Financial management system and method based on ESG dynamic weight distribution

Abstract

The invention discloses a financial management system and method based on ESG dynamic weight distribution. The financial management system comprises a data acquisition module, a preprocessing module, a dynamic matrix construction module, an online weight updating module and a reinforcement learning optimization module. Through multi-source data acquisition and standardization processing, a dynamic ESG material matrix is constructed, and weight adjustment is carried out in combination with market public opinion and policy changes. Through an online gradient descent algorithm and a reinforcement learning mechanism, the system can update and optimize the ESG weight model in real time, thereby providing ESG weight suggestions with timeliness and responsiveness in financial investment decisions. According to the invention, the application efficiency and the intelligent level of the ESG index in financial management are improved.

Description

Technical Field

[0001] The present invention relates to the field of financial management technology, and in particular to a financial management system and method based on ESG dynamic weight allocation. Background Art

[0002] As the concepts of environmental, social, and corporate governance (collectively referred to as ESG) continue to gain importance in the field of financial investment, ESG evaluation indicators have gradually become an important reference for measuring a company's sustainable development capabilities and long-term investment value. At present, the ESG scoring system provided by third-party rating agencies is widely used in the market as an important parameter for financial decision-making and asset allocation. However, most of the existing ESG scores are static scores with long update cycles, making it difficult to timely reflect the dynamic changes in market sentiment, macroeconomic policies, and the actual performance of enterprises. In addition, existing financial management systems generally lack the ability to integrate ESG indicators with multi-source information such as market data, financial data, and policy changes, and to make real-time dynamic adjustments. As a result, ESG scores have problems such as delayed response, poor adaptability, and weak personalization capabilities in actual applications.

[0003] In actual portfolio management, ESG indicators should dynamically adjust their weightings as the external market environment changes, to more accurately reflect a company's current sustainability performance and future potential. However, traditional approaches lack the technical mechanisms for real-time modeling, dynamic adjustment, and strategy optimization of ESG indicators, limiting their in-depth application in scenarios such as quantitative investing and robo-advisory. Therefore, a financial management system that can integrate multi-source data, perform real-time modeling, and support dynamic weighting optimization is urgently needed to enhance the real-time, accuracy, and application value of ESG assessments. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a financial management system and method based on ESG dynamic weight allocation, which is used to improve the real-time and accuracy of ESG indicators.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a financial management system and method based on ESG dynamic weight allocation, comprising:

[0006] The data collection module is used to receive real-time raw data from multiple sources, including third-party ESG rating agencies, market information sources, news and public opinion, and macroeconomic indicators, through a distributed message queue.

[0007] A preprocessing module, connected to the data acquisition module, for cleaning, standardizing, interpolating missing values, and extracting time series window features from the multi-source raw data to generate a standardized feature vector;

[0008] A dynamic matrix construction module, connected to the preprocessing module, is used to construct a dynamic materiality matrix based on preset financial industry materiality indicators, and trigger rules based on data related to market sentiment and policy changes in the standardized feature vector to dynamically adjust the materiality weights of the ESG dimension in the dynamic materiality matrix;

[0009] an online weight updating module, connected to the dynamic matrix building module, configured to take the financial data and ESG raw data in the standardized feature vector and the updated materiality weight as input, and incrementally train the ESG weight model in real time based on an online gradient descent algorithm to generate and publish the latest ESG weight vector and the latest gradient vector;

[0010] a reinforcement learning optimization module, connected to the online weight update module, for iteratively training the policy network according to a preset compound reward function in a custom simulated portfolio environment, and using the latest ESG weight vector, the latest gradient vector, and the market characteristics and ESG status characteristics in the standardized feature vector as environmental inputs to the policy network, so that the policy network outputs ESG weight recommendations;

[0011] The online weight update module dynamically updates the current weight vector, learning rate, regularization coefficient and loss function in the ESG weight model according to the ESG weight recommendation to obtain an optimized ESG weight model.

[0012] Furthermore, it also includes:

[0013] A compliance audit module, connected to the online weight update module, is configured to generate a feature contribution explanation for each weight update based on the received latest ESG weight vector and the latest gradient vector by calling the SHAP algorithm, and automatically fill in the regulatory report template to generate a compliance model document;

[0014] An operation and maintenance management module, connected to the data acquisition module, the preprocessing module, the dynamic matrix construction module, the online weight update module, the reinforcement learning optimization module, and the compliance audit module, for collecting performance indicators of the remaining modules in the financial management system, displaying each performance indicator in real time on a dashboard, and triggering elastic scaling or rollback operations in response to abnormal situations;

[0015] The performance indicators include data delay, training loss, RL cumulative reward and model version information.

[0016] Furthermore, the online weight updating module includes:

[0017] a feature fusion unit, configured to perform ESG dimension weighted fusion processing on the financial data and the ESG raw data in the input standardized feature vector according to the dynamic materiality matrix, whereby the weighted fused vector constitutes the final input feature matrix of the ESG weight model;

[0018] an online incremental training unit, connected to the feature fusion unit, configured to perform a round of incremental update on the current model according to the online gradient descent algorithm each time a new data packet is received;

[0019] The model output unit is connected to the online incremental training unit and is used to output the latest ESG weight vector and the latest gradient vector, and write them into the cache pool and the preset historical database.

[0020] Furthermore, the compound reward function is configured as:

[0021] R t =α·Return t -β·Risk t +γ·ESG_Score t -δ·Penalty t ;

[0022] Among them, ReLurn t represents the return of the portfolio at time step t, Risk t Represents the risk indicator of the investment portfolio, ESG_Score t represents the weighted ESG score of the portfolio at time step t, Penalty t It represents the penalty item for violating compliance requirements, exceeding transaction limits or other abnormal operations. α, β, γ, and δ represent the preset first coefficient, second coefficient, third coefficient, and fourth coefficient, respectively.

