of different types are used to separately extract quantitative calculation information from the quantitative research document. For example, a large model good at logical reasoning and a large model good at long texts are used simultaneously to extract quantitative calculation information from the quantitative research document. For example, two to three SOTA large models are used to separately extract quantitative calculation information from the quantitative research document. This method can solve the hallucination problem of large models to a certain extent.

[0088] In addition, when the difference between the quantitative calculation information extracted by each large model reaches a preset threshold, a prompt message is generated to prompt the user to correct the quantitative calculation information. If the difference between the quantitative calculation information extracted by different large models is serious (for example, by judging whether the difference between the quantitative calculation information exceeds the preset threshold), user intervention can be introduced. After the user solves the problem of the difference between the quantitative calculation information, the subsequent steps of the quantitative research are continued. This method can also incorporate user interaction into the process of quantitative research, avoid a completely black-box mechanism, and give the user a certain sense of participation and dominance.

[0089] Furthermore, after the information extraction agent extracts the quantitative calculation information from the quantitative research document according to each prompt word, it can also perform a completeness verification on the extracted quantitative calculation information, that is, judge whether the currently extracted quantitative calculation information is complete and effective enough to meet the requirements of subsequent quantitative calculations. Please refer to Figure 3 , Figure 3 which is a flowchart of the operation of an information extraction agent provided by the present invention. Specifically, first judge whether the currently extracted quantitative calculation information meets the preset completeness verification conditions. The preset completeness verification conditions are the conditions corresponding to when the quantitative calculation information meets the information required for the code agent to generate quantitative calculation code. If so, enter the step of inputting the quantitative calculation information into the code agent so that the code agent can generate quantitative calculation code according to the quantitative calculation information and the data source information obtained by calling the professional domain knowledge base, data warehouse, and professional calculation tools through the interface agent. If not, enter the step of extracting quantitative calculation information from the quantitative research document according to each prompt word, and continue to extract quantitative calculation information until the quantitative calculation information meets the completeness requirements. For example, the information extraction agent starts from the beginning of the quantitative logic following the first principle and generates a calculation logic plan that is dependent, self-consistent, and rigorous before and after, so as to completely extract the quantitative calculation information for the subsequent code agent to generate professional and rigorous quantitative calculation code.

[0090] The process of the code agent generating quantitative calculation code will be described below.

[0091] As an alternative embodiment, the code agent generates a quantitative calculation code based on the quantitative calculation information and the data source information obtained by invoking the professional domain knowledge base, data warehouse, and professional calculation tools through the interface agent, including:

[0092] The code agent generates a quantitative calculation code according to the quantitative calculation information according to the preset prompt words, and invokes the professional domain knowledge base through the interface agent when generating the quantitative calculation code to adjust the quantitative calculation code in combination with the content in the professional domain knowledge base.

[0093] The code agent invokes the data warehouse through the interface agent to obtain relevant data required for sample testing of the quantitative calculation code, and invokes the professional calculation tool through the interface agent to test the quantitative calculation code using the test sample.

[0094] When the quantitative calculation code passes the sample test, the quantitative calculation code is stored in the database as the final quantitative calculation code; when the quantitative calculation code fails the sample test, the error and exception information that appears during the execution of the quantitative calculation code is incorporated into the professional domain knowledge base, and it re-enters the step of generating the quantitative calculation code according to the quantitative calculation information according to the preset prompt words, and invoking the professional domain knowledge base through the interface agent when generating the quantitative calculation code to adjust the quantitative calculation code in combination with the content in the professional domain knowledge base.

[0095] Please refer to Figure 4 , Figure 4 which is a workflow diagram of a code agent provided by the present invention. In order to generate a professional and format-compliant quantitative calculation code, preset prompt words are set in advance, and the preset prompt words include but are not limited to requirements for code generation, output code format, and domain knowledge booth information, etc. The code agent generates a quantitative calculation code according to the quantitative calculation information according to the preset prompt words, that is, Figure 4 the LLM (Large Language Model Code) code in

[0096] To ensure the correctness and executability of the quantitative calculation code, the present invention also sets up a professional domain knowledge base. Specifically, based on RAG (Retrieval-Augmented Generation), the present invention pre-collects general materials in the field of financial quantification and stores them in the professional domain knowledge base. When the code agent generates quantitative calculation code, it will use the interface agent to query relevant professional information from the professional domain knowledge base first, so as to generate quantitative calculation code that meets industry requirements. At the same time, a number of cases where code errors are likely to occur in the scenarios of financial data processing and quantitative calculation are also collected and stored in the professional domain knowledge base. These cases are used as supplementary information to assist the code agent in generating correct and effectively executable quantitative calculation code, and to avoid general errors when the code agent generates quantitative calculation code. The professional domain knowledge base also includes enterprise private domain data and technical documents of computing tool interfaces, which are used to incorporate enterprise private domain data or call business interfaces in the generated quantitative calculation code, such as obtaining relevant data from the enterprise data warehouse, etc.

