Quantitative strategy research and development system, method and device based on multi-agent collaboration

Abstract

The invention relates to the field of artificial intelligence and financial science and technology, and discloses a quantitative strategy research and development system, method and device based on multi-agent collaboration.The system comprises a data agent, a factor agent, a model agent, a scheduling agent, a knowledge graph module and a strategy generation module, and all the modules cooperate to form a research and development closed loop. The data agent processes 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 reuse knowledge, and the strategy generation module outputs deployable strategies. The system solves the problems of insufficient automation and poor collaboration of existing quantitative research and development, and improves the research and development efficiency and strategy stability.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and financial technology, and in particular to a quantitative strategy development system, method and apparatus based on multi-agent collaboration. Background Technology

[0002] In the current field of quantitative investment, there are still significant technical bottlenecks in the development process of quantitative strategies. Specifically, existing technologies mostly rely on manual, phased operations. Researchers need to complete data cleaning, factor construction, model training, and backtesting sequentially. This serial process leads to long development cycles and makes dynamic adjustments difficult. At the same time, data processing is separated from factor and model development. Data changes cannot be fed back to the factor construction and model optimization processes in real time, resulting in poor factor and model adaptability and insufficient strategy generalization ability. In addition, although existing AutoML tools can automatically search for model structures, they cannot adapt to the non-stationarity and high noise characteristics of financial time series data, and lack task decomposition and collaboration mechanisms between multiple modules. Factor results and model tuning experience accumulated during the development process are difficult to be structured and accumulated, leading to redundant development and waste of resources.

[0003] Based on the above problems, there is an urgent need for a quantitative strategy development solution that can achieve automated collaboration across multiple stages, dynamic optimization, and support knowledge reuse, in order to solve the problems of low development efficiency and poor strategy stability in existing technologies. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a quantitative strategy development system based on multi-agent collaboration, including a data agent, a factor agent, a model agent, a scheduling agent, a knowledge graph module, and a strategy generation module. The data agent, factor agent, model agent, and scheduling agent sequentially establish data interaction connections. The knowledge graph module establishes data storage and retrieval connections with the data agent, factor agent, model agent, and scheduling agent respectively. The strategy generation module establishes data interaction connections with the model agent and the knowledge graph module. The data agent is used for the collection, cleaning, and labeling of multi-source financial data. The system standardizes and generates a shared feature space; a factor agent automatically constructs and filters candidate factors based on the shared feature space and outputs a high-quality factor set; a model agent performs multi-model training and hyperparameter optimization based on the high-quality factor set and outputs the model training results; a scheduling agent dynamically allocates computing resources and task priorities based on the factor performance indicators output by the factor agent and the model performance indicators output by the model agent; a knowledge graph module records the experimental parameters, results, and dependencies of each agent and forms a searchable R&D knowledge network; and a policy generation module outputs deployable quantifiable policies based on the model training results and the R&D knowledge network.

[0005] Preferably, the data intelligence agent performs the following operations: collects market data, fundamental data, macroeconomic data, and alternative financial data from multiple sources to form raw data in the form of a three-dimensional tensor; processes missing values ​​in the raw data using pre- and post-padding and asset grouping median interpolation; aligns the processed raw data to the closing time of the trading day; removes extreme values ​​from the time-aligned data using quantile truncation or absolute median difference methods, and then standardizes the data using cross-sectional Z-score or rank standardization; constructs a sliding window mapping based on a preset window length to generate new factor features and form a shared feature space.

[0006] In a further preferred embodiment, the factor agent performs the following operations: combining symbolic regression, genetic algorithm, reinforcement learning, and large model prompt generation mechanism, an initial candidate factor pool is generated based on the shared feature space; the information coefficient, rank information coefficient, and information coefficient-to-information ratio of each candidate factor in the initial candidate factor pool are calculated; redundant factors in the initial candidate factor pool are removed based on the maximum information coefficient threshold; and a high-quality factor set is selected and output to the model agent according to the preset rank information coefficient threshold and information coefficient-to-information ratio threshold.

[0007] In a further preferred embodiment, the model agent comprises traditional machine learning models and deep learning time series models. The model agent performs the following operations: receiving a high-quality factor set and dividing it into training and validation datasets; training each model using time series rolling cross-validation, with pairwise ranking loss as the objective function during training; calculating the information coefficient, rank information coefficient, annualized return, information ratio, and maximum drawdown of each model after training; selecting the optimal model structure based on the performance metrics of each model and performing hyperparameter optimization; and outputting the model training results to the scheduling agent and policy generation module.

[0008] In a further optimized approach, the factor agent calculates the comprehensive score for each candidate factor using a factor quality comprehensive scoring formula, which is as follows: ; in, The overall score of the candidate factors is dimensionless. Information coefficient, information ratio weighting coefficient, dimensionless, with a value range of 0.3-0.5; The mean information coefficient of the candidate factors is dimensionless and reflects the degree of linear correlation between the predicted value of the factors and the actual return. The standard deviation of the information coefficient of the candidate factor is dimensionless and reflects the degree of fluctuation of the information coefficient. The rank information coefficient and weight coefficient are dimensionless and range from 0.2 to 0.4. The mean of the rank information coefficients of the candidate factors is dimensionless and reflects the correlation between the rank of the predicted factor value and the rank of the actual return. This is the stability weighting coefficient for returns, dimensionless, with a value range of 0.2-0.3; This represents the maximum drawdown corresponding to the candidate factor, which is dimensionless and expressed as a decimal, reflecting the maximum decline in factor returns. This represents the annualized return of the candidate factor, dimensionless and expressed as a decimal, reflecting the factor's annual return level.

[0009] In a further optimized approach, the model agent calculates the fitness of each model with the high-quality factor set using the model fitness formula, which is: ; in, For the model The fit is dimensionless; The factor-model correlation weight coefficient is dimensionless and ranges from 0.6 to 0.8. It is a set of high-quality factors, dimensionless, representing the collection of high-quality factors; The overall score of the candidate factors is dimensionless. For the model Based on factors The correlation coefficient between the predicted results and the actual revenue data is dimensionless and reflects the accuracy of the prediction. The number of factors contained in a high-quality factor set, dimensionless, representing the number of factors in the factor set; This is the model efficiency weight coefficient, dimensionless, with a value range of 0.2-0.4; For the model The information ratio is dimensionless and reflects the excess return per unit of risk in the model. For the model The training duration, expressed in hours, reflects the time cost of model training.

[0010] In a further optimized approach, the scheduling agent calculates the resource allocation ratio between factor exploration and model optimization using a resource allocation weighting formula. The resource allocation weighting formula is as follows: ; in, The resource weights assigned to factor exploration are dimensionless and range from 0 to 1. The resource weights assigned to the model optimization are dimensionless. The priority coefficient for factor exploration is dimensionless and ranges from 0.4 to 0.6. The maximum factor composite score of the high-quality factor set, dimensionless; The variance of the factor composite score of the high-quality factor set is dimensionless and reflects the dispersion of the factor composite score. The priority coefficients for model optimization are dimensionless and range from 0.4 to 0.6. The maximum fit of all models, dimensionless; represents the standard deviation of the fitness of all models, which is dimensionless and reflects the degree of fluctuation in model fitness.

