In recent years, corporate management has shifted its focus from solely financial indicators to also include ESG indicators, which assess the environmental (E), social (S), and governance (G) aspects. However, the impact of ESG indicators on company management varies depending on several factors and is closely tied to the specific social issues each company prioritizes. Consequently, to develop a comprehensive business analysis model that incorporates both ESG and financial indicators, we developed a method for analyzing ESG-financial indicators using a causality hypothesis graph with a structural equation modeling. This method enables us to examine (1) the interrelationships between diferent indicators and (2) the validation of hypotheses regarding the issues that a company should prioritize. We also developed a simulator predicts future financial indicator values by comprehensively combining multiple measures. We evaluated this technology by applying it to our corporate data and present the comparative results of predicted financial indicator values with and without the implementation of a measure.
In recent years, there has been a growing emphasis on corporate social responsibility in business activities. This includes considerations such as the environmental impact of energy and resource consumption, as well as the appropriateness of employment practices. In corporate management, not only financial indicators but also ESG (Environmental, Social, and Governance) indicators, which reflect the status of the environment, society, and governance, are being given increased importance [1].
There have been numerous studies examining the relationship between ESG indicators and corporate performance [2, 3], we are working on the development of a support system to examine how companies can incorporate ESG activities into their management. When applying ESG support systems to actual management, the following challenges have been identified: 1. In cases where the interpretation of statistical analysis results does not align with the actual business situation, managers are unable to make appropriate judgments. 2. Without providing a rationale for the significance of engaging in ESG initiatives, the priority given to ESG eforts in the field may decline. 3. Merely focusing on the relevance of indicators without ofering specific implementation measures does not lead to improvements in management.
To address the specific challenges mentioned above (
Furthermore, in order to address the challenges of (
In this paper, we discuss the efectiveness of this method by providing examples of analysis results from actual companies.
As mentioned above, in order to give interpretability to the analysis of the relationship between ESG indicators and financial indicators, it is necessary to analyze the causal relationship between the indicators. Although it is possible to use statistical methods to estimate causal relationships, we considered the following points and conducted an analysis method in which experts created causal relationship hypotheses and verified them with data: • ESG data is often available for a period of about 10 to 20 years per fiscal year, and it is dificult to obtain statistically suficient data. • When formulating ESG-related policies, there is a demand from a management side to verify the intended causal relationships. • There are constraints on the causal relationships between ESG indicators and financial indicators, and it is necessary to incorporate these constraints into the analysis model.
In particular, the constraints on causal relationships mentioned above often have cases where the positive or negative coeficients are assumed in advance, even if the size of the impact relationship is unknown, such as "if the water usage increases in a company, the cost will increase accordingly". Therefore, we conducted an analysis using structural equation modeling, which incorporates manually created causal relationship hypotheses within the constraints.
To construct causal relationship hypotheses, a workshop was conducted with approximately 20 stakeholders, including the authors, as well as departments within the company such as finance, human resources, and sustainability management. During the workshop, over 130 ESG and financial-related indicators and more than 300 hypotheses regarding their relationships were extracted. The hypotheses of causal relationships that were created as a result were used within the scope of available empirical data. In cases where alternative data for a particular indicator existed, they were also included in the causal relationship hypothesis model. Additionally, multiple structural equation modeling iterations were performed, making adjustments to the model, such as incorporating factor analysis, to address statistical issues like multicollinearity.
Structural Equation Modeling (SEM) can be regarded as a generalization and systematization of linear regression and factor analysis, and is useful for analyzing causal relationships. In this study, we particularly used factor analysis as a partial solution to the problem of multicollinearity in datasets. We also applied linear regression to incorporate constraints between factors. Additionally, to model the cyclic relationships between indicators, we introduced time delays for some indicators. For example, although research and development investment is an expenditure in a single fiscal year, if it contributes to sales and generates profits through business creation, it is assumed that the contribution to sales was made by past research and development investment, and this was represented by a time delay.