[0097] In addition, the professional domain knowledge base also stores information such as error correction suggestions for abnormal errors that occur during the execution of the quantitative calculation code, so as to adopt a self-feedback iteration mechanism to enable the code agent to self-diagnose and repair the current problem code, and ensure the executability and correctness of the quantitative calculation code.

[0098] The present invention further uses the interface agent to call the data in the data warehouse and professional computing tools, and uses test samples to test the generated quantitative calculation code, test the actual execution status of the quantitative calculation code, and verify the correctness of the execution result of the quantitative calculation code from aspects such as format and type. When the quantitative calculation code passes the sample test, it is considered that the quantitative calculation code passes the inspection and the quantitative calculation code is saved; when the quantitative calculation code fails the sample test, the self-feedback iteration mechanism is used to adjust the quantitative calculation code again according to the problems that occur, and the adjusted quantitative calculation code is retested with samples until the quantitative calculation code passes the sample test. The quantitative calculation code that passes the sample test is saved as the final quantitative calculation code, which further ensures the correctness of the quantitative calculation code.

[0099] As an optional embodiment, after the code agent executes the quantitative calculation code to generate quantitative factor data, it further includes:

[0100] The information extraction agent extracts the specified backtesting method in the quantitative research document, and the interface agent is used to call the backtesting tool corresponding to the backtesting method. The backtesting tool is used to perform quantitative testing on the quantitative factor data in combination with the current real market data to obtain the current test result, so as to determine the performance of the quantitative factor data in the current market.

[0101] In this embodiment, considering that quantitative tests are all based on historical data, and the quantitative ideas that were effective in the past may become invalid currently, it is very important whether quantitative research is applicable to the current market environment. Therefore, the present invention supports the user to deploy the recognized and required quantitative research results to the managed simulation with one key. Please refer to Figure 2 , Figure 2 FIG. Figure 2 is the second flowchart of an intelligent financial engineering quantitative research and reproduction method provided by the present invention. The present invention can intelligently screen out effective quantitative factor data for real-time hosting and simulation. For example, the quantitative factors can be evaluated based on real-time data to obtain the current performance, that is, using the backtesting tool corresponding to the backtesting method specified in the quantitative research document, and combining the current real market data to perform a new quantitative test on the quantitative factor data, so that the user can obtain the performance of the quantitative research under the current market conditions, and more effectively assist the user in applying the quantitative research in real trading.

[0102] As an optional embodiment, the interface agent is used to call the professional domain knowledge base, data warehouse, professional computing tool, and backtesting tool, including:

[0103] Taking each first alternative tool in the first layer of the target access object as the current alternative tool, and writing each current alternative tool into the function call parameters of the interface agent, so that the interface agent can select the current target tool from each current alternative tool; wherein, the target access object includes the professional domain knowledge base, data warehouse, professional computing tool, and backtesting tool, and the professional domain knowledge base, data warehouse, professional computing tool, and backtesting tool are all pre-reconstructed into a tree structure according to a preset layering standard.

[0104] Taking each alternative tool in the next layer of the node where the current target tool is located as the new current alternative tool, and entering the step of writing each current alternative tool into the function call parameters of the interface agent, so that the interface agent can select the current target tool from each current alternative tool, until the current target tool is the actual tool, and return the structured parameters required for calling the actual tool to the interface agent, so that the interface agent can call the target access object to complete the corresponding function.

[0105] In the present invention, the generation of quantitative calculation code, the execution of quantitative calculation code to generate quantitative factor data, and the quantitative backtesting all require the support of the interface agent, and it is necessary to call the professional domain knowledge base, data warehouse, professional computing tool, and backtesting tool through the interface agent.

