Carbon Risk Premium: Fact or Fallacy? A study of the relationship between carbon emission and stock returns

(1)

Carbon Risk Premium: Fact or Fallacy?

A study of the relationship between carbon emission and stock returns

Kevin van Dam

Student number: 12478938 University of Amsterdam

Faculty of Economics and Business – Finance (6013B0520) Supervised by: Dr. Patrick Behr

Bachelor’s Thesis June 2022

Abstract

This study investigates the relationship between carbon emission and stock returns using panel data consisting of 1,351 distinct companies from the regions Africa/Middle East, Asia/Pacific, Europe, Latin America and Caribbean and United States and Canada. The dataset represents 42 industries from the second quarter of 2016 up to and including the third quarter of 2019. Using OLS regression techniques, it is found that there is a significant

relationship between carbon intensity and stock returns. However, controlling for fixed effects removes this significant relationship, which is in line with prior literature. Furthermore, when a distinction between regions is made, it can be concluded that in three of the four regions studied the relationship between stock returns and carbon intensity is insignificant. However, this relationship is negative and significant when only observations that belong to the region Asia/Pacific are considered.

(2)

Statement of Originality

This document is written by Student Kevin van Dam who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no

sources other than those mentioned in the text and its references have been used in creating it.

UvA Economics and Business is responsible solely for the supervision of completion of the work and submission, not for the contents.

(3)

1.

Introduction

In his famous New York Times essay, Milton Friedman stated that the sole “social

responsibility of business is to increase its profits.” (Friedman, 1970). Furthermore, economic theory states that firms should not internalize negative externalities they exert on employees, government, or society as a whole (Pigou & Aslanbeigui, 2017). However, Environmental, Social and Governance (ESG) measurement, Socially Responsibility Investing and Corporate Social Responsibility activities are becoming more important topics in academic research and professional areas (Larcker & Watts, 2020). Although it is very difficult to put an exact number on the amount spend on Corporate Social Responsibility initiatives, Hong, Kubik and Scheinkman (2012) concluded in their research that the amount spend on CSR initiatives by large U.S. corporates can end up in the hundred of millions of dollars.

After the Paris Agreement in 2015, initiatives to reduce greenhouse gas emissions such as carbon emissions accelerated (by necessity). The nearly 200 participating countries agreed on a binding climate accord that aims to limit global warming to no more than 1.5 degrees Celsius. On November 4, 2016, this agreement entered into force. Now the question arises whether a company's reduction in carbon emissions also affect its stock prices, which is also the main research question in this paper: Are a firm’s carbon emissions correlated with its stock return?

To answer this question, panel data is gathered on a sample of 1,351 distinct companies from the regions Africa/Middle East, Asia/Pacific, Europe, Latin America and Caribbean and the United States and Canada. The observations belong to the period from the second quarter of 2016 up to and including the third quarter of 2019. By doing multiple regression analyses, both with and without controlling for fixed effects, an answer to the research question can be given. It can be concluded that the relationship between carbon emissions and stock returns is negative and significant when fixed effect are not included. These findings are in line with findings of In et al. (2017).

However, when you control for fixed effects, this relationship turns insignificant, which is in line with prior literature of Aswani et al. (2022) and Bolton and Kacperczyk (2021). This finding also answers the first sub question of this paper: Is there a change in relationship between carbon emissions and stock return when controlled for fixed effects compared to when no fixed effects are included?

(5)

In their paper, Brandon et al. (2021) conclude that institutions that publicly commit to Responsible Investing also show better portfolio-level scores. However, this relationship only holds for institutions outside the United States. In other words, non-U.S. investors value

‘green investing’ more than U.S. investors, so there must be a significant difference in the relationship between ‘green investing’ and stock performance across regions. Therefore, in contrast to prior literature, this study makes a distinction between 5 different regions:

Africa/Middle East, Asia/Pacific, Europe, Latin America and Caribbean and United States and Canada. It can be concluded that there is a significant difference in the relationship between stock return and carbon emissions across regions. It is found that this relationship is negative and significant when observations that belong to the region Asia/Pacific are

considered. However, when only observations are considered that belong to the regions Africa/Middle East, Europe and Latin America & Caribbean, this relationship is insignificant.

This answers the second sub question of this paper: Does the relationship between carbon emissions and stock return differ across different regions?

This paper is structured as follows: To understand the relationship between environmental performance and financial performance, and in more detail, the relationship between carbon emission and stock returns, a theoretical background is provided. Based on this theoretical framework, several hypotheses are put forward. This is followed by a description of the data selection, cleaning and preparation process and research methodology. The described methodology is applied, and the results are given. Finally, the conclusions of this paper are summarized, and several limitations are described. Based on these limitations, valuable recommendations for further research are given.

2. Related Literature

In this section, a theoretical background on the relationship between environmental and financial performance is given. Thereafter, this relationship is narrowed down to the relationship between carbon emissions and stock returns. The conclusions of the leading papers in this field are described and compared and based on these papers several hypotheses are put forward. This section ends with a description of the most important variables used in existing papers.

2.1. Environmental and Financial Performance

The debate on the relationship between environmental performance (EP) and financial performance (FP) has been going on for years. Theoretical and empirical arguments can be used for both a negative and positive relationship (In et al., 2017). For example, Friedman

(6)

(1970), Aupperle et al. (1985), Luken (1997), McWilliams and Siegel (1997), Clift and Wright (2000) and Jensen (2002) claim that improvement on environmental aspects are not compatible with maximization of profit. In other words, these researchers claim a negative relationship between EP and FP. On the other side, Schmidheiny (1992) and Porter and Linde (1995) suggest a positive relationship between EP and FP. They see environmental

improvements as cost-saving instruments that increase a firm’s competitiveness in a changing world (In et al., 2017).