[0011] Further optimized, the knowledge graph module includes knowledge nodes and relationship types. Knowledge nodes include identifier, type, title, description, task features, code, performance indicators, status, error trajectory, dependency, tag, source, and test information. Relationship types include implementation, derived from, dependent on, evaluation, repair, and current best. The knowledge graph module uses a retrieval algorithm that combines SBERT embedding, FAISS vector retrieval, and metadata filtering to calculate a combined score and return the Top-5 retrieval results. The combined score is a weighted sum of cosine similarity and indicator scores.

[0012] A further preferred embodiment is a method for developing quantization strategies based on multi-agent collaboration, applied to the quantization strategy development system based on multi-agent collaboration described above, characterized by comprising the following steps: S1: The data intelligence agent collects market data, fundamental data, macroeconomic data, and alternative financial data from multiple sources. It performs missing value processing, time alignment, extremum removal, and standardization on the collected raw data. Based on a preset window length, it constructs a sliding window mapping to generate new factor features and forms a shared feature space. S2: The factor agent combines symbolic regression, genetic algorithm, reinforcement learning and large model prompt generation mechanism to generate an initial candidate factor pool based on the shared feature space, calculates the information coefficient, rank information coefficient and information coefficient information ratio of each candidate factor, eliminates redundant factors and selects a high-quality factor set. S3: The model agent receives a high-quality factor set, divides it into a training dataset and a validation dataset, and uses time series rolling cross-validation to train the traditional machine learning model and the deep learning time series model. The pairwise ranking loss is used as the objective function to complete the hyperparameter optimization and output the model training results. S4: The scheduling agent obtains the factor performance index output by the factor agent and the model performance index output by the model agent, calculates the resource allocation ratio through the resource allocation weight formula, and dynamically allocates computing resources and task priorities. S5: The knowledge graph module records the experimental parameters, results, and dependencies from S1 to S4, forming a research and development knowledge network, and provides knowledge reuse support through retrieval algorithms; S6: The strategy generation module encapsulates the optimal factor-model combination based on model training results and R&D knowledge network, and automatically generates deployable quantitative strategy code and documentation.

[0013] A quantitative strategy development device based on multi-agent collaboration, applied to the quantitative strategy development system based on multi-agent collaboration as described above, is characterized by comprising a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the following functional modules: a data processing module for collecting, cleaning, time-aligning, removing extrema, standardizing, and sliding window mapping of multi-source financial data to generate a shared feature space; a factor processing module for generating candidate factors, removing redundancy, and screening high-quality factors to output a high-quality factor set; a model processing module for training, validating, optimizing hyperparameters, and selecting the optimal model for multiple models to output model training results; a scheduling module for dynamically allocating computing resources and task priorities, and adjusting the development direction based on factor performance indicators and model performance indicators; a knowledge storage and retrieval module for recording and retrieving experimental parameters, results, and dependencies to form a reusable development knowledge network; and a strategy output module for encapsulating the optimal factor-model combination to generate deployable quantitative strategy code and documentation.

[0014] Technical effects: This invention constructs a closed-loop R&D system based on multi-agent collaboration. Its core innovative technologies lie in the dynamic collaborative mechanism among data agents, factor agents, model agents, and scheduling agents, as well as the knowledge reuse design of the knowledge graph module. This system addresses the main problems in existing technologies, such as insufficient automation of the entire quantitative R&D process, poor collaboration between factors and models, and difficulty in knowledge reuse. It automates data processing and strategy generation, improving R&D efficiency and strategy stability. Attached Figure Description

[0015] Figure 1 This is a connection block diagram of the quantitative strategy development system based on multi-agent collaboration in this application; Figure 2 This is a flowchart of the quantitative strategy development method based on multi-agent collaboration proposed in this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Existing technologies suffer from technical problems such as fragmented data processing, factor construction, model training, and resource scheduling in the quantitative strategy development process, lack of automated collaboration mechanisms, resulting in long development cycles, poor strategy stability, and the inability to effectively reuse development knowledge.

[0018] Based on this, please refer to Figure 1 and Figure 2 This embodiment provides a quantitative strategy development system based on multi-agent collaboration, including a data agent, a factor agent, a model agent, a scheduling agent, a knowledge graph module, and a strategy generation module. The data agent, factor agent, model agent, and scheduling agent establish real-time data interaction connections sequentially through message queues. The message queues adopt a distributed architecture, supporting high-throughput, low-latency transmission. Data transmission uses a standardized format, with timestamps and checksums to ensure data integrity and temporal consistency. The knowledge graph module establishes bidirectional data storage and retrieval connections with the data agent, factor agent, model agent, and scheduling agent through a shared storage system. The shared storage system integrates distributed file storage and a graph database, supporting unified storage of structured and unstructured data, and has data version management capabilities, allowing traceability of historical data from different development stages. The strategy generation module establishes data interaction connections with the model agent and knowledge graph module through standardized data interfaces. These interfaces support real-time data transmission and batch synchronization, and have built-in data verification logic to automatically filter invalid or abnormal data.

[0019] The data intelligence agent focuses on the standardized processing of multi-source financial data throughout the entire process. Specifically, it connects to four types of data sources—market data, fundamental data, macro data, and alternative data—through pre-defined standardized API interfaces. Market data sources cover real-time trading data such as opening prices, closing prices, highest prices, lowest prices, and trading volumes for financial products like stocks and futures. Fundamental data sources include core financial indicators such as return on equity, price-to-earnings ratio, price-to-book ratio, and net profit growth rate for listed companies. Macro data sources involve economic data such as GDP, consumer price index, and producer price index. Alternative data sources include news and public opinion data, market microstructure data, etc. The data collection frequency dynamically adapts to the data source type: market data is collected at the minute level, fundamental data at the daily level, macro data at the weekly level, and alternative data at the hourly level. After collection, all heterogeneous data are uniformly converted into raw data in the form of a three-dimensional tensor. The tensor dimension is defined as asset quantity × time series length × feature dimension, ensuring structural consistency in subsequent processing. The data agent performs multi-step cleaning operations on the raw data, using a combination of fore-and-after padding and asset grouping median imputation to handle missing values. For data with no more than three consecutive missing data points in the time series, forward or backward padding is used. For data with more than three consecutive missing data points, the data is grouped by industry, and the median of the corresponding feature for that industry is calculated for imputation, ensuring that the data integrity meets the requirements of subsequent processing. Then, the timestamps of all data are uniformly calibrated to the closing time of the trading day. Data from non-trading days is marked and filtered, and extreme values ​​are removed using quantile truncation or absolute median difference methods. Quantile truncation uses a fixed quantile value as the threshold, while the absolute median difference method dynamically calculates the threshold based on the data distribution characteristics. After removing extreme values, data standardization is completed using cross-sectional Z-score or rank standardization to eliminate the influence of dimensions. Finally, a sliding window mapping is constructed based on a preset window length to generate new factor features, forming a structured shared feature space. The shared feature space adopts a partitioned storage strategy, dividing the data by time dimension and asset type, supporting concurrent reading by multiple modules.