[0106] In this embodiment, the professional domain knowledge base, data warehouse, professional computing tools, and backtesting tools are used as target access objects, and the professional domain knowledge base, data warehouse, professional computing tools, and backtesting tools are all pre - reconstructed into a tree structure according to a preset hierarchical standard. This requires human experts to classify and layer the existing tools in advance, which is equivalent to endowing the intelligent agent with expert opinions and is suitable for scenarios of intelligent tool invocation in complex business scenarios with long business chains. The target access objects can be classified and layered according to partial systems, functional granularity, etc. For each node in the summary abstract class layer, a category tool needs to be constructed (the function of this tool is to return the position of the current node) until the actual tools in the last layer. The tool consists of two parts. One is the text description of the tool, including the actual function of the tool, the tool name, the parameters included in the tool, and the parameter type and parameter meaning. The other is the implementation of the tool, including obtaining the tool parameters, which can be secondarily processed through a general parameter extractor to obtain the required parameters. The constructed tools are registered in the project for subsequent invocation (i.e., written into the variable space).

[0107] The essence of the interface intelligent agent provided in this embodiment to access the target access object is to gradually access the actual tool in the target access object through multiple inquiries, and each inquiry only returns the key information of that layer. This design can overcome the limitations of the capabilities of large - language models and the context length, and the practical difficulty that the language model cannot process a large number of tools at one time.

[0108] Please refer to Figure 5 , Figure 5 which is the flowchart of the operation of an interface intelligent agent provided by the present invention. Specifically, the interface intelligent agent accesses from the top layer of the tree structure downwards. First, each first alternative tool in the first layer of the target access object (that is, Figure 5 each tool corresponding to the first - level classification in) is used as the current alternative tool, and each current alternative tool is written into the function call parameters of the interface intelligent agent so that the interface intelligent agent can select the current target tool from each current alternative tool. Then, each alternative tool in the next layer of the node where the current target tool is located is used as the new current alternative tool, and the step of writing each current alternative tool into the function call parameters of the interface intelligent agent is entered so that the interface intelligent agent can select the current target tool from each current alternative tool. Until the current target tool is an actual tool, the structured parameters required for calling the actual tool are returned to the interface intelligent agent so that the interface intelligent agent can call the target access object to complete the corresponding function.

[0109] Generally speaking, the interface agent first branches to the next level according to the overall intention, and the next level then continues to branch to the second level according to the refined intention. Continuing to branch in this classification idea until the final actual tool can be reached, a tree-shaped organizational structure is constructed. The interface agent can accurately hit the required actual tool through multiple rounds of requests, successfully call the target access object, and complete the corresponding function. For example, when wanting to obtain the market information of a certain stock in the past ten years, the stock market query interface of the data warehouse is hit through multiple requests.

[0110] Please refer to Figure 6 , Figure 6 FIG. is a schematic structural diagram of an intelligent financial engineering quantitative research and reproduction system provided by the present invention. The system includes:

[0111] A target workflow determination unit 601, configured to use an information extraction agent to extract meta-information related to the workflow structure from a quantitative research document provided by a user, and determine and execute a target workflow in combination with the meta-information and a workflow template.

[0112] A quantitative calculation information extraction unit 602, configured to input the quantitative research document into the information extraction agent when executing the target workflow, and obtain the quantitative calculation information extracted by the information extraction agent from the quantitative research document.

[0113] A quantitative factor data generation unit 603, configured to input the quantitative calculation information into a code agent, so that the code agent generates quantitative calculation code according to the quantitative calculation information and the data source information obtained by calling a professional domain knowledge base, a data warehouse, and a professional calculation tool through an interface agent, and obtains professional data through the interface agent and executes the quantitative calculation code to generate quantitative factor data.

[0114] A backtesting unit 604, configured to extract a specified backtesting method from the quantitative research document by an information extraction agent, and call a backtesting tool corresponding to the backtesting method through an interface agent, and perform quantitative backtesting on the quantitative factor data by using the backtesting tool to obtain a backtesting result.

[0115] A verification unit 605, configured to input the backtesting result and the quantitative research result extracted by the information extraction agent from the quantitative research document into a verification agent, and obtain a quantitative research conclusion obtained by the verification agent comparing the backtesting result with the quantitative research result.

[0116] For a detailed introduction to the intelligent financial engineering quantitative research and reproduction system provided by the present invention, please refer to the embodiments of the above-mentioned intelligent financial engineering quantitative research and reproduction method, and the present invention will not be elaborated herein.