Also empirically speaking there are arguments for both a positive and negative relationship between EP and FP. For example, Hart and Ahuja (1996), Russo and Fouts (1997) and Dowell et al. (2000) claim, after an empirical study, a positive relationship between EP and EF, while studies of Filbeck and Gorman (2004), Telle (2006) and Ziegler and Nogareda (2009) suggest a negative or insignificant relationship between EP and FP.

There are many ways to measure EP and FP. EP is measured using Environmental Performance Indicators that can be divided into Management Performance Indicators (MPI) and Operational Performance Indicators (OPI). MPI’s are measures undertaken by a firm’s top-level management to influence the firm’s environmental impact. Examples include percentage of employees with environmental training, percentage of environmentally friendly suppliers and emission quotas. OPI’s are indicators that provide information about the

environmental performance of a firm’s operations. For example, OPI’s can be measured by the amount of raw materials used, amount of total waste, electricity consumption per product produced or total carbon emission (Jasch, 2000).

There are a few ways to measure FP. Examples include stock return, Tobin’s q, Return on Assets (ROA), Return on Equity (ROE) and Return on Investment (ROI). Tobin’s q is the firm’s market valuation over the replacement value of assets. ROE and ROA are the ratios of income to the firm’s equity and assets, respectively. ROI is measured by dividing total operating income with the book value of assets (King & Lenox, 2001). In this paper, the relationship between EP and FP is narrowed down to the relationship between carbon emission and stock returns.

2.2. Carbon Emission and Stock Returns

Carbon emissions can be grouped in three main categories: Scope 1, Scope 2, and Scope 3. Scope 1 entails direct carbon emission from a firm’s production process. Scope 2, however, entails indirect emission from consumption of heat, electricity and/or steam. Finally, Scope 3

(7)

entails all other indirect emissions from production of purchased materials, waste disposal, outsourced activities, etcetera. (Bolton & Kacperczyk, 2021). Bolton and Kacperczyk (2021) made in their paper a distinction between these three scopes, and they researched the

relationship between carbon emissions and stock returns. They used data from Trucost EDX Database consisting of around 1,000 listed companies since fiscal year 2005 and over 2,900 listed U.S. companies since fiscal year 2017. They merged this data with corresponding balance sheet data from FactSet. They found in their research a statistically significant positive relationship between carbon emissions and stock returns for all three scopes.

Furthermore, Bolton and Kacperczyk (2021) researched how this carbon premium is related with three different measures for carbon emissions. In their research they used the total level of carbon emission, the year-by-year change in carbon emissions and a firm’s carbon intensity. Many other related research focusses solely on the correlation between unscaled carbon emission and stock performance (Aswani et al., 2022). However, because the level of carbon emission is highly determined by a firm’s product output, it is not clear whether these correlations can be used to draw conclusions about the relationship between carbon emission and a firm’s stock performance. A better measure to draw meaningful conclusions may be carbon intensity: the ratio of carbon emission to net sales. Using this measure avoids correlation with measures of firm size. Bolton and Kacperczyk (2021) concluded that when the total level of carbon emission is used as measurement for carbon emissions, the positive relationship between carbon emissions and stock returns is significant. The same conclusion holds when the year-by-year change in emission is used as the independent variable of interest. However, surprisingly, when carbon intensity is used as measurement for carbon emissions, this relationship turns out to be insignificant. One way to explain why this carbon premium is significant when the total level of emissions is used as measurement for carbon emissions is that regulations limiting emissions are more likely to focus on firms which have the highest level of emissions (Bolton and Kacperczyk, 2021). In other words, firms with a higher level of total emissions face more ‘carbon regulatory risk’, so investors require a significant carbon risk premium. On the other hand, since carbon intensity is a ratio, two firms with the same carbon intensity ratio may have very different levels of emissions and therefore may vary in carbon regulatory risk.

It is important to know that Bolton and Kacperczyk (2021) in their research assume that there is no significant difference between vendor-estimated carbon emission and actual firm- disclosed carbon emission numbers. Data providers as TruCost, Sustainalytics and MSCI

(8)

provide a mix of estimated carbon emissions for firms that did not disclose actual numbers and actual disclosed carbon emission data (Aswani et al., 2022). Busch et al. (2020) researched the correlation between actual disclosed carbon emission across several data providers. In their paper, they documented a correlation of around 0.97, suggesting that when firms disclose carbon emission figures, it is correctly implemented across all commercial data providers. However, when looking at the correlation between estimated carbon emissions across these several providers, Busch et al. (2020) documented a correlation of 0.66. It can be concluded that different data providers use different estimation methods to estimate carbon emissions. If these methods are based on firm characteristics and industry characteristics, it is not possible to capture within-industry differences using vendor-estimated numbers (Aswani et al., 2022). Furthermore, when vendor-estimated numbers are used to research the

correlation between carbon emission and stock returns, these results reflect only the correlation between firm fundamentals as firm size and sales growth and stock returns.

In contrast to Bolton and Kacperczyk (2021), Aswani et al. (2022) did make a distinction between vendor-estimated and actual disclosed carbon emission. Using a sample of 2,729 U.S. firms from 2005 until 2019, they found a significant difference in the correlation between vendor-estimated Scope 1 emission and stock returns and the correlation between actual Scope 1 emission and stock returns. More specifically, they found a significant relation between vendor-estimated emissions and stock returns, but the relationship between actual disclosed emissions and stock returns was insignificant in their paper. This suggests that the described correlation between carbon emissions and stock returns by Bolton and Kacperczyk (2021) is because of the relationship between firm fundamentals and stock returns.

Brandon et al. (2021) argue in their paper that European investors place more value on responsible investing compared to American investors. This suggests that even if Aswani et al. (2022) did not find a significant relationship between actual disclosed emissions and stock returns in the United States, such a relationship could exist in Europe. Using OLS regression techniques, Aswani et al. (2022) indeed found a positive significant relationship between unscaled carbon emissions and stock returns for European firms. However, controlling for industry fixed effects and time fixed effects removes this significant relationship. Also, when carbon intensity is used as a measure for carbon emissions, no significant relationship

between carbon emissions and stock returns is found after controlling for industry and time fixed effects.