[0023] Furthermore, the reinforcement learning optimization module includes:

[0024] An environment simulation unit, used to construct the custom simulated investment portfolio environment that is highly similar to the real financial market, and the custom simulated investment portfolio environment supports the simulation of dynamic changes in asset prices, transaction constraints, policy shocks, and public opinion events;

[0025] a strategy learning unit connected to the environment simulation unit, configured to use the initial strategy network as the core and adopt a preset reinforcement learning algorithm to learn historical weight recommendations under different historical environment states to obtain a preliminary strategy network;

[0026] A reward unit, connected to the strategy learning unit, is used to iteratively train the preliminary strategy network according to the composite reward function to obtain the strategy network after the iteration, and input the latest ESG weight vector, the latest gradient vector and the market characteristics and ESG status characteristics in the standardized feature vector into the strategy network to output the corresponding ESG weight recommendation.

[0027] Furthermore, the expression of the policy network is configured as:

[0028]

[0029] in, represents the ESG weight recommendation, σ(·) represents the normalization function, Used to compress the output to the (0, 1) interval; represents the kth market characteristic or the ESG characteristic value, ζ(·) represents the Riemann Zeta function, which is used to express the complex series relationship between characteristics and weights, Here s=∈ k +1; J n (·) represents a Bessel function of order n, which represents the nonlinear response of the gradient change to the local weight of the policy network; Represents the current k-th gradient component, ∈ m 、 Respectively represent the components in the direction of the maximum modulus in the current weight and gradient, which are used to construct the elliptic function input; cn(u, k) represents the Jacobi elliptic cosine function, which is used to capture the nonlinear periodic feature changes. θ represents the projection angle of the feature rotation dimension, ∈ j represents the jth item of the current ESG weight vector, Indicates that the current eigenvalue is the same as ∈ j Normalized features of the pair, represents the gradient value of the i-th item;

[0030] K(·) represents the elliptic integral of the first kind, which is defined as: Used to comprehensively reflect the nonlinear behavior of the coupling between the overall ESG and characteristics.

[0031] Furthermore, the online incremental training unit includes:

[0032] A first updating subunit is configured to construct a fusion function based on the current weight vector, the ESG weight suggestion, and the current gradient vector, and compress the update result through a normalization function to complete a progressive update of the ESG weight;

[0033] a second updating subunit, configured to dynamically update the learning rate in response to a nonlinear trend suggested by the ESG weight;

[0034] a third updating subunit, configured to adjust the regularization coefficient to suppress model overfitting by constructing a high-order offset response function;

[0035] The fourth updating subunit is used to construct the composite loss function, which recommends deviation from the strategy, L1 regularization term and nonlinear penalty mechanism based on Jacobi elliptic function integral term to perform multi-dimensional optimization control on the model updating process.

[0036] Furthermore, the updating process of the current weight vector is configured as follows:

[0037]

[0038] in, represents the updated current weight vector, σ(·) represents the normalization function, α t The fusion coefficient vector representing the recommended weight ratio of the ESG weight, β t Represents the gradient adjustment coefficient at the current moment, represents the current gradient vector;

[0039] The updating process of the learning rate is configured as follows:

[0040]

[0041] Among them, η t represents the current learning rate, η t+1 represents the updated learning rate, γ′ represents the preset sensitivity adjustment coefficient, represents the standardized feature, μ i represents the value of the i-th dimension in the ESG weight recommendation;

[0042] The updating process of the regularization coefficient is configured as follows:

[0043]

[0044] Among them, λ t Represents the current regularization coefficient, λ t+1 represents the regularization coefficient after update, represents the gradient value of the jth dimension, ρ represents the preset adjustment factor, and n represents the number of feature dimensions;

[0045] The update process of the loss function is configured as follows:

[0046]

[0047] Among them, ∈ i represents the current weight vector of the i-th dimension, and cn(·) represents the Jacobi elliptic cosine function.

[0048] A financial management method based on ESG dynamic weight allocation, applied to the above-mentioned financial management system based on ESG dynamic weight allocation, comprises:

[0049] Step S1: The data collection module receives real-time raw data from multiple sources, including third-party ESG rating agencies, market information sources, news and public opinion, and macroeconomic indicators, through a distributed message queue.

[0050] Step S2: The preprocessing module cleans, standardizes, interpolates missing values, and extracts time series window features from the multi-source raw data to generate a standardized feature vector;

[0051] Step S3: The dynamic matrix construction module constructs a dynamic materiality matrix based on preset financial industry materiality indicators, and triggers rules based on data related to market sentiment and policy changes in the standardized feature vector to dynamically adjust the materiality weights of the ESG dimension in the dynamic materiality matrix;

[0052] Step S4: The online weight update module takes the financial data and ESG raw data in the standardized feature vector and the updated materiality weight as input, and incrementally trains the ESG weight model in real time based on the online gradient descent algorithm to generate and publish the latest ESG weight vector and the latest gradient vector;

[0053] Step S5: The reinforcement learning optimization module iteratively trains the policy network in a custom simulated portfolio environment according to a preset compound reward function, and uses the latest ESG weight vector, the latest gradient vector, and the market characteristics and ESG status characteristics in the standardized feature vector as environmental inputs of the policy network, so that the policy network outputs ESG weight recommendations;

[0054] Step S6: The online weight update module dynamically updates the current weight vector, learning rate, regularization coefficient and loss function in the ESG weight model according to the ESG weight recommendation to obtain an optimized ESG weight model.

