Advanced TRST01 ESG Scoring Model with Beta Based Financial Metrics and Machine Learning Techniques

: In the current corporate world, assessing a company's sustainability performance is very important for investors, stakeholders, and policymakers. The TRST01’s ESG (Environmental. Social and Governance) Scoring Model introduces an innovative approach integrating beta-based financial metrics with advanced machine learning techniques to comprehensively evaluate ESG credentials. This study demonstrates the development and application of the TRST01’s ESG scoring model, which leverages data from the most reputable sources such as MSCI and S&P Global to ensure its reliability and accuracy. The model's unique methodology involves calculating country-specific beta values to normalize carbon emission data, thereby providing a standardized metric for meaningful comparisons across countries. Further, ESG scores are adjusted using both country and company beta values to reflect specific risk exposures, enhancing the precision and relevance of the assessments. The model ensures robust input data quality, by taking Market capital, Scope 1, Scope 2, industry wise data and beta values as predictors through extensive data preprocessing and encoding categorical variables for top 1000 listed companies. A comparative analysis of Traditional model such as Simple Linear Regression (SLR) and multiple Machine Learning (ML) models, including Gradient Boosting (GB), Support Vector Regression (SVR), and Random Forest (RF), demonstrates that the Gradient Boosting model achieves superior performance with minimal overfitting and consistent prediction accuracy. The study employs a comprehensive evaluation framework using various metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and R-squared, supplemented by detailed visualizations of actual vs. predicted values, residuals, and error distributions. This research underscores the significance of incorporating advanced financial metrics and machine learning techniques in ESG assessments, providing a reliable, accurate, and holistic framework for understanding corporate sustainability. The TRST01 ESG Scoring Model sets a new standard in sustainability evaluation, offering valuable insights for stakeholders committed to integrating sustainability into core business strategies.


