The effect of portfolio performance using social responsibility screens

Transcription

1 The effect of portfolio performance using social responsibility screens Master Thesis Author: Donny Bleekman BSc. (927132) Supervisor: dr. P. C. (Peter) de Goeij Study program: Master Finance December 2013

2 Abstract This study investigates the effect of portfolio performance between July 2002 and December 2011 using social responsibility scores from ASSET4. Based on ASSET4 scores equal- and value-weighted portfolios were created and tested for abnormal returns using the Carhart (1997) four-factor model. Although the estimated alphas are statistically insignificant, I observe systematic differences in the factor loadings between high-rated and low-rated socially responsible companies. On average, highrated socially responsible companies have lower market betas than low-rated socially responsible companies. High-rated socially responsible companies have a negative loading on the SMB factor, while low-rated socially responsible companies have a positive loading on the SMB factor. Low-rated socially responsible companies have a greater negative loading on momentum than high-rated socially responsible companies. 1

3 Contents 1. Introduction Literature Survey & Hypothesis Development Literature Survey Hypothesis Development Data & Methodology Data Methodology Empirical findings Descriptive Statistics Portfolio Results Conclusion References Appendix

4 1. Introduction Socially responsible investment (SRI) is becoming widely spread among institutional global investors and has drawn much attention among researchers in recent years. According to the US SIF Foundation in 2012 the assets engaged in sustainable and responsible investing practice represent 11.3 percent of the $33.3 trillion in total assets under management tracked by Thomson Reuters Nelson. 1 The last decade the number of SRI related studies have increased substantially. However, the results from these studies do not give a clear picture, whether it pays off to invest socially responsible. Derwall et al. (2005) report that a best-in-class portfolio that score high on eco-efficiency scores will return a four-factor alpha of 4.15 percent per year over the period Kempf and Osthoff (2007) show that environment, employee, and community screening using the KLD dataset will earn statistically significant higher four-factor alphas for high-rated portfolios. However, they do not observe statistical significant outperformance of high-rated companies over low-rated companies for the diversity, human rights, and product screen. Statman and Glushkov (2009) used the same dataset as Kempf and Osthoff (2007) and found that stocks with high social responsibility ratings performed generally better than stocks with low social responsibility ratings. However, none of the results were statistically significant except for the employee screen. Galema et al. (2008) who also used the KLD dataset only found outperformance for the high-rated community portfolio. Studies that covered the European market even tell a less convincing story. Van de Velde et al. (2005) as well as Brammer et al. (2006) did not find any statistically significant outperformance of high-rated socially responsible companies over low-rated companies. In order to extent the literature on SRI I used the ASSET4 dataset from Thomson Reuters for my thesis. To my knowledge the ASSET4 dataset has not been used for SRI related studies. The advantage of using the ASSET4 dataset is that it covers the global market instead of just one particular region. The ASSET4 dataset assigns scores to companies based on four pillars: corporate governance, economic, environment, and social. Moreover, ASSET4 provides an overall indicator that combines all four pillar scores into one score that measures overall SRI performance. With these five scores I have constructed several equal- and value-weighted portfolios. The first portfolio consists of the top 10 percent of companies that perform the best based on one of the five scores. The second portfolio consists of the bottom 10 percent of companies that perform the worst based on one of the five scores. In addition, for each pillar score and the overall score high-low portfolios were constructed in which I go long in the top ten percent and short in the bottom ten percent for each Report on Sustainable and Responsible Investing Trends in the United States, see 3

5 score. These portfolios were tested for abnormal return using the Carhart (1997) four-factor model covering the period July 2002 till December The results are further tested by performing some robustness checks. First, I test for any regional bias by considering the US and European market separately. Second, I check whether the estimated alphas remain the same, if I apply different cutoffs. Third, I apply best-in-class screening to overcome a possible bias towards some economic sectors. Results from the equal-weighted portfolios show that investors can earn positive abnormal returns, if they invest in companies that score high on corporate governance and economic. The opposite is true for screening companies on environment and social scores. Low-rated environment and social companies earn positive abnormal returns and high-rated environment and social companies earn negative abnormal returns. However, the estimated alphas are statistically insignificant. Using valueweighted portfolios, regional samples, different cut-offs, and best-in-class screening do not result in any statistically significant outperformance using SRI. In this study I do observe statistically significant differences in the factor loadings of the high- and low-rated portfolios. On average high-rated socially responsible companies have lower market betas than low-rated socially responsible companies. Also, high-rated socially responsible companies have a negative loading on the SMB factor, while low-rated socially responsible companies have a positive loading on the SMB factor. Low-rated socially responsible companies have a greater negative loading on momentum than high-rated socially responsible companies. These findings are robust for regional and sector bias. The remainder of this thesis proceeds as follows. In section 2, previous literature on SRI is covered and I develop the hypotheses I want to test. Section 3 gives a description of the data and present the methodology for portfolio formation as well as the model to test the hypotheses. In section 4, I present the descriptive statistics and the empirical results from my regression analyses. I conclude in section 5. 4