(9)

The conclusions of Aswani et al. (2022) are in line with the conclusions in the paper of Larcker and Watts (2020). They researched the relationship between ESG (environmental, social and governance) investments and the price of securities. More specifically, they

researched if investors are willing to pay more for high-ESG securities compared to low-ESG securities with identical risk and maturity. In their paper, Larcker and Watts (2020) concluded that the premium that green assets trade compared to non-green assets, the so-called

‘greenium’, is equal to zero. In other words, an average investor is not willing to give up returns to invest in environmentally friendly securities. The concern for investors of greenwashing -using the funding proceeds for investments that have little environmental value- could be a possible explanation for the lack of a greenium (Larcker & Watts, 2020).

However, after two analyses, they found no significant relationship between a greenium and the environmental impact of projects associated with the funding proceeds of green bonds.

In et al. (2017) researched if investors achieve a higher return with low-emission

portfolios. Instead of doing a regression analysis, In et al. (2017) formed portfolios by carbon intensity tertiles and formed a carbon efficient-minus-inefficient portfolio, which is a zero- cost portfolio. This means they used an investment strategy where you go long on carbon- efficient stocks and short on carbon-inefficient stocks. In their research, they used carbon- intensity as the only key variable of interest. In et al. (2017) found a positive cumulative return since 2010, which implies that low-emission firms outperform high-emission firms.

However, for very small firms, there is an exception. Despite In et al. (2017) include risk factors as High-minus-Low (HML) and Winners-minus-Losers (WML) in their model, small firms still have characteristics that cannot be fully explained. Their result contrasts with Bolton and Kacperczyk (2021), which did not find a significant relationship between carbon emissions and stock returns when carbon intensity is used as a measure for a firm’s carbon emissions. There are three key differences that may explain this difference between results.

First, Bolton and Kacperczyk (2021) used a different time period and firm sample. Second, In at al. (2017) did not control for industry fixed effects. It seems that, when controlling for industry-fixed effects, the relationship between carbon intensity and stock returns becomes insignificant. The third difference between the paper of In et al (2017) and Bolton and

Kacperczyk (2021) is that Bolton and Kacperczyk divided carbon emissions in three different scopes. They analyzed the effects of carbon emissions on stock returns for each scope

separately to avoid double counting.

(10)

2.3. Hypotheses

Firms with high carbon emissions have a greater risk of reduction in future sales and cash flows because of the increasing pressure of stakeholders to ‘go green’. Capturing these risk factors must result in a risk-premium for high-emission firms, which implies a positive relationship between carbon emission and stock returns (Aswani et al., 2022). Bolton and Kacperczyk (2021) use the same argument for this ‘carbon risk premium hypothesis’: Firms with high carbon emissions are more exposed to carbon pricing risk and other disrupting carbon regulations to limit carbon emission. Relatedly, firms that mainly rely on fossil energy are more exposed to technology risk from new ‘green’ energy. Investors may want to be compensated for holding stocks of these high-emission firms, which results in a positive relationship between carbon emission and stock returns.

However, Cheng et al. (2020) and Krueger (2015) suggest another argument to explain the relationship between carbon emission and stock returns: efforts to reduce carbon emission reflect agency conflicts. It may be that firms reduce their carbon emissions because top-level executives desire them to do so to boost their public image. A reduction in carbon emission may signal operational underperformance and therefore lead to stock market

underperformance, which suggest a negative relationship between carbon emission and stock returns.

A third argument to explain the relationship between carbon emission and stock return relies on the efficient market hypothesis (Bolton & Kacperczyk, 2021). Carbon risk is inefficiently priced in by financial markets and the risk associated with carbon emissions is underpriced. This is because most investors are biased and ignore unrepresentative

information about global warming and its related risk. This argument suggests a negative relationship between carbon emissions and stock returns.

Based on these arguments, the following hypotheses can be put forward:

(1) Hypotheses 1a: A firm’s carbon emissions and its stock returns are positively related.

(2) Hypotheses 1b: A firm’s carbon emissions and its stock returns are negatively related.

Furthermore, it can be argued whether the relationship between carbon emissions and stock returns is different when you control for fixed effects. According to Aswani et al. (2022) and Bolton and Kacperczyk (2021), the relationship between carbon and stock returns turns insignificant when controlling for industry and time fixed effects.

(11)

(3) Hypotheses 2a: The relationship between carbon emissions and stock return does not change when you include fixed effects.

(4) Hypotheses 2b: The relationship between carbon emissions and stock return does change when you include fixed effects.

Additionally, as described in section 2.2, it can be argued whether there is a difference in relationship between carbon emissions and stock returns across the globe. Following the conclusions of Brandon et al. (2021), there is significant difference in relationship between carbon emissions and stock returns in Europe compared to the United States. However, Aswani et al. (2022) suggest that the relationship between carbon emissions and stock return is insignificant in Europe, as well as in the United States.

(5) Hypotheses 3a: The relationship between carbon emissions and stock return is not different across regions.

(6) Hypotheses 3b: The relationship between carbon emissions and stock return is different across regions.

2.4. Variables

Based on leading papers of Bolton and Kacperczyk (2021) and Aswani et al. (2022), the following (control) variables are used in this paper: Return, Carbon Intensity, Leverage, ROE, LOGSales, CapEx/Assets, LOGPP&E, SalesGR and EPSGR. Return is the depended variable in the empirical regression model and is defined as the quarterly stock return (in percentages).

Carbon Intensity is the main independent variable of interest. It is defined as the ratio of carbon emission to sales. Leverage is defined as the ratio of long-term debt to total assets and ROE is defined as the ratio of net profit to shareholders equity. LOGSales is the natural logarithm of yearly sales and CapEx/Assets is defined as the ratio of capital expenditures to the total value of year-end assets. LOGPP&E is equal to the natural logarithm of plant, property, and equipment and SalesGR and EPSGR are equal to the change in annual sales and change in annual earnings per share, respectively.

