Enterprise ESG score prediction method based on machine learning

Abstract

The invention discloses an enterprise ESG score prediction method based on machine learning, and the method comprises the steps: collecting multi-dimensional ESG, financial and risk data, carrying out the normalization and missing / abnormal value processing, and dividing a data set; a multi-model system is constructed, and the optimal LSTM is selected by using five-dimensional indexes to predict ESG rating; in combination with an LSTM result, by taking maximization of an ESG mean value-minimization of a fluctuation ratio as a target, GRU is used for early warning of a short-term risk, LSTM controls a long-term trend, after constraint is added, an optimal investment portfolio weight is output through PSO, and a model updating and portfolio rebalance mechanism is set. According to the method, the ESG prediction precision is improved, the ESG-investment optimization integration is realized, and the data timeliness demand is adapted.

Description

Technical Field

[0001] This invention relates to the intersection of machine learning technology and financial investment optimization, specifically to a machine learning-based method for predicting corporate ESG scores. Background Technology

[0002] With the deepening of the global concept of sustainable development, Environmental, Social, and Governance (ESG) has become a core framework for measuring corporate non-financial performance, and its correlation with corporate long-term financial performance and risk management capabilities has been confirmed by numerous studies. In recent years, the scale of ESG investment has grown rapidly, and institutional investors and regulators have become significantly more reliant on ESG data. Accurately predicting corporate ESG ratings and optimizing investment portfolios based on them has become a key need in the financial sector.

[0003] However, existing ESG rating forecasting and portfolio optimization techniques have the following core shortcomings:

[0004] Predictive models are often limited in scope and performance: Existing research often relies on a single machine learning model or traditional statistical model, making it difficult to capture the nonlinear dynamic characteristics of ESG data. Some studies only compare two models, failing to cover multiple algorithms applicable to different data characteristics; while others apply multiple algorithms, they do not fully integrate multi-dimensional dynamic ESG indicators, resulting in limited prediction accuracy.

[0005] The ESG impact factor system is incomplete: most studies rely solely on historical ESG scores or basic financial data, ignoring key external factors such as environmental policies and social sentiment. Some studies depend on a single data source and lack sustainability attribute data in the environmental and social dimensions; others are based only on corporate data from specific regions, failing to incorporate global environmental and social indices and thus failing to comprehensively reflect the multidimensional impact factors of ESG.

[0006] Portfolio optimization is disconnected from ESG forecasting: Existing portfolio optimization focuses primarily on the risk-return balance, failing to incorporate short-term fluctuations and long-term trends in ESG ratings into tiered management. Current technologies do not combine the strengths of different models in capturing short-term characteristics and modeling long-term dependencies, making it difficult to achieve dynamic early warning of ESG risks and continuous portfolio optimization.

[0007] Insufficient model interpretability and generalization: Some studies use deep learning models that lack an interpretation mechanism for the prediction results, and the datasets have limited spatiotemporal scope, resulting in weak generalization ability of the models in different regions and different market cycles; at the same time, no dynamic update mechanism is set up, which cannot cope with the performance degradation of the models caused by the timeliness and policy sensitivity of ESG data.

[0008] To address the aforementioned issues, this invention proposes a method that integrates multi-dimensional ESG factors, a multi-model prediction system, and a hierarchical risk optimization strategy. It achieves high-precision prediction of ESG ratings through an LSTM model and dynamically optimizes the investment portfolio by combining a GRU model with a particle swarm optimization (PSO) algorithm, thus filling the gap in the existing technology for the integrated ESG prediction-optimization process. Summary of the Invention

[0009] The core objective of this invention is to overcome the problems of low ESG rating prediction accuracy, incomplete factor system, disconnect between portfolio and ESG prediction, and poor dynamic adaptability of the model in the existing technology, and to provide a machine learning-based method for predicting corporate ESG scores.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A machine learning-based method for predicting enterprise ESG scores includes the following steps:

[0012] S1. Data Acquisition and Preprocessing:

[0013] S11. Data Collection: Collect multi-dimensional input data, including enterprise environmental indicators, social indicators, governance indicators, and historical ESG rating data.

