The Issuer-Pays Rating Model and Ratings Inflation: Evidence from Corporate Credit Ratings

The Issuer-Pays Rating Model and Ratings Inflation: Evidence from Corporate Credit Ratings Günter Strobl and Han Xia November 2011 Abstract This paper provides evidence that the conflict of interest caused by the issuer-pays rating model leads to inflated corporate credit ratings. Comparing the ratings issued by Standard & Poor s Ratings Services (S&P) which follows this business model to those issued by the Egan-Jones Rating Company (EJR) which adopts the investor-pays model, we demonstrate that the difference between the two is more pronounced when S&P s conflict of interest is particularly severe: firms with more short-term debt, a newly appointed CEO or CFO, and a lower percentage of past bond issues rated by S&P are significantly more likely to receive a rating from S&P that exceeds their rating from EJR. However, we find no evidence that these variables are related to corporate bond yield spreads, which suggests that investors may be unaware of S&P s incentive to issue inflated credit ratings. JEL classification: D82, G24 Keywords: Corporate credit ratings; Issuer-pays rating model; Ratings inflation We thank Paolo Fulghieri, Paige Ouimet, Anil Shivdasani, as well as seminar participants at the Fordham University, University of New South Wales, the University of North Carolina, the University of Texas at Dallas, Washington University,the 7th Annual Conference on Corporate Finance at Washington University in St. Louis and the 2011 European Finance Association Annual Conference for comments on an early draft. We thank the Egan-Jones Rating Company for sharing their data with us. We particularly thank Peter Arnold and Chris Bauman for their help on data collection. Any errors belong to the authors. Günter Strobl, Kenan-Flagler Business School, University of North Carolina at Chapel Hill, McColl Building, C.B. 3490, Chapel Hill, NC 27599-3490. Tel: 1-919-962-8399; Fax: 1-919-962-2068; Email: strobl@unc.edu Corresponding author: Han Xia, School of Management, University of Texas at Dallas, 800 West Campbell Rd., SM31, Richardson, TX 75080-3021. Tel: 1-972-883-6385; Fax: 1-972-883-2799; Email: Han.Xia@utdallas.edu

1 Introduction Many market observers have accused credit rating agencies of having contributed to the recent financial crisis by having been too lax in the ratings of some structured products. The picture that has emerged in the press is one in which rating agencies compromised the quality of their activities to facilitate the selling of their services. This behavior has been attributed to an inherent conflict of interest in the agencies business model: rating agencies receive their principal revenue stream from issuers whose products they rate ( issuer-pays model ). Rating agencies have responded to this accusation by arguing that such an attitude would put their reputation, which is arguably their most valuable asset, at risk and would therefore irrevocably damage their business in the long run. The objective of this paper is to empirically investigate whether reputational concerns are sufficient to prevent rating agencies from issuing inflated credit ratings. We address this question by comparing the credit ratings issued by two different rating agencies, Standard & Poor s Ratings Services (S&P) and Egan-Jones Rating Company (EJR). Unlike S&P (and other major rating agencies) whose ratings are paid for by issuers, EJR relies on subscription fees paid by investors as its principal source of income ( investor-pays model ). To the extent that investors value accurate ratings, this alternative business model eliminates EJR s incentive to shade its ratings upwards so as to please their clients, making them an ideal benchmark for identifying any bias in S&P s ratings. Although S&P and EJR use the same categories for their credit ratings (AAA to D for long-term ratings and A-1 to D for short-term ratings), one might argue that their meaning is potentially different. A higher (i.e., closer to AAA) rating from S&P may thus not be an indication of ratings inflation, but simply reflect differences in the agencies rating methodology. Rather than comparing S&P s ratings directly to those of EJR, we rely on the cross-sectional analyses. Specifically, we identify circumstances under which the conflict of interest caused by the issuer-pays model is particularly severe and examine whether the difference in ratings 1

between S&P and EJR is especially pronounced under these circumstances. We employ several proxies to measure the severity of S&P s conflict of interest. Our first proxy exploits the amount of the issuer s short-term debt. The greater the firm s short-term liquidity needs, the more likely it is to issue large amounts of debt in the near future, and thus the more business S&P can obtain from the firm in the future. The prospect of earning additional rating fees gives S&P an incentive to issue a favorable rating so as to attract the firm s future business and forestall the firm s taking its business to another rating agency. Our second set of results are related to S&P s market share on a per-client basis. In particular, we examine how S&P s revenue share, defined as the volume-weighted percentage of bond issues rated by S&P over the past 2, 4, 6, or 8 quarters, affects the difference between S&P s and EJR s rating. We conjecture that a lower revenue share increases S&P s incentive to produce issuer-friendly ratings so as to attract more business in the future. Our final set of tests are motivated by the observation that newly appointed corporate leaders are more inclined to change the firm s operational and financial strategy (e.g., Giambatista et al., 2005), and may therefore also be more likely to switch credit rating agencies. We hypothesize that S&P may thus have a particularly strong desire to please its customers by issuing favorable ratings for firms that recently appointed a new CEO or a new CFO. In each test, we find evidence that S&P is more likely to issue an inflated rating when its conflict of interest is particularly severe. The effect of the issuer-pays model on rating inflation is of significant economic magnitude. The average difference in the two agents ratings is about 0.2 notch, indicating that S&P issues a rating that is one notch step-up from EJR s for proximately every one out of five issuers. Moreover, one standard deviation increase in the amount of a firm s short-term debt leads to a 0.16 notch increase in the difference between S&P s and EJR s ratings, amounting to a 84% change of the sample mean of the rating difference. Similarly, one standard deviation decrease in S&P s revenue share over the past 4 quarters results in a 0.1 notch increase in the difference in ratings, and the appointment of a new CFO increases the difference by 0.29 notches. 2