2.5. Contribution to existing Literature

This thesis will contribute to the existing literature by expanding the analysis to other regions like Africa/Middle East, Asia/Pacific, Latin America and Caribbean and Canada.

Existing literature solely focused on the United States and Europe, but no research is done using data on other regions. By expanding the dataset to regions like Africa/Middle East,

(12)

Asia/Pacific, Latin America and Caribbean and Canada, the external validity of this paper is higher than when only data on Europe and the United States is used.

3. Data Description and Methodology

In the following sections, the data sample as well as the empirical model used in this paper are discussed. The process of data selection and data cleaning is described, and the choices made in this process are evaluated.

3.1. Data Sample

In this paper, data is retrieved from three different databases from two distinct data providers. Sustainalytics is used to get data on carbon emissions. Sustainalytics is a data provider that collects and processes carbon intensity data on global public and private

companies. The selected data from Sustainalytics is available from August 2009 to December 2019 and the observations are monitored and registered monthly. The day on which the observations are recorded in the database is random but is always in the first seven days of each month. Since this paper investigates the relationship between carbon emissions and stock performance, it was decided to drop the observations from private companies.

In order to merge the different datasets, each company must be described by a unique identifier. In this paper it was decided to use the ISIN Code as the unique identifier. An ISIN Code is an alphanumeric 12-digit code that identifies each unique security. The observations with a missing ISIN Code were dropped. It appears that there is no ISIN Code available for all observations in the dataset prior to June 7, 2016. Therefore, the observations in the dataset are in the range from June 7, 2016 until December 31, 2019. Also, the observations whose carbon intensity is missing have been dropped from the dataset, resulting in a dataset with 95,035 observations from 8,728 distinct public companies.

The associated daily closing prices of a firm’s stock were retrieved from the Compustat database. Observations with missing closing prices and ISIN Codes were dropped from the dataset. Using the ISIN Codes as unique identifiers, this dataset is merged with the dataset on carbon emissions, which results in a new dataset consisting of 1,859 distinct companies with monthly data on carbon intensity and stock prices, registered on one of the first seven days in each month.

For data on firm fundamentals, another database of Compustat is used. Observations in this dataset are registered on the last day of each quarter, which results in a merging problem with the dataset on carbon intensity and closing prices. To solve this problem, the monthly

(13)

observations of the dataset on carbon intensity and closing prices are converted to quarterly data by taking the average of three consecutive months for each numeric variable. By doing so, the quarterly average is being calculated and the dataset on carbon emissions and closing prices can be merged with quarterly data on firm fundamentals, again using ISIN Codes as unique identifiers.

After merging data from three databases, a dataset consisting of 11,748 observations of 1,351 distinct companies from the regions Africa/Middle East, Asia/Pacific, Europe, Latin America and Caribbean and United States and Canada is remained. Furthermore, 42 different industries are represented. A more detailed description about the distribution of regions and industries can be found in the Appendix in Table 9 and 10.

Given the summary statistics in Table 1, more data cleaning and preparation is required to do a valid analysis. For example, given EPSGR is defined as the change in annual earnings per share, a minimum value of -225.097 and a maximum value of 165.066 seems very

unrealistic. Also, given SalesGR is the (percentage) change in annual sales, a maximum value of 19,322.538 seems not realistic.

Table 1: Summary Statistics before winsorizing

This table provides summary statistics for all variables used in this paper before they are winsorized.

Variable Obs Mean Std. Dev. Min Max

Return 10,413 .048 .605 -.973 27.663

Carbon Intensity 11,748 .182 .794 0 8.8

Leverage 10,848 .202 .145 0 1.001

ROE 9,385 .029 .333 -23.296 12.598

LOGSales 9,638 9.322 2.467 -.511 19.312

CapEx/Assets 11,201 .005 .006 -.031 .134

LOGPP&E 11,485 8.916 3.151 -2.303 18.89

SalesGR 11,748 5.011 308.749 -1.344 19,322.538

EPSGR 7,065 -.081 9.736 -255.097 165.066

In order to take outliers into account, EPSGR and SalesGR are winsorized at the 2.5%

level. Furthermore, following Aswani et al. (2022), Leverage, CapEx/Assets, ROE, LOGPP&E and Carbon Intensity are winsorized at the 2.5% level. In this paper, to take outliers in the variables Return and LOGSales into account, also these variables are

winsorized at the 2.5% level. Winsorizing these variables results in summary statistics given in Table 2.

(14)

Table 2: Summary Statistics after Winsorizing

This table provides summary statistics after the variables are winsorized at percentage levels described in section 3.1.

Variable Obs Mean Std. Dev. Min Max

Return 10,413 .015 .121 -.239 .331

Carbon Intensity 11,748 .162 .683 0 3.468

Leverage 10,848 .2 .137 .001 .544

ROE 9,385 .033 .041 -.077 .158

LOGSales 9,638 9.339 2.312 5.576 16.769

CapEx/Assets 11,201 .004 .005 0 .019

Table 2 (continued)

LOGPP&E 11,485 8.928 3.02 3.355 15.91

SalesGR 11,748 .048 .159 -.343 .58

EPSGR 7,065 .151 1.524 -4.143 5.023

3.2. Methodology and Empirical Model

This paper aims to research the relationship between carbon emissions and stock returns.

In order to do so, a multivariate panel data linear regression model is used. The dependent variable in the empirical model is Return and it is tested against an independent variable Carbon Intensity that measures carbon emissions as a ratio to sales. As described in section 2.2., there are three ways to measure carbon emissions. Due to limitations, which are described in more detail in section 5.2, carbon intensity is used as a metric for carbon emissions in this paper. Furthermore, several control variables are used in the regression model to isolate the causal relationship between the dependent variable Return and the independent variable of interest Carbon Intensity.