[0020] The factor agent, based on the shared feature space output by the data agent, automatically constructs, filters, and optimizes candidate factors through a multi-mechanism fusion. The factor generation stage combines four mechanisms: symbolic regression, genetic algorithm, reinforcement learning, and large model suggestion generation. Symbolic regression is used to construct linear and simple nonlinear factor expressions, generating initial factors by randomly combining basic features and mathematical operators. The genetic algorithm sets a fixed population size and number of generations, optimizing the factor structure and improving predictive ability through crossover and mutation operations. Reinforcement learning uses factor performance indicators as reward signals to dynamically adjust the factor exploration direction and discover effective factors that conform to market rules. The large model suggestion generation mechanism, based on a pre-trained model in the financial field, takes feature descriptions and market logic prompts as input to generate economically meaningful factor expressions. The initially generated candidate factors constitute a large initial factor pool. The factor agent filters the initial factor pool by calculating performance indicators such as the information coefficient, rank information coefficient, and information coefficient-to-information ratio for each candidate factor. Highly redundant factors are removed based on a maximum information coefficient threshold, and high-performing factors are retained according to preset rank information coefficient and information coefficient-to-information ratio thresholds, forming the final high-quality factor set, which is then output to the model agent.

[0021] The model agent incorporates two main model libraries: traditional machine learning models and deep learning time-series models. Traditional machine learning models include gradient boosting tree models such as LightGBM, XGBoost, and CatBoost, while deep learning time-series models include neural network models such as Transformer, LSTM, and GAT. After receiving a high-quality factor set, the model agent first divides it into training and validation datasets according to time series. The division ratio is dynamically adjusted based on the amount of data to ensure the sufficiency of training data and the reliability of validation results. The training process employs rolling cross-validation of time series data to avoid data leakage, using pairwise ranking loss as the objective function to adapt to the return ranking requirements of quantization strategies. During training, a hyperparameter optimization mechanism is activated, setting up a dedicated hyperparameter search space for different model types. The optimal hyperparameter combination is found through a combination of grid search and random search. Simultaneously, performance metrics such as information coefficient, rank information coefficient, annualized return, information ratio, and maximum drawdown are calculated in real time. Based on the metric evaluation results, the optimal model structure is selected. After training is complete, the model training results are output to the scheduling agent and policy generation module.

[0022] The scheduling agent monitors the factor performance metrics output by the factor agent and the model performance metrics output by the model agent in real time, constructs a comprehensive performance evaluation vector, and dynamically calculates the resource allocation ratio between factor exploration and model optimization using a preset resource allocation algorithm. When the improvement in factor performance metrics exceeds a threshold, the proportion of computing resources allocated to the factor exploration stage is increased; when the improvement in model performance metrics meets the target, resources are allocated primarily to the model optimization stage, ensuring that computing resources are tilted towards high-potential stages. The scheduling agent achieves dynamic resource scheduling through a distributed computing framework, supporting flexible allocation of CPU and GPU resources, and automatically selects appropriate computing resources based on task type to improve resource utilization.

[0023] The knowledge graph module records the experimental parameters, results, and dependencies of each agent during the R&D process, forming a searchable R&D knowledge network. Knowledge nodes include complete fields such as identifier, type, title, description, task features, code, performance metrics, state, error trajectory, dependencies, tags, source, and testing information, comprehensively covering the entire R&D process. Relationship types include implementation, derived from, dependent on, evaluated, repaired, and current best, clearly defining the logical connections between nodes. The knowledge graph module employs a retrieval algorithm combining SBERT embedding, FAISS vector retrieval, and metadata filtering. During retrieval, the task description is first converted into an embedding vector. FAISS vector retrieval quickly matches similar knowledge nodes, and metadata filtering filters effective results. When calculating the combined score, cosine similarity reflects the semantic matching degree, and the metric score reflects the performance fit. Finally, the top-5 retrieval results are returned, providing knowledge reuse support for subsequent R&D.

[0024] After receiving the optimal model training results output by the model agent, the strategy generation module retrieves historical best factor-model combination cases from the knowledge graph module for comparative analysis. A weighted fusion mechanism is then used to determine the final factor-model combination scheme, automatically generating deployable quantitative strategy code and accompanying documentation. The strategy code includes complete functional modules such as data reading, factor calculation, model inference, and trade execution, and is compatible with the deployment requirements of mainstream quantitative trading platforms. The documentation details strategy parameter configuration, operating environment requirements, and precautions to ensure the strategy can be directly implemented and applied.

[0025] Existing technologies suffer from problems such as inconsistent formats of multi-source financial data, missing and extreme values, and inconsistent time dimensions, resulting in low data quality and affecting the accuracy of subsequent factor construction and model training.

[0026] Based on this, the data intelligence agent performs the following operations: It collects market data, fundamental data, macroeconomic data, and alternative financial data from multiple sources. The collection process is implemented through standardized API interfaces, automatically connecting to various data sources. After collection, data in different formats is converted into raw data in the form of three-dimensional tensors. The tensor dimensions are uniformly set as asset quantity × time series length × feature dimension to ensure data structure consistency. Missing value handling employs a combination of fore-and-after padding and asset grouping median imputation. For data with no more than three consecutive missing data points in the time series, forward or backward padding is used. For data with more than three consecutive missing data points, the data is grouped by industry, and the median of the corresponding feature for that industry is calculated for imputation, ensuring data integrity. Time alignment operations convert the timestamps of all data sources to the closing time of the trading day. Data from non-trading days is marked as invalid and filtered in subsequent processing to eliminate the impact of time dimension differences. Extreme value removal offers two optional methods: quantile truncation and absolute median difference. Quantile truncation uses a fixed quantile value as the truncation threshold, while the absolute median difference method calculates a dynamic threshold based on the median value and absolute deviation of the data. Both methods can effectively remove outliers. After extremum removal, data standardization is achieved using either cross-sectional Z-score or rank standardization. Z-score standardization calculates the deviation of the data from the mean and divides it by the standard deviation, while rank standardization converts the data into rank values ​​after sorting. Both methods eliminate the influence of dimensions, ensuring the data meets the requirements for subsequent processing. Finally, a sliding window mapping is constructed based on a preset window length to generate new factor features. The sliding window mapping supports various forms such as differencing, moving average, symbolic regression, or lightweight neural networks. By extracting time-series features through the sliding window, the dimensionality of the feature space is enriched, ultimately forming a standardized shared feature space that provides high-quality data input for the factor agent.

[0027] Existing technologies suffer from technical problems such as reliance on human experience in factor construction, high redundancy of generated candidate factors, and lack of objective quality evaluation standards, resulting in low efficiency in screening high-quality factors and affecting strategy performance.