While our findings indicate that the conflict of interest associated with the issuer-pays model leads to a significant amount of ratings inflation, it is not clear whether it leads to any misallocation of resources. If investors anticipate the bias in S&P s ratings, the firm s debt will be correctly priced and inflated credit ratings may not be harmful to society. To address this issue, we examine whether our variables that predict the extent of S&P s ratings inflation can also predict a bond s yield spread at issuance. We find no evidence that this is the case. The amount of a firm s short-term debt, S&P s past revenue share, and the appointment of a new CEO/CFO have no significant effect on the bond s yield spread. These findings are consistent with the view that investors are unaware of S&P s incentive to issue inflated credit ratings. Hence, rating inflation can lead to misallocation of resources. These results further imply that regulators intervention in the credit rating industry would be beneficial to investors who use credit ratings to guide their investment decisions, and can potentially increase social welfare. Our study adds to the literature that examines different properties in ratings from issuerpaid and investor-paid rating agencies. As one of the first studies to compare ratings issued by different types of agencies, Beaver, Shakespeare, and Soliman (2006) find that EJR reacts more timely in changing its ratings than Moody s, and EJR s rating changes are followed by a stronger market reaction. These results suggest that EJR seems to provide more informative ratings than the issuers-paid major rating agencies. The authors attribute this finding to the two agencies NRSRO designation; while Moody s ratings serve for the regulatory and contractual purposes due to Moody s NRSRO certification, the non-certified EJR only servers as an additional information provider. 1 In a more recent study by Cornaggia and Cornnagia (2011), the authors also document that Financial Health Ratings, an independent rating company compensated by investors, provide more timely information than Moody s. They provide a market-based explanation for this result; regulatory reliance on sluggish 1 The sample in Beaver, Shakespeare, and Soliman (2006) ends in 2002, while EJR was designated as one of the NRSROs at the end of 2007. 3

ratings from NRSROs can benefit regulated investors by allowing them to hold riskier fixedincome securities for higher returns. Different from these studies, in this paper we focus on another angle of the different properties in ratings from the two types of agencies and explicit exploit whether this difference is a result of their compensation structures. Our study is also related to a growing literature on incentive problems of credit rating agencies raised from the issuer-pays model. Relying on historical rating data between 1971 and 1978, Jiang, Stanford, and Xie (2011) examines changes in bond ratings surrounding the date when S&P began to adopt the issuer-pays business model. In a difference-in-difference setting and using Moody s rating for the same bond as a benchmark, the authors find that S&P increases its rating levels once it switches to collecting fees from issuers. While consistent with the results in our paper, their finding leaves an question regarding the persistency of the conflict of interest resulting from the issuer-pays model. As documented in Blume, Lim, and MacKinlay (1998) and Baghai, Servaes, and Tamayo (2011), rating agencies become more conservative and adopt a stricter rating standard over the period 1978-2006. This increased conservatism makes it nontrivial whether S&P s incentive problems is merely a short-term phenomenon immediately following the change of its compensation structure. In this paper, we fill in this gap by showing that this incentive problem is still present in the more recent corporate credit ratings, despite the rising rating standard. Analyzing a sample of collateralized debt obligations (CDO) issued between 1997 and 2007, Griffin and Tang (2009) report that rating agencies frequently made adjustments to their quantitative model that, on average, amounted to a 12% increase in the AAA tranche size. Ahcraft, Goldsmith-Pinkham, and Vickery (2010) document a progressive decline in rating standards for mortgage-backed securities (MBS) between the start of 2005 and mid-2007. He, Qian, and Strahan (2011) provide evidence that Moody s and S&P rewarded large issuers of MBS by granting them unduly favorable ratings during the boom years of 2004 through 2006. Becker and Milbourn (2010) argue that the increased competition from Fitch over the past decade resulted in more issuer-friendly and less informative ratings from S&P and Moody s. Our approach differs 4