[0014] Among them, the environmental dimension indicators should at least include the CSI Metasequoia Environmental Patent 50 Index (SEP), the CSI Nuclear Energy Index (NEI), the CSI Green Electricity Index (GPI), and the CSI Green Bond Index (GBI); the social dimension indicators should at least include the CSI 300 ESG Benchmark Index (ESG) and the CSI Green Industry Quality Strategy Index (GIQS); and the governance dimension indicators should at least include the CSI Water Environment Governance Theme Index (WEG) and the SSE Corporate Governance Index (CGI).

[0015] S12. Data Preprocessing: The collected input data is standardized using the Min-Max normalization method, mapping all variables to the [0,1] interval. The normalization process is expressed as follows:

[0016]

[0017] In the formula, The original data, , These are the maximum and minimum values ​​of the dataset, respectively.

[0018] The data is after normalization;

[0019] S13. Divide the preprocessed data into training and testing sets in a 7:3 ratio for subsequent model training and validation.

[0020] S2. Model Building and Training:

[0021] S21. Construct a multi-model prediction system, including Long Short-Term Memory Network (LSTM model), Backpropagation Neural Network (BP model), Multilayer Perceptron (MLP model), Recurrent Neural Network (RNN model), Gated Recurrent Unit (GRU model), and Convolutional Neural Network (CNN model).

[0022] S22. For the LSTM model, the timing data is processed using the memory units and gating mechanisms (forget gate, input gate, output gate) according to the following formula:

[0023]

[0024]

[0025]

[0026] in, These are the output values ​​of the forget gate, input gate, and output gate, respectively. It is the sigmoid activation function. The hyperbolic tangent activation function is used. This is the weight matrix. For bias terms, This represents the hidden layer state at the previous moment. Input data for the current moment. This represents the cell state at the previous moment. Candidate cell state, This represents the current state of the cell. This is the output of the hidden layer at the current moment;

[0027] S23. The predictive performance of the six models is evaluated using a five-dimensional evaluation index, which includes the coefficient of determination (COP). The mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are calculated as follows:

[0028]

[0029]

[0030] in, This is the actual ESG rating value. To predict ESG rating values, This is the average of the actual ESG rating values. The number of samples;

[0031] S24. Select the LSTM model with the best evaluation results as the final ESG rating prediction model. The optimal criterion is: The maximum value is found when MSE, MAE, RMSE, and MAPE are the minimum.

[0032] S3, Portfolio Optimization:

[0033] S31. Based on the enterprise ESG rating prediction results output by the LSTM model, and combined with enterprise financial data (including debt-to-equity ratio and company size) and market risk indicators, construct an investment portfolio optimization objective function. The objective function has two objectives: "maximizing the mean of portfolio ESG rating and minimizing portfolio return volatility".

[0034] S32. A two-stage integrated strategy is adopted to manage portfolio risk in a tiered manner: Short-term risk warning uses the GRU model, which captures short-term ESG rating volatility characteristics through its update gate and reset gate. The update gate and reset gate are calculated as follows:

[0035]

[0036] in, To reset the gate output, To update the gate output, , The weight matrix is ​​used; long-term trend control is combined with LSTM model prediction results to identify companies with stable and positive long-term ESG ratings.

[0037] S33. Introduce constraints to limit the optimization process. The constraints include that the investment weight of a single enterprise does not exceed 10% of the total portfolio size, the average ESG rating of enterprises in the portfolio is not lower than a preset threshold, and the industry distribution concentration of the portfolio does not exceed 30%.

[0038] S34. Solve the objective function using the Particle Swarm Optimization (PSO) algorithm and output the optimal portfolio weight allocation scheme;

[0039] S4. Model Updates and Portfolio Rebalancing: Through quarterly model updates and monthly portfolio rebalancing mechanisms, the model performance degradation caused by ESG data timeliness, market environment changes, and policy adjustments is addressed, ensuring the long-term effectiveness of the method.

[0040] As a further optimization of a machine learning-based method for predicting enterprise ESG scores.