INTRODUCTION
In today's dynamic corporate environment, evaluating a company's sustainability performance is crucial for investors, stakeholders, and policymakers.The TRST01 ESG Scoring Model is an innovative tool that offers a thorough assessment of a business's Environmental, Social, and Governance (ESG) credentials.This model distinguishes itself from traditional assessment tools by employing a beta-based scoring methodology, which integrates financial metrics with ESG evaluations to provide a nuanced and statistically robust analysis.The demand for a sophisticated and reliable ESG assessment tool is higher than ever.Companies face increasing scrutiny over their sustainability practices, and investors are eager to incorporate ESG factors into their decision-making processes.
The TRST01 model is designed to meet various needs.It provides reliable data by sourcing information from reputable sources such as MSCI and S&P, ensuring that its ESG scores are accurate and dependable.The model also enhances investment decisions by incorporating financial metrics, offering a balanced view of a company's risk profile and market dynamics, which is essential for making informed investment decisions.With a strong statistical foundation and a significant R-squared value, the model demonstrates statistical precision, ensuring consistency and precision in ESG scoring.
The TRST01 ESG Scoring Model thus represents a critical advancement in ESG evaluation, addressing the growing need for reliable, comprehensive, and precise sustainability assessments.Environmental, social, and governance (ESG) concepts in investing strategies are increasingly important in today's financial world.Investors  ideals help or hurt financial performance.Our research examines how ESG ratings affect financial performance using the Extreme Gradient method and cutting-edge machine learning.ESG score is one of the top five predictors of mutual fund performance, according to our research.As sustainable investing becomes more important in financial markets, this study shows how ESG aspects affect investment outcomes.Socially responsible investors and financial advisors can use our information to make informed decisions that fit their financial goals with their ESG beliefs (Momparler et al., 2024).
Random Forests (ML model) is to examine the correlation between betas and 13 financial and non-financial characteristics for S&P 500 stocks from 2015 to 2019.The Relative Importance criterion indicates that ESG Scores are the primary factor in beta generation, determination, and sign.They confirm a correlation between betas and industrial sectors, revealing some variation across the investigated betas.Using the Random Forests methodology ensures relevant results by preventing plausible correlations between variables(Martín-Cervantes and Valls Martínez, 2023).
For private enterprises or non-traded assets, equity betas are often estimated using comparable company research.Test if the Random Forest method may improve projections.Every year from 1992 to 2018, Random Forest forecasts have fewer average and mean absolute errors in out-of-sample tests (Alanis, 2022).(Avramov et al., 2022) examined the impact of uncertainty surrounding ESG scores for publicly traded securities on the NYSE, AMEX, and Nasdaq.High uncertainty led to increased beta risk of stocks, altering the risk-return profiles of the evaluated companies.
This study examines how the global financial crisis affected multihorizon systematic risk and market risk using daily data from eight major European equities markets from 2005 to 2018.Wavelet multiscale is used in a capital asset pricing model.Beta coefficients grow at greater scales (lower frequencies) according to empirical data.During a crisis, betas and R2s tend to grow larger than before.According to scale-dependent value at risk (VaR), market risk is concentrated at lower time scales (greater frequencies) of data.Finally, our method accurately predicts time-dependent betas and VaR (Alexandridis & Hasan, 2020).This paper introduces Weighted Forward Search (FSW) for outlier detection in asset pricing data.It down weights the most anomalous observations and tests using simulated and empirical data.It assesses outliers' impact on asset pricing model estimation and introduces statistical tests based on this new approach.It also presents an alternative robust portfolio beta estimation approach for comparing asset pricing models.(Aronne et al., 2020).This study uses ESG data to create an automated trading strategy and assess a company's ESG premium and ESG alpha.It involves establishing an ESG investing universe, conducting feature engineering on ESG data, training machine learning models to forecast stock prices, and automatically managing portfolios using an ensemble method.Trades are executed and the portfolio adjusted weekly based on forecasted stock prices.(Chen & Liu, 2020).
ESG information affected organizations valuation and performance through their systematic (lower capital costs and higher valuations) and idiosyncratic (increased profitability and reduced tail risk) risk profiles.Research suggests that ESG reforms may benefit a company's finances.ESG ratings in policy benchmarks and financial evaluations may work (Giese et al., 2019).
The empirical evidence on the performance of responsible investing (RI) that combines environmental, social, and governance (ESG) into investment decisions is mixed.Fama and French's five-factor model and an ESG-related factor apply to 1,425 US openend equity fund returns from April 2009 to December 2016.The data suggests that RI protects against ESG-related systematic risk that diversification cannot minimize.(Jin, 2018).
The beta anomaly, or low (high) abnormal returns of equities with high (low) beta, is one of the most persistent anomalies in empirical asset pricing research.Investor desire for lottery-like equities drives the beta anomaly, as shown in this article.The beta anomaly disappears when beta-sorted portfolios are neutralized to lottery demand, regression specifications adjust for lottery demand, or factor models include a lottery demand factor.The beta anomaly only occurs in low-institutional-owned equities when lottery demand affects high-beta stocks (Bali et al., 2017).
The use of ESG scores and a European panel dataset with 8752 company-year observations from 2002 to 2014 to evaluate how Corporate Social Performance (CSP) strategies affect corporate financial performance.Thus, they show that only the second of the three ESG (Environmental, Social, and Governance) scores corporate governance activities related to the social aspect offers firms under analysis a significant opportunity to increase their value at the expense of reduced systematic risk (Sassen et al., 2016).
This study demonstrated how business sustainability affects processes and performance.They found in a matched sample of 180 U.S. companies that high sustainability companies, which voluntarily adopted sustainability policies by 1993, had distinct organizational processes by 2009, compared to low sustainability companies, which adopted almost none.Finally, high sustainability companies beat their peers in stock market and accounting success over time (Eccles et al., 2014).
This study examines CSR and corporate risk in controversial industries.After controlling for company characteristics in a large 1991-2010 U.S. sample of problematic industries like alcohol, tobacco, gambling, and others, CSR activity inversely affects business risk.They find that CSR participation in problematic industries lowers business risk using system equations and difference regressions to address endogeneity.The comparison of non-controversial and controversial firm samples shows that CSR engagement decreases risk more economically and statistically in controversial enterprises (Jo & Na, 2012).

