The influence of rating conservatism after the Dodd-Frank Act on the Quality of credit ratings

(1)

“The Influence of Rating Conservatism after the

Dodd-Frank Act on the Quality of Credit Ratings”

Master thesis by Claudia van Galen

University of Amsterdam, Amsterdam Business School MSc Business Economics, Finance track

Supervisor: Dhr. Prof. dr. A.W.A Boot

July 7, 2016

ABSTRACT

This thesis analyses the impact of the Dodd-Frank Act on rating conservatism and credit rating quality and examines if there are certain circumstances under which the Dodd-Frank Act had more severe consequences. This thesis demonstrates that credit rating agencies issue lower ratings after Dodd-Frank, ceteris paribus, and that the degree of rating conservatism is stronger when: (i) competition between rating agencies is low; (ii) the issued ratings are high and (iii) information asymmetry is high. The stricter ratings after Dodd-Frank seem to be unwarranted as the incidence of false warnings increased. However, the lower ratings for bonds that are exposed to high information asymmetry do seem to be warranted. The average rating experienced a decline in the information content of downgrades and an increase in the information content of upgrades. When reputation concerns are expected to be strongest, the information content of downgrades did not change and the information content of upgrades increased. These findings are consistent with the reputation hypothesis of Goel and Thakor (2011). When information asymmetry is high, the information content of both downgrades and upgrades increases, which is consistent with the disciplining hypothesis. However, the robustness analysis suggests that these results should be interpreted with caution. When industry fixed effects are used in the empirical analysis, the presence of rating conservatism and false warnings decreases and the significance levels decrease as well. This gives reason to belief that part of the findings in the original analysis are driven by unobserved issuer characteristics.

(2)

1

Statement of Originality

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

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text an d its references have been used in creating it.

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

(3)

2

1. Introduction

Credit rating agencies assess the creditworthiness of debt instruments and disclose information about the likelihood of defaults and the recovery rates of securities. Credit ratings are based on both public and private information and during the rating process quantitative models and subjective judgement are used. In the early years of the twenty-first century, credit rating agencies were heavily scrutinized for being strategically sluggish in downgrading. At the end of 2001, Enron suffered one of the largest bankruptcies in history. However, six days before default Enron was still rated at investment grade. A similar story holds for WorldCom, which was rated at BB/Ba2 until three months preceding its bankruptcy (Greenbaum, Thakor and Boot, 2016). Given the large influence that credit rating agencies have in capital markets and given their importance to regulation, serious reforms in the credit rating industry followed. Two important reforms that were applicable to the credit rating industry were the Sarbanes–Oxley Act (SOX) of 2002 and the Credit Rating Agency Duopoly Relief Act of 2006. The SOX required the Securities and Exchange Commission (SEC) to investigate the role and function that credit rating agencies had in the operation of securities markets. The Credit Rating Agency Duopoly Relief Act aimed to increase competition by facilitating rating agencies to achieve the Nationally Recognized Statistical Rating Organization (NRSRO) status (Cheng and Neamtiu, 2009). Despite these reforms credit rating agencies again came under fire during the 2007-2009 financial crisis. This time, the criticism focused mainly on the ratings of structured finance products because of the mass downgrade of the mortgage-backed securities in 2008. After these downgrades the debate about the functioning of the credit rating industry flared up again. Rating agencies were accused of acting as oligopolists, issuing inflated ratings and having a lax attitude in issuing timely ratings (Manso, 2013).

As a response to the financial crisis, the U.S. Congress passed the Dodd-Frank Act which came into effect from 21 July 2010. The Dodd-Frank Act is a financial reform regulation aiming to decrease various risks in the U.S. financial system in order to protect consumers. In subtitle C, broad new regulation for the credit rating industry is outlined. However, the SEC and other federal agencies had the responsibility to develop more detailed regulation. For this reason, not all reforms were put into place directly after Dodd-Frank. Nevertheless, two important provisions became effective at the passage of the law. First, diminished pleading standards for private actions against credit rating agencies significantly increased their liability for inaccurate ratings. Second, the law made it easier for the SEC to sanction credit rating agencies for material misstatements and fraud. The overall objective of section C is to improve the quality of credit ratings. The rationale behind the increased legal liability is that it will discipline credit rating agencies to exert more effort in due diligence, to develop and use better rating methodologies and to strictly monitor their analysts. In the literature

(5)

4 this is also referred to as the disciplining hypothesis. The extra effort and strengthened internal control and corporate governance mechanisms should subsequently improve credit rating quality. However, earlier research predicts and demonstrates that the new regulation might not have had the desired result. Goel and Thakor (2011) contend that because of the asymmetric penalties, credit rating agencies will protect their reputation by issuing “conservative” ratings. By rating more conservative, the ex post revealed creditworthiness is more likely to lie in the range that the reported credit rating determined ex ante. So by issuing conservative ratings with a downward bias, the chance of ex-post reputational damage decreases. For this reason, the prediction that Dodd-Frank will lead to more conservative ratings is referred to as the reputation hypothesis. Dimitrov, Palia and Tang (2015) empirically examine the effect of Dodd-Frank on the quality of credit ratings and test if Dodd-Frank has indeed disciplined credit rating agencies. They do not find evidence for more accurate and informative ratings. Instead they find proof that is consistent with the reputation hypothesis of Goel and Thakor. Credit rating agencies have become more conservative which manifests itself in the issuance of worse ratings, more false alarms and less informative downgrades. When credit rating agencies reputation concerns are more important, even stronger evidence for the reputation hypothesis is found. Studies that have examined rating conservatism in different contexts, give insights in the circumstances under which rating conservatism is likely to be most prevalent. Bannier, Behr and Güttler (2010) demonstrate that stricter ratings are more pronounced when rating agencies have to rate bonds of opaque firms. A similar finding is detected by Atilgan, Ghosh, Yan and Zhang (2015) who argue that credit rating agencies will rate more conservative when information asymmetry is high. Furthermore, Atilgan et al. show that rating conservatism is more apparent for investment grade bonds than for speculative grade bonds.