[0041] Preferably, in step S21, for the BP model, the weights and biases are iteratively adjusted using the backpropagation algorithm, and the output layer is calculated as follows:

[0042]

[0043] in, The weights from the input layer to the hidden layer. For input features, For bias terms, The activation function is used; for the MLP model, the input-to-output mapping is achieved through multiple nonlinear transformations:

[0044]

[0045] in, , This is the weight matrix. , For bias terms, For output of the hidden layer, For activation function, This is the activation function for the output layer.

[0046] Preferably, in step S31, the enterprise financial data also includes return on equity (ROE) and price-to-earnings ratio (P / E), and market risk indicators include stock volatility and beta coefficient. Indicators with an absolute correlation value greater than 0.7 with ESG rating are selected through Pearson correlation analysis and included in the optimization model, while multicollinearity indicators are eliminated.

[0047] Preferably, in step S34, the parameters of the particle swarm optimization algorithm are set as follows: population size 50-100, maximum number of iterations 100-200, initial inertia weight 0.9, linearly decreasing to 0.4, and learning factor... The fitness function is set to 2, and the quality of each particle is evaluated by a fitness function that is a linear weighted combination of the objective function.

[0048] Preferably, in step S21, the CNN model extracts local features of the data by alternately using convolutional layers and pooling layers, and the size of the output feature map of the convolutional layer is calculated as follows:

[0049]

[0050] in, , These are the height and width of the input feature map, respectively. , These are the height and width of the convolution kernel, respectively. Step size, This is the number to fill in.

[0051] Preferably, in step S32, the time window for short-term risk warning is set to 1-3 months. The short-term ESG rating fluctuation is predicted by the GRU model, and a risk warning is triggered when the predicted fluctuation exceeds 20%. The time window for long-term trend control is set to 1-3 years. Based on the prediction results of the LSTM model, companies whose ESG ratings have improved for three consecutive quarters are selected and included in the core investment pool.

[0052] If the collected data contains missing values, multiple imputation is used to fill them in, with five imputations performed and the average value taken as the final filled value. For outliers, a... The criteria are identified and replaced with this indicator. quantiles or Quantiles ensure data quality.

[0053] The beneficial effects of this invention are:

[0054] (1) Improve the accuracy and reliability of ESG rating prediction: By constructing a multi-model prediction system containing 6 models such as LSTM, BP, and MLP, and combining the five-dimensional evaluation index to select the optimal LSTM model, it can accurately capture the nonlinear dynamic characteristics of ESG data, significantly reduce the prediction error compared with a single model, and continuously maintain high prediction performance through quarterly data updates and model retraining, thus solving the problems of "low prediction accuracy and weak generalization" of existing technologies.

[0055] (2) Improve the ESG impact factor system and cover more dimensions: break through the limitations of existing research that only relies on historical ESG scores or basic financial data. It not only includes the core indices of the three dimensions of environment, society and governance, but also combines financial data such as ROE and price-earnings ratio, as well as risk indicators such as stock volatility and Beta coefficient. Through Pearson correlation analysis, key factors are screened to comprehensively reflect the multi-dimensional impact factors of ESG and avoid decision-making bias caused by "incomplete factor system".

[0056] (3) Achieve deep coupling between ESG forecasting and portfolio optimization, and make risk stratification management more efficient: Based on the LSTM forecasting results, construct a dual objective function of "maximizing the average ESG rating of the portfolio and minimizing the volatility of returns", and combine it with a two-stage strategy of "GRU short-term risk warning + LSTM long-term trend control". Combine the particle swarm optimization algorithm to solve the optimal weight, and at the same time, solve the pain point of "disconnect between portfolio and ESG forecasting" through monthly portfolio rebalancing and dynamic adjustment, and take into account both ESG performance and risk control.