[0055] Beneficial effects of the present invention:

[0056] By integrating a distributed message queue architecture, the system can receive data from multiple sources such as third-party ESG rating agencies, market conditions, news and public opinion, and macroeconomics in real time, thereby enhancing the breadth and timeliness of data processing and ensuring the comprehensiveness and timeliness of the ESG analysis foundation.

[0057] At the same time, data-triggered rules based on market sentiment and policy changes are introduced to dynamically adjust the weights of the ESG materiality matrix. This effectively overcomes the problem of the traditional static ESG weight model's delayed response to changes in the market environment, thereby improving the flexibility and adaptability of the ESG assessment model.

[0058] In addition, an online gradient descent algorithm is used to conduct real-time incremental training of the ESG weight model, allowing the model to be continuously optimized based on changing data, thereby improving the timeliness and accuracy of ESG scoring results.

[0059] By building a strategy network and training the ESG weight recommendation model in a customized simulated investment environment, we can achieve intelligent optimization of investment strategies, provide high-quality ESG dimension support for financial investment management, and enhance the sustainability and profitability of strategies.

[0060] In summary, the present invention integrates multiple intelligent modules such as data preprocessing, dynamic adjustment, online optimization and reinforcement learning to build a highly automated and sustainably optimized ESG financial management tool, which significantly improves the financial institutions' sensitive response capabilities and management efficiency to environmental, social and governance factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a schematic diagram of the structure of the financial management system based on ESG dynamic weight allocation in the present invention;

[0062] Figure 2 Schematic diagram of the structure of the online incremental training unit in the present invention;

[0063] Figure 3 This is a flowchart of the steps of the financial management method based on ESG dynamic weight allocation in the present invention.

[0064] Figure numerals: 1. Data acquisition module; 2. Preprocessing module; 3. Dynamic matrix construction module; 4. Online weight update module; 41. Feature fusion unit; 42. Online incremental training unit; 421. First update subunit; 422. Second update subunit; 423. Third update subunit; 424. Fourth update subunit; 43. Model output unit; 5. Reinforcement learning optimization module; 51. Environment simulation unit; 52. Strategy learning unit; 53. Reward unit; 6. Compliance audit module; 7. Operation and maintenance management module. DETAILED DESCRIPTION

[0065] The present invention will be described in further detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom," "top," "inner," and "outer" refer to directions toward or away from the geometric center of a particular component, respectively.

[0066] Example 1, reference Figure 1 , which is the first embodiment of the present invention, provides a financial management system based on ESG dynamic weight allocation, which can significantly improve the dynamic responsiveness and accuracy of ESG indicators, the rationality of weight allocation, and the foresight of investment strategies, including:

[0067] Data collection module 1 is used to receive real-time raw data from multiple sources, including third-party ESG rating agencies, market information sources, news and public opinion, and macroeconomic indicators, through a distributed message queue.

[0068] A preprocessing module 2, connected to the data acquisition module 1, is used to clean, standardize, interpolate missing values, and extract time series window features from the multi-source raw data to generate a standardized feature vector;

[0069] Dynamic matrix construction module 3, connected to the pre-processing module 2, is used to construct a dynamic materiality matrix based on preset financial industry materiality indicators, and trigger rules based on data related to market sentiment and policy changes in the standardized feature vector to dynamically adjust the materiality weights of the ESG dimension in the dynamic materiality matrix;

[0070] An online weight updating module 4, connected to the dynamic matrix construction module 3, is configured to take the financial data and ESG raw data in the standardized feature vector and the updated materiality weight as input, and incrementally train the ESG weight model in real time based on an online gradient descent algorithm to generate and publish the latest ESG weight vector and the latest gradient vector;

[0071] a reinforcement learning optimization module 5, connected to the online weight updating module 4, for iteratively training the policy network according to a preset compound reward function in a custom simulated portfolio environment, and using the latest ESG weight vector, the latest gradient vector, and the market characteristics and ESG status characteristics in the standardized feature vector as inputs to the policy network environment, so that the policy network outputs ESG weight recommendations;

[0072] The online weight update module 4 dynamically updates the current weight vector, learning rate, regularization coefficient and loss function in the ESG weight model according to the ESG weight recommendation to obtain an optimized ESG weight model.

[0073] Working principle of embodiment 1:

[0074] Data Collection Module 1 is deployed on a distributed architecture that supports high-concurrency processing. Leveraging distributed message queuing technologies like Kafka, it receives real-time raw data from multiple sources, including third-party ESG rating agencies (such as MSCI and Sustainalytics), financial market information providers (such as Wind and Bloomberg), news and public opinion platforms, and macroeconomic indicator libraries. This data includes corporate ESG ratings, stock prices, trading volumes, relevant news summaries, policy and regulatory changes, and economic indicators.

[0075] The preprocessing module 2 cleans, denoises, standardizes, and fills missing values in the received multi-source raw data, and uses a sliding time window mechanism to extract time series features, and finally generates a standardized feature vector in a unified format for subsequent processing.

[0076] Dynamic Matrix Construction Module 3 constructs a three-dimensional dynamic materiality matrix (e.g., multiple sub-indicators under the three dimensions of environment, society, and governance) based on the ESG materiality framework specific to the financial industry, and defines triggering rules: when market sentiment or macroeconomic policies change drastically (such as negative news or the promulgation of green finance policies), the weights of the corresponding ESG dimensions in the matrix are automatically adjusted to achieve the ability to respond to real-time events.