Building on these ideas, the Dodd-Frank Act may have had more severe consequences under certain circumstances. The research question of this thesis exists out of two parts. First it verifies if the results of Dimitrov et al. (2015) hold by performing a replication of their research. The second part is supplementary to the research of Dimitrov et al. as it examines if there are certain circumstances under which Dodd-Frank had a stronger impact on rating conservatism and credit rating quality. The sample consists of all corporate bond ratings between 2006 and 2012, excluding ratings in the financial industry and ratings without available Compustat data. The pre Dodd-Frank period runs from January 2006 to the 21th of July 2010. The post Dodd-Frank period runs from the 22th of July 2010 until December 2012. After Dodd-Frank the odds of being rated one notch lower increase by 56.4%. When reputation concerns are more important to credit rating agencies, the odds of being rated one notch lower increase by 192.9%. When the lower ratings after Dodd-Frank are unwarranted, the default probabilities in each rating category should decrease because firms with

(6)

5 equal risk characteristics as before Dodd-Frank are assigned lower ratings after Dodd-Frank. To test if this is the case the occurrence of false warnings (i.e. speculative grade bonds that do not default within one year) in the pre and post Dodd-Frank period is compared. After Dodd-Frank the odds of a false warning increase by 63.9% and when reputation concerns matter the most to credit rating agencies, the odds show an increase of 355.4%. In case the conservative ratings are unwarranted, the reaction of investors to credit rating changes might have changed. If the reputation hypothesis holds, investors are likely to respond less to credit rating downgrades. They know that rating agencies require less evidence of decreased credit quality to issue a downgrade than they required before Dodd-Frank because they do not want to risk issuing a rating that is too high. On the other side, investors are likely to respond more to credit rating upgrades since they know that after Dodd-Frank, credit rating agencies will only issue an upgrade when they have very convincing evidence that the credit quality has improved. The empirical results show no statistically significant difference between the bond market returns after Dodd-Frank. On the contrary, the stock market does show a different reaction to rating changes pre and post Dodd-Frank. The median return after a downgrade becomes 0.263% less negative and is significant at the 1% significance level. The mean return after an upgrade increases by 0.355% and is significant at the 1% significance level. In economic terms this means that the information content of downgrades decreased and the information content of upgrades increased. This finding is in line with the reputation hypothesis. When reputation concerns are most important to credit rating agencies, the difference of both the mean and median excess returns after upgrades are significant and the difference of the mean return is almost twice as large as the difference of the mean return in the entire sample. This provides even stronger evidence for the reputation hypothesis. However, the statistically significant difference in the stock market’s reaction to downgrades that is found in the entire sample is absent when reputation concerns are most important to credit rating agencies.

The most important contribution of this thesis is the fact that it provides a more nuanced view on the impact that the Dodd-Frank Act had on credit rating quality. This thesis is the first to shed light on the circumstances under which the Dodd-Frank Act could have had a more pronounced impact. The choice of subsamples that will be examined is based on the insights provided by previous research on rating conservatism. Furthermore, the research of Dimitrov et al. (2015) will be critically checked which is useful as they are the first that empirically examine rating conservatism in the context of the Dodd-Frank Act. The results are roughly in line with the reputation hypothesis, but for the subsamples some evidence is found for the disciplining hypothesis. Compared to Dimitrov et al. this thesis finds a larger impact of Dodd-Frank on the rating levels but a smaller impact on the chance of a false warning. As regards the information content of credit rating changes, there are both similarities

(7)

6 and differences. This thesis cannot confirm the finding of Dimitrov et al. that the information content of downgrades to bond holders has decreased. Furthermore, this thesis finds that the information content of upgrades to stockholders has increased. This finding is absent in the paper of Dimitrov et al. and it strengthens the evidence for the reputation hypothesis. The first additional subsample that is analyzed is based on the presumption of Atilgan et al. (2015) that credit rating agencies will rate investment grade bonds more conservative than speculative grade bonds. For the five highest rating levels, i.e. AAA/Aaa to A+/A1, the odds of being rated one notch lower increase significantly after Dodd-Frank. As regards the information content, this subsample shows a statistically significant difference between the pre and post Dodd-Frank period for both the mean and median stock market returns to upgrades and downgrades. However, the results are ambiguous. The pre Dodd-Frank stock market reaction to downgrades is positive and the reaction to upgrades is negative. This contradicts with the idea that the stock market responds negatively to downgrades and positively or neutral to upgrades. Furthermore, this is the only subsample that displays a statistically significant difference in the bond market reaction. Before Dodd-Frank the bond market responded positively to upgrades but after Dodd-Frank the statistically significant reaction disappears. This suggests that upgrades after Dodd-Frank do not contain any information to bond holders anymore. This finding contradicts the belief of both the disciplining and reputation hypothesis. The second additional subsample that is examined looks at ratings for which information asymmetry is high. The results show that these bonds get stricter ratings after Dodd-Frank but that the incidence of false warnings remains equal. This gives reason to belief that the stricter ratings might be warranted by other factors for which the model did not controlled. Both stock market downgrades and upgrades become more informative to equity holders which is in line with the disciplining hypothesis. Finally, the entire sample is split up according to the Fama and French 12 industry classification. After separately analyzing all industries, there is no direct evidence that there are certain industries that were more strongly affected by the Dodd-Frank Act than other industries. The results of this thesis are interesting for policy makers that focus on the regulation in the credit rating industry because this thesis shows that regulation had a different effect under different circumstances. When it is possible to determine the underlying reason for the different effect, the government and the SEC can use this knowledge when (re)designing regulation for the credit rating industry.

De remainder of this thesis is organized as follows. Section 2 discusses the current literature on credit ratings, rating conservatism and the Dodd-Frank Act. Section 3 states the hypothesis and methodology. Section 4 explains how the data is collected and defined and contains descriptive statistics. Section 5 demonstrates the results and section 6 performs several robustness tests. Lastly, section 7 concludes and discusses the limitations, implications and recommendations.

(8)

7

2. Institutional background and literature review

2.1 Institutional background

2.1.1 The Sarbanes-Oxley Act and the Credit Rating Agency Duopoly Relief Act

After the defaults of Enron and WorldCom in 2001 and 2002, the U.S. Congress faced the important task to restore the confidence of investors in capital markets. As a response the Sarbanes-Oxley Act passed on the 25th of July 2002 by both the Senate and the House. Section 702 (b) of the Act required the SEC to extensively examine the role and function that credit rating agencies had in the operation of securities markets. This resulted in the issuance of a series of reports by the SEC which contained information about the role of credit rating agencies as financial intermediaries. Furthermore, a series of hearings were conducted by the U.S. Congress in April 2003. These hearings focused on the structure of the credit rating industry. Succeeding the tedious review process, the “Credit Rating Agency Duopoly Relief Act of 2006” became active in September 2006. After this act the SEC was removed from the process of approving rating agencies as nationally recognized. From that moment, any rating agency with more than three years of experience that met certain standards could register to become a nationally recognized rating agency. In this way the SEC’s opaque designation process disappeared, which facilitated increased competition in the credit rating industry (Cheng and Neamtiu, 2009).