[0057] (4) Ensuring data quality and optimizing compliance, with stronger feasibility: The data preprocessing stage adopts multiple imputation to fill missing values ​​and the 3σ criterion to handle outliers, ensuring the reliability of input data; the portfolio optimization sets constraints such as "individual enterprise weight ≤ 10%, ESG mean ≥ 1.1 times the industry mean, and industry concentration ≤ 30%", and the rebalancing transaction cost is controlled within 0.5%, which meets the compliance requirements of financial investment and can directly provide institutional investors with feasible ESG investment decision-making solutions. Attached Figure Description

[0058] Figure 1 Flowchart of the technical solution of this invention;

[0059] Figure 2 Average performance of 50 companies;

[0060] Figure 3 A graph comparing the predicted and actual values ​​of six models;

[0061] Figure 4 Portfolio optimization results diagram;

[0062] Figure 5 A bar chart comparing performance before and after model update;

[0063] Figure 6 Comparison chart of ESG rating prediction errors across different industries;

[0064] Figure 7 A graph showing the change in fitness function values ​​during the PSO algorithm iteration process;

[0065] Figure 8 Comparison chart of the impact of model update frequency on prediction performance;

[0066] Figure 9 Scatter plots of ESG mean and volatility of portfolios under different constraints. Detailed Implementation

[0067] To make the above-mentioned objectives, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to examples. The following content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0068] Example

[0069] A machine learning-based method for predicting enterprise ESG scores includes the following steps:

[0070] S1. Data Acquisition and Preprocessing:

[0071] S11. Data Collection: ESG Indicator Data: Monthly data of 8 key ESG indicators from July 1, 2019 to August 1, 2023 were collected from the Wind database. The data covers 50 companies in 5 industries: finance, industry, technology, consumer goods and energy, with 10 companies in each industry. Descriptive statistics for some indicators are as follows: CSI 300 ESG Benchmark Index: Mean 1219.74, Maximum 1636.50, Minimum 962.86, Standard Deviation 150.2; SSE Corporate Governance Index: Mean 1060.35, Maximum 1226.91, Minimum 892.65, Standard Deviation 80.5; CSI Green Power Index: Mean 929.23, Maximum 1256.39, Minimum 690.64, Standard Deviation 110.8; ESG Rating Data: Monthly ESG rating data of 50 companies collected from the DJSI database during the same period, Mean 3.8, Maximum 4.8, Minimum 2.5, Standard Deviation 0.6; Financial and Risk Data: Corporate financial data and market risk indicators collected from the S&P Global database.

[0072] S12. Data preprocessing, including: Missing value handling: The CSI Green Bond Index has missing values ​​for four months: February 2021, May 2022, January 2023, and July 2023. An imputation model was constructed based on the Pearson correlation between GBI and GPI. After five imputations, the filled values ​​were 121.5, 122.1, 120.8, and 121.7, respectively, which are close to the overall mean of GBI; Outlier handling: CSI Core... The energy index has three outliers. The calculated NEI mean is 2339.42, standard deviation is 300.5, and the 3σ range is 1437.92-3240.92. 3100.50 (exceeding 3σ) is replaced with the 99th quantile of 3000.0, and 1846.22 (below 1σ) is replaced with the 1st quantile of 1900.0. Standardization: Taking the ESG index as an example, the normalization formula calculates the standardized value of a monthly data point of 1200 as... After standardization, the mean of all indicators is concentrated in the range of 0.4-0.6, and the standard deviation is concentrated in the range of 0.15-0.25. Data split: the training set is from July 2019 to June 2022, and the test set is from July 2022 to August 2023. The training set sample size is about 2520, and the test set sample size is about 1800.

[0073] S2. Model Building and Training:

[0074] S21. Construct six models based on the TensorFlow deep learning framework, with the following parameters: LSTM: 2 hidden layers, learning rate 0.006, batch size 32, time step 12, ReLU activation function, 100 iterations; BP: 1 hidden layer, learning rate 0.012, Sigmoid activation function, 80 iterations; MLP: 3 hidden layers, learning rate 0.008, ReLU activation function, 100 iterations; RNN: 1 hidden layer, learning rate 0.007, Tanh activation function, 90 iterations; GRU: 2 hidden layers, learning rate 0.005, batch size 32, time step 6, ReLU activation function, 100 iterations; CNN: 2 convolutional layers, 1 max pooling layer, 1 fully connected layer, learning rate 0.006, 80 iterations. All models are trained using 5-fold cross-validation, and an early stopping strategy is adopted during training. Training is stopped when the validation set MSE does not decrease for 10 consecutive rounds.