6 2. Literature Survey & Hypothesis Development 2.1 Literature Survey There has been an increase in the attention towards social responsible investment (SRI) related academic studies. Renneboog et al. (2008) provide a full-scale overview of some of the most recent studies on SRI found in literature. This section will cover the most recent studies on portfolio performance. Van de Velde et al. (2005) conducted a study about the return of SRI portfolios using Vigeo corporate social responsibility scores. The sample consisted of European firms covering the period from 2000 to Portfolio construction was based on a company s relative sustainability performance within the sector it operates in. A sustainability score greater than one standard deviation above the sector mean meant it would be placed in the Best portfolio. A sustainability score between the sector mean and one standard deviation meant it would be placed in the Good portfolio. In similar fashion, Bad and Worst portfolios were constructed. Although they observed that firms with lowsustainability scores under-performed the market and firms with high-sustainability scores outperformed the market, their findings are not statistically significant most likely due to the relative short time horizon. Brammer et al. (2006) have studied the relationship between corporate social performance and stock returns for a sample of UK listed companies. They used three measures of social performance (community, environmental, and employee) from the Ethical Investment Research Service (EIRIS) database. The indicator of employee responsibility relates to six measures based on health and safety systems, systems for employee training and development, equal opportunities policies, equal opportunities systems, systems for good employee relations, and systems for job creation and security. The environment indicator is based on three measures: quality of environmental policies, environmental management systems, and environmental reporting. The community indicator is measured as one single variable. Each measure is normalized to a score that runs from zero to three, and then summed to generate an overall CSR score of zero to nine. Their results show that firms with higher social performance scores have significantly lower average returns than the FTSE benchmark. Moreover, the portfolio that consisted of 17 firms with the lowest score on every social performance measure yielded a positive return of 8 percent, outperforming 20 percent of the FTSE benchmark. Brammer et al. (2006) concluded that standard risk models (CAPM and Fama-French (1993)) and industry effects do not explain the low returns of firms that score the highest with respect to social performance. However, their findings might be due to the fact that the sample period is rather small. In conclusion, they argue that, in line with Navarro (1988), companies that have high expenditures on 5

7 corporate social activities underperform compared to companies that have low expenditures on corporate social activities. In contrast to Brammer et al. (2006), Derwall et al. (2005) show evidence that SRI can be profitable for investors. They obtained data on eco-efficiency scores from Innovest and combined this with the CRSP database. Eco-efficiency is defined as the ratio of the value that a firm adds to the waste that it generates. On the basis of these eco-efficiency scores two yearly-rebalanced portfolios were constructed: the low-ranked (high-ranked) portfolio consisted of companies making up the 30 percent of total capitalization lowest rated (highest) by Innovest. The high-ranked portfolio performed substantially and statistically significant better than the low-ranked portfolio, even after controlling for differences in market sensitivity, investment style, or industry-specific factors. They also constructed portfolios based on best-in-class analysis for practical portfolio construction. The best-in-class portfolio outperformed the worst-in-class portfolio by about 3 percent with the best-inclass portfolio having a lower volatility than the worst-in-class portfolio. This difference persisted when using the Carhart (1997) four-factor model. Furthermore, when transaction costs are incorporated, the return gap widened even further because the worst-in-class portfolio turnover rate was higher than the best-in-class portfolio. The best-in-class portfolio outperformed the worst-inclass portfolio annually by 2.05 percent in a CAPM framework (not statistically significant) and by 4.25 percent in a Carhart (1997) four-factor framework (statistically significant at the ten percent level). Edmans (2011) investigated the relationship between employee satisfaction and long-run stock returns. A value-weighted portfolio was constructed consisting of companies that were included in the 100 Best Companies to Work For in America. This portfolio earned a four-factor alpha of 3.5 percent per year from 1984 to 2009 in excess of the risk-free rate. Moreover, the Best Companies portfolio still earned a statistically significant alpha of 2.1 percent annually while controlling for industry effects. The results remained the same when controlling for firm characteristics, different weighting methodologies, and the exclusion of outliers. According to Edmans (2011) these results suggest that the market fails to incorporate intangible assets (in this case employee satisfaction) fully into stock valuations. Hong and Kacperczyk (2009) studied another aspect of SRI screening, namely the exclusion of sin stocks (also called negative screening). Sin stocks are publicly traded firms that are involved in producing alcohol, tobacco, and gambling. Hong and Kacperczyk (2009) studied the performance of sin stocks on the American market over the period An equal-weighted portfolio (long in 6

