Search

CN-122023013-A - Quantization strategy research and development system, method and device based on multi-agent cooperation

CN122023013ACN 122023013 ACN122023013 ACN 122023013ACN-122023013-A

Abstract

The invention relates to the field of artificial intelligence and financial science and technology, and discloses a quantization strategy research and development system, method and device based on multi-agent cooperation. The data agent processes the multi-source financial data, the factor agent generates a high-quality factor set, the model agent trains an optimization model, the scheduling agent dynamically allocates resources, the knowledge graph module stores multiplexing knowledge, and the strategy generation module outputs a deployable strategy. The system solves the problems of insufficient automation and poor coordination of the existing quantitative research and development, and improves the research and development efficiency and strategy stability.

Inventors

  • LI MIN
  • HUANG SHUQUAN

Assignees

  • 江苏九州云数智科技有限公司

Dates

Publication Date
20260512
Application Date
20251225

Claims (10)

  1. 1. A quantitative strategy research and development system based on multi-agent cooperation is characterized by comprising a data agent, a factor agent, a model agent, a scheduling agent, a knowledge graph module and a strategy generation module, wherein the data agent, the factor agent, the model agent, the scheduling agent and the strategy generation module are sequentially connected with each other, the knowledge graph module is respectively connected with the data agent, the factor agent, the model agent and the scheduling agent in a data storage and retrieval mode, the strategy generation module is connected with the model agent and the knowledge graph module in a data interaction mode, the data agent is used for collecting, cleaning and standardizing multi-source financial data and generating a shared feature space, the factor agent is used for automatically constructing and screening candidate factors based on the shared feature space and outputting a high-quality factor set, the model agent is used for performing multi-model training and super-parameter optimization based on the high-quality factor set and outputting model training results, the scheduling agent is used for dynamically distributing calculation resources and task priorities based on factor performance indexes output by the factor agent and model performance indexes output by the model agent, the knowledge graph is used for recording experimental parameters, results and dependence relations of each agent and forming a retrievable knowledge research and development network, and strategy generation module is used for developing a strategy research and development network based on the model development and development network.
  2. 2. The multi-agent collaboration-based quantization strategy development system according to claim 1, wherein the data agents perform the operations of collecting quotation, basic surface, macro and other multi-source financial data to form three-dimensional tensor-form original data, processing missing values in the original data by front-back filling and asset grouping median interpolation, unifying the processed original data to transaction day-to-disc time for time alignment, performing depolarization processing on the time-aligned data by a split-bit truncation or absolute median difference method, completing data standardization by a cross section Z-score or rank standardization method, constructing a sliding window map based on a preset window length, generating new factor features and forming a shared feature space.
  3. 3. The multi-agent synergy-based quantization strategy development system of claim 1, wherein the factor agents perform the operations of generating an initial candidate factor pool based on a shared feature space in combination with a symbolic regression, a genetic algorithm, reinforcement learning and a large model hint generation mechanism, calculating an information coefficient, a rank information coefficient and an information coefficient information ratio of each candidate factor in the initial candidate factor pool, eliminating redundant factors in the initial candidate factor pool based on an information coefficient maximum value threshold, screening a high-quality factor set according to a preset rank information coefficient threshold and information coefficient information ratio threshold, and outputting the high-quality factor set to the model agents.
  4. 4. The multi-agent collaboration-based quantization strategy development system according to claim 1, wherein the model agents comprise a traditional machine learning model and a deep learning time sequence model, the model agents receive a high-quality factor set and divide the high-quality factor set into a training data set and a verification data set, train each model in a time sequence rolling cross verification mode, take pairwise ordering loss as a target function in the training process, calculate information coefficients, rank information coefficients, annual yield, information ratios and maximum withdrawal of each model after training, select an optimal model structure based on performance indexes of each model and complete super-parameter optimization, and output model training results to the scheduling agent and strategy generation module.
  5. 5. The multi-agent collaboration-based quantization strategy development system of claim 3, wherein the factor agents calculate a composite score for each candidate factor by a factor quality composite score formula: ; Wherein, the The comprehensive score of the candidate factors is dimensionless; the information coefficient is information, the weight coefficient is compared with the information coefficient, the dimensionless value is 0.3-0.5; The information coefficient mean value of the candidate factors is dimensionless, and the linear correlation degree of the factor predicted value and the actual benefits is reflected; the standard deviation of the information coefficient which is the candidate factor has no dimension and reflects the fluctuation degree of the information coefficient; The value range is 0.2-0.4 for the rank information coefficient weight coefficient without dimension; The rank information coefficient mean value of the candidate factors is dimensionless, and the correlation degree of the factor predicted value rank order and the actual gain rank order is reflected; The value range is 0.2-0.3 for the gain stability weight coefficient without dimension; The maximum withdrawal corresponding to the candidate factors is dimensionless, expressed in decimal form and reflects the maximum withdrawal amplitude of the factor benefits; The annual income ratio corresponding to the candidate factors is expressed in the form of decimal, and the annual income level of the factors is reflected.