[0075] Performance Validation: The performance of the six models was evaluated on the test set, and the results are as follows: Figure 1 As shown. By Figure 1 As can be seen, the LSTM model performs best across all metrics. The highest value was found in MAE, while the lowest values ​​were found in MSE, RMSE, and MAPE. The MSE was 0.875 times that of the BP model, 0.778 times that of the MLP and RNN models, 0.636 times that of the CNN model, and 0.039 times that of the GRU model, which falls within the range of 0.14-25.14 times the advantage. Therefore, LSTM was determined as the final ESG rating prediction model.

[0076] S3, Portfolio Optimization:

[0077] S31. Combined screening and parameter setting, divided into the following processes: Enterprise screening: Based on the LSTM prediction results, the top 20 companies in ESG rating among 50 companies are selected as candidate companies for the combined screening; Objective function parameters: weight coefficients. That is, 60% weight is assigned to ESG performance and 40% to risk control; constraint parameter settings: individual enterprise weight [0.005, 0.1], combined ESG mean threshold 3.9, single industry proportion [0.05, 0.3]; PSO parameter settings: population size 80, maximum number of iterations 150, inertia weight 0.9→0.4, learning factor .

[0078] S32. Optimization Results and Risk Adjustment: Initial Optimization Results: After solving the PSO, the top 5 weighted companies and their weights, and ESG ratings are as follows: Company A: 9.4%, Company B: 9.1%, Company C: 8.7%, Company D: 8.4%, Company E: 7.9%. The portfolio's average ESG is 4.1, the return volatility is 12.2%, and the industry distribution is: Technology 28%, Consumer 25%, Financials 22%, Industrials 15%, Energy 10%, all of which meet the constraints.

[0079] S33. Short-term risk warning adjustment: The GRU model predicts that the ESG rating of company E will fluctuate by 23% next month, triggering a warning. Its weight is reduced from 7.9% to 5.1%, while the weight of company F is increased from 6.2% to 8.9%. After the adjustment, the portfolio volatility drops to 11.7%, and the ESG mean remains at 4.1.

[0080] S4. Model Update and Combinatorial Rebalancing:

[0081] S41. Quarterly Model Update, including: Data Update: Collecting ESG data from July to September 2023, supplementing the dataset to cover July 2019 to September 2023; Retraining: Retraining the LSTM model with hyperparameters adjusted to 120 hidden units, a learning rate of 0.006, and a batch size of 32. The model then performs better on the test set. MSE=0.00065, performance remains stable; Forecast adjustment: LSTM predicts that the ESG rating of company C will drop from 4.3 to 4.0 in the next 6 months, marking it as a company requiring attention.

[0082] S42. Monthly portfolio rebalancing, including: Trigger condition check: Enterprise C's actual weight is 8.7%, the optimal weight is adjusted to 7.5%, and the deviation is 1.2%; Enterprise G's actual weight is 6.8%, the optimal weight is adjusted to 9.5%, and the deviation is 2.7%; Enterprise H's volatility increases from 18% to 24%.

[0083] S43. Rebalancing Operation: Reduce the weight of company C to 7.5%, reduce the weight of company H from 7.1% to 5.8%; increase the weight of company G to 9.5%, increase the weight of company I from 6.5% to 8.2%; the transaction cost of this rebalancing is 0.3%, the adjusted portfolio has an ESG mean of 4.2 and a volatility of 11.5%, meeting all constraints.

[0084] Figure 1 This is a flowchart of the technical solution of the present invention.

[0085] Figure 2This is the average performance of 50 companies: It compares the five-dimensional evaluation metrics (R², MAE, MSE, RMSE, MAPE) of six models (LSTM, BP, MLP, RNN, CNN, GRU) on the test set. The LSTM model performs best on all metrics, validating its suitability as the final ESG rating prediction model.