8 sin stocks and short in comparable industry stocks) yields a CAPM alpha of 25 basis points that is statistically significant at the 10 percent level. The Fama-French (1993) model and Carhart (1997) model both yield an alpha of 26 basis points per month. They also performed cross-sectional regressions which controlled for firm characteristics and well-known determinants of expected returns (market size, past return, and market-to-book ratio). Again, they found that sin stocks outperformed their comparables by 29 basis points per month. They mention two reasons for their findings. First, following along the lines of Merton (1987) prices of neglected stocks will be lower relative to their fundamental value because of limited risk sharing. Second, sin stocks face higher litigation risk. Due to the fact that these sin stocks are shunned by large institutional investors they showed that there is a significant price effect in the range of percent. Fabozzi et al. (2008) also investigated sin stocks in an international context and found similar results. Their sample consisted of companies from 21 countries (including the US). They classified companies that obtained more than 30 percent of their revenue from six sin product categories (alcohol, tobacco, defense, biotech, gambling, and adult services) as sin stocks. The average company from the sin portfolio produced an average annual return of percent compared to 7.87 percent return on the market. Similar to Hong and Kacperczyk (2009) they conclude that this outperformance of sin stocks is due to the fact that they are shunned away by the average investor. Moreover, since sin industries have high barriers to entry (strict ordinances, rules, regulations, and multi-jurisdictional laws that control sin industries), companies operating in these industries, which have managed to survive despite these barriers, should be compensated with monopolistic rents. Salaber (2007) conducted another study about the performance of sin stocks. She used a sample of 18 European countries with different legal and cultural environments and she wanted to test whether these aspects have any influence on the return of sin stocks. First, she tested whether sin stocks exhibit higher risk-adjusted returns than other stocks in Protestant countries only. This is because Protestants are less willing to promote sin compared to Catholics. She also constructed a portfolio that was long in sin stocks and short in non-sin stocks. In Catholic countries she observed no abnormal returns whereas in Protestant countries the long-short portfolio had a statistically significant alpha of 5.7 percent using the Carhart (1997) four-factor model. Second, she hypnotized that in countries with high litigation risk you should observe higher risk-adjusted returns. She showed that a long-short sin portfolio returns a significant alpha of percent annually in countries with high litigation risk. Her third hypothesis is whether sin stocks in countries with high excise taxation on beer have higher risk-adjusted returns than both non-sin stocks and sin stocks in countries with lower excise taxes. When looking at the sin portfolio there is a statistically significant difference in the risk-adjusted return between the low-excise group and the high-excise group. 7

9 Many researchers have used ratings from KLD Research & Analytics for analyzing social responsibility on stock returns. This is mainly because KLD has the longest track record when it comes to social ratings. KLD uses two broad criteria to evaluate a company: qualitative and exclusionary criteria. Qualitative criteria (community, corporate governance, diversity, employee relations, environment, human rights, and products) are used for positive and best-in-class screening policies. Exclusionary criteria (involvement in alcohol, tobacco, gambling, military, nuclear power, and firearms) are used for negative screening policies. Kempf and Osthoff (2007) created several portfolios based on negative and positive screening policies from KLD. Similar to Hong and Kacperczyk (2009) and Fabozzi et al. (2008) they found that sin stocks perform better than non-sin stocks in their sample. However, the difference in return is not statistically significant, because Hong and Kacperczyk (2009) and Fabozzi et al. (2008) used a longer data sample and a narrower definition of sin stocks. Kempf and Osthoff (2007) also used positive screens; high-rated portfolios consisted out of the top 10 percent with respect to a screen (community, diversity, employee relations, environment, human rights, and products) and low-rated portfolios consisted of the bottom 10 percent. Their results show that a strategy long in the highrated portfolio and short in the low-rated portfolio, when using the community or the employee relations screens, yielded a statistically significant alpha of 4.55 percent and 5.98 percent respectively. When all positive screens were combined, the portfolio yielded an alpha of 4.46 percent and when the negative screen was added alpha increased to 4.8 percent. The best-in-class method (selecting stocks in the top 10 percent per industry) showed the highest performance with an alpha of about 5 percent per year. As in Derwall et al. (2005) including transaction costs did not change the performance of the long-short strategies. Increasing the cut-off points lowered the abnormal return leading to the conclusion that investors should only focus on the best stocks with respect to socially responsible screens. According to the view of Galema et al. (2008) the observed difference between theory and empirical research on SRI performance comes from misinterpretation of the results stemming from two possible errors. First, financial performance is calculated while controlling for systematic risk. However, the empirical measure used by researchers does not fully capture systematic risk. SRI and non-sri firms with equal risk levels may have different book-to-market ratios, because there is excess demand for SRI stocks. This would imply that exposure to the book-to-market ratio factor is independent of the risk profile of the underlying cash flows. As a consequence, the trade-off between SRI and financial performance is only partly captured by the book-to-market ratio. Second, the use of aggregate social responsibility ratings may confound existing relationships between individual dimensions of SRI and returns. They formed portfolios based on positive and negative 8