[0028] Based on this, the factor agent performs the following operations. Combining symbolic regression, genetic algorithms, reinforcement learning, and a large model suggestion generation mechanism, an initial candidate factor pool is generated based on a shared feature space. The symbolic regression mechanism generates linear and simple nonlinear factor expressions by randomly combining basic features and mathematical operators; the genetic algorithm sets a fixed population size and number of generations, optimizing the factor structure and improving factor diversity through genetic operations such as crossover and mutation; reinforcement learning uses factor performance indicators as reward signals to dynamically adjust the factor exploration direction and discover high-potential factors; the large model suggestion generation mechanism, based on a pre-trained model in the financial field, inputs feature descriptions and market logic suggestions to generate economically meaningful factor expressions. These four mechanisms work together to ensure the breadth and quality of the initial candidate factor pool. After generating the initial candidate factor pool, the information coefficient, rank information coefficient, and information coefficient-to-information ratio of each candidate factor are calculated. The information coefficient is obtained by calculating the correlation coefficient between the factor's predicted value and the actual return, reflecting the factor's linear predictive ability; the rank information coefficient is obtained by calculating the correlation coefficient between the rank of the factor's predicted value and the rank of the actual return, reflecting the factor's nonlinear predictive ability; and the information coefficient-to-information ratio is obtained by the ratio of the mean to the standard deviation of the information coefficient, reflecting the stability of the factor's prediction. Redundant factors in the initial candidate factor pool are removed based on the maximum information coefficient threshold. When the maximum information coefficient of two factors exceeds the threshold, they are considered highly redundant, and the factor with better performance is retained. A high-quality factor set is then selected based on preset rank information coefficient thresholds and information coefficient-to-information ratio thresholds. Only factors that simultaneously meet both threshold requirements are included in the high-quality factor set, ensuring that the selected factors possess good predictive ability and stability. Finally, the high-quality factor set is output to the model agent.

[0029] Existing technologies suffer from technical problems such as low model prediction accuracy and poor generalization ability, due to model training not being optimized for the characteristics of financial time series data, unreasonable verification methods, and a lack of targeted hyperparameter optimization.

[0030] Based on this, the model agent includes traditional machine learning models and deep learning time-series models. Traditional machine learning models include gradient boosting tree models such as LightGBM, XGBoost, and CatBoost, while deep learning time-series models include neural network models such as Transformer, LSTM, and GAT, covering different types of data adaptation scenarios. The model agent performs the following operations: It receives a high-quality factor set and divides it into training and validation datasets. This division uses a time-series splitting method to ensure the temporal consistency of training and validation data, avoiding data leakage that could affect the model's generalization ability. It trains each model using a time-series rolling cross-validation method, setting a fixed rolling window length and step size. Each training iteration uses data within the window as the training set and continuous data outside the window as the validation set, completing multiple rounds of training and validation sequentially to fully utilize the characteristics of time-series data and improve model stability. During training, it uses pairwise ranking loss as the objective function. Pairwise ranking loss is constructed by calculating the prediction difference between positive and negative sample pairs, adapting to the core requirement of profit ranking in quantification strategies and guiding the model to learn the ranking relationship between factors and profits. After training, the information coefficient, rank information coefficient, annualized return, information ratio, and maximum drawdown of each model are calculated to comprehensively evaluate the model's predictive ability, return level, and risk control capability. Based on the performance metrics of each model, the optimal model structure is selected, and a hyperparameter optimization process is initiated. A dedicated hyperparameter search space is set for different model types, and the optimal hyperparameter combination is found through a combination of grid search and random search. Model performance changes are monitored in real time during hyperparameter optimization to ensure the optimization direction is correct. After hyperparameter optimization, the model training results are output to the scheduling agent and policy generation module to support subsequent resource scheduling and policy generation.

[0031] Existing technologies suffer from the problem that factor quality assessment relies on a single indicator, which fails to comprehensively reflect the predictive power, stability, and profitability of factors, leading to a high misjudgment rate of high-quality factors.

[0032] Based on this, the factor agent calculates the comprehensive score of each candidate factor using the factor quality comprehensive scoring formula, which is: ; The design logic of this formula is based on the core needs of factor evaluation, comprehensively considering three key dimensions: predictive stability, nonlinear predictive ability, and risk-reward ratio, and achieving a comprehensive evaluation through weighted summation. The score represents the overall score of the candidate factors. It is dimensionless, and a higher score indicates a better factor quality. The information coefficient / information ratio weighting coefficient is dimensionless and ranges from 0.3 to 0.5. It is used to adjust the importance of the information coefficient / information ratio in the overall score and can be flexibly adjusted according to different market environments and R&D needs. It can be appropriately increased in markets with strong trends. The value is chosen to emphasize the predictive stability of the factor; The mean information coefficient of the candidate factors is dimensionless. It is obtained by calculating the correlation coefficient between the daily predicted value of the factor and the actual return on the next day, and then taking the mean value. It reflects the degree of linear correlation between the predicted value and the actual return of the factor. The closer the value is to 1 or -1, the stronger the linear correlation. The standard deviation of the information coefficient of the candidate factor is dimensionless and is obtained by calculating the daily standard deviation of the information coefficient, reflecting the degree of fluctuation of the information coefficient. The smaller the value, the higher the stability of the factor prediction; The information coefficient and information ratio constitute the comprehensive reflection of the linear predictive ability and stability of the factor, and are one of the core indicators for evaluating the quality of the factor. This is the rank information coefficient weighting coefficient, dimensionless, with a value range of 0.2-0.4. It is used to adjust the proportion of the rank information coefficient in the comprehensive score, and can be improved when the market structure is complex and the nonlinear characteristics are obvious. Values; The mean of the rank information coefficients of the candidate factors is dimensionless. It is obtained by calculating the correlation coefficient between the daily predicted rank of the factor and the actual return rank of the next day and then taking the mean. It reflects the degree of correlation between the predicted rank of the factor and the actual return rank, which makes up for the insufficiency of the information coefficient in reflecting only a linear relationship and captures the nonlinear predictive ability of the factor. This is the return stability weighting coefficient, dimensionless, with a value range of 0.2-0.3, used to balance the impact of return stability in the overall score; This represents the maximum drawdown corresponding to the candidate factor. It is dimensionless and expressed as a decimal. It is obtained by calculating the maximum drawdown during the cumulative return of the factor and reflects the risk level of the factor's return. The annualized rate of return corresponding to the candidate factor is dimensionless and expressed in decimal form. It is obtained by converting the cumulative return of the factor into the annual return level, reflecting the profitability of the factor. A return stability indicator is constructed by the ratio of maximum drawdown to annualized return. The smaller the ratio, the higher the return stability. This indicator combines the profitability of the factor with the risk level, avoiding the pursuit of high returns while ignoring risks.

[0033] In practical applications, the first step is to calculate the value of each candidate factor. , , , and Based on these basic indicators, further determinations will be made according to current R&D needs. , , The specific value is obtained by substituting it into the formula. Finally, according to The numerical values ​​of the factors are used to rank candidate factors, and the factors with the best overall quality are selected. This formula comprehensively evaluates factor quality through the synergistic effect of multiple dimensions, effectively reduces the misjudgment rate of high-quality factors, and provides an objective and scientific basis for factor selection.

[0034] Existing technologies suffer from the problem that model and factor fit assessment focuses only on prediction accuracy while ignoring model training efficiency. This results in selected models having high performance but high training costs, failing to meet the technical requirements of real-time research and development.