[0086] Figure 3 The data was plotted based on the test set data of Company A in the example: the LSTM prediction curve had the highest overlap with the actual curve, with an average deviation of 0.021; the GRU prediction curve fluctuated the most, with some nodes having a deviation of over 0.035; the prediction deviations of other models were between the two, which is consistent with the performance ranking of the six models.

[0087] Figure 4 The chart shows the portfolio optimization results: Based on the optimization results of 20 candidate companies in the example: The optimal portfolio has an average ESG of 4.1 and a volatility of 11.7%. Among the top 10 companies, the technology industry has the highest proportion, followed by the financial and consumer industries, and the energy industry has the lowest proportion. The weight of all companies is ≤10%, and the industry concentration is ≤30%.

[0088] Figure 5 The bar chart shows the performance comparison before and after the model update: Based on the model update data from January to September 2023 in the example, the R² of the model remained between 0.70 and 0.71 after the quarterly update, and the MSE remained stable between 0.00065 and 0.00075. During the period without update, R² did not decrease significantly, and MSE did not increase significantly, which verifies that the quarterly update mechanism can effectively maintain the stability of model performance and avoid performance degradation.

[0089] Figure 6 This is a comparison chart of ESG rating prediction errors across different industries: Using the financial, industrial, technology, consumer, and energy industries as the horizontal axis, it displays the MAE and MAPE of each industry under the LSTM model, with two sets of bar charts for each industry. Based on test set data from 50 companies in the example, the calculations show: the technology industry has the lowest MAE and a MAPE of 0.10%; the energy industry has the highest MAE and a MAPE of 0.15%; the financial, industrial, and consumer industries have MAEs between 0.020 and 0.022, and MAPEs between 0.11% and 0.13%.

[0090] Figure 7 The graph shows the change in fitness function value during the PSO algorithm iteration process. Based on the PSO parameter calculation in the example, the fitness value decreases rapidly in the first 20 iterations, tends to stabilize after 50 iterations, and the optimal fitness value stabilizes at 0.69 after 150 iterations, while the average fitness value of the population stabilizes at 0.71. The difference between the two is less than 0.02, proving the convergence of the algorithm.

[0091] Figure 8This is a comparison chart of the impact of model update frequency on prediction performance. Based on data from January to September 2023 in the example: quarterly updates have the highest R² and the lowest MSE; semi-annual updates have R² of 0.69 and MSE of 0.00072; annual updates have R² of 0.67 and MSE of 0.00081; no updates have R² that drops to 0.65 and MSE that rises to 0.00089, and the error bar length of the no-update group is twice that of the quarterly update group.

[0092] Figure 9 This is a scatter plot of the ESG mean and volatility of the portfolio under different constraints: Based on the portfolio optimization data in the example, the scatter plot moves from the "low ESG-high volatility" quadrant (unconstrained) to the "high ESG-low volatility" quadrant (fully constrained) as the constraints increase. The fully constrained group has a 13.9% higher ESG mean and a 22.9% lower volatility than the unconstrained group.