10 screens on six SRI dimensions of KLD during the period First, they tested if these portfolios could deliver excess return by using the Carhart (1997) four-factor model in a GMM framework as in Mackinlay and Richardson (1991) and Clare et al. (1997). Estimating portfolio returns in a GMM framework has the advantage of relying on weaker assumptions compared to OLS. In contrast to Kempf and Osthoff (2007), they do not find any risk-adjusted out- or underperformance. Second, Galema et al. (2008) ran cross-sectional regressions to investigate the impact of SRI scores on excess returns. Only the employee relations screen had a statistically significant positive effect on excess return of 84 basis points annually. Moreover, they noticed when they looked at the subscreens of employee relations that only concern subscreens have a statistically significant effect on excess return, but they were not consistent in sign. Therefore, they concluded that adding the subscores of screens leads to confounding effects. Lastly, they ran book-to-market regressions to measure whether a stock that has a low book-to-market ratio scores high on one of the SRI ratings. They found that diversity and environment have a negative relationship with the book-to-market ratio and governance a positive relationship with the book-to-market ratio. Therefore, Galema et al. (2008) argue that SRI is reflected in demand differences between SRI and non-sri stocks. Statman and Glushkov (2009) also studied the performance of companies that are listed on the S&P 500 using KLD ratings. They also used the best-in-class method to control for industry effects. Unlike Kempf and Osthoff (2007) they excluded companies that had no strength or concern indicators. Per KLD rating they ranked all companies by their best-in-class scores. Then they divided the companies into three groups of approximately the same number of stocks. An equal-weighted yearly rebalanced long-short portfolio was constructed that was long in top-third group and short in the bottom-third group. The returns were benchmarked against CAPM, Fama-French (1993) three factor model, and the four-factor model of Carhart (1997). Stocks with high social responsibility ratings performed generally better than stocks with low social responsibility ratings. However, most of the results were statistically insignificant except for the employee screen. As in Hong and Kacperczyk (2009) shunning controversial stocks is bad for performance. A portfolio long in accepted stocks and short in controversial stocks yielded a yearly excess return of percent (CAPM), percent (3-factor model), and percent (4-factor model). As in Galema et al. (2008), they found that applying both positive and negative screenings the effect of positive screens on return are offset by the effect of negative screens. Derwall et al. (2011) constructed two yearly rebalanced portfolios using the KLD data running from 1992 till The first portfolio consisted of only sin stocks as in Hong and Kacperczyk (2009). The second portfolio consisted of the top 30 percent of stocks that KLD ranks the highest with an employee-relations score. The abnormal return of both portfolios were measured by estimating the 9

11 Carhart (1997) four-factor model over four different time periods ( , , , and ). The annualized abnormal return of the first portfolio ranges from 2.58 percent to 2.86 percent over the different time horizons and is always statistically significant. The annualized abnormal return of the second portfolio is 5.62 percent over the period and 4.55 percent over the period. Both abnormal returns are statistically significant at the 10 percent level. The annualized abnormal return over the and period are 2.94 percent and 2.81 percent respectively. However, they are not statistically significant at the 10 percent level. 2.2 Hypothesis Development My thesis focuses on the risk-adjusted performance of portfolios based on SRI screening. In particular, to my knowledge, the ASSET4 database has not been used for this topic. Past studies have focused on mutual fund performance or used other databases. The advantage of using the ASSET4 dataset is that it has ESG (Environmental, Social and Corporate Governance) scores of multiple markets. The KLD dataset only has information from the US market and the Vigeo corporate social responsibility scores used by Van de Velde et al. (2005) only covers the European market. In contrast, the ASSET4 dataset covers both these markets and this increases the sample size. ASSET4 has four pillar scores which are updated annually. These pillar scores are aggregated into one score that reflects overall ESG performance. I am interested in whether portfolios based on ESG criteria can earn abnormal return after adjusting for common risk factors. Therefore, the topic of my thesis is: The effect of portfolio performance using social responsibility screens I investigate this topic by means of 5 hypotheses. Kempf and Osthoff (2007) and Statman and Glushkov (2009) showed that applying multiple screens yields a positive abnormal return. The aggregated score of the ASSET4 database is a combination of all four pillar scores and based on this score I test whether SRI will result in a positive abnormal return. Hypothesis 1: High social responsible stocks have higher risk-adjusted returns than low socially responsible stocks based on the overall ASSET4 rating. The first pillar in the ASSET4 database concerns the economic performance of a company. Economic performance is measured by client loyalty, performance, and shareholders loyalty. Fornell et al. (2006) show that companies that score high on customer satisfaction outperform the market. Performance as described by ASSET4 relates to the financial performance (for example, return on equity and profit margin) of a company. Shareholder loyalty relates to the ability of management to 10