[0035] Based on this, the model agent calculates the fitness of each model with the high-quality factor set using the model fitness formula, which is: ; The core design principle of this formula is to balance the predictive performance and training efficiency of the model, achieving a comprehensive evaluation of its fit. For the model The fit is dimensionless; the larger the value, the better the model fits the high-quality factor set. The factor-model correlation weighting coefficients are dimensionless and range from 0.6 to 0.8. Since predictive performance is central to model fit, therefore... Values ​​higher than This ensures dominant predictive performance and can be further improved in R&D scenarios where predictive accuracy is extremely critical. Values; It is a high-quality factor set, dimensionless, representing a collection of high-quality factors, including all high-performance factors after screening; The comprehensive score of the candidate factors is dimensionless, meaning it is the result calculated using the factor quality comprehensive score formula, reflecting the quality level of the factors themselves. For the model Based on factors The correlation coefficient between the predicted results and the actual return data is dimensionless. It is obtained by calculating the correlation coefficient between the model's predicted value and the actual return, and reflects the accuracy of the model's prediction using this factor. The closer the correlation coefficient is to 1, the higher the prediction accuracy. As a weighted summation term, it combines factor quality with model prediction accuracy. The higher the quality of a factor, the greater its corresponding prediction accuracy weight, thus comprehensively reflecting the model's overall ability to utilize high-quality factor sets. The number of factors contained in a high-quality factor set is dimensionless and represents the number of factors in the factor set. It is used to normalize the weighted summation result and eliminate the influence of the factor set size on the evaluation result. The mean of factor-model correlation reflects the overall accuracy of the model's predictions using a high-quality factor set. This is a dimensionless model efficiency weighting coefficient, ranging from 0.2 to 0.4. It is used to adjust the impact of model efficiency on fitness evaluation and can be appropriately increased in R&D scenarios with high real-time requirements. Values; For the model The information ratio, a dimensionless value, is the ratio of the mean excess return to the standard deviation of excess return of a model. It reflects the excess return generated per unit of risk of the model. The higher the information ratio, the better the risk-adjusted return of the model. For the model The training duration, measured in hours, is obtained by recording the total time from the start of training to the completion of training, reflecting the time cost of model training. This is a model efficiency metric that comprehensively reflects the model's risk-reward ratio and training efficiency. The higher the metric, the better the risk-adjusted return the model can achieve per unit of time.

[0036] In practical applications, the model is calculated first. For high-quality factor sets Each factor of Combining the factors Calculate the weighted summation term and normalize it to obtain the mean factor-model correlation; then calculate the model... of and The model efficiency index is obtained; finally, it is determined based on the R&D needs. and The value of is substituted into the formula to calculate . This formula can comprehensively evaluate the model's suitability for high-quality factor sets, ensuring both predictive performance and training efficiency, thus meeting the needs of different R&D scenarios.

[0037] Existing technologies suffer from the problem that the allocation of computing resources relies on human experience and cannot be dynamically adjusted based on the real-time performance of factor exploration and model optimization, leading to resource waste or insufficient resources in key areas and affecting R&D efficiency.

[0038] Based on this, the scheduling agent calculates the resource allocation ratio between factor exploration and model optimization using the resource allocation weight formula, which is: ; The design logic of this formula is based on the real-time performance potential of factor exploration and model optimization, dynamically allocating resources to maximize resource utilization. The resource weights assigned to factor exploration are dimensionless and range from 0 to 1. The larger the value, the more resources are allocated to factor exploration; The resource weights assigned to the model optimization are dimensionless and... Complementarity ensures full allocation of resources; This is the factor exploration priority coefficient, dimensionless, ranging from 0.4 to 0.6. It is used to adjust the priority of factor exploration in resource allocation and can be flexibly adjusted according to the research and development stage. It can be appropriately increased in the early stages of research and development. Value selection focuses on factor exploration; The maximum factor composite score of the high-quality factor set, dimensionless, i.e., the highest factor composite score in the high-quality factor set. The maximum value reflects the best results of the current factor exploration; the larger the value, the higher the potential of factor exploration. The variance of the comprehensive factor scores of the high-quality factor set is dimensionless. This variance is calculated by examining all factors in the high-quality factor set. The variance is obtained, reflecting the dispersion of the factor composite score. The larger the variance, the greater the quality difference within the factor set, indicating that there is room for further optimization. This is a potential assessment value for factor exploration, which comprehensively reflects the current achievements and room for optimization in factor exploration. The larger the value, the more worthwhile it is to invest resources in the factor exploration stage. A priority coefficient for model optimization, dimensionless, ranging from 0.4 to 0.6, is used to adjust the priority of model optimization in resource allocation. It can be increased in the later stages of development. The value selection focuses on model optimization. The maximum fitness of all models, dimensionless, meaning the highest fitness among all models. The maximum value reflects the best result of the current model optimization; the larger the value, the better the model optimization effect. The standard deviation of the fit for all models is dimensionless, calculated by all models. The standard deviation is obtained, reflecting the degree of fluctuation in model fitness. The larger the standard deviation, the greater the difference in fitness between models, and the potential for further optimization. This is an evaluation value for the model's optimization potential, which comprehensively reflects the current results and optimization space of the model optimization. The larger the value, the more worthwhile it is to invest resources in the model optimization process.

[0039] In practical applications, the scheduling agent collects high-quality factor sets in real time. With all models Calculations yielded , , , Determined based on the current research and development stage and The value of is substituted into the formula to calculate . And then according to Numerical allocation of computational resources. When the factor exploration potential assessment value is higher than the model optimization potential assessment value... When the value approaches 1, resources are tilted towards factor exploration; conversely, Approaching zero, resources are tilted towards model optimization. This formula achieves optimal resource allocation by dynamically assessing the potential of both parties, thereby improving resource utilization and R&D efficiency.

[0040] Existing technologies suffer from the problem of unstructured storage of experimental parameters, results, and dependencies in the quantitative research and development process. This leads to the inability to retrieve and reuse research and development knowledge, serious duplication of development, and waste of research and development resources.

[0041] Based on this, the knowledge graph module includes knowledge nodes and relationship types. Knowledge nodes include identifiers, types, titles, descriptions, task features, code, performance metrics, status, error paths, dependencies, tags, sources, and testing information, comprehensively covering key information throughout the entire R&D process. The identifier is a unique number for each knowledge node, used to distinguish different nodes; the type defines node attributes, including data nodes, factor nodes, model nodes, task nodes, etc., with different types of nodes containing corresponding exclusive fields; the title is a brief name for the node, facilitating quick identification; the description details the core content of the node, such as describing the data source type and data range for data nodes, and describing the factor expression and generation mechanism for factor nodes; task features record the R&D task attributes corresponding to the node, including task type, execution time, etc.; the code stores the program code related to the node, such as factor calculation code, model training code, etc.; performance metrics record the core performance data of the node, such as the completeness rate of data nodes, the completeness rate of factor nodes, etc. Model nodes The system includes: status indicators for node execution (not executed, executing, successful, failed, etc.); error tracks recording error information and troubleshooting logs during node execution for easy problem tracing; dependencies clearly defining the relationship between nodes and other nodes (e.g., data nodes depend on data nodes, model nodes depend on factor nodes, etc.); tags for classifying and labeling nodes for easy retrieval; sources recording the generating entity of the node (e.g., data nodes from data agents, factor nodes from factor agents, etc.); and test information recording the verification process and results of the node to ensure the reliability of node information.