[0077] Online Weight Update Module 4 uses preprocessed standardized financial data, ESG data, and a dynamically adjusted materiality matrix as input. It uses the Online Gradient Descent (OGD) algorithm to incrementally update the ESG weight model in real time, generating the current ESG weight vector and its corresponding gradient vector. This model supports lightweight deployment and can be embedded into investment strategy engines for automatic push notifications.

[0078] The reinforcement learning optimization module 5 constructs a simulated investment portfolio environment, takes the ESG weight vector, gradient vector, market characteristics and ESG status characteristics as the environment state, and uses a preset compound reward function (including yield, volatility, risk exposure, ESG performance and other indicators) to iteratively train the strategy network, such as the reinforcement learning algorithm based on Proximal Policy Optimization (PPO), to obtain better ESG weight recommendations.

[0079] The system further feeds back the optimized ESG weight recommendations to the online weight update module 4, dynamically adjusting hyperparameters such as the learning rate, regularization coefficient, and loss function, so that the entire model has adaptive capabilities and maintains sensitivity and stability to market changes during long-term operation.

[0080] Through modular architecture design and multi-level algorithm integration, this embodiment significantly improves the dynamic responsiveness of ESG indicators, the rationality of weight distribution and the forward-looking nature of investment strategies, helping to support a more intelligent and sustainable financial decision-making system.

[0081] Example 2 is the second embodiment of the present invention. Unlike the previous embodiment, this embodiment provides a compliance audit module 6 and an operation and maintenance management module 7, which can enhance the interpretability, compliance, and operational stability of the system. It includes:

[0082] The compliance audit module 6 is connected to the online weight update module 4 and is used to call the SHAP algorithm to generate a feature contribution explanation for each weight update based on the latest ESG weight vector and the latest gradient vector received, and automatically fill in the regulatory report template to generate a compliance model document;

[0083] Operation and maintenance management module 7, which connects to data collection module 1, preprocessing module 2, dynamic matrix construction module 3, online weight update module 4, reinforcement learning optimization module 5, and compliance audit module 6, is used to collect performance indicators of other modules in the financial management system, display each performance indicator in real time on the dashboard, and trigger elastic scaling or rollback operations in response to abnormal situations;

[0084] Performance indicators include data latency, training loss, RL cumulative reward, and model version information.

[0085] Working principle of embodiment 2:

[0086] Compliance Audit Module 6 is used to perform interpretable analysis and generate compliance documents for each ESG weight update. Its workflow is as follows:

[0087] After each weight update is completed, the latest ESG weight vector and the corresponding gradient vector are used as input;

[0088] Call the feature importance interpretation algorithm based on TreeSHAP or DeepSHAP to quantify the contribution of each ESG factor change in the current weight vector;

[0089] The results are output in charts and tables, automatically matched to pre-defined regulatory report templates;

[0090] This template includes fields such as explanations of ESG factor changes, model change records, and a summary of risk exposure changes. It ultimately generates a structured compliance report document (e.g., PDF or HTML format) that can be used for filing with auditors or regulators. This module enables "black box transparency" in the model decision-making process, improving system compliance, regulatory responsiveness, and corporate credibility.

[0091] The operation and maintenance management module 7 serves as the operation monitoring center of the entire system. It maintains connections with multiple core modules in the system (including the data acquisition module 1, preprocessing module 2, dynamic matrix construction module 3, online update module, reinforcement learning optimization module 5, and compliance audit module 6) to achieve comprehensive control of the system's operating status. Its functions include the following aspects:

[0092] Performance indicator collection: Periodically collect the following key performance indicators:

[0093] Data collection delays (such as Kafka message accumulation and network delays);

[0094] Model training loss (such as MSE, cross entropy);

[0095] Cumulative Reward for reinforcement learning

[0096] A snapshot of the currently running model version and its parameters.

[0097] The Operation and Maintenance Management Module 7 also integrates Grafana or self-developed dashboards to display the above indicators in real time and conduct multi-dimensional visual analysis using various forms such as charts and heat maps.

[0098] Abnormal detection and fault tolerance mechanism of operation and maintenance management module 7:

[0099] When a performance indicator (such as a surge in training loss or an abnormal decrease in cumulative rewards) exceeds a preset threshold, the elastic strategy is automatically triggered: training resources can be expanded (such as GPU nodes); roll back to the last stable model version; and an alert notification (such as email / SMS) is sent to maintenance personnel.

[0100] This module significantly enhances the system's availability, monitorability and fault tolerance, and helps build a stable, reliable and easy-to-maintain intelligent financial system operating platform.

[0101] Preferably, the online weight updating module 4 includes:

[0102] A feature fusion unit 41 is configured to perform ESG dimension weighted fusion processing on the financial data and ESG raw data in the input standardized feature vector according to the dynamic materiality matrix. The weighted fusion vector constitutes the final input feature matrix of the ESG weight model.

[0103] The online incremental training unit 42 is connected to the feature fusion unit 41 and is used to perform a round of incremental update on the current model according to the online gradient descent algorithm each time a new data packet is received;

[0104] The model output unit 43 is connected to the online incremental training unit 42 and is used to output the latest ESG weight vector and the latest gradient vector, and write them into the cache pool and the preset historical database.

[0105] Specifically, in this embodiment, the feature fusion unit 41 receives the standardized feature vector output from the preprocessing module 2, including the company's financial data (such as profitability, debt ratio, cash flow, etc.) and the original ESG indicator data from the ESG rating agency. Combined with the dynamic materiality matrix generated by the dynamic matrix construction module 3, the above financial data and ESG original data are weighted according to the three dimensions of ESG (environment, society, and governance) to generate a weighted fusion vector. This process uses weighted vector multiplication, so that the weight of each type of data in the final feature input is dynamically determined by the materiality matrix, thereby enhancing the sensitivity of the input features to current market and policy changes.