12 retain its shareholders. For example, one of the measures used is the dividend payout ratio. If a company decides to lower its dividend payout ratio, this may result in shareholders selling their stake in the company. A lower dividend payout ratio thus negatively affects a company s economic performance and return. Therefore, my second hypothesis is: Hypothesis 2: Stocks that score high on economic performance have higher risk-adjusted returns than stocks that score low on economic performance based on the economic ASSET4 rating. The second pillar in the ASSET4 database is the environmental performance of a company. Environmental performance is measured by resource reduction, emission reduction, and product innovation. Derwall et al. (2005) show that companies with high eco-efficiency scores substantially and statistically significant outperform those companies with low eco-efficiency scores. Moreover, Kempf and Osthoff (2007) show that a portfolio consisting out of 10 percent of all stocks with the highest rating on environment generates an abnormal return of 3.6 percent annually. Therefore, my third hypothesis is: Hypothesis 3: Stocks that score high on environmental performance have higher risk-adjusted returns than stocks that score low on environmental performance based on the environment ASSET4 rating. The third pillar in the ASSET4 database is the social performance of a company. Social performance is measured by employment quality, health & safety, training & development, diversity, human rights, community, and product responsibility. Studies by Kempf and Osthoff (2007), Statman and Glushkov (2009), and Edmans (2011) have showed that stocks with high employee scores have an abnormal return that is positive and statistically significant. Empirical studies on diversity suggest that there is no statistically significant relation between stock return and diversity. Kempf and Osthoff (2007) did find a positive relation between diversity and stock return. However, their results were not statistically significant. Empirical studies on human rights show different results. Kempf and Osthoff (2007) show that there is a positive relation between human rights and stock return, while Statman and Glushkov (2009) show a negative relation. In both studies the relation between human rights and stock return was statistically non-significant. Studies by Kempf and Osthoff (2007), Statman and Glushkov (2009), and Galema et al. (2008) have showed that stocks with best community scores have an abnormal return that is positive. However, only the results of Kempf and Osthoff (2007) are statistically significant. Kempf and Osthoff (2007), Statman and Glushkov (2009), Galema et al. (2008), and Van de Velde et al. (2005) found no statistically significant relation between stock return and product quality. All in all, most empirical studies show that it is possible to achieve positive abnormal returns by screening stocks on social performance. 11

13 Hypothesis 4: Stocks that score high on social performance have higher risk-adjusted returns than stocks that score low on social performance based on the social ASSET4 rating. The fourth pillar in the ASSET4 database is the corporate governance performance of a company. Corporate governance performance is measured by board structure, compensation policy, board functions, shareholders rights, and vision & strategy. Empirical studies by Statman and Glushkov (2009), Galema et al. (2008), and Van de Velde et al. (2005) found no statistically significant relation between good corporate governance and stock return. Therefore, my fifth hypothesis is: Hypothesis 5: Stocks that score high on corporate governance performance have the same riskadjusted returns as stocks that score low on corporate governance performance based on the corporate governance ASSET4 rating. First, I will test these hypotheses using the raw sample returns. Next, I will run Carhart (1997) regressions to estimate the risk-adjusted returns in order to test the hypotheses more formally. 12

14 3. Data & Methodology 3.1 Data My main data resource is the ASSET4 database. ASSET4 rates companies against over 750 individual data points 2, which are combined into over 250 key performance indicators (KPIs). These KPIs are aggregated into a framework of 18 categories grouped within 4 pillars (Economic, Environmental, Social, and Corporate Governance) that are integrated into a single overall score. 3 Figure 1: The ASSET4 ESG framework KPIs, Categories, Pillars and Overall Score are equally weighted calculations of relative company performance, the benchmark being the ASSET4 company universe. These ratings are normalized to position the score between 0 and 100 percent. It expresses the value in units of standard deviation of that value from the mean value of all companies in the ASSET4 universe. The data comes from publicly available information, including sustainability/csr reports, company websites, annual reports, proxy filings, NGO as well as news of all major providers. In addition, CO2 data is sourced from the Carbon Disclosure Project. The ESG data is typically updated on an annual basis Using Thomson Reuters Datastream I have extracted companies that have available data on each of the four pillar scores (ECNSCORE, ENVSCORE, SOCSCORE, and CGVSCORE) and the overall score 2 For an overview of each individual data point see the ASSET4 ESG Data Glossary which is available on the website of Thomson Reuters: sx

15 (A4IR) over the period Next to the scores I have downloaded the monthly return and the market capitalization of each stock from Datastream. As in Fama and French (1992) I have excluded financials from the sample. The total sample consists out of 562 companies from 25 countries, 10 economic sectors, and 45 industries. For each score I have created two equal-weighted and two value-weighted portfolios. The following formula calculates the return for equal-weighted portfolios: represents the return of an portfolio p in month t. k is the total number of companies in portfolio p. is the return of company i in month t. The following formula calculates the return for valueweighted portfolios: Again, represents the return of an portfolio p in month t. k is the total number of companies in portfolio p. is the weight of company i in portfolio p in July of year τ. is the return of company i in month t. The first equal-weighted portfolio consists of the top 10 percent of companies that perform the best based on one of the five scores. The second equal-weighted portfolio consists of the bottom 10 percent of companies that perform the worst based on one of the five scores. The two valueweighted portfolios were created in similar fashion. All portfolios are yearly rebalanced in July. In addition, for each pillar score and the overall score high-low portfolios were constructed in which I go long in the top ten percent and short in the bottom ten percent for each score. 3.2 Methodology To measure the performance of the high-rated, low-rated, and the high-low portfolios, I employ the Carhart (1997) four-factor model. It controls for the impact of the market risk, the size factor, the book-to-market factor, and the momentum factor on returns. To control for these common factors, I estimate the following regression: The dependent variable is the monthly return of portfolio i in month t in excess of the risk-free rate. The independent variables are the returns of four factor portfolios. denotes the excess 14