[0042] Relationship types include Realization, Derived From, Dependent From, Evaluation, Repair, and Current Best. Realization relationships are used to associate task nodes with execution result nodes, describing the specific implementation results of the task. Derived From relationships are used to associate new nodes with original nodes, describing the origin of the node's derivation, such as the optimized factor node being derived from the initial factor node. Dependent From relationships are used to clarify the dependency logic between nodes, such as the model node depending on the factor node and the data node. Evaluation relationships are used to associate evaluation nodes with the evaluated node, describing the evaluation process and results. Repair relationships are used to associate erroneous nodes with repaired nodes, recording the problem repair trajectory. Current Best relationships are used to mark the best node among nodes of the same type, facilitating rapid retrieval.

[0043] The knowledge graph module employs a retrieval algorithm combining SBERT embedding, FAISS vector retrieval, and metadata filtering. During retrieval, the user-input task description is first converted into a 512-dimensional embedding vector using the SBERT model. The SBERT model, pre-trained on financial domain data, accurately captures the semantic information of the task description. Then, the FAISS vector retrieval algorithm quickly searches for similar nodes in the knowledge graph's embedding vector library. FAISS supports efficient retrieval of massive vectors, ensuring fast retrieval response times. Next, metadata filtering filters valid results, setting selection criteria based on node type, tags, performance metrics, etc., to eliminate nodes irrelevant to the task. Finally, a combined score is calculated, which is a weighted sum of cosine similarity and metric scores. Cosine similarity reflects the semantic matching degree between the task description and the node, while the metric score reflects the node's performance level. The weighting coefficients are set according to retrieval requirements to ensure that the retrieval results are both semantically matching and performance-compliant. The final result is a Top-5 retrieval list, providing knowledge reuse support for developers.

[0044] Existing technologies suffer from problems such as sequential steps in quantitative strategy development methods, reliance on manual intervention, and a lack of data feedback mechanisms between steps, leading to rigid development processes, inability to dynamically adjust, and low development efficiency.

[0045] Based on this, this embodiment provides a method for developing a quantization strategy based on multi-agent collaboration, including the following steps: S1, the data intelligence agent collects market data, fundamental data, macroeconomic data, and alternative financial data from multiple sources. Collection is achieved through standardized API interfaces, automatically connecting to various data sources. After collection, the data is converted into raw data in three-dimensional tensor form. Missing value processing is performed on the raw data, using a combination of pre- and post-patch padding and asset group median imputation to ensure data integrity. Data timestamps are then aligned to the closing time of the trading day to filter out invalid data from non-trading days. Outlier removal is performed using quantile truncation or absolute median difference methods. Data standardization is achieved using cross-sectional Z-score or rank standardization to eliminate the influence of dimensions. A sliding window mapping is constructed based on a preset window length to generate new factor features, ultimately forming a shared feature space stored in a distributed shared storage system for subsequent use.

[0046] S2. The factor agent reads shared feature space data through the shared storage system and generates an initial candidate factor pool by combining symbolic regression, genetic algorithm, reinforcement learning, and large model prompt generation mechanisms. It calculates the information coefficient, rank information coefficient, and information coefficient-to-information ratio of each candidate factor in the initial pool. Redundant factors are removed based on the maximum information coefficient threshold. Then, a high-quality factor set is selected according to preset rank information coefficient thresholds and information coefficient-to-information ratio thresholds. This high-quality factor set is output to the model agent, and information such as factor generation parameters, the selection process, and performance indicators are uploaded to the knowledge graph module.

[0047] S3. The model agent receives a high-quality factor set, divides it into training and validation datasets according to time series, and trains traditional machine learning models and deep learning time series models using rolling cross-validation, with pairwise ranking loss as the objective function. During training, a hyperparameter optimization process is initiated, setting up a dedicated hyperparameter search space for different model types, and finding the optimal hyperparameter combination through a combination of grid search and random search. After training, performance metrics such as information coefficient, rank information coefficient, annualized return, information ratio, and maximum drawdown are calculated for each model. Based on these metrics, the optimal model structure is selected, and the model training results are output to the scheduling agent and policy generation module. Simultaneously, information such as model type, hyperparameter configuration, training process, and performance metrics are uploaded to the knowledge graph module.

[0048] S4. The scheduling agent acquires the factor performance indicators output by the factor agent and the model performance indicators output by the model agent in real time, calculates the factor exploration potential assessment value and the model optimization potential assessment value, and calculates the resource allocation ratio through the resource allocation weight formula. and Computational resources are dynamically allocated based on the resource allocation ratio, and the task priorities of factor exploration and model optimization are adjusted. When the change in factor performance indicators or model performance indicators exceeds the preset threshold, resource reallocation is triggered to ensure that resources are tilted towards high-potential stages. At the same time, information such as resource allocation strategies, allocation results, and adjustment logs are uploaded to the knowledge graph module.

[0049] S5, the knowledge graph module receives experimental parameters, results, and dependencies uploaded by data agents, factor agents, model agents, and scheduling agents. It stores these in a knowledge node format, establishes relationship types between nodes, and forms a research and development knowledge network. When subsequent research and development tasks are initiated, it receives retrieval requests and returns Top-5 retrieval results using a retrieval algorithm combining SBERT embedding, FAISS vector retrieval, and metadata filtering. This provides knowledge reuse support for research and development tasks, while simultaneously recording the retrieval process and reuse results, and updating the knowledge graph.

[0050] S6. The strategy generation module receives the optimal model training results output by the model agent, retrieves historical best factor-model combination cases from the knowledge graph module, and performs comparative analysis. A weighted fusion mechanism is used to determine the final factor-model combination scheme. This mechanism sets weights based on factor performance, model fit, and historical reuse effects to ensure the optimality of the combination scheme. Deployable quantitative strategy code is automatically generated, including complete modules for data reading, factor calculation, model inference, and trade execution, compatible with mainstream quantitative trading platforms. Simultaneously, accompanying documentation is generated, detailing strategy parameter configuration, operating environment requirements, deployment steps, and precautions to facilitate rapid application. Finally, the generated strategy code and documentation are stored in a shared storage system, and relevant information is uploaded to the knowledge graph module.

[0051] Existing technologies suffer from problems such as chaotic functional module division and unclear data interaction between modules in quantitative strategy R&D devices, resulting in low device operating efficiency and an inability to stably support automated R&D processes.

[0052] Based on this, it includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the following functional modules.

[0053] The data processing module is used to collect, clean, time-align, remove extrema, standardize, and perform sliding window mapping on multi-source financial data to generate a shared feature space. The module has a built-in standardized API interface library that supports integration with market data, fundamental data, macroeconomic data, and alternative data sources, automatically adapting to different data formats for batch data collection. After collection, data processing operations are performed according to a preset process. The missing value handling module uses an algorithm combining pre- and post-filling and asset group median interpolation to ensure data integrity; the time alignment module calibrates data timestamps to the closing time of the trading day, filtering out invalid data; the extrema removal module provides two algorithms: quantile truncation and absolute median difference, which can be automatically selected or manually configured according to data characteristics; the standardization module supports two methods: cross-sectional Z-score and rank standardization to eliminate the influence of dimensions; the sliding window mapping module generates new features based on a preset window length through various forms such as differencing and moving averages, ultimately organizing the processed feature data into a shared feature space, storing it in memory, and providing data access interfaces to other modules.