[0106] The online incremental training unit 42 is connected to the feature fusion unit 41. When the system receives a new data packet through the data acquisition module 1, an online learning process is triggered. The system performs a round of incremental updates to the current ESG weight model based on an online gradient descent algorithm (such as SGD or Adam variants). During the training process, only the current data packet is needed to complete local fine-tuning of model parameters, thus avoiding retraining the entire model, significantly improving processing efficiency and maintaining the model's responsiveness to real-time data.

[0107] The updated model outputs the latest ESG weight vector and gradient vector through the model output unit 43. After output, the system writes these vectors into a fast cache pool (such as Redis or an in-memory database) for immediate invocation by the subsequent policy network module, and also writes them into a historical database (such as PostgreSQL or a time-series database) for model performance tracking and back-analysis.

[0108] Through the design of this embodiment, the system can achieve real-time optimization of the ESG weight model, improve its adaptability to the dynamic market environment, and improve the computing efficiency and call response speed of subsequent modules through the caching mechanism, thereby enhancing the practicality and intelligence level of the system as a whole.

[0109] Preferably, the compound reward function is configured as:

[0110] R t =α·Relurn t -β·Risk t +γ·ESG_Score t -δ·Penalty t ;

[0111] Among them, Return t represents the return of the portfolio at time step t, Risk t Represents the risk indicator of the investment portfolio, ESG_Scoret represents the weighted ESG score of the portfolio at time step t, Penalty t It represents the penalty item for violating compliance requirements, exceeding transaction limits or other abnormal operations. α, β, γ, and δ represent the preset first coefficient, second coefficient, third coefficient, and fourth coefficient, respectively.

[0112] α (return weight) encourages intelligent agents to pursue positive returns and improve the profitability of the investment portfolio; β (risk penalty weight) prevents strategies from taking excessive risks due to excessive pursuit of high returns and maintains the stability of the investment portfolio; γ (ESG score incentive weight) strengthens the emphasis on corporate sustainable development performance and guides funds to targets with good ESG performance, in line with the requirements of green finance and socially responsible investment; δ (violation penalty weight) punishes illegal operations (such as excessive concentration of holdings, ESG scores below the threshold, excessive trading, etc.) to promote strategy compliance and long-term healthy operation.

[0113] Specifically, in this embodiment, α = 1.0, β = 0.5, γ = 0.8, δ = 0.3; at a certain time t, if the portfolio return rate is 1.5%, that is, Return t =0.015, the risk (volatility) is 2.0%, that is, Risk t =0.020, the ESG weighted score is 85 (out of 100), and due to excessive concentration of holdings, a penalty of 0.050.050.05 points is imposed. The compound reward is calculated as:

[0114] R t =1.0×0.015-0.5×0.020+0.8×0.85-0.3×0.05=0.67

[0115] Preferably, the reinforcement learning optimization module 5 includes:

[0116] Environment simulation unit 51, used to build a customized simulated investment portfolio environment that is highly similar to the real financial market. The customized simulated investment portfolio environment supports the simulation of dynamic changes in asset prices, trading constraints, policy shocks, and public opinion events;

[0117] The strategy learning unit 52 is connected to the environment simulation unit 51 and is used to use the initial strategy network as the core and adopt a preset reinforcement learning algorithm to learn historical weight suggestions under different historical environment states to obtain a preliminary strategy network;

[0118] The reward unit 53 is connected to the strategy learning unit 52 and is used to iteratively train the preliminary strategy network according to the composite reward function to obtain the strategy network after the iteration, and input the market characteristics and ESG status characteristics in the latest ESG weight vector, the latest gradient vector and the standardized feature vector into the strategy network to output the corresponding ESG weight recommendation.

[0119] Specifically, in this embodiment, the environment simulation unit 51 is responsible for constructing a customized simulated investment portfolio environment that closely resembles the behavioral characteristics of real financial markets. This environment not only simulates the characteristics of basic asset price fluctuations over time, but also supports the following key elements:

[0120] Transaction constraint simulation: including transaction costs, minimum transaction units, buying and selling limits, etc.

[0121] Policy shock simulation: supports the generation of structured policy disturbances based on macroeconomic indicators, industry regulatory changes, etc.

[0122] Public Opinion Event Simulation: This simulation uses data-driven mechanisms such as text sentiment analysis results and the sudden public opinion index to trigger event-driven changes in market behavior. This design allows the simulation environment to provide a more realistic, multi-dimensional, and dynamic input context for policy network training.

[0123] Strategy learning unit 52 operates on the basis of environment simulation unit 51, with the initial policy network as the core of the learning model. It uses reinforcement learning algorithms such as PPO (Proximal Policy Optimization) and DDPG (Deep Deterministic Policy Gradient) to learn from multiple periods of historical ESG data and market status data. The system performs supervised or semi-supervised simulations of ESG weighting recommendations in historical investment scenarios and outputs a preliminary policy network capable of making preliminary weighting decisions based on environmental conditions.

[0124] The reward unit 53 is connected to the policy learning unit 52 and iteratively optimizes and trains the policy network based on a composite reward function. During each training round, the system adjusts the policy network parameters based on the composite reward function feedback, gradually optimizing the policy. The final policy network takes the latest ESG weight vector, the latest gradient vector, and the market and ESG status features in the standardized feature vector as inputs, and outputs corresponding ESG weight recommendations for dynamic adjustment of the ESG weight model by the online weight update module 4.