METHODOLOGY 2.1. Data Collection
The carbon emission factors are sourced from Carbon Footprint, which provides opensource data suitable for organizations reporting their carbon emissions.These factors are collated from numerous sources and reflect the carbon emissions per kWh of electricity for different countries.The average industry scores of MSCI and S&P Global for top 1000 listed companies have taken as a base reference for calculation of adjusted scores.In addition, to this we have collected scope 1 and scope 2 emissions data from BRSR reports, Market cap, and beta value of stock for the past years 2023 and 2022 data and segregated all into respective industries.The TRST01 Adjusted Scores are calculated in a two-step process: first, adjusting the average scores based on country specific beta values, and then further adjusting these scores based on the company's beta value.Link: https://www.carbonfootprint.com/international_electricity_factors.html

Calculation of Country Beta
The calculation of country beta is an essential part of evaluating the environmental performance of countries based on their grid electricity carbon emissions.This process involves the use of data from Grid Electricity Conversion Factors to create a standardized metric that allows for meaningful comparisons of carbon emissions across different countries.Here's a detailed explanation of the scope and calculation of the country beta.

Scope
The country beta calculation involves:  Collecting carbon emission factors for grid electricity from various countries over the past five years (2018 -2023) and have taken two years data for future calculation in analysis to maintain uniformity across the years. Using these emission factors to determine a relative measure (beta) that reflects each country's carbon intensity compared to a baseline. Adjusting the beta values to account for variations across countries and provide a standardized metric for comparison.

Calculation Process
The process of calculating the country beta is as follows: 2.3.2.1.Data Collection  Gather carbon emission factors (in kg CO2e per kWh) for grid electricity for all countries from 2018 to 2023. Ensure the data is aligned and consistent across the five-year period.

Baseline Emission Factor
 Set a baseline emission factor, Ebaseline, which is a reference value for comparison.In this case, Ebaseline=0.4.This average reference is taken to involve as many countries as possible.

Calculate Average Emission Factor
 Compute the average emission factor for all countries for a specific year.This is done by averaging the emission factors of all countries:

𝑛
Where Ei is the emission factor of the ith country and n is the total number of countries.

Determine Country Specific Ratios
 For each country, compare its emission factor to the baseline.If the country's emission factor is greater than the baseline, use the emission factor; otherwise, use zero. Calculate the ratio of this adjusted emission factor to the average emission factor Ri =

Calculate Adjusted Beta for Each Country
 The country adjusted beta is determined by normalizing the country specific ratios by the average of these ratios across all countries: where  ̅ is the average of all country specific ratios Ri.
The country beta provides a relative measure of a country's carbon intensity in grid electricity production compared to a baseline and an average emission factor.By normalizing these values, the TRST01 model creates a standardized metric that allows for meaningful comparisons of carbon emissions across different countries.This process ensures that stakeholders can accurately assess and report on the environmental performance of organizations in various regions.The adjustment of ESG scores using the company beta is grounded in financial risk management principles.Here's a deeper look into the rationale.

Understanding Beta
Beta is a measure of a company's volatility or risk relative to the overall market. A beta greater than 1 implies the company is more volatile and riskier than the market.To reflect this risk, the ESG score should be reduced, indicating the company must do more to achieve the same level of ESG performance. A beta less than 1 implies the company is less volatile and hence less risky than the market.The ESG score should be increased to reflect the lower risk, indicating the company's ESG performance is more credible due to its lower risk profile.

Final Adjusted S Score=
Adjusted S Score from Country Beta x (

MODELS PERFORMED 3.1. Simple Linear Regression
Regression is a statistical term used for describing models that estimate the relationships among variables.Linear regression models study the relationship between a single dependent variable Y and one or more independent variables, denoted by X.If there is only one independent variable X, it is called simple linear regression; if there is more than one independent variable, it is called multiple linear regression.This Accidental Note discusses the statistical and geometric interpretation of simple linear regression models (Bangdiwala, 2018).