𝑅𝑒𝑡𝑢𝑟𝑛_𝑖𝑡 = 𝛽₀+ 𝛽₁𝐶𝑎𝑟𝑏𝑜𝑛 𝐼𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦_𝑖𝑡 + 𝛽₂ 𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒_𝑖𝑡+ 𝛽₃ 𝑅𝑂𝐸_𝑖𝑡+ 𝛽₄ 𝐿𝑂𝐺𝑆𝑎𝑙𝑒𝑠_𝑖𝑡+ 𝛽₅ 𝐶𝑎𝑝𝐸𝑥/𝐴𝑠𝑠𝑒𝑡𝑠_𝑖𝑡 + 𝛽₆ 𝐿𝑂𝐺𝑃𝑃&𝐸_𝑖𝑡+ 𝛽₇ 𝑆𝑎𝑙𝑒𝑠𝐺𝑅_𝑖𝑡+ 𝛽₈ 𝐸𝑃𝑆𝐺𝑅_𝑖𝑡 + 𝜀_𝑖𝑡 (1)

In the linear regression model (1), 𝑅𝑒𝑡𝑢𝑟𝑛_𝑖𝑡 is the quarterly stock return of firm i at time t. 𝐶𝑎𝑟𝑏𝑜𝑛 𝐼𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦_𝑖𝑡 is the main independent variable of interest and the remaining terms are the control variables as described in section 2.4. Finally, the idiosyncratic error term of firm i at time t is denoted by 𝜀_𝑖𝑡. To estimate the coefficient of interest 𝛽₁, ordinary least squares regression techniques are used. The interpretation of 𝛽₁ is as follows: a one unit increase in 𝐶𝑎𝑟𝑏𝑜𝑛 𝐼𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦_𝑖𝑡, results in a 𝛽₁ change in 𝑅𝑒𝑡𝑢𝑟𝑛_𝑖𝑡, ceteris paribus (all other things being equal). After doing the regression analysis, the estimated 𝛽₁ will be used to evaluate hypotheses (1a) and (1b). A positive and significant 𝛽₁ evaluates hypothesis (1a) as true, and a negative significant 𝛽₁ will evaluate hypothesis (1b) as true.

To evaluate hypothesis (2a) and hypothesis (2b), industry and time fixed effects must be added to regression model (1). This results in regression model (2), as stated below.

(15)

Industry fixed effects are captured in the term 𝛿_{𝐼𝑛𝑑𝑢𝑠𝑡𝑟𝑦} and 𝛾_𝑡 denotes the time fixed effects.

The meaning and interpretation of the other terms in regression model (2) are the same as described above for regression model (1). To evaluate the significance of the fixed effects model compared to a random effect model, a Hausman test will be used where the null

hypothesis is that a random effect model is preferred, and the alternative hypothesis is that the fixed effect model is preferred (Hausman, 1978).

𝑅𝑒𝑡𝑢𝑟𝑛_𝑖𝑡 = 𝛽₀+ 𝛽₁𝐶𝑎𝑟𝑏𝑜𝑛 𝐼𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦_𝑖𝑡 + 𝛽₂ 𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒_𝑖𝑡+ 𝛽₃ 𝑅𝑂𝐸_𝑖𝑡+ 𝛽₄ 𝐿𝑂𝐺𝑆𝑎𝑙𝑒𝑠_𝑖𝑡+ 𝛽₅ 𝐶𝑎𝑝𝐸𝑥/𝐴𝑠𝑠𝑒𝑡𝑠_𝑖𝑡 + 𝛽₆ 𝐿𝑂𝐺𝑃𝑃&𝐸_𝑖𝑡+ 𝛽₇ 𝑆𝑎𝑙𝑒𝑠𝐺𝑅_𝑖𝑡+ 𝛽₈ 𝐸𝑃𝑆𝐺𝑅_𝑖𝑡 + 𝛿_{𝐼𝑛𝑑𝑢𝑠𝑡𝑟𝑦} + 𝛾_𝑡+

𝜀_𝑖𝑡 (2)

To evaluate hypothesis (3a) and (3b), an if-statement and a dummy variable for each region must be added to regression model (2). The regression model itself will remain the same. By creating a dummy variable for each region and using only the observations whose region dummy equals one, an OLS regression analysis can be performed using observations which belong to one and the same region. After running multiple regression analyses, each analysis using observations that belong to the same region, the estimated coefficients of interest 𝛽₁ of each regression can be compared with each other.

Because it is very likely that unobserved characteristics within industry groups are correlated with each other, the standard errors in all regression models are clustered at industry level.

4. Analysis and Results

In this chapter, the regression analyses according to the methodology described in chapter 3 will be performed and the estimated results will be displayed in regression tables.

Furthermore, the outcomes will be linked and compared to earlier research. Finally, two robustness checks of the regression outcomes will be performed.

4.1. Regression Analyses

Table 3 present results of estimating Equation (1). It is showed how Carbon Intensity changes when several control variables are added. In all 4 columns, there are no dummies for industry and time fixed effects. Furthermore, in all 4 regressions displayed in Table 3, the error terms are clustered at industry level. In column (1) the baseline regression without any control variables is displayed. It appears that the estimated variable of interest is -0.002.

However, for all significance levels, this coefficient is insignificant. Adding more variables such as Leverage, LOGSales and EPSGR, decreases this coefficient to -0.005, as displayed in

(16)

column (2). This estimated coefficient is significant at a 5% level. Columns (3) and (4) show the effect of Carbon Intensity on Return when even more control variables are added. When also ROE and SalesGR are added as control variables, the estimated coefficient of Carbon Intensity remains at -0.005, significant at a 1% level. In the full model, as displayed in column (4), the estimated independent variable of interest equals -0.005, significant at a 1% level.