16 return of the global market portfolio over the risk-free rate. denotes the return difference between a small and a large capitalization portfolio in month t. denotes the return difference between a high and a low book-to-market portfolio in month t. A stock with a low book-to-market ratio is referred to as a growth stock, while a high book-to-market ratio is referred to as a value stock. denotes the return difference between portfolios of stocks with high and low returns over the past twelve months. α denotes the abnormal return of the portfolio i.,,, and are the factor loadings and stands for the idiosyncratic return. The risk-free rate and the excess return of the market portfolio, the size factor, the value factor, and the momentum factor were taken from the Kenneth R. French data library using the global factors. Robustness checks In order to further test my hypotheses I perform three robustness checks. First, I test whether the results in previous regressions differ, if I look at different regions. Fama and French (2012) show that researchers can use the global four-factor model to explain the returns on global portfolios as long as the portfolios do not have a strong tilt toward stocks with very small market capitalization or toward the stocks of a particular region. My sample is heavily tilted toward the US (260 stocks) and Europe (267 stocks). As before, I employ the Carhart (1997) four-factor model only now using the US factors for the US sample and the European factors for European sample. Again, the US factors and the European factors were taken from the Kenneth R. French data library. Second, I analyze if the riskadjusted returns of the portfolios differ choosing various cut-offs. I recreate the portfolios by selecting the top 20 (30) percent stocks and the bottom 20 (30) percent stocks for both the equalweighted and the value-weighted portfolios. Third, due to the nature of their business some companies have lower score than other companies and create bias towards some industries. For example, oil companies tend to have lower scores relative to other companies with respect to environment. I recreate portfolios using best-in-class screening based on Thomson Reuters Business Classification. Thomson Reuters classifies companies in ten sectors: Basic Materials, Consumer Cyclicals, Consumer Non-Cyclicals, Energy, Financials, Healthcare, Industrials, Technology, Telecommunications Services, and Utilities. For the best-in-class approach the companies are divided into the ten sectors and then ranked according to their ASSET4 scores. Next, for each score the top 10 percent of companies that perform the best within a sector based on one of the five scores are selected and put in the high-rated portfolio. Similar, for each score the bottom 10 percent of companies that perform the worst within a sector based on one of the five scores are put in the lowrated portfolio. 15

17 4. Empirical findings 4.1 Descriptive Statistics Table 1 presents the descriptive statistics of the equal-weighted and the value-weighted portfolios created from the 562 companies that I consider. The monthly mean returns of the high-rated portfolios are lower than the low-rated portfolios in all screens except for the corporate governance screen (both equal- and value-weighted) and the value-weighted economic screen. The monthly mean returns are not corrected for any exposure to common risk factors. Not surprisingly, the monthly standard deviations of return of the high-rated portfolios are lower than the low-rated portfolios in all screens except for the corporate governance screen (both equal- and valueweighted) and the value-weighted economic screen. The high-low strategy has negative monthly mean returns for the aggregate, environment, and social screen for both the equal- and valueweighted portfolios. Only the corporate governance and economic screen have positive monthly mean returns for both the equal- and value-weighted portfolios. The monthly standard deviations of return of high-low portfolios are systematically lower than both the high-rated and low-rated portfolios. Furthermore, table 1 presents the t-statistics and the corresponding p-values of the difference in mean test between the high- and low-rated portfolios. Only the social screen of the value-weighted portfolios shows a statistically significant difference between the high-rated and lowrated portfolio mean returns. For all the other screens I cannot conclude that the difference between the mean returns is greater than zero. Table 2 presents the distribution of companies per country and economic sector. As noted in part 3.2 my sample consists mostly out of US companies (260) and European companies (267). The sectors consumer cyclicals and industrials are the two largest sectors in my sample and therefore a possible bias as described in part 3.2 could be present. 16

18 Table 1: Summary Statistics Portfolio Mean return (%) St. dev. (%) Minimum (%) Maximum (%) Difference of means t-statistic p-value Panel A: Equal-weighted portfolios Aggregate High-rated Low-rated High-Low (0.404) Corporate Governance High-rated Low-rated High-Low (0.465) Economic High-rated Low-rated High-Low (0.759) Environment High-rated Low-rated High-Low (0.171) Social High-rated Low-rated High-Low (0.082) Panel B: Value-weighted portfolios Aggregate High-rated Low-rated High-Low (0.405) Corporate Governance High-rated Low-rated High-Low (0.471) Economic High-rated Low-rated High-Low (0.867) Environment High-rated Low-rated High-Low (0.169) Social High-rated Low-rated High-Low ** (0.036) 17