[0117] Please refer to Figure 7 , Figure 7The structural schematic diagram of an intelligent financial engineering quantitative research and replication device provided by the present invention, the device includes:

[0118] A memory 701 for storing computer programs;

[0119] A processor 702 for implementing the steps of any of the above-mentioned intelligent financial engineering quantitative research and replication methods when executing the computer program.

[0120] For a detailed introduction to the intelligent financial engineering quantitative research and replication device provided by the present invention, please refer to the embodiments of the above-mentioned intelligent financial engineering quantitative research and replication method, and the present invention will not be elaborated herein.

[0121] To solve the above technical problems, the present invention also provides a storage medium with a computer program stored thereon, and the steps of any of the above-mentioned intelligent financial engineering quantitative research and replication methods are implemented when the computer program is executed by a processor.

[0122] For a detailed introduction to the storage medium provided by the present invention, please refer to the embodiments of the above-mentioned intelligent financial engineering quantitative research and replication method, and the present invention will not be elaborated herein.

[0123] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. It should also be noted that in this specification, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover a non-exclusive inclusion, so that a process, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, article or device including the said element.

[0124] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent financial engineering quantitative research and replication method, characterized in that, Including: Using an information extraction agent to extract meta-information related to the workflow structure from the quantitative research document provided by the user, and determining and executing a target workflow in combination with the meta-information and the workflow template; When executing the target workflow, inputting the quantitative research document into the information extraction agent to obtain the quantitative calculation information extracted by the information extraction agent from the quantitative research document; Inputting the quantitative calculation information into the code agent, so that the code agent generates quantitative calculation code according to the quantitative calculation information and the data source information obtained by calling the professional domain knowledge base, data warehouse and professional calculation tools through the interface agent, and obtaining professional data through the interface agent and executing the quantitative calculation code to generate quantitative factor data; According to the backtesting method specified in the quantitative research document extracted by the information extraction agent, and calling the corresponding backtesting tool of the backtesting method through the interface agent, and using the backtesting tool to perform quantitative backtesting on the quantitative factor data to obtain a backtesting result; Inputting the backtesting result and the quantitative research result extracted by the information extraction agent from the quantitative research document into the verification agent to obtain the quantitative research conclusion obtained by the verification agent comparing the backtesting result with the quantitative research result; 2. The intelligent financial engineering quantitative research and replication method according to claim 1, wherein The information extraction agent extracts quantitative calculation information from the quantitative research document, including: The information extraction agent sets each prompt word related to the quantitative calculation information in a step-by-step manner from macro to detail, and sequentially extracts the quantitative calculation information from the quantitative research document according to each prompt word in the order from macro to detail; And / or, the information extraction agent sets prompt words related to the quantitative calculation information from a preset plurality of dimensions, and extracts the quantitative calculation information from the quantitative research document according to each prompt word; 3. The intelligent financial engineering quantitative research and reproduction method according to claim 2, wherein The information extraction agent extracts quantitative calculation information from the quantitative research document according to each prompt word, including: The information extraction agent extracts quantitative calculation information from the quantitative research document by using two or more different types of large models respectively; Deduping and integrating the quantitative calculation information extracted by each large model from the quantitative research document to obtain the final quantitative calculation information, and generating a prompt message when the difference between the quantitative calculation information extracted by each large model reaches a preset threshold to prompt the user to correct the quantitative calculation information; 4. The intelligent financial engineering quantitative research and replication method according to claim 2, characterized in that After the information extraction agent extracts the quantitative calculation information from the quantitative research document according to each prompt word, it further includes: Judging whether the currently extracted quantitative calculation information meets the preset completeness verification condition, where the preset completeness verification condition is the condition corresponding to when the quantitative calculation information meets the information required for the code agent to generate the quantitative calculation code; If so, enter the step of inputting the quantitative calculation information into the code agent, so that the code agent generates quantitative calculation code according to the quantitative calculation information and the data source information obtained by calling the professional domain knowledge base, data warehouse and professional calculation tools through the interface agent; If not, then enter the step of extracting the quantitative calculation information from the quantitative research document according to each of the above-mentioned prompt words.