2.1.2 Subtitle C – Improvements to the regulation of credit rating agencies

After the 2007-2009 financial crisis it turned out that the increased legislation after SOX was not sufficient to stop rating agencies from issuing inflated ratings. Where in 2001 and 2002 the problems were caused by inflated ratings of corporate bonds, in the 2007-2009 crisis it were the structured finance products that were inaccurately rated. The inflated ratings on structured finance products had a large adverse impact on the health of the economy of both the United States and the rest of the world. The U.S. Congress responded to the financial crisis by imposing the Dodd-Frank Act, a financial reform regulation that aims to protect consumers by mitigating various risks in the financial system. In SEC. 931. FINDINGS, part of Subtitle C of the Dodd-Frank Act, the SEC explains the motivation for the increased regulation in the credit rating industry. The SEC perceives the activities and performance of credit rating agencies as a matter of national public interest. Credit rating agencies fulfil a critical “gatekeeper” role in the debt markets. This role is similar to the role of security analysts and auditors and hence a similar level of public oversight and accountability is justified. Credit rating agencies are facing conflicts of interest and therefore they should be carefully supervised. To be able to give clearer authority to the SEC, the credit rating industry needed to be explicitly addressed in regulation.

(9)

8 One of the most important changes is the increased legal liability of credit rating agencies. Sec. 933. STATE OF MIND IN PRIVATE ACTIONS, describes the regulation around accountability. The enforcement of penalty provisions applies to the statements that credit rating agencies make in the same way as it applies to statements made by registered public accounting firms or security analysts. A credit rating agency is obliged to conduct a reasonable investigation of the rated security combining factual elements with its own risk evaluation methodology. Furthermore, they should obtain reasonable verification of the factual elements from other sources that are independent of the issuer and the underwriter. To make the ratings as transparent as possible the SEC requires credit rating agencies to fill in a form that discloses a wide range of information used in the rating process. To claim economic damages, Dodd-Frank made it sufficient for plaintiffs to show that the above stated rules were knowingly or recklessly violated. In the period before Dodd-Frank, credit rating agencies could appeal to the fact that their ratings were forward-looking statements, which meant that they only needed to be true at the time of issuance. By that time, they were not obliged to change the ratings when conditions changed. After the passage of the law this special protection disappeared and credit rating agencies now have the same legal liability as auditors and security analysts. This makes it easier for the SEC to punish credit rating agencies for fraud and material misstatements.

2.2 Literature review

The primary task of credit rating agencies is to evaluate the creditworthiness of debt instruments. Credit rating agencies act as intermediaries between investors and issuers and produce information that is widely available to investors. They have a large influence on the issuer’s cost of capital and access to capital, and their opinion is relied upon by many investors worldwide. They allow investors to rapidly check the global risk properties of many individual securities by means of a clear and well-known rating scale. Besides their information function to investors, credit ratings are also used in regulation and private contracting. For example, many institutional investors are only allowed to invest in investment grade securities and credit ratings are often included in the clauses of bond covenants. Furthermore, credit rating agencies also have an influence on the future credit quality of firms. By placing firms on their negative watch list for example, they provide firms with an incentive to undertake actions that prevent their credit quality from deterioration. Since credit ratings are an important channel of information dissemination in financial markets and are highly important to many parties, high quality credit ratings are relevant for a well-functioning financial system (Greenbaum et al., 2016). However, incidents in the past show that that the credit rating industry is not functioning as it is supposed to.

(10)

9 Since the adoption of the issuer-pays model in 1974, there have been a lot of accusations against credit rating agencies for issuing upward biased ratings. Jiang, Stanford and Xie (2012) show that after Moody’s adoption of the issuer-pays model, Moody’s ratings increased compared to Standard Poor’s (S&P) ratings who still used the investor-pays model. They also find that higher ratings are assigned when there is a greater conflict of interest, measured in terms of lower credit quality or higher expected rating fees. Skreta and Veldkamp (2009), Bolton, Freixas and Shapiro (2012) and Opp, Opp and Harris (2013) examine several failures in the credit rating process and confirm the finding of rating inflation. One of the aims of the Dodd-Frank Act is to temper credit rating agencies’ incentive to issue upward biased ratings. To predict the impact of the increased legal liability after Dodd-Frank on the behavior of credit rating agencies, Goel and Thakor (2011) designed a theoretical model. They forecast that in a situation without reputation concerns and conflicts of interest, increased legal liability will result in greater due-diligence effort and more informative ratings. When reputation concerns do matter, two opposing influences on the information content of ratings will play a role. On the one hand, credit rating agencies have the incentive to exert effort to build a reputation of being a high-ability credit rating agency. On the other hand, they are tempted to bias upwards investors’ beliefs about their ability by reporting strategically. If investors anticipate the latter, they will not be fooled. Nevertheless, rating agencies will report strategically, resulting in a weakening of the informational efficiency of credit ratings. Another adverse effect of the increased legal liability predicted by Goel and Thakor is caused by the fact that the legal liability increases asymmetrically, i.e. credit rating agencies will only be sued for ratings that ex post turn out to be too high. As a result, rating agencies will become more conservative and bias their ratings downwards. These predictions give reason to belief that the new regulatory framework of the Dodd-Frank Act might have failed to increase credit rating informativeness. The importance of reputation concerns as demonstrated by Goel and Thakor is in accordance with the theory of Morris (2001). One of the important insights of Morris’ model is that there will be no information conveyed in equilibrium if reputation concerns are sufficiently important. According to the model, credit rating agencies may start to lower their ratings beyond a level that is justified by the fundamentals of an issuer in order to protect their reputation. In the context of the Dodd-Frank Act this means that there will be less informative downgrades because credit rating agencies partly issue downgrades to protect their reputation. The opposite is expected for upgrades, which should become more informative as rating agencies will exert more effort because wrongly estimating default risk becomes costlier to them.

Dimitrov et al. (2015) are the first to empirically examine the effect of the Dodd-Frank Act on the quality of credit ratings by examining two hypotheses. The disciplining hypothesis predicts that the Dodd-Frank Act will achieve its objective of improving the quality of credit ratings. The rationale is

(11)

10 that the increased legal and regulatory penalties will encourage rating agencies to improve their due diligence, methodologies and the performance monitoring of their credit analysts. The second hypothesis, the reputation hypothesis, predicts an adverse effect on the quality of credit rating downgrades. The latter hypothesis is consistent with the theory of Goel and Thakor (2011) and Morris (2001) who state that credit rating agencies will issue lower ratings after Dodd-Frank because of the asymmetric penalties. Dimitrov et al. find that after the Dodd-Frank Act the odds that a bond is rated as B+ are 1.19 times greater than before Dodd-Frank, ceteris paribus.1_{After examining the} entire sample, they look at a subsample of ratings for which the ex-ante reputation costs are likely to be highest. Low Fitch market share is used as a proxy for high reputation costs since Becker and Milbourn (2011) have shown that rating agencies invest more in reputation when competition is low. Becker and Milbourn try to shed light on the issue whether or not increased competition in the rating industry will change the quality of credit ratings. They exploit the industry-level variation in Fitch market share to test if the ratings from Moody’s and S&P increase in quality when competition is high, i.e. when Fitch market share is high. They find evidence that is unequivocally consistent with lower rating quality when competition is higher. When competition from Fitch increases, the ratings issued by S&P and Moody’s increase and the ability of these ratings to explain bond yields decreases. Furthermore, they find that the ability of credit ratings to predict defaults decreases when competition increases. An explanation they propose is that increased competition leads to lower future expected rents, which decreases the incentive to invest in a reputation for accurate ratings. These findings together point out that the proposed plan to increase competition might not work out as well as expected. When Dimitrov et al. examine the ratings of S&P and Moody’s for industries in which Fitch market share is lowest, they find strong evidence for the reputation hypothesis. Ratings in this subsample see the odds of being rated one notch lower increase by 2.27 times. Additionally, the Dodd-Frank dummy becomes insignificant, suggesting that S&P and Moody’s only issue lower ratings after Dodd-Frank at times when their reputation concerns are most important.