Claims

1. A machine learning-based method for predicting enterprise ESG scores, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: S11. Data Collection: Collect multi-dimensional input data, including enterprise environmental indicators, social indicators, governance indicators, and historical ESG rating data. S12. Data Preprocessing: The collected input data is standardized using the Min-Max normalization method, mapping all variables to the [0,1] interval. The normalization process is expressed as follows: In the formula, The original data, , These are the maximum and minimum values ​​of the dataset, respectively. The data is after normalization; S13. Divide the preprocessed data into training and testing sets in a 7:3 ratio; S2. Model Building and Training: S21. Construct a multi-model prediction system, including LSTM model, BP model, MLP model, RNN model, GRU model, and CNN model; S22. For the LSTM model, the timing data is processed using the memory units and gating mechanism according to the following formula: in, These are the output values ​​of the forget gate, input gate, and output gate, respectively. It is the sigmoid activation function. The hyperbolic tangent activation function is used. This is the weight matrix. For bias terms, This represents the hidden layer state at the previous moment. Input data for the current moment. This represents the cell state at the previous moment. Candidate cell state, This represents the current state of the cell. This is the output of the hidden layer at the current moment; S23. The predictive performance of the six models is evaluated using a five-dimensional evaluation index, which includes... MAE, MSE, RMSE, and MAPE are calculated as follows: in, This is the actual ESG rating value. To predict ESG rating values, This is the average of the actual ESG rating values. The number of samples; S24. Select the LSTM model with the best evaluation results as the final ESG rating prediction model. The optimal criterion is: The maximum value is found when MSE, MAE, RMSE, and MAPE are the minimum. S3, Portfolio Optimization: S31. Based on the enterprise ESG rating prediction results output by the LSTM model, and combined with enterprise financial data and market risk indicators, construct an investment portfolio optimization objective function. The objective function has two objectives: "maximizing the mean of portfolio ESG rating and minimizing portfolio return volatility". S32. A two-stage integrated strategy is adopted to manage portfolio risk in a tiered manner: Short-term risk warning uses the GRU model, which captures short-term ESG rating volatility characteristics through its update gate and reset gate. The update gate and reset gate are calculated as follows: in, To reset the gate output, To update the gate output, , This is the weight matrix; S33. Introduce constraints to limit the optimization process. The constraints include that the investment weight of a single enterprise does not exceed 10% of the total portfolio size, the average ESG rating of enterprises in the portfolio is not lower than a preset threshold, and the industry distribution concentration of the portfolio does not exceed 30%. S34. Solve the objective function using the particle swarm optimization algorithm and output the optimal portfolio weight allocation scheme; S4. Model Updates and Portfolio Rebalancing: Through quarterly model updates and monthly portfolio rebalancing mechanisms, the model performance degradation caused by ESG data timeliness, market environment changes, and policy adjustments is addressed, ensuring the long-term effectiveness of the method.

2. The enterprise ESG score prediction method based on machine learning according to claim 1, characterized in that, In step S21, for the BP model, the weights and biases are iteratively adjusted using the backpropagation algorithm, and the output layer is calculated as follows: in, The weights from the input layer to the hidden layer. As input features, For bias terms, The activation function is used; for the MLP model, the input-to-output mapping is achieved through multiple nonlinear transformations: in, , This is the weight matrix. , For bias terms, For output of the hidden layer, For activation function, This is the activation function for the output layer.

3. The enterprise ESG score prediction method based on machine learning according to claim 1, characterized in that, In step S31, before constructing the portfolio optimization objective function, the corporate financial data also includes return on net assets and price-to-earnings ratio. Market risk indicators include stock volatility and Beta coefficient. Pearson correlation analysis is used to screen indicators with an absolute correlation value greater than 0.7 with ESG ratings and include them in the optimization model, while multicollinearity indicators are removed.

4. The enterprise ESG score prediction method based on machine learning according to claim 1, characterized in that, In step S34, the parameters of the particle swarm optimization algorithm are set as follows: population size 50-100, maximum number of iterations 100-200, initial inertia weight 0.9, linearly decreasing to 0.4, and learning factor... The fitness function is set to 2, and the quality of each particle is evaluated by a fitness function that is a linear weighted combination of the objective function.

5. The enterprise ESG score prediction method based on machine learning according to claim 1, characterized in that, In step S21, the CNN model extracts local features of the data by alternately using convolutional layers and pooling layers. The size of the feature map output by the convolutional layer is calculated as follows: in, , These are the height and width of the input feature map, respectively. , These are the height and width of the convolution kernel, respectively. Step size, This is the number to fill in.

6. The enterprise ESG score prediction method based on machine learning according to claim 1, characterized in that, In step S32, the time window for the short-term risk warning is set to 1-3 months. The short-term ESG rating volatility is predicted using the GRU model. When the predicted volatility exceeds 20%, a risk warning is triggered. The time window for long-term trend control is set at 1-3 years, and companies whose ESG ratings have improved for three consecutive quarters are selected based on the prediction results of the LSTM model and included in the core investment pool.

Images