[0125] Preferably, the expression configuration of the policy network is:

[0126]

[0127] in, represents the ESG weight recommendation, σ(·) represents the normalization function, Used to compress the output to the (0, 1) interval; represents the kth market characteristic or ESG characteristic value, ζ(·) represents the Riemann Zeta function, which is used to express the complex series relationship between characteristics and weights. Here s=∈ k +1; J n (·) represents a Bessel function of order n, which represents the nonlinear response of the gradient change to the local weight of the policy network; Represents the current k-th gradient component, ∈ m 、 Respectively represent the components in the direction of the maximum modulus in the current weight and gradient, which are used to construct the elliptic function input; cn(u, k) represents the Jacobi elliptic cosine function, which is used to capture the nonlinear periodic feature changes. θ represents the projection angle of the feature rotation dimension, ∈j represents the jth item of the current ESG weight vector, represents the normalized feature paired with ∈.j in the current feature value, represents the gradient value of the i-th item;

[0128] K(·) represents the elliptic integral of the first kind, which is defined as: Used to comprehensively reflect the nonlinear behavior of the coupling between the overall ESG and characteristics.

[0129] Specifically, in this embodiment, the recommended value range of ESG weight is when When , it means that the importance of the ESG sub-dimension has significantly decreased, and it is recommended to reduce the investment proportion; when When , it means that the dimension weight is in the normal acceptable range; when , it means that this dimension is currently the most sensitive to the overall model or policy shock response, and it is recommended to enhance the impact of this dimension indicator on decision-making.

[0130] Preferably, reference Figure 2 , the online incremental training unit 42 includes:

[0131] The first updating subunit 421 is configured to construct a fusion function based on the current weight vector, the ESG weight suggestion, and the current gradient vector, and compress the update result using a normalization function to progressively update the ESG weight.

[0132] a second updating subunit 422 for dynamically updating the learning rate in response to the nonlinear trend suggested by the ESG weights;

[0133] The third updating subunit 423 is used to adjust the regularization coefficient to suppress model overfitting by constructing a high-order offset response function;

[0134] The fourth updating subunit 424 is used to construct a composite loss function, which recommends deviation from the strategy, L1 regularization term, and a nonlinear penalty mechanism based on the Jacobi elliptic function integral term to perform multi-dimensional optimization control on the model updating process.

[0135] Specifically, in this embodiment, a joint update mechanism involving four core parameters is constructed, comprising a first update subunit 421, a second update subunit 422, a third update subunit 423, and a fourth update subunit 424. This mechanism ensures that the model can adapt to changing trends in the external environment and policy feedback while maintaining responsiveness and stability, and has good iterative efficiency and control capabilities.

[0136] First, for the current ESG weight vector, the first update subunit 421 uses a fusion update approach, nonlinearly integrating historical weight information with the policy network's recommended vector while also incorporating the current model gradient information as a correction term. This update process is normalized using a sigmoid function to ensure that the updated weight values always fall within a reasonable range. This process balances historical stability with policy responsiveness, allowing the model to gradually evolve towards optimization.

[0137] Secondly, regarding learning rate updates, the second update subunit 422 dynamically adjusts the learning rate based on a nonlinear combination of standardized market characteristics and ESG weight recommendations. Specifically, the hyperbolic tangent function activates the trend of strategy changes, amplifying the model's sensitivity to weight fluctuations, thereby achieving a balance between accelerating strategy convergence and controlling oscillations. This automatic learning rate adjustment mechanism avoids the problem of unstable convergence speed that can occur with a fixed learning rate under different market conditions.

[0138] Regarding the regularization coefficient, the third update subunit 423 performs cubic root control adjustments based on the statistical fluctuations of the gradient norm. If the model experiences severe parameter fluctuations or gradient instability at a certain stage, the regularization coefficient is automatically increased to suppress model complexity and improve generalization. This update mechanism ensures that the model maintains a reasonable structural complexity during dynamic training, reducing the risk of overfitting.

[0139] Finally, regarding the loss function structure, the fourth update subunit 424 employs a composite loss function design that integrates a strategy recommendation deviation metric, an L1 regularization constraint, and a nonlinear penalty term based on an elliptic function integral. This loss function not only reflects the discrepancy between the model and the strategy recommendation but also incorporates information about the periodic coupling of market fluctuations and gradient direction, further enhancing the model's ability to perceive and adapt to complex financial environments.

[0140] In summary, through the dynamic update mechanism of the above four parameters, the system realizes closed-loop response and adjustment from strategy recommendation to model update, effectively supports the intelligent, real-time and stability optimization goals of ESG weight modeling under the reinforcement learning framework, and has good practical application prospects and innovation.

[0141] Preferably, the update process of the current weight vector is configured as follows:

[0142]

[0143] in, represents the updated current weight vector, σ(·) represents the normalization function, α t The fusion coefficient vector representing the recommended weight ratio of ESG weight, β t Represents the gradient adjustment coefficient at the current moment, Represents the current gradient vector;

[0144] The learning rate update process is configured as:

[0145]

[0146] Among them, ηt represents the current learning rate, ηt1 represents the updated learning rate, and γ′ represents the preset sensitivity adjustment coefficient. represents the standardized feature, μ i represents the value of the i-th dimension in the ESG weight recommendation;

[0147] The update process of the regularization coefficient is configured as:

[0148]

[0149] Among them, λ t represents the current regularization coefficient, λ t+1 represents the updated regularization coefficient, represents the gradient value of the jth dimension, ρ represents the preset adjustment factor, and n represents the number of feature dimensions;

[0150] The update process of the loss function is configured as:

[0151]

[0152] Among them, ∈ i represents the current weight vector of the i-th dimension, and cn(·) represents the Jacobi elliptic cosine function.