Although Dimitrov et al. (2015) are the first to examine rating conservatism after the Dodd-Frank Act, the presence of rating conservatism is demonstrated by more authors. One of the first authors that looked into rating conservatism are Blume, Lim and Mackinlay (1998). During the time when their paper is written, there was a widespread belief among practioners that the credit quality of U.S. corporate debt was deteriorating. This belief was fueled by a trend in bond ratings that showed an increasing number of downgrades. Blume et al. examined the alternative view that there was no real decline in credit quality but that instead the rating standards were becoming more stringent. They

1_{The rating of B+ is used to demonstrate the general idea, but this rating could also have been any of the other}

(12)

11 concluded that there was reason to belief that from 1978 to 1995 the ratings became more stringent. However, their study does not rule out that the informational content of some variables might have changed over time. The paper of Alp (2013) complements the research of Blume, Lim and Mackinlay by examining structural shifts in credit rating standards from 1985 to 2007. Alp investigates if the systematic pattern of slackening ratings observed for mortgage-backed securities in the course of the financial crisis was also present in the corporate bond market. His analysis shows two main patterns in rating standards. Between 1985 and 2002 there is a divergent pattern between investment grade and speculative grade rating standards. The standards for investment grade ratings tighten while the standards for speculative grade ratings loosen. The second pattern shows a shift towards more stringent ratings for all credit ratings in 2002. The tightening standards until 2002 are in line with the finding of Blume, Lim and Mackinlay, but the loosening standards for speculative grade bonds are seen as puzzling. The finding of tightening rating standards after 2002 is in line with the idea that rating agencies began to employ more conservative rating practices after the SOX came into place. The research of Baghai, Servaes and Tamayo (2014) is complementary to the research of Alp as they examine the changes in rating standards over time and additionally look at the consequences of these changes for corporate behavior and debt pricing. Their base-case analysis shows a three notch decrease in credit rating levels from 1985 to 2009. The degree of conservatism is measured explicitly by the difference between the actual rating and the rating that is predicted by the estimated rating model over the period from 1985 to 1996. The difference between the estimated and predicted value seems to explain capital structure decisions. If the actual rating is one notch worse than the predicted rating, debt issuance decreases by 8%. For the sample average this is only 2.6%. Furthermore, the likelihood of obtaining a bond rating declined, cash holdings increased and growth slowed down and again these results are stronger for issuers that were exposed most to stricter ratings. If the increased stringency is warranted by a changing macroeconomic environment, the default rates should not have changed over time for bonds holding the same credit rating. In case the stricter ratings are unwarranted, the default rates for each rating category should have decreased over time. Baghai et al. find evidence that the rating conservatism is unwarranted for both investment grade and speculative grade bonds.

Besides research on the impact of the Dodd-Frank Act on credit ratings, research has also been done on the impact that the SOX had on credit ratings. Cheng and Neamtiu (2009) investigate whether and how credit rating agencies responded to the increased regulatory pressure and investor criticism after the SOX. The SOX made is costlier for rating agencies to give an investment grade rating to a bond that subsequently would default within one year (type I error). Given these increased costs, Cheng and Neamtiu examine if the timeliness, accuracy and volatility of credit ratings changed. They

(13)

12 state that there is a trade-off between rating accuracy and the timeliness of ratings. Rating agencies can tighten their credit standards so that they will not miss any defaults. In this case timeliness will increase, i.e. less type I errors, but accuracy will decrease because the ratings will be too harsh for companies with a lower risk profile. If ratings become too harsh, more speculative grade ratings will be given to issuers that will not default within one year (type II error). So the aim to reduce type I errors goes at the cost of an increase in type II errors. Another way to increase the timeliness of ratings is by shortening the information collection period and responding faster to new information. The drawback is that ratings may become more volatile since it becomes more likely that ratings will have to be reversed when new information comes out. Cheng and Neamtiu find that after the SOX, rating agencies start to downgrade defaulting bonds earlier and assign closer-to-default ratings. To examine rating accuracy, they look at the frequency of type I and II prediction errors. Contrary to the expectations of the trade-off theory, they find that both timeliness and accuracy improved at the same time. They also find a reduction in rating volatility, i.e. the standard deviation of credit rating levels. The finding that both timeliness and accuracy could improve while volatility did not increase, supports the criticism that in the past credit rating agencies did not put enough effort to assure high quality credit rating. Like Cheng and Neamtiu, Dimitrov et al. (2015) look at the accuracy of ratings and use it to verify if the lower ratings after Dodd-Frank are warranted by subsequent outcomes. If the stricter ratings after Dodd-Frank are unwarranted, two predictions should hold. First, less investment grade bonds should default within one year (type I error). In case the lower ratings are unwarranted, firms with for example a BBB– rating after Dodd-Frank are in a better condition than firms that got a BBB– rating before Dodd-Frank. So after Dodd-Frank the chance that an investment grade bond defaults within one year decreases compared to the pre Dodd-Frank period. Second, the likelihood that a speculative grade bond does not default within one year (type II error) should increase. Firms with for example a BB+ rating after Dodd-Frank are less risky than firms with a BB+ rating before Dodd-Frank because credit rating agencies issue downward biased ratings. When BB+ rated firms are less risky, the chance that they will default within one year will decrease which means that likelihood of false warnings (type II error) increases. As the sample of Dimitrov et al. does not contain investment grade bonds that default within one year, they focus on type II errors which are also referred to as false warnings. Dimitrov et al. find evidence that is consistent with the lower ratings being unwarranted. The chance that a rating will be a false warning is 1.84 times greater after Dodd-Frank in the general sample and 8.21 times greater when credit rating agencies face highest reputation concerns.