[0054] The factor processing module is used to generate candidate factors, remove redundancies, and select high-quality factors, outputting a high-quality factor set. The module includes a multi-mechanism factor generation unit, a factor performance evaluation unit, and a factor selection unit. The multi-mechanism factor generation unit integrates four mechanisms: symbolic regression, genetic algorithm, reinforcement learning, and large model suggestion generation. It can flexibly select a single mechanism or a combination of mechanisms to generate the initial candidate factor pool according to R&D needs. The factor performance evaluation unit has built-in algorithms for calculating indicators such as information coefficient, rank information coefficient, and information coefficient-to-information ratio, calculating the performance indicators of each candidate factor in real time. The factor selection unit, based on preset thresholds, first removes redundant factors, then selects high-quality factors to form a high-quality factor set, which is output to the model processing module through a data interaction interface, while simultaneously synchronizing relevant information to the knowledge storage and retrieval module.

[0055] The model processing module is used to perform training, validation, hyperparameter optimization, and optimal model selection for multiple models, and outputs the model training results. The model processing module includes a model library, a data partitioning unit, a training unit, a hyperparameter optimization unit, and a model evaluation unit. The model library contains traditional machine learning models and deep learning time-series models, covering various types such as LightGBM, XGBoost, CatBoost, Transformer, LSTM, and GAT. The data partitioning unit divides the high-quality factor set into training and validation datasets according to time series, ensuring time series consistency. The training unit uses time-series rolling cross-validation with pairwise ranking loss as the objective function to perform model training. The hyperparameter optimization unit sets up a dedicated hyperparameter search space for different model types, finding the optimal hyperparameter combination through a combination of grid search and random search. The model evaluation unit calculates performance metrics such as the model's information coefficient, rank information coefficient, annualized return, information ratio, and maximum drawdown, selects the optimal model structure based on these metrics, outputs the model training results to the scheduling and policy output modules, and simultaneously synchronizes relevant information to the knowledge storage and retrieval module.

[0056] The scheduling module dynamically allocates computing resources and task priorities, adjusting the research and development direction based on factor performance indicators and model performance indicators. The scheduling module includes an indicator acquisition unit, a potential assessment unit, and a resource allocation unit. The indicator acquisition unit receives factor performance indicators from the factor processing module and model performance indicators from the model processing module in real time, performing data preprocessing and format conversion. The potential assessment unit calculates factor exploration potential assessment values ​​and model optimization potential assessment values ​​based on the indicator data; these potential assessment values ​​are obtained by integrating current performance results and optimization space. The resource allocation unit calculates the resource allocation ratio using a resource allocation weight formula, dynamically allocates computing resources such as processors and memory according to the ratio, adjusts the task priorities of the factor processing module and the model processing module, ensuring resources are tilted towards high-potential stages, and simultaneously synchronizes the resource allocation information to the knowledge storage and retrieval module.

[0057] The knowledge storage and retrieval module is used to record and retrieve experimental parameters, results, and dependencies, forming a reusable R&D knowledge network. This module includes a data storage unit, a relationship building unit, and a retrieval unit. The data storage unit stores experimental parameters, results, and dependencies uploaded by each module in a knowledge node format, supporting unified storage of structured and unstructured data. The relationship building unit establishes relationship types such as implementation, derived from, and dependent on each node based on their association logic, forming the R&D knowledge network. The retrieval unit uses a retrieval algorithm combining SBERT embedding, FAISS vector retrieval, and metadata filtering. Upon receiving a retrieval request, it quickly matches similar knowledge nodes, filters valid results, calculates a combined score, and returns the Top-5 retrieval results, providing knowledge reuse support for each module.

[0058] The strategy output module is used to encapsulate the optimal factor-model combination, generating deployable quantitative strategy code and documentation. The strategy output module includes a combination optimization unit, a code generation unit, and a documentation generation unit. The combination optimization unit receives the optimal model training results output by the model processing module, retrieves historical best cases from the knowledge storage and retrieval module, and uses a weighted fusion mechanism to determine the final factor-model combination scheme. The code generation unit automatically generates quantitative strategy code based on the combination scheme. The code includes complete functional modules such as data reading, factor calculation, model inference, and trade execution, and is compatible with mainstream quantitative trading platforms. The documentation generation unit generates accompanying documentation, detailing strategy parameter configuration, operating environment requirements, deployment steps, and precautions to facilitate rapid deployment. Finally, the strategy code and documentation are stored in the storage device, and relevant information is synchronized to the knowledge storage and retrieval module.

[0059] The processor, as the core computing unit of the device, adopts a multi-core architecture and supports multi-threaded parallel processing to ensure the efficient operation of each functional module. The memory employs a distributed storage architecture, including cache and persistent storage. The cache stores real-time processed data to improve access speed, while persistent storage stores R&D data, knowledge graphs, strategy codes, and other information to ensure data security and reliability. The processor and memory are connected via a high-speed bus to achieve high-speed data transmission. Each functional module interacts through standardized data interfaces to ensure the accuracy and real-time performance of data transmission. The entire device has a clear structure and complete functionality, capable of stably supporting the fully automated operation of the quantitative strategy R&D process.

[0060] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A quantitative strategy development system based on multi-agent collaboration, characterized in that, It includes a data agent, a factor agent, a model agent, a scheduling agent, a knowledge graph module, and a strategy generation module. The data agent, factor agent, model agent, and scheduling agent establish data interaction connections sequentially. The knowledge graph module establishes data storage and retrieval connections with the data agent, factor agent, model agent, and scheduling agent respectively. The strategy generation module establishes data interaction connections with the model agent and the knowledge graph module. The data agent is used for the collection, cleaning, and standardization of multi-source financial data and the generation of a shared feature space. Factor agents are used to automatically construct and screen candidate factors based on a shared feature space and output a high-quality factor set. The model agent is used to perform multi-model training and hyperparameter optimization based on a high-quality factor set and output the model training results. The scheduling agent is used to dynamically allocate computing resources and task priorities 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 to record the experimental parameters, results, and dependencies of each agent and form a searchable R&D knowledge network; The strategy generation module is used to output deployable quantifiable strategies based on model training results and the R&D knowledge network.

2. The quantitative strategy development system based on multi-agent collaboration according to claim 1, characterized in that, The data intelligence agent performs the following operations: collects market data, fundamental data, macroeconomic data, and alternative financial data from multiple sources to form raw data in the form of a three-dimensional tensor; processes missing values ​​in the raw data using pre- and post-padding and asset grouping median interpolation; aligns the processed raw data to the closing time of the trading day; removes extreme values ​​from the time-aligned data using quantile truncation or absolute median difference methods, and then standardizes the data using cross-sectional Z-score or rank standardization; constructs a sliding window mapping based on a preset window length to generate new factor features and form a shared feature space.