[0153] A financial management method based on ESG dynamic weight allocation is applied to the above-mentioned financial management system based on ESG dynamic weight allocation, referring to Figure 3 ,include:

[0154] Step S1: Data collection module 1 receives multi-source raw data from third-party ESG rating agencies, market information sources, news and public opinion, and macroeconomic indicators in real time through a distributed message queue;

[0155] Step S2, the preprocessing module 2 cleans, standardizes, interpolates missing values, and extracts time series window features from the multi-source raw data to generate a standardized feature vector;

[0156] Step S3: Dynamic Matrix Construction Module 3 constructs a dynamic materiality matrix based on preset financial industry materiality indicators, and triggers rules based on data on market sentiment and policy changes in the standardized eigenvectors to dynamically adjust the materiality weights of the ESG dimensions in the dynamic materiality matrix.

[0157] Step S4: The online weight update module 4 takes the financial data and ESG raw data in the standardized feature vector and the updated materiality weights as input, and incrementally trains the ESG weight model in real time based on the online gradient descent algorithm to generate and publish the latest ESG weight vector and the latest gradient vector;

[0158] Step S5: The reinforcement learning optimization module 5 iteratively trains the policy network in a custom simulated portfolio environment according to a preset compound reward function, and uses the latest ESG weight vector, the latest gradient vector, and the market characteristics and ESG status characteristics in the standardized feature vector as the environment input of the policy network, so that the policy network outputs ESG weight recommendations;

[0159] In step S6, the online weight update module 4 dynamically updates the current weight vector, learning rate, regularization coefficient and loss function in the ESG weight model according to the ESG weight recommendation to obtain an optimized ESG weight model.

[0160] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that improvements and modifications that do not depart from the principles of the present invention are within the scope of protection of the present invention.

Claims

1. A financial management system based on ESG dynamic weight allocation, characterized by: include: The data collection module (1) is used to receive multi-source raw data provided by third-party ESG rating agencies, market information sources, news and public opinion, and macroeconomic indicators in real time through a distributed message queue; A preprocessing module (2), connected to the data acquisition module (1), is used to clean, standardize, interpolate missing values, and extract time series window features from the multi-source raw data to generate a standardized feature vector; A dynamic matrix construction module (3), connected to the pre-processing module (2), is used to construct a dynamic materiality matrix based on preset financial industry materiality indicators, and trigger rules based on data of market public opinion and policy changes related to the standardized feature vector to dynamically adjust the materiality weight of the ESG dimension in the dynamic materiality matrix; an online weight updating module (4), connected to the dynamic matrix building module (3), for taking the financial data and ESG raw data in the standardized feature vector and the updated materiality weight as input, and incrementally training the ESG weight model in real time based on an online gradient descent algorithm to generate and publish the latest ESG weight vector and the latest gradient vector; a reinforcement learning optimization module (5), connected to the online weight updating module (4), for iteratively training the policy network according to a preset compound reward function in a custom simulated portfolio environment, and using the latest ESG weight vector, the latest gradient vector, and the market characteristics and ESG state characteristics in the standardized feature vector as environment inputs of the policy network, so that the policy network outputs ESG weight recommendations; The online weight updating module (4) dynamically updates the current weight vector, learning rate, regularization coefficient and loss function in the ESG weight model according to the ESG weight recommendation to obtain an optimized ESG weight model.

2. The financial management system based on ESG dynamic weight allocation according to claim 1 is characterized in that: Also includes: A compliance audit module (6), connected to the online weight update module (4), is configured to generate a feature contribution explanation for each weight update based on the received latest ESG weight vector and the latest gradient vector by calling the SHAP algorithm, and automatically fill in a regulatory report template to generate a compliance model document; An operation and maintenance management module (7) is connected to the data acquisition module (1), the preprocessing module (2), the dynamic matrix construction module (3), the online weight update module (4), the reinforcement learning optimization module (5) and the compliance audit module (6), and is used to collect performance indicators of the remaining modules in the financial management system, and to display the performance indicators in real time on a dashboard, and to trigger elastic scaling or rollback operations in response to abnormal situations; The performance indicators include data delay, training loss, RL cumulative reward and model version information.

3. The financial management system based on ESG dynamic weight allocation according to claim 1 is characterized by: The online weight updating module (4) comprises: A feature fusion unit (41) is used to perform ESG dimension weighted fusion processing on the financial data and the ESG raw data in the input standardized feature vector according to the dynamic materiality matrix, and the weighted fusion vector constitutes the final input feature matrix of the ESG weight model; An online incremental training unit (42), connected to the feature fusion unit (41), configured to perform a round of incremental update on the current model according to the online gradient descent algorithm each time a new data packet is received; A model output unit (43) is connected to the online incremental training unit (42) and is used to output the latest ESG weight vector and the latest gradient vector, and write them into a cache pool and a preset history database.

4. The financial management system based on ESG dynamic weight allocation according to claim 1, characterized in that: The compound reward function is configured as: R t =a·Return t -b·Risk t +c·ESG_Score t -d·Penalty t ; Among them, ReLurn t represents the return of the portfolio at time step t, Riskt represents the risk indicator of the portfolio, ESG_Score t represents the weighted ESG score of the portfolio at time step t, Penalty t It represents the penalty item for violating compliance requirements, exceeding transaction limits or other abnormal operations. α, β, γ, and δ represent the preset first coefficient, second coefficient, third coefficient, and fourth coefficient, respectively.