After concluding that rating standards have tightened and that the conservative ratings are likely to be unwarranted, Dimitrov et al. (2015) examine if there has been an impact on the information

(14)

13 content of credit ratings. They do this by comparing the mean and median (excess) returns after credit rating changes in the pre and post Dodd-Frank period. In their analysis they make a distinction between the reaction of the bond and the stock market. The disciplining hypothesis contends that the Dodd-Frank Act should improve the quality of credit ratings, resulting in more informative upgrades and downgrades. The reputation hypothesis argues that downgrades should become less informative because part of the downgrades are issued because rating agencies try to protect their reputation by issuing stricter ratings. On the contrary, upgrades are predicted to be more informative because upgrades expose credit rating agencies to legal and regulatory penalties, so rating agencies will require more convincing evidence before issuing an upgrade after Dodd-Frank. Their analysis provides evidence that is in line with the reputation hypothesis. The response of the bond and stock market to downgrades decreased as the mean announcement return became less negative than before Dodd-Frank. The mean return for upgrades in the bond and stock market increased but the increase was insignificant. This suggests that the information content of upgrades is not affected by the Dodd-Frank Act. The finding that the stock market does not react to upgrades is consistent with the findings of earlier research. Holthausen and Leftwich (1986) find little evidence of abnormal performance after upgrade announcements. Hand, Holthausen and Leftwich (1992) find no average excess stock returns at the time firms are put on the S&P Credit Watch list as an “indicated upgrade” and only very little evidence of average excess stock returns when an actual upgrade takes place. May (2010) finds a significant but economically small bond market reaction to upgrades and a statistically insignificant stock market reaction to upgrades.

The existing literature on rating conservatism and rating informativeness gives some insights about circumstances that could induce rating agencies to use tighter rating standards. According to Alp (2013) there are several explanations for tightening rating standards. First, rating agencies might use tighter ratings to prevent additional regulation in their industry. Second, accounting scandals could have led to the discounting of favorable information by ratings agencies because of increased skepticism about the financial reports and forecasts of companies. Lastly, he points to reputation concerns which could have tightened rating standards. Many other papers also point to reputation concerns as a cause for increased conservatism. Dimitrov et al. (2015) explain the rating conservatism after Dodd-Frank by the increased value that rating agencies attach to their reputation. Additionally, the results of Bedendo, Cathcart and El-Jahel (2016) are in line with the prediction that credit rating agencies become more conservative at times when their reputation is at stake. Atilgan et al. (2015) build on the literature about reputation concerns by contending that ratings are more conservative for issuers with relatively high information asymmetry. Holding information-collection effort constant, rating agencies are more likely to overestimate or underestimate the

(15)

14 creditworthiness of issuers when information is opaque. Bedendo et al. assume that credit rating agencies have an asymmetric loss function because the costs of an issuer’s default are higher than the costs when an issuer has a credit rating that is too low compared to its characteristics. Furthermore, they show that conservatism is higher for investment grade bonds than for speculative grade bonds because the reputation costs of underestimating the default risk are substantially higher for investment grade bonds than for speculative grade bonds. Bannier, Behr and Güttler (2010) examine rating conservatism in unsolicited ratings. Their results provide tentative evidence that stricter ratings are more pronounced in case of higher opaqueness. Lastly Opp et al. (2013) state that the introduction of the Dodd-Frank Act would lead to a reduction of the regulatory advantage of higher ratings. Consequently, their model predicts a systematic downward shift in the distribution of the ratings of Moody’s, S&P and Fitch. They expect that the increase in conservatism will be strongest for companies that benefited the most from the regulatory advantages of higher ratings. Given these insights the following subsamples will be examined in this thesis: (i) Ratings for which rating agencies have the highest reputation concerns; (ii) Ratings that belong to the highest rating categories; (iii) Ratings for which information asymmetry is high; and (iv) All Fama and French industries.

To verify the results of Dimitrov et al. (2015) and to examine if the impact of the Dodd-Frank Act on credit ratings is more pronounced in the subsamples mentioned in the previous paragraph, the methodologies of several papers that are discussed in this section are relevant for both the initial analysis and the robustness analysis. The general set-up of this thesis is similar to the set-up of Dimitrov et al. First, the rating levels before and after Dodd-Frank will be compared. Many papers that examine the difference in rating levels over time use the credit rating model of Blume et al. (1998) as their basis. They perform an ordered probit analysis using panel data. Accounting variables used by S&P and the equity risk measures that have been proved to have explanatory power are used to verify if firms get lower ratings over time while firm characteristics remain constant. The ordered probit model relates the rating categories to observed explanatory variables by means of an unobserved continuous linking variable. The rating categories of the independent variable are mapped into a partition of the range of the unobservable linking variable. This linking variable is in turn a linear function of the explanatory variables included in the model. More recent research, by amongst others Dimitrov et al. and Jiang et al. (2012), uses the same explanatory variables as Blume et al. but they exchanged the ordered probit model for an ordered logit model as the interpretation of this model is easier. The ordered logit model, also known as the proportional odds model, is often used when the dependent variable is coded on a discrete or ordinal scale and can have more than two outcomes (Baetschmann, Staub and Winkelmann, 2011). Ordered logit estimates an underlying

(16)

15 score as a linear function of the independent variables and a set of cutpoints. A positive ordered logit coefficient means that the likelihood of the dependent variable increases when the value of the independent variable increases. A negative coefficient means that the likelihood of the dependent variable decreases when the value of the independent variable increases. Since the ordered logit coefficient is relatively difficult to interpret, it is most convenient to translate it into an odds ratio, which is calculated as the natural logarithm of the logit coefficient. The odds ratio gives the change in odds for a unit increase in a continuous predictor or for a one level change in a categorical predictor. The difference between the ordered logit and ordered probit model is that the ordered logit model is based on a cumulative standard logistic distribution and the ordered probit model on a cumulative standard normal distribution. Besides ordered probit and logit models, Jian et al. (2012) and Baghai et el. (2014) additionally estimate a linear probability model using ordinary least squares (OLS). The advantage is that the interpretation of OLS coefficients is more straightforward. OLS does provide an unbiased estimate of the coefficients when the dependent variable is ordinal but the standard errors are biased because the error terms are non-homogeneous (Jiang et al., 2012).

The second part of the analysis exists out of an examination of the accuracy of credit ratings pre and post Dodd-Frank. Cheng and Neamtiu (2009) use several measures to test how rating accuracy changes over time. First, they compare the frequency of type I and type II errors in the pre and post SOX period by means of a univariate analysis. Second, they examine the changes in the frequency of type I and II errors using an extensive multivariate model including control variables for issue characteristics, firm-specific financial characteristics and macroeconomic conditions. They use three different cut-off points to classify defaulting issues; issues rated CC/Ca and worse, issues rated CCC+/Caa1 and worse and issues rated BB+/Ba1 and worse. They conclude that the multivariate analysis presents a more consistent picture than the univariate analysis. Additionally, they perform a relative accuracy test by plotting the fraction of firms with equal or lower ratings that defaulted within one year against the fraction of all firms that had an equal or lower rating. Finally, they also use two different bankruptcy prediction models as benchmark models to test if the accuracy increased after the SOX. To compare the accuracy before and after Dodd-Frank, Dimitrov et al. (2015) estimate a logit model of false warnings (type II errors) as a function of firm characteristics and bond market conditions. A logit model is a non-linear regression model that forces the predicted values to be either zero or one. The model is appropriate when the dependent variable can only take two possible values, representing the presence or absence of an attribute of interest. A logit model estimates the probability of the dependent variable to be one, i.e. the probability that the event that is being examined will happen.