Factor analysis, which is included in structural equation modeling is a statistical analysis method for
estimating the factors behind observed variables obtained through experiments or observations. For
example, for multiple observed variables, if there are common factors that are common factors, an
observed variable can be expressed by the factor loadings and the common factor and the unique
factor as follows:
= ∑︁ +
(
Here, the unique factor is a component specific to each item, and it only afects one observed variable, and the factor loading represents the strength of the relationship between the factor and the observed variable. In this analysis, factors were specifically set for datasets that tend to have problems with multicollinearity. For example, the results of survey questionnaires in companies have high correlations between each question, and if each question in the questionnaire is individually added to the causal relationship hypothesis model, multicollinearity problems occur in the regression analysis of structural equation modeling. In such cases, we considered the intention of the questionnaire and added factors.
For example, in cases such as "an increase in water usage in a company leads to a corresponding increase in costs," it is believed that there is an environmental indicator of water usage and a usage fee (cost) that is expected to increase proportionally. A positive proportionality constant is assumed to exist between them. However, since accounting items that only disclose water usage fees are rare, it becomes necessary to conduct linear regression with a positive coeficient for accounting items that include water usage fees.
In this case, the analysis is performed by setting the range of possible coeficient values to be non-negative. Similarly, constraints were applied between indicators where a positive or negative relationship was assumed in advance for the purpose of conducting the analysis.
The results of the analysis using structural equation modeling can be used to read the direct impact of indicators, but the impact of indicators that have causal relationships through other indicators cannot be interpreted. In this study, in particular, it is important to evaluate the impact of non-financial indicators on management indicators, but non-financial indicators rarely have direct causal relationships with management indicators. Therefore, in this technique, the impact of each indicator on financial indicators is calculated based on the causal relationship paths.
Assuming that the impact of indicator on indicator is to be calculated, the calculation formula is explained. In this case, there may be a direct causal relationship between indicator and indicator , or they may be related through other indicators. Moreover, there may be multiple causal paths between indicator and indicator .
In this case, the impact of indicator on indicator is defined as the sum of the impact through all causal paths from to . If there are causal paths from indicator to indicator , and the direct impact of a certain indicator on another indicator in a certain path is represented as , the impact from indicator to indicator in path is calculated as the product of the impact of indicators on that path, based on the characteristics of linear regression models. Therefore, if path consists of indicators, it can be expressed as follows:
Here, , represents the impact of the th indicator on path . Then, the total impact from indicator to indicator is defined as the sum of all impact through the paths, so if there are paths, it can be expressed as follows:
= ∏︁ , =1
= ∑︁
=1
Additionally, financial indicators have defined calculation formulas based on indicators. For example, operating profit is calculated by subtracting selling, general, and administrative expenses and cost of goods sold from sales. The impact on such indicators is calculated by performing calculations based on the predefined formulas for each indicator and then calculating the impact after removing the standardization of the data, such as averaging 0 and standard deviation 1.
Future predictions in the simulator are based on the analysis results of causal hypotheses and structural equation modeling. The prediction of each indicator is executed by propagating the predicted values along the causal relationships, starting from the predicted value of the indicator that serves as the cause, to predict the value of the indicator that is influenced by the cause. This allows the non-financial indicators that change due to the applied measures to afect the financial indicators according to the intended causal relationships. The following formulas formalize the prediction methods for the five types of indicators that appear in the causal relationships.
(
For indicators (
Here, , represents the value of indicator in year , and , is the standard deviation of the data
for indicator over years from the latest year.
(
Then, using the already obtained predicted value , of indicator , an optimization calculation is
performed to minimize the squared error between ˆ, and ,. The obtained factor score , is used
as the predicted value of indicator (
ˆ, = , · , + , min ∑︁(ˆ, − ,)2
For indicator (
Here, the regression coeficients , and the intercept are the regression coeficients and intercepts for indicator in the regression equation obtained from the structural equation modeling analysis.
The prediction of indicator (
Indicator (5) assumes financial indicators that can be calculated based on formulas, such as ROA and operating profit. For such indicators, the calculation method is defined before conducting the simulation, and the predicted values of the relevant indicators are used to calculate them.
Using the above methods, the predicted values of each indicator in the causal relationships can be obtained for a single year. By repeating this process, predictions for future years can be made.
To apply the efects of policy measures in this simulator, the settings of the measures need to be defined in advance. The settings of the measures refer to the indicators to which the measure efects are applied, the strength of the efects, and the start year of the measures, for example. The simulation is then conducted based on the settings of the measures, applying the measure efects to the predicted values of the target indicators. When a measure efect is applied to a certain indicator, the efects are propagated to other indicators that have causal relationships with that indicator. As a result, it is possible to calculate how the financial indicators have changed due to the measures.