5. The financial management system based on ESG dynamic weight allocation according to claim 1, characterized in that: The reinforcement learning optimization module (5) includes: An environment simulation unit (51) is used to construct the custom simulated investment portfolio environment that is highly similar to the real financial market, and the custom simulated investment portfolio environment supports the simulation of dynamic changes in asset prices, transaction constraints, policy shocks and public opinion events; A strategy learning unit (52) is connected to the environment simulation unit (51) and is used to learn historical weight suggestions under different historical environment states using a preset reinforcement learning algorithm with an initial strategy network as the core to obtain a preliminary strategy network; A reward unit (53) is connected to the strategy learning unit (52) and is used to iteratively train the preliminary strategy network according to the composite reward function to obtain the strategy network after the iteration, and input the latest ESG weight vector, the latest gradient vector and the market characteristics and ESG status characteristics in the standardized feature vector into the strategy network to output the corresponding ESG weight recommendation.

6. The financial management system based on ESG dynamic weight allocation according to claim 3 is characterized by: The expression configuration of the policy network is: in, represents the ESG weight recommendation, σ(·) represents the normalization function, Used to compress the output to the (0, 1) interval; represents the kth market characteristic or the ESG characteristic value, ζ(·) represents the Riemann Zeta function, which is used to express the complex series relationship between characteristics and weights, Here s=∈ k +1; J n (·) represents a Bessel function of order n, which represents the nonlinear response of the gradient change to the local weight of the policy network; Represents the current k-th gradient component, ∈m, Respectively represent the components in the direction of the maximum modulus in the current weight and gradient, which are used to construct the elliptic function input; cn(u, k) represents the Jacobi elliptic cosine function, which is used to capture the nonlinear periodic feature changes. k=1 / π;θ represents the projection angle of the feature rotation dimension,∈ j represents the jth item of the current ESG weight vector, Indicates that the current eigenvalue is the same as ∈ j Normalized features of the pair, represents the gradient value of the i-th item; K(·) represents the elliptic integral of the first kind, which is defined as: Used to comprehensively reflect the nonlinear behavior of the coupling between the overall ESG and characteristics.

7. The financial management system based on ESG dynamic weight allocation according to claim 6, characterized in that: The online incremental training unit (42) comprises: A first updating subunit (421) is configured to construct a fusion function based on the current weight vector, the ESG weight suggestion, and the current gradient vector, and compress the update result through a normalization function to complete a progressive update of the ESG weight; a second updating subunit (422) for dynamically updating the learning rate in response to the nonlinear trend suggested by the ESG weight; A third updating subunit (423) is configured to adjust the regularization coefficient to suppress model overfitting by constructing a high-order offset response function; The fourth updating subunit (424) is used to construct the composite loss function, which is based on the strategy suggestion deviation, L1 regularization term and nonlinear penalty mechanism based on Jacobi elliptic function integral term to perform multi-dimensional optimization control on the model update process.

8. The financial management system based on ESG dynamic weight allocation according to claim 7, characterized in that: The updating process of the current weight vector is configured as follows: in, represents the updated current weight vector, σ(·) represents the normalization function, α t The fusion coefficient vector representing the recommended weight ratio of the ESG weight, β t Represents the gradient adjustment coefficient at the current moment, represents the current gradient vector; The updating process of the learning rate is configured as follows: Among them, η t represents the current learning rate, η t+1 represents the updated learning rate, γ′ represents the preset sensitivity adjustment coefficient, represents the standardized feature, μ i represents the value of the i-th dimension in the ESG weight recommendation; The updating process of the regularization coefficient is configured as follows: Among them, λ t Represents the current regularization coefficient, λ t+1 represents the regularization coefficient after update, represents the gradient value of the jth dimension, ρ represents the preset adjustment factor, and n represents the number of feature dimensions; The update process of the loss function is configured as follows: Wherein, ∈i represents the current weight vector of the i-th dimension, and cn(·) represents the Jacobi elliptic cosine function.

9. A financial management method based on ESG dynamic weight allocation, applied to the financial management system based on ESG dynamic weight allocation according to any one of claims 1 to 8, characterized in that: include: Step S1, the data collection module (1) receives in real time through a distributed message queue multi-source raw data provided by third-party ESG rating agencies, market information sources, news and public opinion, and macroeconomic indicators; Step S2, the preprocessing module (2) cleans, standardizes, interpolates missing values, and extracts time series window features from the multi-source raw data to generate a standardized feature vector; Step S3, the dynamic matrix construction module (3) constructs a dynamic materiality matrix based on the preset financial industry materiality indicators, and triggers rules based on the data of the standardized feature vector related to market public opinion and policy changes to dynamically adjust the materiality weights of the ESG dimension in the dynamic materiality matrix; Step S4, the online weight updating module (4) takes the financial data and ESG raw data in the standardized feature vector and the updated materiality weight as input, and incrementally trains the ESG weight model in real time based on the online gradient descent algorithm to generate and publish the latest ESG weight vector and the latest gradient vector; Step S5, the reinforcement learning optimization module (5) iteratively trains the policy network in a custom simulated portfolio environment according to a preset compound reward function, and uses the latest ESG weight vector, the latest gradient vector, and the market characteristics and ESG state characteristics in the standardized feature vector as environment inputs of the policy network, so that the policy network outputs ESG weight recommendations; In step S6, the online weight updating module (4) dynamically updates the current weight vector, learning rate, regularization coefficient and loss function in the ESG weight model according to the ESG weight recommendation to obtain an optimized ESG weight model.

Images