(17)

16 Lastly the information content of credit rating changes is examined. Holthausen and Leftwich (1986) are one of the first to investigate this subject. They examine the impact of rating changes on the stock price by estimating the prediction errors from the market model. To measure the price response to a rating announcement, they look at the two-day window mean return over day 0 and day 1. They pick a two-day window because the press release might occur after the trading day is over. Hand et al. (1992) use a similar method but also look at the median return as the excess returns are slightly skewed. Besides the excess stock returns, Hand et al. (1992) also examine the excess bond returns. Since bonds are traded infrequently, they use ‘window-spanning’ excess returns. The window-spanning raw return is measured by taking the last transaction price in the period -11 to -1 and the first transaction price on or after day +1. If the window-span is larger than twenty days, the observation is eliminated. To calculate the return, accrued interest is added to the trading price. An excess bond return is defined as a raw bond return minus a risk free return, i.e. a long-term U.S. Treasury bond. The estimation of variances and covariances of returns between individual bonds is difficult because of the lack of frequent trades. For this reason, relatively simple tests are used. To test if the mean excess return is statistically different from zero a t-test based on the cross-sectional standard deviation is performed. To test if the median excess return is different from zero a binomial test, i.e. a Z-statistic, is used. May (2010) uses a relatively similar method as Hand et al. do. But instead of using window-spanning when a bond is not traded at the days surrounding a rating announcement, he simply uses the accrued interest as the raw return. Furthermore, he does not use the last price of the day but he estimates the volume-weighted average price. Bedendo et al. (2016) ass well use the market model to calculate excess stock price returns but instead of using a two-day window, they examine a three-day window ranging from day -1 until day +1. Dimitrov et al. use a similar method as Hand et al. do, but they use a raw bond return instead of an excess bond return because of the difference in maturity, credit quality and characteristics of the bonds in the sample.

This thesis complements the existing literature on rating conservatism after a change in regulation by examining if there are certain circumstances under which rating agencies apply stricter rating standards. Additionally, it examines if there is a larger impact on the information content of ratings under these circumstances. This is done by testing if factors that are marked by previous research as enhancing conservatism, also enhance conservatism in the context of the Dodd-Frank Act.

3. Hypothesis and Methodology

3.1 Hypothesis

The aim of this thesis is to test if the Dodd-Frank Act has led to rating conservatism and if this subsequently has changed the quality of credit ratings. The existing literature often explains rating

(18)

17 conservatism by the asymmetric loss function of credit rating agencies. Since their ratings play an important role in regulation and debt contracting, the costs from losses due to overvaluation are perceived as more important than the foregone gains in case of undervaluation (Beaver, Shakespeare and Soliman, 2006). The Dodd-Frank Act has increased the legal and regulatory penalties, which further increased the costs of overvaluation. This leads to the first hypothesis:

H1: Credit rating agencies have become more conservative after the Dodd-Frank Act.

It might be the case that there are unobservable or unknown factors driving the rating conservatism. This alternative explanation cannot be ruled out by examining the first hypothesis alone. To provide more direct evidence, the idea of Cheng and Neamtiu (2009) and Dimitrov et al. (2015) that is explained in section 2 is used. If rating conservatism is warranted by unknown factors, the average credit rating level after Dodd-Frank will be worse but the chance of default within each rating category will remain equal. On the other hand, if the lower ratings are solely caused by the increased legal and regulatory penalties and not by decreasing credit quality, the average rating level after Dodd-Frank will be lower so the default probability in each rating category will decline. When there is an unwarranted decrease in the average rating, bonds that had the highest default risk in the investment grade category now become bonds with the lowest default risk in the speculative grade category. Since the average credit quality of speculative grade bonds increases, the likelihood of a default in this group decreases, i.e. there will be more false warnings. Therefore, the first part of the second hypothesis states that:

H2a: Credit ratings issued post Dodd-Frank are more likely to be a false warning (type II error).

Since the bonds that had the highest default risk in the investment grade category move to the speculative grade category, the average default risk in the investment grade category decreases. When the average credit quality in the investment grade category increases it becomes less likely that an investment grade bond will default. Therefore, the second part of the second hypothesis states that:

H2b: The likelihood that an investment grade bond defaults within one year (type I error) decreases.2

If hypotheses 1 and 2 are true and investors anticipate on the rating conservatism, the bond and stock market response to credit rating changes might have changed after Dodd-Frank. If investors know that credit rating agencies downgrade bonds much faster after Dodd-Frank, market participants will discount credit rating downgrades. This discounting results in a loss of some of the

2_{As the sample that is examined in this thesis does not contain any investment grade bonds that default within}

(19)

18 credit rating agencies’ private information. According to this idea about the following hypothesis about the information content of downgrades is set:

H3a: The bond and stock market response to downgrades is weaker after Dodd-Frank.

Since rating misstatements have become costlier, a credit rating agency will require more convincing evidence before it issues an upgrade because it does not want to risk getting a penalty. Additionally, rating agencies are more likely to have obtained private information to make sure that the upgrade is justified. This leads to the following hypothesis about the information content of upgrades:

H3b: The bond and stock market response to upgrades is stronger after Dodd-Frank.

The above hypotheses will be tested using the entire sample and using several subsamples based on the literature in section 2. The first subsample exists of ratings for which rating agencies had strongest reputation concerns (Dimitrov et al., 2015). In industries with lower Fitch market share, Moody’s and S&P have a stronger incentive to protect their reputation, so Fitch market share is used as a proxy for reputation concerns (Becker and Milbourn, 2011). Second, firms with different rating levels could be differently impacted by rating conservatism. Atilgan et al. (2015) contend that the reputation costs of underestimating the default risk of an investment grade bond are substantially higher than the costs of underestimating the default risk of a speculative grade bond. Since the credit quality within the investment grade category and the speculative grade category also differs substantially, this analysis is performed by using additional subcategories as indicated in column 2 of appendix A. Thirdly, rating conservatism is more present in the financial and insurance industry (Bannier et al., 2010). To examine if there are other industries in which rating conservatism increased after Dodd-Frank, the entire sample will be divided in multiple subsamples using the Fama and French 12 industry classification. The final subsample will look into credit ratings for issuers with relative high information asymmetry. Earlier research3_{finds that issuers with relative high} information asymmetry or issuers in opaque industries get more conservative ratings.

3.2 Empirical methodology

To test the hypotheses formulated in section 3.1, three different tests will be employed. Section 3.2.1 outlines the model that will test if ratings issued after Dodd-Frank are lower than those issued before Dodd-Frank. Section 3.2.2 describes the model that will examine if the incidence of false warnings in the post Dodd-Frank period has increased. Finally, section 3.2.3 explains the methodology that is used to investigate if the information content of credit rating changes has changed.