First, let’s consider the case of applying a single measure. If we denote the value of indicator in year before and after applying the measure as ′, and ,, respectively, the equation for applying the measure efect is as follows: , = (′,, Φ ,) (for ≤ < + ) (8)
Here, Φ , represents the measure efect that measure has on indicator in year , and represents the function for applying the measure efect. represents the start year of measure . For example, if a measure increases indicator by 10 in year , then Φ , = 10, and (′,, Φ ,) = ′, + Φ ,. represents the duration of the measure efect that measure has on indicator . The reason for defining is to consider that the duration of the impact from the measures difers depending on the indicator.
In this study, only indicators (
= (∑︁ , · ,) · Φ , + ( + ,) · Φ ,
The first term in the above equation can be interpreted as the values of each indicator or the measure
efects applied to the values or regression coeficients of each indicator. Therefore, this equation is
considered inappropriate, and the application to indicator (
The reason for not considering the application to indicator (
Next, let’s consider the case of applying multiple measures simultaneously. When multiple measures
have efects on a certain indicator, using Equation 8 would result in diferent prediction results depending
on the order of applying the measures. Therefore, the following equation is used to calculate the change
in the indicator due to each measure, and the sum of these changes is added to the value of the indicator
to obtain the predicted value when the measures are applied:
not defined in the measure settings would be involved in the calculation formula. For example, if the
measure efect is applied to indicator (
[ (′,, Φ ,) − ′,] + =1,=′ ∑︁ [∑︁( (′,, Φ ,+− 1) − ′,)] =1,=′′ =1 ∑︁ =1,=′ , = ′, +
[ (′,, Φ ,) − ′,] (if ′ ∈ and ′ ≤ < ′ + ′ )
Here, is the set of measures . ′ refers to the measure in that is applied within the target year for the prediction. This allows the efects of each measure to be applied regardless of the order of application.
Indicators (
(if ′ ∈ and ′ ≤ < ′ + ′ , if ′′ ∈ and = ′′ + ′′ and flag ′′ = False)
3.1. Data Table 1 shows the actual authors’ afiliation corporate data to which this technique was applied. The applied data includes ESG disclosure indicators, financial indicators, and survey data within the company. Financial indicators have been disclosed every year from 2004 to 2021, which is generally the range of data acquisition. Environmental disclosure indicators have also been obtained over a relatively long period of about 10 years. Social data such as average overtime hours can be obtained for the same period as the financial indicators, and the female and foreign oficer ratios belonging to governance can be obtained from 2013. The survey data, which started in 2018, is relatively new and asks whether companies and workplaces value diversity and whether the work environment is conducive to smooth operations. The average rating given by employees was obtained. The survey used in the data analysis consisted of 14 questions, including 3 questions related to diversity and 11 questions related to the work environment.
As mentioned earlier, time delayed indicators were added to the variables to incorporate temporal causality into the structural equation modeling. In particular, time delayed indicators were set for the survey results on the work environment and research and development investment to model the time relationship between sales and sales forecasts. Furthermore, a delay of 1-2 years was set for environmental indicators, which are alternative variables for the cost of capital.
When applying structural equation modeling, the above data was normalized to have a mean of 0 and a variance of 1.
In this study, due to the high correlation among the survey results within the company, two factors were calculated: One representing diversity and the other representing the work environment, based on the aforementioned survey data. The factor coeficients ranged from 0.77 to 1.11 for diversity and from 0.87 to 1.01 for the work environment.
The calculation of the impact on financial indicators was performed using the method described in Section 2.3.1. In this case, the impact on two financial indicators, operating profit and ROA, was calculated. Operating profit was defined as "sales minus selling and administrative expenses and cost of goods sold". For ROA, it was calculated by dividing net income by total assets. However, net income was derived through statistical impact calculation using operating profit, interest expense ratio, and ordinary income, and then calculated based on the actual value of total assets.
The model was implemented using the Python implementation of structural equation modeling, semopy1. Structural equation modeling was applied multiple times with diferent seed values, and the average impact and its variance were calculated.