The mathematical simple linear regression model
The simple linear model is exactly that, a simple straight line that relates the one independent variable X to the dependent variable Y.It is given by the mathematical formula for a straight line.
where  0 is called the intercept and  1 is called the slope.In the standard x-y Cartesian plane, the intercept is the point on the y-axis that is intersected by the line, and the slope is the amount of change in the y-axis for a 1 unit change in the x-axis.The intercept and slope are the two values that completely characterize a straight line, i.e. by providing these two values, one knows exactly the only straight line that they describe.The intercept is interpreted as 'positioning' the regression, while the slope is interpreted as 'quantifying the association' between X and Y.If the slope is 0, there is no association; if the slope is negative, the association is negative (for a unit increase in X, Y decreases by  1 ); and if the slope is positive, the association is positive (for a unit increase in X, Y increases by  1 ).The larger the absolute value of the slope, the 'stronger' is the association.

Random Forest (Regression)
Random forest (RF) is a flexible, easy-to-use machine learning method that, in most cases, delivers good results even without hyperparameter tuning.Because of its simplicity and diversity, it is also one of the most often used algorithms.Random forest is a supervised machine learning algorithm.It builds a "forest" out of an ensemble of decision trees, which are generally trained using the "bagging" process.A Random Forest is made up of several trees that are built in a specific "random" manner.Each tree is made up of a distinct sample of rows, and each node is split up into a different set of features.Each tree has its prediction.After then, the average of these predictions is used to create a single result.The bagging method's general concept is that combining several learning models improves the outcome.Form of the regression trees model Where, R1, R2, …, RM represent a partition of feature space.

Gradient Boost
Gradient boosting regression tree algorithms employ an ensemble learning approach to create robust forecasting models by integrating multiple individual regression trees, also known as weak learners.These algorithms aim to minimize the error rate of weakly learned models, which have high bias and low variance on the training dataset.Boosting algorithms typically consist of three components: an additive model, weak learners, and a loss function.This approach can represent non-linear relationships and utilizes various differentiable loss functions.Gradient boosting machines (GBM) work by identifying the limitations of weak models through gradients and use an iterative approach to gradually improve the base learners and reduce forecast errors.This is achieved by combining decision trees through an additive model while minimizing the loss function using gradient descent.

Support Vector Regression (SVR)
Support Vector Regression (SVR), the supervised learning algorithm is used to predict discrete values.SVR's main aim is to locate the line of best fit.The best-fit line in SVR is the hyperplane with the maximum number of points.The SVR, unlike other regression models, aims to fit the best line inside a threshold value, rather than minimizing the error between the real and projected value.The distance between the hyperplane and the boundary line is the threshold value.The goal of Support vector regression is to develop a function that approximates mapping from an input domain to real numbers by using training sample data.The key goal here is to choose a decision boundary that is a distance from the original hyperplane and contains data points closest to the hyperplane or support vectors.
Fitting of Support Vector regression () =  0 +  1  1 + ⋯ +     can be expressed as The main idea behind SVR is to transform the input space into a high-dimensional variable space and then build the regression or time series model in this transformed feature space.A vector of data set says  = {    } =1  , where   ∈   is the input vector,   is the scalar output, and N is the size of the data set.The general equation SVR can be written as follows: where, W is the weight vector, b is the bias term, and superscript T denotes the transpose.The coefficients W and b are estimated from data by minimizing the following regularized risk function: This regularized risk function minimizes both the empirical error and the regularized term simultaneously, helping to avoid both underfitting and overfitting of the model.The first term 1/2|| w ||2 is called the 'regularized term', which measures the flatness of the function.Minimizing 1/2|| w ||2 will make a function as flat as possible.
The second term 1/∑ =1    (  , (  ) is called the 'empirical error', which was estimated by Vapnik ε-insensitive loss function Where n is number of data points,  is the observed values,  ̂ is the predicted values.