  6. 6. The multi-agent collaboration-based quantization strategy development system of claim 5, wherein the model agents calculate the fitness of each model to the set of merit factors by a model fitness formula, the model fitness formula being: ; Wherein, the Is a model Is dimensionless; The value range is 0.6-0.8 for the factor-model correlation weight coefficient without dimension; is a set of quality factors, is dimensionless, and represents a set of quality factors; the comprehensive score of the candidate factors is dimensionless; Is a model Based on factors The correlation coefficient of the prediction result and the actual income data is dimensionless and reflects the prediction accuracy; The number of factors contained in the high-quality factor set is dimensionless, and represents the number of factors in the factor set; The model efficiency weight coefficient is dimensionless, and the value range is 0.2-0.4; Is a model Is dimensionless and reflects excess benefits brought by model unit risks; Is a model The training time length of (2) is expressed in hours, and reflects the time cost of model training.
  7. 7. The multi-agent collaboration-based quantization strategy development system of claim 6, wherein the scheduling agent calculates a factor exploration and model optimization resource allocation ratio through a resource allocation weight formula, and the resource allocation weight formula is: ; Wherein, the The value range is 0-1 in order to allocate the resource weight for factor exploration, without dimension; The resource weight optimized for the model is dimensionless; Searching priority coefficients for factors, wherein the value range is 0.4-0.6 without dimension; The maximum factor comprehensive score of the high-quality factor set is dimensionless; the variance of the factor comprehensive score of the high-quality factor set is dimensionless, and the discrete degree of the factor comprehensive score is reflected; optimizing priority coefficient for the model, wherein the value range is 0.4-0.6 without dimension; maximum adaptation degree of all models is zero; and the standard deviation of the adaptation degree of all the models is dimensionless, and the fluctuation degree of the adaptation degree of the models is reflected.
  8. 8. The multi-agent collaboration-based quantization strategy development system of claim 1, wherein the knowledge graph module comprises knowledge nodes and relationship types, the knowledge nodes comprise identification, type, title, description, task features, codes, performance indexes, states, error trajectories, dependency relationships, labels, sources and test information, the relationship types comprise implementation, derivation, dependence, evaluation, restoration and current optimization, the knowledge graph module adopts a search algorithm combining SBERT embedding, FAISS vector search and metadata filtering, calculates a combined score and returns a Top-5 search result, and the combined score is a weighted sum of cosine similarity and index score.
  9. 9. A quantization strategy development method based on multi-agent cooperation, which is applied to the quantization strategy development system based on multi-agent cooperation as claimed in any one of claims 1 to 8, and is characterized by comprising the following steps: S1, collecting quotation, basic surface, macroscopic and other multi-source financial data by a data agent, carrying out missing value processing, time alignment, extremum removal and standardization processing on the collected original data, and constructing sliding window mapping based on a preset window length to generate new factor characteristics so as to form a shared characteristic space; S2, combining a factor agent with a symbolic regression, a genetic algorithm, reinforcement learning and large model prompt generation mechanism, generating an initial candidate factor pool based on a shared feature space, calculating an information coefficient, a rank information coefficient and an information coefficient information ratio of each candidate factor, removing redundant factors and screening a high-quality factor set; s3, the model intelligent body receives the high-quality factor set, divides the high-quality factor set into a training data set and a verification data set, trains a traditional machine learning model and a deep learning time sequence model by adopting a time sequence rolling cross verification mode, takes pairwise ordering loss as an objective function, completes super-parameter optimization and outputs a model training result; S4, scheduling the agent to acquire factor performance indexes output by the factor agent and model performance indexes output by the model agent, calculating resource allocation proportion through a resource allocation weight formula, and dynamically allocating calculation resources and task priorities; S5, the knowledge graph module records experimental parameters, results and dependency relations in the S1-S4 process to form a research and development knowledge network, and knowledge multiplexing support is provided through a search algorithm; And S6, the strategy generation module is used for automatically generating a deployable quantization strategy code and a document based on the model training result and the research and development knowledge network and packaging the optimal factor-model combination.
  10. 10. The multi-agent cooperation based quantization strategy research and development device is applied to the multi-agent cooperation based quantization strategy research and development system according to any one of claims 1-8, and is characterized by comprising a processor, a memory and a computer program which is stored in the memory and can run on the processor, wherein the processor is used for realizing the following functional modules when executing the computer program, namely a data processing module for executing acquisition, cleaning, time alignment, extremum removal, standardization and sliding window mapping processing of multi-source financial data to generate a shared feature space, a factor processing module for executing generation of candidate factors, redundancy elimination and high-quality factor screening to output a high-quality factor set, a model processing module for executing training, verification, super-parameter optimization and optimal model selection of a multi-model to output model training results, a scheduling module for executing dynamic allocation of computing resources and task priorities to adjust research and development directions based on factor performance indexes and model performance indexes, a knowledge storage and retrieval module for executing recording and retrieval of experimental parameters, results and dependency relations to form a reusable research and development knowledge network, and factor processing module for executing packaging strategy of optimal factor-combination and strategy deployment of a strategy and a model.