19 Table 1 (continued) Portfolio Mean return (%) St. dev. (%) Minimum (%) Maximum (%) Panel C: Fama-French Portfolios Market SML HML MOM Note: This table summarizes for each screen the monthly mean return, the monthly standard deviation of return, the lowest and highest observed return in a month and the difference of means of the high- and low-rated portfolios. The difference of means test represents the t-statistics and the corresponding p-values are in parentheses. *, **, *** indicate statistical significance at the 10%, 5%, and 1% level. Panel A shows the summary statistics of equal-weighted portfolios. Panel B shows the summary statistics of value-weighted portfolios. Panel C shows the summary statistics of the Fama-French global portfolios. Table 2: Country and Sector Composition Country Number of Companies Economic Sector Number of Companies Australia 3 Basic Materials 58 Austria 7 Consumer Cyclicals 126 Belgium 9 Consumer Non-Cyclicals 60 Brazil 1 Energy 40 Canada 7 Financials 0 Denmark 11 Healthcare 55 Finland 13 Industrials 113 France 35 Technology 56 Germany 25 Telecommunications Services 22 Greece 6 Utilities 32 Hong Kong 4 Ireland 6 Israël 1 Italy 14 Japan 17 Mexico 1 Netherlands 13 Norway 9 Portugal 2 Spain 13 Sweden 25 Switzerland 24 Taiwan 1 United Kingdom 55 United States 260 Note: This table shows the number of companies from each country and economic sector as specified by Thomson Reuters Business Classification. 18

20 4.2 Portfolio Results Table 3 gives the results of the monthly abnormal return of the equal-weighted portfolios for each screen. The estimated alphas are positive for the high-rated portfolios of the aggregate, corporate governance, and economic screen. In contrast, the estimated alphas of the high-rated portfolios of the environment and social screen are negative. The low-rated portfolios of the aggregate, environment, and social screen show positive estimated alphas, while the low-rated portfolios of the corporate governance and economic screen show negative estimated alphas. The estimated alphas of the high-low portfolios of the corporate governance and the economic screen are positive, while the estimated alphas of the environment and social screen are negative. The estimated alpha of the aggregate high-low portfolio is positive, but very close to zero. In line with Galema et al. (2008) and Statman and Glushkov (2009), since the aggregate score is a combination of all four scores, it is possible that the positive abnormal returns of the corporate governance and economic screen are offset by the negative abnormal returns of the environment and social screens in the aggregate screen. To give an indication of economic significance: if an investor had gone long in the high-rated corporate governance portfolio and short in the low-rated corporate governance portfolio, the investor would have earned a monthly abnormal return of percent which is equal to percent on an annual basis. The high-low economic portfolio would have earned an abnormal return of percent annually over the period In contrast, the high-low portfolios of the environment and social screen would have returned an annual abnormal return of and percent respectively. The aggregate high-low portfolio only delivered an abnormal return of percent annually. However, all estimated alphas are not statistically significant. Looking at the market factor loadings, I observe that on average high-rated socially responsible companies have market betas lower than one. The low-rated portfolios of the aggregate, environment, and social screens have on average slightly higher market betas than one. The low-rated corporate governance portfolio has a similar market beta as the high-rated corporate governance portfolio. The low-rated economic portfolio has a rather high market beta of Furthermore, looking at the SMB factor I observe that on average high-rated socially responsible companies have a negative loading on the SMB factor. In contrast, low-rated socially responsible companies have a positive loading on the SMB factor. Thus, the high-rated socially responsible companies tend to be bigger with respect to market capitalization than low-rated socially responsible companies. Also, the HML factor loadings are positive for all the high- and low-rated portfolios indicating that these portfolios consist mostly out of value stocks. High-rated socially responsible companies tend to have lower loadings on the HML factor compared to low-rated socially responsible companies. Remarkably, all portfolios have a negative loading on the momentum factor. The negative momentum factor loading is greater for the low-rated portfolios than for the high-rated portfolios.

21 Table 4 presents the estimation results of the four-factor Carhart (1997) model of the value-weighted portfolios for each screen. The estimated alphas are negative for all of the high-rated portfolios. The high-low portfolios show the same direction as with the equal-weighted portfolios except that the aggregate high-low portfolio is negative. As with equal-weighted portfolios, all the estimated alphas are not statistically significant except for the high-low portfolio of the social screen which we also observed in table 1. Looking at the market factor loadings of the value-weighted portfolios I observe that these loadings are lower compared to the equal-weighted portfolios. The market factor loadings of the high-rated portfolios of the aggregate, economic, and social screens are lower than the lowrated portfolios which I also observed in the equal-weighted portfolios. As with the equal-weighted portfolios high-rated socially responsible companies tend to be large companies. In contrast to the equal-weighted portfolios, the exposure to the HML factor is much lower and in most portfolios not statistically significant. The momentum factor loading loses much of its statistical significance across all portfolios. For all the high-rated portfolios I cannot conclude that the momentum factor loading is statistically different from zero. I observe that there is only a difference between the momentum factor loading of the high- and low-rated portfolios of the aggregate and economic screen. Robustness checks Table 5 and 6 (see Appendix) show the performance of the equal-weighted portfolios of the US and European sample respectively. In both samples the estimated alphas of all the portfolios are statistically insignificant. Also, high-rated socially responsible companies have higher market risk than low-rated socially responsible companies. When I compare the US sample with the European sample, I observe that the European portfolios have much lower market risk exposure than the US portfolios. With respect to the SMB factor loadings I observe that the low-rated portfolios in the US sample are tilted towards small companies. The same holds on average for the European sample, but not as strong as in the US sample. The observed difference between the SMB factor loadings of the low- and high-rated portfolios in both samples are the same as in the total sample (see table 3). Moreover, in Europe low-rated portfolios have higher factor loadings on HML than the high-rated portfolios and thus the European low-rated portfolios consist of companies that have higher book-to-market ratios than the European high-rated portfolios. The negative momentum factor loadings observed in table 3 persist in both the US and European sample. Table 7 (see Appendix) shows the estimated alphas of equal-weighted and value-weighted portfolios for various cut-offs. In case of the equal-weighted portfolios increasing the cut-off from 20 to 30 percent, increases the monthly abnormal return of the high-rated aggregate portfolio by percent, the high-rated environment portfolio by percent, and the high-rated social portfolio 20