R Square
The coefficient of determination,  2 , is a measure to goodness of fit of a model.In the context of regression, it is a statistical measure of how well the regression line approximates the actual data.The Simple Linear Regression model shows poor performance across all scenarios (E, S, G) with significant prediction errors and residuals.Scenario E and S have more pronounced issues with skewed and bimodal residual distributions, respectively, while scenario G shows a more balanced error distribution but still with significant prediction inaccuracies.Table 1 clearly explains SLR is poor fit due to least r square and high errors.

Combined Actual vs.
Predicted Plots Let's analyze these combined actual vs. predicted plots for the three models (Random Forest, Gradient Boosting, and Support Vector Regressor) across the three targets (E, S, G).
Figure 4 clearly shows Gradient Boosting and Support Vector Regression models generally show better performance, with points more closely aligned with the diagonal line across different scenarios (S and G in particular).Random Forest shows slightly more spread in its predictions, indicating more variance in its accuracy, especially in the E scenario.Scenarios S and G seem to yield more consistent predictions across all models compared to scenario E, which shows more deviation in the plots.

Residual Plots
These plots show the residuals (errors) vs. the predicted values.Ideally, residuals should be randomly scattered around zero without any discernible pattern.Figure 6 stated that Gradient Boosting and Support Vector Regression models generally show tighter clustering of residuals around the zero line, indicating higher accuracy and less variance in predictions.Random Forest shows more spread in residuals, particularly in scenario E, indicating more variability in its predictions.

Learning Curves
The Gradient Boosting and Support Vector Regression models generally show better performance with training and cross-validation scores closely aligned, indicating good generalization.Figure 7 clearly shows that the cross-validation scores improve significantly and get closer to the training scores as more data is added, reducing overfitting.This model shows higher variance in its accuracy, with consistently high training scores indicating overfitting.The cross-validation scores, while improving with more data, remain consistently lower than the training scores, especially in the E scenario.

Fig 1 .
Fig 1. Procedure for Calculation of Scores

Fig 2 :
Fig 2: Schematic representation for Model Building

ISSN: 2581-8341 Volume 07 Issue 06 June 2024 DOI: 10.47191/ijcsrr/V7-i6-09, Impact Factor: 7.943 IJCSRR @ 2024 www.ijcsrr.org 3609 * Corresponding Author: Gurucharan Kottapalli Volume 07 Issue 06 June 2024 Available at: www.ijcsrr.org Page No. 3598-3611 4.6. Comparison of models Table 2: Evaluation Criteria across models ML Models Mean Absolute Error Mean Squared Error
Shows a good balance between training and cross-validation scores, indicating it generalizes well with more data.The cross-validation scores improve significantly and approach the training scores, reducing overfitting. Actual vs Predicted Plots: Points closely align with the diagonal line, indicating strong prediction accuracy. Residuals Plots: Residuals are more evenly distributed around zero with minimal spread, indicating low prediction errors. Residuals Distribution Histograms: Residuals are normally distributed, suggesting the model does not have significant biases and performs consistently well across different scenarios.Demonstrates a steady improvement in both training and cross-validation scores with more data.The narrow gap between these scores indicates good generalization capabilities. Actual vs Predicted Plots: Points align well with the diagonal line, showing good prediction accuracy. Residuals Plots: Residuals are fairly evenly distributed around zero, with some minor spread indicating slight prediction errors. Residuals Distribution Histograms: Residuals are fairly normally distributed, showing the model's consistent performance and balanced error distribution.Training scores remain high, but cross-validation scores are lower, indicating overfitting.Despite some improvement in cross-validation scores with more data, the consistent high training scores suggest overfitting issues. Actual vs Predicted Plots: Points show more spread from the diagonal line, indicating higher variance in prediction accuracy, especially in scenario E.  Residuals Plots: Residuals show larger spreads and patterns, indicating inconsistent predictions and overfitting.