Description

Quantization strategy research and development system, method and device based on multi-agent cooperation Technical Field The invention relates to the field of artificial intelligence and financial science and technology, in particular to a quantization strategy research and development system, method and device based on multi-agent cooperation. Background In the current quantitative investment field, a significant technical bottleneck still exists in a quantitative strategy research and development process, specifically, the prior art relies on manual staged operation, researchers need to complete data cleaning, factor construction, model training and back test in sequence, a serial process leads to long research and development period and difficult dynamic adjustment, meanwhile, data processing and factor and model development links are split, data change cannot be fed back to the factor construction and model optimization process in real time, adaptability of factors and models is poor, strategy generalization capability is insufficient, in addition, the prior AutoML tool can automatically search a model structure, non-stationarity and high noise characteristics of financial time sequence data cannot be adapted, task decomposition and cooperation mechanisms among multiple modules are lacked, and factor achievement and model tuning experience accumulated in the research and development process are difficult to be subjected to structural precipitation, so that repeated development and resource waste are caused. Based on the above-mentioned problems, a quantization strategy research and development scheme capable of realizing multi-loop automation collaboration, dynamic optimization and supporting knowledge multiplexing is needed to solve the problems of low research and development efficiency and poor strategy stability in the prior art. Disclosure of Invention The invention aims to solve the defects existing in the prior art, and provides a quantization strategy research and development system based on multi-agent cooperation, which comprises a data agent, a factor agent, a model agent, a scheduling agent, a knowledge graph module and a strategy generation module, wherein the data agent, the factor agent, the model agent and the scheduling agent are sequentially connected in a data interaction mode, the knowledge graph module is respectively connected with the data agent, the factor agent, the model agent and the scheduling agent in a data storage and retrieval mode, the strategy generation module is connected with the model agent and the knowledge graph module in a data interaction mode, the data agent is used for collecting, cleaning and standardizing multi-source financial data and generating a shared feature space, the factor agent is used for automatically constructing and screening candidate factors based on the shared feature space and outputting a high-quality factor set, the model agent is used for multi-mode training and super-parameter optimization based on the high-quality factor set and outputting a model training result, the scheduling agent is used for dynamically distributing computing resources and task levels based on the factor performance index output by the factor agent and the model performance index output by the model agent, the knowledge graph module is used for recording parameters and relations of each agent and forming a research and development strategy, and development network strategy can be used for developing and developing a network strategy. The data agent preferably collects market, basic surface, macroscopic and alternative multi-source financial data to form three-dimensional tensor original data, processes missing values in the original data by front and back filling and asset grouping median interpolation, time aligns the processed original data to trade day closing time, degerming the time aligned data by adopting a split-bit truncation or absolute median difference method, and then completes data standardization by a cross section Z-score or rank standardization mode, builds sliding window mapping based on preset window length, generates new factor characteristics and forms a shared characteristic space. Further preferably, the factor agent performs the operations of combining a symbolic regression, a genetic algorithm, reinforcement learning and a large model prompt generation mechanism, generating an initial candidate factor pool based on a shared feature space, calculating an information coefficient, a rank information coefficient and an information coefficient information ratio of each candidate factor in the initial candidate factor pool, eliminating redundant factors in the initial candidate factor pool based on an information coefficient maximum value threshold, screening a high-quality factor set according to a preset rank information coefficient threshold and information coefficient information ratio threshold, and outputting the high-quality factor set t