Figure 1 shows the analysis results of the relationship between ESG-financial indicators based on the hypothesis of causal relationships. In the figure, only the indicators of the environmental category are shown with a one-year time lag, but data from two years ago are also used. However, the influence from two years ago was almost zero. Regarding the interest-bearing debt interest rate, which is used as a substitute variable for the cost of capital, it can be seen that the coeficient of the energy consumption and waste generation one year ago is positive, indicating that the cost of capital increases as these values increase (environmental burden increases). Therefore, it can be concluded that reducing energy consumption and waste generation is efective in obtaining funds at a lower cost from external sources. In addition, the diversity factor node has a negative coeficient with respect to the interest-bearing debt interest rate, indicating that improving diversity is beneficial for obtaining funds.
Sales forecast were analyzed using multivariate analysis with factors such as working environment and research and development investment from 3-5 years ago, aiming to represent new businesses and business development potential. As a result, it was found that the working environment and research and development investment from 4 years ago have a positive coeficient on sales. This suggests that improving the working environment and investing in research and development are efective for improving company performance. Furthermore, water usage and energy usage have positive coeficients on sales forecast, indicating that they have an increasing environmental impact on business expansion. However, it is also found that these factors do not have a significant impact on sales costs.
It was also shown that research and development investment and overtime hours in the current year lead to an increase in costs and selling, general, and administrative expenses.
When applying structural equation modeling, the aforementioned data was normalized to have a mean of 0 and a variance of 1.
As can be seen from Figure 1, the improvement in the ratio of renewable energy in the past contributes to a decrease in the interest rate on interest-bearing debt and ultimately leads to an improvement in ROA. Therefore, in this experiment, we set "improvement in the ratio of renewable energy through investment in solar power generation facilities" as a policy and examined its impact on future ROA.
To set the policy, we calculated the cost-efectiveness of investment in solar power generation facilities. According to the materials from the Agency for Natural Resources and Energy2, the capital cost for introducing solar power generation facilities in 2023 is estimated to be 22.3 [thousand yen/kW]. 1https://pypi.org/project/semopy/ 2https://www.meti.go.jp/shingikai/santeii/pdf/091_01_00.pdf (a) Renewable Energy Ratio (b) ROA
Based on this, we calculated the impact on the target company using data from the fiscal year 2021, and obtained the result that "an investment of 20 billion yen will increase the ratio of renewable energy by approximately 5.26%". In addition, the useful life of a solar power generation system is specified as 9 years according to the National Tax Agency’s website3.
Based on the cost-efectiveness calculation results, we selected the ratio of renewable energy, depreciation expenses, and total assets as the indicators afected by the policy, and set the policy accordingly. The ratio of renewable energy is assumed to increase by 5.26% in the year the policy is applied, and its impact continues thereafter ( = 1, flag =True). Depreciation expenses and total assets are assumed to include 20 billion yen in costs for 9 years from the year the policy is applied, and the impact disappears from the 10th year ( = 9, flag =False). In addition, we set in Equation 4 to 8.
In the next section, we will discuss the results of simulating the application of this policy in 2023 and conducting the simulation until 2030.
Figure 2 shows the predicted changes in the renewable energy ratio and ROA through simulations. The simulations were run 100 times with diferent seed values. Figure 2 shows the average predicted values and standard deviations.
It can be observed that the renewable energy ratio increases by the start of the policy in the fiscal year 2023. Looking at the predicted values of ROA, it decreases from 4.59% in the case of no policy implementation to 4.53% when the policy is applied. This can be attributed to the fact that the increase in the renewable energy ratio has not yet afected the interest coverage ratio and at the same time, the investment costs have increased depreciation expenses and total assets.
However, by the fiscal year 2030, it is predicted that the ROA will increase to 4.14% when the policy is implemented, compared to 4.02% in the case of no policy implementation. This suggests that the positive efect of the increase in the renewable energy ratio on ROA outweighs the negative efect of investment costs.
Based on these results, it can be considered that this policy has the potential to improve ROA and is an efective measure for the target companies.
Furthermore, this method suggests that by modeling the relationship between non-financial and ifnancial indicators and simulating how measures to improve non-financial indicators afect financial indicators, it is possible to propose ESG indicators and measures that companies should promote while considering long-term returns.
Additionally, this method suggests that it is possible to propose ESG indicators and measures that companies should promote within the balance of various financial and ESG indicators.
We developed a method for analyzing ESG-financial indicators using a causality hypothesis graph with
a structural equation modeling. This method enables us to examine (