Table 3: Regression results Return on Carbon Intensity without controlling for fixed effects This table provides the regression results of four different regressions, each with different control variables included. Fixed effects are excluded, and the error terms are clustered at industry level in each regression.

(1) (2) (3) (4)

Return Return Return Return

Carbon Intensity -.002 -.005** -.005*** -.005***

(.002) (.002) (.002) (.002)

Leverage -.01 -.014 -.027*

(.014) (.013) (.013)

LOGSales -.001 -.001 -.006***

(.001) (.001) (.002)

EPSGR .008*** .006*** .005***

(.002) (.002) (.002)

ROE .246*** .279***

(.043) (.042)

SalesGR .075*** .08***

(.015) (.015)

CapEx/Assets -1.536***

(.564)

LOGPP&E .005***

(.002)

Constant .015*** .023** .011 .021**

(.002) (.009) (.009) (.01)

Observations 10,413 6,660 6,543 6,320

R-squared 0 .014 .035 .04

Adj R² 0 .014 .034 .038

Industry Fixed Effects NO NO NO NO

Time Fixed Effects NO NO NO NO

Clustered YES YES YES YES

Standard errors are in parentheses

*** p<.01, ** p<.05, * p<.1

These findings are in line with the findings of In et al. (2017), which also found a significant negative relationship between Return and Carbon Intensity, when no fixed effects are included in the regression model. However, the statistical negative relationship between Return and Carbon Intensity found in column (4) can not be compared with the findings of Aswani et al. (2022) and Bolton and Kacperczyk (2021), because in each regression model in their papers, time fixed effects are included.

(17)

Based on the findings in Table 3, hypothesis (1b) can be accepted: when including all control variables, Return and Carbon Intensity are negatively related when fixed effects are not considered.

When deciding to include random effects or fixed effects in the regression analyses, a Hausman test is performed. The results of this test are displayed in Table 11 in the Appendix.

Following the methodology described in section 3.2., the p-value tells whether a fixed effect or random effect model is preferred. Because the p-value of this test equals 0.001, the null hypothesis is rejected, and the fixed effect model is preferred compared to a random effect model. After running the regression model including the time fixed effects, we can run a joint hypothesis test on the dummy time variables to see whether we have to include time fixed effects in our model. It appears that the time dummies are jointly significant, and we therefore have to include time fixed effect in our regression. The results of this joint hypothesis test can be found in Table 12 in the Appendix.

In Table 4, the estimated regression results of regressing Return on Carbon Intensity are displayed. In each regression, industry fixed effects as well as time fixed effects are included based on the argumentation described above. In column (1), the estimated results of the baseline model where Return is only regressed on Carbon Intensity are given. The

estimated coefficient equals -0.004 and is insignificant at all percentage levels. When

including more control variables like Leverage, LOGSales and EPSGR as displayed in column (2), the coefficient of the variable of interest remains at -0.004 and insignificant. When also ROE and SalesGR are added to the model in column (3), the variable of interest Carbon Intensity remains at -0.004 and insignificant at all significance levels. In the regression model in column (4) where all the control variables are included, the variable of interest decreases to -0.005 and remains insignificant at all significant levels.

Table 4: Regression results Return on Carbon Intensity with controlling for fixed effects This table provides the regression results of four different regressions, each with different control variables included. In all four columns, the regression models include industry and time fixed effects, and the error terms are clustered at the industry level.

(1) (2) (3) (4)

Carbon Intensity -.004 -.004 -.004 -.005

(.003) (.004) (.004) (.004)

Leverage -.008 -.018 -.023

(.014) (.013) (.015)

LOGSales -.001 -.001 -.005**

(.001) (.001) (.002)

EPSGR .005*** .003* .003

(18)

(.002) (.002) (.002)

ROE .253*** .269***

(.041) (.042)

SalesGR .066*** .072***

(.013) (.014)

CapEx/Assets -1.101**

(.506)

LOGPP&E .004**

(.002)

Constant 0 .02 .017 .021

(.005) (.015) (.016) (.018)

Observations 10,413 6,657 6,540 6,317

R-squared .12 .117 .136 .141

Adj R² .119 .115 .134 .138

Industry Fixed Effects YES YES YES YES

Time Fixed Effects YES YES YES YES

*** p<.01, ** p<.05, * p<.1

These findings are in line with earlier literature. Both Aswani et al. (2022) and Bolton and Kacperczyk (2021) conclude in their papers that the relationship between Return and Carbon Intensity is insignificant when controlling for industry fixed effects and time fixed effects. However, it is not possible to compare the estimated results as displayed in column (4) of Table 4 to the results of In et al. (2017), because they did not control for fixed effects in their regression models.

Based on the findings in Table 4, hypotheses (2a) and (2b) can be evaluated. When the results of column (4) in Table 4 are compared with the estimated results in column (4) in Table 3, it can be concluded that the relationship between Return and Carbon Intensity changes when fixed effects are included. The relationship remains at -0.005 but turns insignificant at all significance levels when fixed effects are included. Based on this, hypothesis (2b) is accepted.

In contrast to earlier research, in this study a distinction is made between five different regions: Africa/Middle East, Asia/Pacific, Europe, Latin America and Caribbean and United States and Canada. Following the methodology described in section 3.2., several regressions are performed, each using observations that belong to one and the same region. The estimated results are displayed in Table 5. In all regressions in Table 5, industry fixed effects and time fixed effects are included. Furthermore, in all regressions the standard errors are clustered at industry level. It is important to note that no regression is performed using observations that

(19)

belong to the region United States and Canada. As displayed in Table 9 in the Appendix, after the data cleaning process as described in section 3.1., only 17 observations in the dataset belong to this specific region. These are too few to draw meaningful conclusions, so it is decided to perform regression analyses only on the other four regions.