(20)

19

3.2.1 Conservatism test

Like previous research, this thesis will use the credit rating model of Blume et al. (1998) to test if the first hypothesis is true. The example of Dimitrov et al. (2015) is followed and an ordered logit model instead of an ordered probit model is chosen because it makes the interpretation of the coefficients easier and it facilitates the comparison with Dimitrov et al. An important advantage of the ordered probit and ordered logit model is that it makes the exact magnitude of the rating levels irrelevant. The dependent variable 𝑅𝑎𝑡𝑖𝑛𝑔 𝑙𝑒𝑣𝑒𝑙𝑖,𝑡 represents the rating recoded into a numeric rating, where the best rating is coded as one and higher numeric ratings are considered to be worse ratings. The independent variable of interest is 𝐷𝐹𝑖,𝑡, which is a dummy variable with the value zero pre Dodd-Frank and the value one post Dodd-Dodd-Frank. To control for the determinants of credit ratings the following control variables (𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡) are added: operating margin, long-term debt leverage, total debt leverage, the logarithm of the market value of equity, stock beta, idiosyncratic stock return volatility and interest coverage. Additional to the determinants of credit ratings, two dummy variables are added to capture the effect of the type of rating agency that issued the rating. The definitions of all variables are discussed in section 4.2 and appendix B. Standard errors are clustered at the firm level to account for the fact that the standard errors are only uncorrelated between firms but not within firms. It is important to cluster at the firm level as most firms have multiple ratings in the sample. The result is the following base-case model:

(1.1) 𝑅𝑎𝑡𝑖𝑛𝑔 𝑙𝑒𝑣𝑒𝑙𝑖,𝑡 = 𝛽1∗ 𝐷𝐹𝑖,𝑡+ ∑ 𝛽𝑖∗ 𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡 8

𝑖=2

+ 𝜀𝑖,𝑡

Hypothesis 1 states that credit rating agencies have become more conservative so the expectation is that, after controlling for the determinants of credit ratings, the average rating becomes worse after Dodd-Frank. This leads to a positive expected sign on the Dodd-Frank coefficient. As regards the controls, firms are expected to have a better credit rating4_{when they have a higher operating} margin, are larger in terms of the market value of equity and have higher interest coverage. So the expected sign for operating margin, market value of equity and interest coverage is negative. Firms are expected to get a worse credit rating when they have higher leverage and when the stock beta and idiosyncratic return volatility are higher as both are sources of risk. Beforehand there is no expectation on the sign of the rating agency dummy variables. To test if rating conservatism is stronger in the subsamples discussed in section 3.1, several extended models (model 1.2) are estimated using the independent dummy variable 𝑆𝑢𝑏𝑖,𝑡, that has the value one for all ratings that fall in the subsample and zero otherwise. To be able to test if rating conservatism increased within

4_{Pay attention that a better credit rating means a lower value for the dependent variable as the highest credit}

(21)

20 this subsample after Dodd-Frank, model 1.1 is appended by an interaction term(𝐷𝐹𝑖,𝑡∗ 𝑆𝑢𝑏𝑖,𝑡). The result is the following extended model:

(1.2) 𝑅𝑎𝑡𝑖𝑛𝑔 𝑙𝑒𝑣𝑒𝑙_𝑖,𝑡 = 𝛽1∗ 𝐷𝐹𝑖,𝑡+ 𝛽2∗ 𝑆𝑢𝑏𝑖,𝑡+ 𝛽3∗ (𝐷𝐹𝑖,𝑡∗ 𝑆𝑢𝑏𝑖,𝑡) + ∑ 𝛽𝑖∗ 𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡 10

𝑖=4

+ 𝜀𝑖,𝑡

The first subsample dummy (𝑆𝑢𝑏𝑖,𝑡) is one for all issuers that are part of an industry that falls into the lowest quartile of Fitch market share. The expected sign on the interaction term is positive because stronger reputation concerns are likely to lead to more conservative ratings. There is no theory or hypothesis that predicts the effect of low Fitch market share on the credit rating level regardless of Dodd-Frank, so there is a neutral expectation for the independent Fitch market share dummy. The second subsample makes a distinction between the different rating categories. The “highest” rating category contains rating levels 1 and 2, the “high” category levels 3 to 5, the “medium” category levels 6 to 10, the “non-investment grade” category levels 11 to 14 and the “low” category levels 15 to 21.5_{As the reputation costs of underestimating default risk are expected to be higher for} investment grade bonds than for speculative grade bonds (Atilgan, 2015), the expected signs on the interaction terms with the highest, high and medium category dummy variables are positive and the expected signs on the interaction terms with the non-investment grade and low category dummy variables are negative. The dummy variables are also added to the model as independent variables because of statistical purposes. However, their coefficients should not be interpreted as causal. As regards the third sub-analysis, the examination of the Fama and French industries, there is no presumption about which industries in the sample are most strongly affected. Previous research indicates that issuers in the financial and insurance industry are rated more conservative, but this industry is excluded from the sample. For the other eleven industries there is no presumption about the expected sign. In the final subsample information asymmetry is proxied by the number of analysts that are covering a firm. The choice of this proxy is based on research by Dahiya, Iannotta, and Navone (2011) who examine the effectiveness of various opacity proxies. They conclude that the number of analysts and the Amihud ratio are the most reliable proxies. The information asymmetry dummy variable is one if an issuer falls in the bottom decile of analyst coverage and zero otherwise. The expected signs for the independent variable and the interaction variable are positive as information asymmetry is likely to make rating agencies more cautious subsequently resulting in increased conservatism.

(22)

21

3.2.2 Rating accuracy test

Due to time constraints it is not feasible to use multiple tests to test whether the second hypothesis is true. The logit model of Dimitrov et al. (2015) is performed because this facilitates comparison. For the subsamples in which the logit model causes a statistical problem, an OLS model will be estimated. As explained in section 3.1, the empirical research will only focus on hypothesis 2a since hypothesis 2b cannot be examined as the sample does not contain any investment grade rated bonds that default within one year. The dependent variable in the model is the binary variable 𝐹𝑎𝑙𝑠𝑒 𝑤𝑎𝑟𝑛𝑖𝑛𝑔𝑖,𝑡, which is one if an issuer received a speculative grade rating and did not defaulted within one year and zero otherwise. The independent variable of interest is 𝐷𝐹𝑖,𝑡 and to control for issuer characteristics and market conditions the following control variables (𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡) are added: bond index return, years to maturity, return on assets, the logarithm of the market value of equity, interest coverage, total stock return volatility, book to market ratio and the long-term debt to equity ratio. Additionally, rating agency dummy variables are added and standard errors are clustered at the firm level. This results in the following model:

(2.1) 𝐹𝑎𝑙𝑠𝑒 𝑤𝑎𝑟𝑛𝑖𝑛𝑔𝑖,𝑡= 𝛽1∗ 𝐷𝐹𝑖,𝑡+ ∑ 𝛽𝑖∗ 𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡 9

𝑖=2

+ 𝜀𝑖,𝑡

If the rating conservatism is unwarranted, the incidence of false warnings should increase since firms with the same default risk characteristics as before Dodd-Frank get lower credit ratings. The odds of a false warning are expected to decrease when the bond market in general performs better, i.e. when the bond index return is high. The odds of a false warning are also expected to decrease when the market value of equity increases, when a company has higher interest coverage and when the book to market ratio increases. The odds of a false warning are expected to increase when return on assets becomes higher, total stock return volatility is higher and when the long-term debt to equity ratio becomes larger. Beforehand there is no expectation on the sign of the rating agency dummy variables, years to maturity and the intercept. To test if the incidence of false warnings is stronger in the subsamples discussed in section 3.1, the logit model is extended by the variables 𝑆𝑢𝑏𝑖,𝑡 and (𝐷𝐹𝑖,𝑡∗ 𝑆𝑢𝑏𝑖,𝑡). The resulting extended model is defined as:

(2.2) 𝐹𝑎𝑙𝑠𝑒 𝑤𝑎𝑟𝑛𝑖𝑛𝑔 = 𝛽₁∗ 𝐷𝐹𝑖,𝑡+ 𝛽2∗ 𝑆𝑢𝑏𝑖,𝑡+ 𝛽3∗ (𝐷𝐹𝑖,𝑡∗ 𝑆𝑢𝑏𝑖,𝑡) + ∑ 𝛽𝑖∗ 𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖,𝑡 11

𝑖=4

+ 𝜀𝑖,𝑡

In model 2.2 the predicted signs on the 𝑆𝑢𝑏𝑖,𝑡 and (𝐷𝐹𝑖,𝑡∗ 𝑆𝑢𝑏𝑖,𝑡) variables are equal to those explained in section 3.2.1.

(23)

22

3.2.3 Informativeness test

Bond and stock return tests can be used to deduce information about the timeliness and information content of rating changes (Beaver et al., 2006). This thesis will use the market model methodology that is outlined at the end of section 2 to estimate the abnormal returns. The bond market returns will be estimated using the window-spanning methodology of Holthausen and Leftwich (1986) but instead of excess bond returns the example of Dimitrov et al. (2015) is followed and raw bond returns are used.6_{For the reasons explained in section 2, a series of simple stock return tests will be} carried out to provide evidence relating to hypothesis 3a and 3b. Before examining the differences between the pre and post Dodd-Frank returns, a one-sample t-test is applied to the mean returns and a one-sample Wilcoxon signed rank test is applied to the median returns. This is done to check if the bond and stock market perceive credit rating upgrades and downgrades as containing relevant information. To verify if hypothesis 3a is true, the mean bond and stock announcement returns around downgrades before and after Dodd-Frank are compared using a two-sample t-test. As the estimated returns are slightly skewed the median announcement returns will also be compared. To test if the medians are significantly different from each other a Wilcoxon two-sample test is used. To verify hypothesis 3b the same tests that are just described are performed for upgrade announcements. Two-tailed critical values are used to determine the significance levels. The usage of the bond market reaction versus the stock market reaction has both its advantages and its disadvantages. Bond prices are more directly affected by rating changes that indicate changing default probabilities but bonds are often traded infrequently. Stocks are traded daily, which makes it easier to isolate the reaction to the rating change from the reaction to other firm relevant news. On the other hand, stock prices are less sensitive to changing default probabilities.

The difference between the reaction of the stock and bond market in the pre and post Dodd-Frank period is measured by subtracting the pre Dodd-Frank return from the post Dodd-Frank return. According to hypothesis 3a the expected sign on the difference in reaction to downgrades will be positive, i.e. the reaction post Dodd-Frank will be less negative than the reaction pre Dodd-Frank. Hypothesis 3b predicts that rating upgrades will become more positive which is expressed by a positive expected sign, i.e. the reaction post Dodd-Frank will be more positive than the reaction pre Dodd-Frank. Though, it might be the case that the stock market reaction to upgrades is weaker than the reaction of the bond market as Hand et al. (1992) and May (2010) find little evidence of positive excess returns for the equity of upgraded firms.

(24)

23

4. Data and descriptive statistics

4.1 Data

From the Mergent Fixed Income Securities Database (FISD), data is gathered on credit ratings issued by Moody’s, S&P and Fitch over the period 2006 to 2012. The sample starts in 2006 to avoid contaminated results because of market adjustments that are caused by the SOX in 2002. The sample ends in 2012 to retain the possibility for further research to examine the potential effects of rating conservatism on long-run investment and firm performance. Since ratings indicating default are issued ex-post, they are excluded from the sample. Additionally, financials (SIC 6000-6999) are removed from the sample. For the empirical analysis, quarterly financial statement data from Compustat and market value data from the Center for Research in Security Prices (CRSP) is needed. To make sure that the quarterly financial statement data is available to the credit rating agencies at the moment they issue a rating, the credit ratings are matched with the financial statement data of the previous fiscal quarter.7_{An important dependent variable in the analysis is the occurrence of} false warnings. As Moody’s does not provide ex-post default ratings it is not possible to construct a false warning variable for bonds that are only rated by Moody’s. Hence, bond issuers that are only rated by Moody’s are excluded from the sample. The remaining dataset contains several bonds that are rated by different rating agencies at the same date. In this case, the rating with the largest rating change compared to the previous rating is kept in the sample. Additional to FISD, Compustat and CRSP, Datastream is used to access the thirty year Barclays United States Treasury Bond Index. The final sample contains 21,337 ratings which are be upgrades, downgrades, affirmed ratings, confirmed ratings or initial ratings. As the data sample contains ratings of Moody’s, S&P and Fitch which use two different types of rating notations, ratings are transformed to numerical categories ranging from one to twenty-two. The first category identifies an AAA/Aaa rating, the second category identifies an AA+/Aa1 rating and so forth. The complete numerical transformation can be found in appendix A. Table 1a contains the frequency distribution of the ratings over time per rating level. Additionally, table 1a shows the average rating for each year and divides the entire sample in a pre and post Dodd-Frank sample to facilitate comparison. The average credit rating has become half a notch better after Dodd-Frank. This points to the economic recovery after the financial crisis. Another remarkable point is that the ratings post Dodd-Frank have a lower standard deviation and are more centred around the average rating where the pre Dodd-Frank period contains more ratings in the tails of the distribution.

The definitions of the main dependent and independent variables and some control variables are explained below and the definitions of the remaining variables can be found in appendix B. Following

The influence of rating conservatism after the Dodd-Frank Act on the Quality of credit ratings