22 by percent. In contrast, for the value-weighted portfolios the estimated alphas of the highrated portfolios remain negative for all cut-offs. Also, for the value-weighted portfolios the estimated alphas of the low-rated portfolios remain positive for all cut-offs. However, changing the cut-off of the top and bottom portfolios do not result in any statistically significant alphas except for the valueweighted social high-low portfolio with a cut-off of 10 percent which was already observed in table 4. Table 8 (see Appendix) gives the results of the performance of the equal-weighted portfolios using best-in-class screening. Again, none of the portfolios show any statistically significant alphas. The market factor loadings are the same as observed in table 3. Overall, in contrast to table 3, the SMB factor loses statistical significance with best-in-class screening of the corporate governance, economic, environment, and social screen. The HML factor loadings are positive for the high- and low-rated portfolios, but only for the economic screen can I conclude that the high-rated portfolio has a statistically significant lower HML factor loading than the low-rated portfolio. As in previous results, the MOM factor loading is negative for the low- and high-rated portfolios. Table 9 (see Appendix) gives the results of the performance of the value-weighted portfolios using best-in-class screening. The estimated alphas of the value-weighted portfolio using best-in-class screening do not differ from the estimated alphas that I have found in table 4. Although the estimated alphas of the low-rated aggregate portfolio and the high-rated economic portfolio are different in sign compared to table 4, the differences are small and not statistically significant. Moreover, as in table 4 all the estimated alphas are statistically insignificant except for the high-low social portfolio. Again, using best-in-class screening results in lower market exposure with value-weighted portfolios than best-inclass screening with equal-weighted portfolios. On average high-rated socially responsible companies have negative loading on SMB. For the aggregate, environment, and social screen the high-rated portfolios have higher HML factor loadings. The momentum factor loadings are the same as observed in table 4. 21

23 Table 3: The Performance of Equal-weighted Portfolio Returns by Social Responsibility Characteristic, August 2002-December 2011 Aggregate Corporate Governance Economic Environment Social Alpha Market SMB HML MOM adj. R-sq High-rated (0.688) 0.818*** (0.000) * (0.064) 0.217* (0.079) * (0.075) Low-rated (0.788) 1.061*** (0.000) 0.485*** (0.005) 0.510*** (0.004) *** (0.000) High-Low (0.965) *** (0.000) *** (0.000) (0.106) 0.274*** (0.000) High-rated (0.456) 0.848*** (0.000) (0.281) 0.316** (0.042) ** (0.012) Low-rated (0.670) 0.857*** (0.000) 0.379** (0.012) 0.293* (0.068) *** (0.001) High-Low (0.286) (0.888) *** (0.006) (0.894) (0.453) High-rated (0.744) 0.901*** (0.000) (0.121) (0.767) (0.385) Low-rated (0.828) 1.233*** (0.000) 0.477* (0.060) 0.774*** (0.001) *** (0.000) High-Low (0.563) *** (0.000) *** (0.002) *** (0.000) 0.639*** (0.000) High-rated (0.803) 0.911*** (0.000) * (0.094) (0.152) ** (0.032) Low-rated (0.653) 1.040*** (0.000) (0.168) 0.493*** (0.007) *** (0.001) High-Low (0.436) ** (0.028) *** (0.003) * (0.092) 0.151* (0.095) High-rated (0.608) 0.852*** (0.000) *** (0.009) (0.286) ** (0.037) Low-rated (0.588) 1.068*** (0.000) (0.138) (0.102) *** (0.000) High-Low (0.244) *** (0.000) *** (0.000) (0.320) 0.177** (0.018) Note: This table summarizes for each screen the monthly abnormal return, factor loadings, and the adjusted R² of a portfolio strategy using the Carhart four-factor model. The high-rated (low-rated) portfolio consists of the 10% of all stocks with the highest (lowest) rating. The high-low portfolio is a trading strategy going long in the high-rated and short in the low-rated portfolio. The portfolios are equal-weighted. P-values are in parentheses. *, **, *** indicate statistical significance at the 10%, 5%, and 1% level. 22