In column (1) in Table 5, Return is regressed on Carbon Intensity including all control variables. The estimated coefficient of interest equals 0.025. However, this result is not significant at all significance levels. It appears that when only observations that belong to the region Africa/Middle East are considered, the relationship between Return and Carbon Intensity is insignificant. This conclusion is the same when it is compared to the estimated results in column (4) in Table 4 where data on all regions is considered. Furthermore, when only observations that belong to the regions Europe and Latin America and Caribbean are considered, the coefficient of interest changes to -0.002 and 0.003, respectively. However, this coefficient remains insignificant at all significance levels. Aswani et al. (2022) also performed a regression analysis where they only toke into account observations that belong to the region Europe. When controlling for fixed effects, they also found an insignificant

relationship between Return and Carbon Intensity. The result in this paper is therefore in line with their results.

Surprisingly, a different and significant result is found when only observations that belong to the region Asia/Pacific are considered. As displayed in column (2) in Table 5, the estimated relationship between Return and Carbon Intensity equals -0.014 and is significant at a 5% level. This result is in contrast with the estimated results in other regions, where this relationship is insignificant at all significance levels.

Table 5: Regression results Return on Carbon Intensity per Region with controlling for fixed effects This table provides the regression results of four different regressions. Each regression includes all control variables, industry and time fixed effects and clustered standard errors at industry level. Column (1) displays the estimated regression results when only observations based in Africa/Middle East are considered. Column (2) only takes into account observations from Asia/Pacific. Column (3) considers European firms and column (4) represents regression results when only observations from Latin America and Caribbean are considered.

Africa / Middle East

Asia / Pacific Europe Latin America and Caribbean

(1) (2) (3) (4)

Carbon Intensity .025 -.014** -.002 .003

(.017) (.006) (.004) (.016)

Leverage -.258* -.01 -.017 -.656***

(.125) (.026) (.018) (.154)

LOGSales -.083 -.004 -.004* -.02

(.046) (.004) (.002) (.023)

EPSGR .017* .002 .002 .019

(20)

(.009) (.003) (.001) (.012)

ROE -.366 .401*** .249*** .214

(.451) (.12) (.05) (.292)

SalesGR .042 .086*** .063*** .05

(.076) (.029) (.013) (.234)

CapEx/Assets -6.502 -.509 -.588 -4.443

(8.468) (.878) (.778) (7.463)

LOGPP&E .053 .002 .004* .015

(.042) (.003) (.002) (.031)

Constant .555*** .034 -.04 .303**

(.163) (.023) (.047) (.135)

Observations 128 1,722 4,328 126

R-squared .282 .127 .168 .16

Adj R² .148 .117 .164 .018

*** p<.01, ** p<.05, * p<.1

Using the estimated regression result in Table 5, hypotheses (3a) and (3b) can be evaluated. Because the relationship between Return and Carbon Intensity differs in

Asia/Pacific compared to the other regions, it can be concluded that the relationship differs across regions. Therefore, hypothesis (3b) is accepted.

4.2. Robustness Checks

To check for robustness of the results, the sample is split in observations before and including the third quarter of 2016 and observations that belong to after the third quarter of 2016. This split is made because the Paris Agreement entered into force on 4 November 2016, as mentioned in the introduction of this paper. The same regressions as performed in Table 4 are performed on these two different samples, so for all regressions industry fixed effects and time fixed effects are included and standard errors are clustered at industry level. In Table 6, the estimated regression results for the sample with observations that belong to before the Paris Agreement are displayed. In all regressions the coefficient of interest is significant, in contrast to the results in Table 4 where the full sample used. In column (4) of Table 6, the regression results when all the control variables are included are displayed. The independent variable of interest in the full model has an estimated coefficient of 14.07 and is significant at all significance level. This estimated coefficient is much larger than the estimated coefficient in column (4) of Table 4. It can be concluded that there is a carbon risk premium for

observations before the Paris Agreement. However, the sample size of the regression performed in column (4) in Table 6 is significant smaller than the sample size of the regression performed in column (4) of Table 4. With only 60 observations, it can be

(21)

questioned whether the conclusions made based on the estimated regression results are meaningful.

Table 6: Regression results Return on Carbon Intensity for observations pre–Paris Agreement This table provides the regression results of four different regressions using observations before the Paris Agreement.

(1) (2) (3) (4)

Carbon Intensity 7.257*** 12.192*** 16.2*** 14.07***

(1.815) (2.456) (3.176) (3.93)

Leverage -.069 -.063 -.06

(.141) (.121) (.136)

LOGSales .027* .023 -.002

(.015) (.019) (.016)

EPSGR .016** .015** .013**

(.008) (.007) (.006)

ROE .779* .987**

(.412) (.455)

SalesGR .046 .073

(.164) (.153)

CapEx/Assets -10.131

(7.457)

LOGPP&E .022

(.014)

Constant -.053*** -.313** -.331** -.26*

(.013) (.147) (.146) (.151)

Observations 663 61 61 60

R-squared .004 .233 .31 .434

Adj R² .002 .178 .234 .345

*** p<.01, ** p<.05, * p<.1

The regression estimates for observations that belong to the period after the Paris Agreement are displayed in Table 7. In all four regressions, industry fixed effects and time fixed effects are included, and standard errors are clustered at the industry level. When the results in column (4) of Table 7 are compared to the estimations in column (4) of Table 4, the conclusion that holds for the results in Table 4 also holds for the regressions in Table 7. It can be concluded that when only observations that belong to the period after the Paris Agreement are considered, the relationship between Return and Carbon Intensity is not significant. This conclusion is the same when the full sample is used. Because the sample size of the regression in column (4) in Table 7 is significantly larger than the sample size used for the regression in column (4) in Table 6, these conclusions are more likely to be valid.

(22)

Table 7: Regression results Return on Carbon Intensity for observations post–Paris Agreement This table provides the regression results of four different regressions using observations after the Paris Agreement.