5. The intelligent financial engineering quantitative research and replication method according to claim 1, wherein The code agent generates a quantitative calculation code based on the quantitative calculation information and the data source information obtained by calling a professional domain knowledge base, a data warehouse, and professional calculation tools through an interface agent, including: The code agent generates a quantitative calculation code according to the quantitative calculation information according to a preset prompt word, and calls the professional domain knowledge base through the interface agent when generating the quantitative calculation code to adjust the quantitative calculation code in combination with the content in the professional domain knowledge base; The code agent calls the data warehouse through the interface agent to obtain relevant data required for sample testing of the quantitative calculation code, and calls the professional calculation tool through the interface agent to test the quantitative calculation code using the test sample; When the quantitative calculation code passes the sample test, the quantitative calculation code is stored in the database as the final quantitative calculation code; When the quantitative calculation code fails the sample test, the error and exception information that appears during the execution of the quantitative calculation code is incorporated into the professional domain knowledge base, and the process re-enters the step of generating a quantitative calculation code according to the quantitative calculation information according to a preset prompt word, and calls the professional domain knowledge base through the interface agent when generating the quantitative calculation code to adjust the quantitative calculation code in combination with the content in the professional domain knowledge base.

6. The intelligent financial engineering quantitative research and replication method according to claim 1, wherein After the code agent executes the quantitative calculation code to generate quantitative factor data, it further includes: The information extraction agent extracts the specified backtesting method in the quantitative research document, and calls the backtesting tool corresponding to the backtesting method through the interface agent, and uses the backtesting tool to perform a quantitative test on the quantitative factor data in combination with the current real market data to obtain the current test result, so as to determine the performance of the quantitative factor data in the current market.

7. The intelligent financial engineering quantitative research and reproduction method according to claim 1, characterized in that Calling a professional domain knowledge base, a data warehouse, professional calculation tools, and a backtesting tool through an interface agent, including: Taking each first alternative tool in the first layer of the target access object as the current alternative tool, and writing each current alternative tool into the function call parameter of the interface agent, so that the interface agent selects the current target tool from each current alternative tool; Among them, the target access object includes the professional domain knowledge base, the data warehouse, the professional calculation tools, and the backtesting tool, and the professional domain knowledge base, the data warehouse, the professional calculation tools, and the backtesting tool are all pre-reconstructed into a tree structure according to a preset layering standard; Taking each alternative tool in the next layer of the node where the current target tool is located as the new current alternative tool, and entering the step of writing each current alternative tool into the function call parameter of the interface agent, so that the interface agent selects the current target tool from each current alternative tool, until the current target tool is the actual tool, and returns the structured parameters required for calling the actual tool to the interface agent, so that the interface agent calls the target access object to complete the corresponding function.

8. An intelligent financial engineering quantitative research and replication system, characterized in that, Including: A target workflow determination unit, configured to use an information extraction agent to extract meta-information related to a workflow structure from a quantitative research document provided by a user, and determine and execute a target workflow in combination with the meta-information and a workflow template; A quantitative calculation information extraction unit, configured to, when executing the target workflow, input the quantitative research document into the information extraction agent, and obtain quantitative calculation information extracted by the information extraction agent from the quantitative research document; A quantitative factor data generation unit, configured to input the quantitative calculation information into a code agent, so that the code agent generates a quantitative calculation code according to the quantitative calculation information and data source information obtained by calling a professional domain knowledge base, a data warehouse, and a professional calculation tool through an interface agent, and obtain professional data through the interface agent and execute the quantitative calculation code to generate quantitative factor data; A backtesting unit, configured to extract a specified backtesting method in the quantitative research document by the information extraction agent, and call a backtesting tool corresponding to the backtesting method through the interface agent, and perform quantitative backtesting on the quantitative factor data by using the backtesting tool to obtain a backtesting result; A verification unit, configured to input the backtesting result and the quantitative research result extracted by the information extraction agent from the quantitative research document into a verification agent, and obtain a quantitative research conclusion obtained by the verification agent comparing the backtesting result with the quantitative research result; 9. An intelligent financial engineering quantitative research and replication device, characterized in that, including: a memory, configured to store a computer program; a processor, configured to implement the steps of the intelligent financial engineering quantitative research and reproduction method according to any one of claims 1 to 7 when executing the computer program; 10. A storage medium, characterized in that, A computer program is stored on the storage medium, and when the computer program is executed by a processor, the steps of the intelligent financial engineering quantitative research and reproduction method according to any one of claims 1 to 7 are implemented.

Images