3. The quantitative strategy development system based on multi-agent collaboration according to claim 1, characterized in that, The factor agent performs the following operations: combining symbolic regression, genetic algorithm, reinforcement learning, and large model prompt generation mechanism to generate an initial candidate factor pool based on the shared feature space; Calculate the information coefficient, rank information coefficient, and information coefficient-to-information ratio for each candidate factor in the initial candidate factor pool; Redundant factors in the initial candidate factor pool are removed based on the maximum information coefficient threshold; a high-quality factor set is selected and output to the model agent based on the preset rank information coefficient threshold and information coefficient information ratio threshold.

4. The quantitative strategy development system based on multi-agent collaboration according to claim 1, characterized in that, The model agent includes traditional machine learning models and deep learning time series models. The model agent performs the following operations: receives a high-quality factor set and divides it into training datasets and validation datasets; trains each model using time series rolling cross-validation, with pairwise ranking loss as the objective function during training. Calculate the information coefficient, rank information coefficient, annualized return, information ratio, and maximum drawdown of each model after training; select the optimal model structure based on the performance indicators of each model and complete hyperparameter optimization; output the model training results to the scheduling agent and policy generation module.

5. The quantitative strategy development system based on multi-agent collaboration according to claim 3, characterized in that, The factor agent calculates the overall score for each candidate factor using the factor quality comprehensive scoring formula, which is: ； in, The overall score of the candidate factors is dimensionless. Information coefficient, information ratio weighting coefficient, dimensionless, with a value range of 0.3-0.5; The mean information coefficient of the candidate factors is dimensionless and reflects the degree of linear correlation between the predicted value of the factors and the actual return. The standard deviation of the information coefficient of the candidate factor is dimensionless and reflects the degree of fluctuation of the information coefficient. The rank information coefficient and weight coefficient are dimensionless and range from 0.2 to 0.

4. The mean of the rank information coefficients of the candidate factors is dimensionless and reflects the correlation between the rank of the predicted factor value and the rank of the actual return. This is the stability weighting coefficient for returns, dimensionless, with a value range of 0.2-0.3; This represents the maximum drawdown corresponding to the candidate factor, which is dimensionless and expressed as a decimal, reflecting the maximum decline in factor returns. This represents the annualized return of the candidate factor, dimensionless and expressed as a decimal, reflecting the factor's annual return level.

6. The quantitative strategy development system based on multi-agent collaboration according to claim 5, characterized in that, The model agent calculates the fitness of each model with the set of high-quality factors using the model fitness formula, which is: ； in, For the model The fit is dimensionless; The factor-model correlation weight coefficient is dimensionless and ranges from 0.6 to 0.

8. It is a set of high-quality factors, dimensionless, representing the collection of high-quality factors; The overall score of the candidate factors is dimensionless. For the model Based on factors The correlation coefficient between the predicted results and the actual revenue data is dimensionless and reflects the accuracy of the prediction. The number of factors contained in a high-quality factor set, dimensionless, representing the number of factors in the factor set; This is the model efficiency weight coefficient, dimensionless, with a value range of 0.2-0.4; For the model The information ratio is dimensionless and reflects the excess return per unit of risk in the model. For the model The training duration, expressed in hours, reflects the time cost of model training.

7. The quantization strategy development system based on multi-agent collaboration according to claim 6, characterized in that, The scheduling agent calculates the resource allocation ratio between factor exploration and model optimization using a resource allocation weighting formula, which is: ； in, The resource weights assigned to factor exploration are dimensionless and range from 0 to 1. The resource weights assigned to the model optimization are dimensionless. The priority coefficient for factor exploration is dimensionless and ranges from 0.4 to 0.

6. The maximum factor composite score of the high-quality factor set, dimensionless; The variance of the factor composite score of the high-quality factor set is dimensionless and reflects the dispersion of the factor composite score. The priority coefficients for model optimization are dimensionless and range from 0.4 to 0.

6. The maximum fit of all models, dimensionless; represents the standard deviation of the fitness of all models, which is dimensionless and reflects the degree of fluctuation in model fitness.

8. The quantitative strategy development system based on multi-agent collaboration according to claim 1, characterized in that, The knowledge graph module includes knowledge nodes and relationship types. Knowledge nodes include identifier, type, title, description, task features, code, performance metrics, status, error trajectory, dependencies, tags, source, and test information. Relationship types include implementation, derived from, dependent on, evaluation, repair, and current best. The knowledge graph module uses a retrieval algorithm that combines SBERT embedding, FAISS vector retrieval, and metadata filtering to calculate a combined score and return the Top-5 retrieval results. The combined score is a weighted sum of cosine similarity and metric scores.

9. A method for developing quantization strategies based on multi-agent collaboration, applied to the quantization strategy development system based on multi-agent collaboration as described in any one of claims 1-8, characterized in that, Includes the following steps: S1: The data intelligence agent collects market data, fundamental data, macroeconomic data, and alternative financial data from multiple sources. It performs missing value processing, time alignment, extremum removal, and standardization on the collected raw data. Based on a preset window length, it constructs a sliding window mapping to generate new factor features and forms a shared feature space. S2: The factor agent combines symbolic regression, genetic algorithm, reinforcement learning and large model prompt generation mechanism to generate an initial candidate factor pool based on the shared feature space, calculates the information coefficient, rank information coefficient and information coefficient information ratio of each candidate factor, eliminates redundant factors and selects a high-quality factor set. S3: The model agent receives a high-quality factor set, divides it into a training dataset and a validation dataset, and uses time series rolling cross-validation to train the traditional machine learning model and the deep learning time series model. The pairwise ranking loss is used as the objective function to complete the hyperparameter optimization and output the model training results. S4: The scheduling agent obtains the factor performance index output by the factor agent and the model performance index output by the model agent, calculates the resource allocation ratio through the resource allocation weight formula, and dynamically allocates computing resources and task priorities. S5: The knowledge graph module records the experimental parameters, results, and dependencies from S1 to S4, forming a research and development knowledge network, and provides knowledge reuse support through retrieval algorithms; S6: The strategy generation module encapsulates the optimal factor-model combination based on model training results and R&D knowledge network, and automatically generates deployable quantitative strategy code and documentation.

10. A quantitative strategy development device based on multi-agent collaboration, applied to the quantitative strategy development system based on multi-agent collaboration as described in any one of claims 1-8, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the following functional modules: a data processing module, which is used to perform multi-source financial data acquisition, cleaning, time alignment, extremum removal, standardization, and sliding window mapping processing to generate a shared feature space; The factor processing module is used to generate candidate factors, remove redundancies, and screen high-quality factors, and output a set of high-quality factors. The model processing module is used to perform training, validation, hyperparameter optimization, and optimal model selection for multiple models, and output the model training results; the scheduling module is used to dynamically allocate computing resources and task priorities, and adjust the research and development direction based on factor performance indicators and model performance indicators; the knowledge storage and retrieval module is used to record and retrieve experimental parameters, results, and dependencies, forming a reusable research and development knowledge network. The strategy output module is used to encapsulate the optimal factor-model combination and generate deployable quantitative strategy code and documentation.

Images