(1) (2) (3) (4)

(.003) (.004) (.004) (.004)

Leverage -.008 -.018 -.023

(.014) (.014) (.015)

LOGSales -.001 -.001 -.005**

(.001) (.001) (.002)

EPSGR .005*** .002 .002

(.002) (.002) (.002)

ROE .254*** .269***

(.04) (.041)

SalesGR .067*** .072***

(.013) (.013)

CapEx/Assets -.989*

(.491)

LOGPP&E .004**

(.002)

Constant .023** .021 .019 .024

(.01) (.016) (.017) (.019)

Observations 9,750 6,596 6,479 6,257

R-squared .125 .118 .137 .141

Adj R² .124 .116 .135 .139

*** p<.01, ** p<.05, * p<.1

Another robustness check can be performed by excluding the three industries with the highest average Carbon Intensity. It is often stated that a few industries produce the biggest fraction of carbon emissions (Bolton & Kacperczyk, 2021). It might be the case that the results in Table 4 are driven by these industries. Therefore, it is essential to check if the conclusions based on the estimation results in Table 4 still hold when these industries are excluded.

When analyzing the full sample, the three industries with the highest average Carbon Intensity are Homebuilders, Traders & Distributors and Automobiles. Therefore, these

industries are excluded when performing the regressions in Table 8. All regressions in Table 8 include industry fixed effects and time fixed effects and standard errors are clustered at

industry level.

Column (4) of Table 8 displays the estimated regression results when all control variables are added. It is found that when the three industries with the highest average Carbon Intensity are excluded, the relationship between Return and Carbon Intensity is still insignificant.

These results are in line with the results in Table 4 where the whole sample is used. Therefore,

(23)

it can be concluded that the results in Table 4 are not driven by the industries with the highest Carbon Intensity.

Table 8: Regression results Return on Carbon Intensity when salient industries are excluded This table provides the regression results of four different regressions, each with different control variables included. The three industries with the highest average Carbon Intensity (Homebuilder, Trader & Distributors and Automobiles) are excluded in this sample.

(1) (2) (3) (4)

(.003) (.004) (.004) (.004)

Leverage -.011 -.02 -.025*

(.014) (.013) (.015)

LOGSales -.001 -.001 -.005**

(.001) (.001) (.002)

EPSGR .005*** .002 .002

(.002) (.002) (.002)

ROE .244*** .259***

(.04) (.041)

SalesGR .066*** .072***

(.014) (.014)

CapEx/Assets -1.135**

(.505)

LOGPP&E .004**

(.002)

Constant .002 .021 .02 .024

(.005) (.015) (.016) (.018)

Observations 10,102 6,511 6,394 6,171

R-squared .114 .115 .134 .138

Adj R² .113 .113 .131 .135

*** p<.01, ** p<.05, * p<.1

5. Discussion and Conclusion

In this chapter the findings and implications of this paper are summarized and several limitations during the process of data preparation and data analyzing are mentioned. Based on these limitations, recommendations for further research are given.

5.1. Conclusions

Based on the regression analyses, answers to the main research question and the sub questions can be formulated. The main research question in this paper is: Are a firm’s carbon emissions correlated with its stock return? Based on the estimated results in column (4) in Table 3, it can be concluded that there is a negative relationship between a firm’s return and its carbon intensity, and this relationship is significant at a 1% significance level. However, as

(24)

displayed in column (4) in Table 4, controlling for fixed effects removes this significant effect. Therefore, the answer on the first sub question can be formulated. The first sub question in this paper is: Is there a change in relationship between carbon emissions and stock return when controlled for fixed effects compared to when no fixed effects are included?

Based on the regression results in Table 4, it can be concluded that there is a change in

relationship between stock returns and carbon intensity when you include fixed effects in your regression model. The answer on this sub question implicates that, when fixed effects are included in the regression model, investors do not require a carbon risk premium.

Finally, in this study it is researched if the relationship between stock returns and carbon intensity is different across different regions. Based on the regression estimates in Table 5, it can be concluded that in three of the four regions studied the relationship between stock returns and carbon intensity is insignificant. However, this relationship is negative and significant at a 5% level when only observations that belong to the region Asia/Pacific are considered. This answers the second sub question of this study: Does the relationship between carbon emissions and stock return differ across different regions?

To check if the findings are robust, two robustness checks are performed. When the same regressions are performed on different sub samples, a different conclusion can be made for the observations pre-Paris Agreement compared to the full sample as well as compared to the post-Paris Agreement sample. It can be concluded that the relationship between stock returns and carbon intensity is positive and significant for pre-Paris Agreement observations.

This is different compared to the full sample or post-Paris Agreement, where an insignificant relationship is estimated. However, because of the very small sample size for the pre-Paris Agreement sample, it can be questioned if these conclusions are valid.

However, a second robustness check reinforces the validity of the previously found results. When the three salient industries are not considered in the regression analysis, it can still be concluded that there is no significant effect of carbon intensity on stock returns, when controlled for fixed effects.

5.2. Limitations and Recommendations

There are a few limitations which (might) influence the results in this study. First of all, in earlier research also a control variable called the Herfindahl-index, which measures the market concentration, is included. However, when calculating the market share per firm for this dataset, the results are very unrealistic when there are only a few firms per industry. This

(25)

dataset does not cover each firm in each industry so calculating the market share for each firm assuming all firms per industry are present in the dataset will provide an unrealistic picture.

Therefore, it is decided not to include this control variable in this research. Second, because of limited licenses the University of Amsterdam provided, Sustainalytics was the only database that could be used for data on carbon emissions. Unfortunately, the only available carbon emission metric was carbon intensity. Also, no distinction is made between the three different carbon emission scopes. Therefore, in further research, I would recommend doing the same regression analyses, but use different carbon emission metrics like total carbon emission and year-by-year growth, as used in earlier papers. Also, I would make a distinction between three different carbon emission scopes to research what the impact of each scope is on stock

returns.

Carbon Risk Premium: Fact or Fallacy? A study of the relationship between carbon emission and stock returns