from these studies in an important way. Rather than relying on changes in the agencies incentive to issue inflated ratings caused by changes in overall market conditions or in the competitive landscape of the rating industry, we directly compare the ratings of two agencies that follow different business models (issuer-pays model versus investor-pays model), and relate the difference in ratings to issuer-level proxies for the severity of the conflict of interest associated with the issuer-pays model. This approach enables us to provide more direct evidence on rating inflation that is not driven by changes in market and industrial factors but rather the inherent incentive problems resulting from rating agencies own compensation structure. The remainder of this paper is organized as follows. Section 2 provides some institutional background of the credit rating industry. Section 3 describes the data and discusses our empirical methodology. Section 4 presents our empirical results and Section 5 discusses their robustness. Section 6 summarizes our contribution and concludes. 2 Institutional Background The credit rating industry has long been dominated by a handful of companies designated as nationally recognized statistical rating organizations (NRSROs) by the Securities and Exchange Commission (SEC). As of 2002, Standard and Poor s, Moody s, and Fitch were the only rating agencies that were granted the NRSRO status. More recently, the SEC arguably as a result of political pressure and/or concern about concentration in the industry added another seven rating agencies to this group. 2 A majority of these rating agencies follow the issuer-pays business model. An exception is the Egan-Jones Rating Company (EJR). EJR is an independent rating agency founded by 2 Dominion Bond Rating Service (a Canadian rating agency) and A.M. Best (highly regarded for its ratings of insurance companies) received their NRSRO designation in 2003 and 2005, respectively. In 2007, the SEC added two Japanese rating agencies (Japan Credit Rating Agency, Ltd. and Ratings and Investment Information, Inc.) and Egan-Jones Rating Company (EJR). More recently, two other rating agencies, LACE Financial and Realpoint LLC, joined this group. 5

Sean Egan and Bruce Jones that started issuing ratings in December 1995. Since its foundation, EJR has rated more than 1,300 companies in the industrial, financial, and service sectors. According to its rating policy, EJR selects an issuer for a credit analysis generally based on developments within issuers and industries, market developments and requests of subscribers. EJR uses the same credit rating scales as S&P, namely from AAA to D (including the modifiers + and - ) for long-term ratings, and from A-1 to D for short-term ratings. Relying on subscription fees paid by investors, EJR claims that it delivers highly accurate ratings with predictive value for equity, debt, and money market portfolios and has no conflicts of interest. With this aim, it successfully predicted the pitfalls of Enron, WorldCom, and more recently, Lehman Brothers through its credit ratings. Beaver et al. (2006) empirically compare EJR s and Moody s credit ratings, and find that EJR provides more informative ratings than the issuers-paid major rating agencies. As an extension of their findings, we compare the ability of EJR s and S&P s ratings to predict defaults, the most important credit events, as the first step of our tests. We obtain information on issuers default from S&P s issuer rating database. S&P identifies an issuer as in default if it misses a payment of any debt obligations and assigns a D rating to this issuer to indicate its default status. 3 We then calculate the 5-year default rates for each of S&P s rating categories based on the static pool methodology used by S&P. 4 Specifically, at the beginning of each fiscal quarter, we group issuers into different pools based on their S&P ratings (AAA, AA, A, etc.). Each pool is followed from that point forward. The five-year default rates is calculated as the percentage of issuers in each rating category that default within the next five years (20 quarters). We present the default rates next to each of S&P s rating categories on the X-axis in Panel A. Consistent with S&P s rating report (Standard & Poor s (2008)), default is a relatively rare event. For example, the average default rate for issuers with an S&P rating 3 More details about this database is discussed in Section 3. 4 In Figure 1, we focus on defaults at the five-year horizon, even though our results are qualitatively unchanged using the three-year or ten-year horizon. 6

of A or above is only 0.1% in the 5-year horizon. This rate goes up as ratings approach the lower end of D and reaches about 27% for ratings of the CCC category or below. To further compare the informativeness of EJR s and S&P s ratings, we divide issuers within each rating category into two subgroups: (1) those whose EJR ratings are less favorable than S&P s, and (2) those whose EJR ratings are equal to or more favorable than S&P s. We can see in Panel A that conditional on a certain S&P s rating category, issuers with a lower EJR rating are indeed significantly more likely to default at the five-year horizon. For example, in the BB category, issuers with an EJR rating equal to or more favorable than S&P s have an average default rate of 0.71%. In contrast, their counterparts whose EJR ratings are less favorable are over ten times more likely to default at the five-year horizon, with an average default rate of 10.3%. 5 Similar results can be found for all other rating categories. Interestingly, when we sort issuers by their EJR ratings in Panel B, we see a different pattern. First, within each EJR rating category, all the firms have very similar default rates, even if they have different S&P ratings. This is consistent with the expectation that firms with the same rating should have similar credit quality and hence, their default rates should not be significantly different. Second, within each rating category, issuers whose S&P ratings are lower than EJR s have even lower default rates than the other group, contradicting the notion that S&P s rating are more informative than EJR s. These findings are consistent with Beaver, Shakespeare, and Soliman (2006). They further alleviates the concern that EJR may issue ratings that are over conservative (and thus, uninformative) in order to differentiate itself from existing issuer-paid rating agencies to attract business from investors. Overall, the above findings justify our use of EJR s ratings as a reasonable benchmark to identify bias in S&P s ratings. 5 In comparison, the average default rates of issuers with a B rating from S&P is only 8.9% (lower than 10.3%). This suggests that EJR s lower ratings are not just capturing issuers whose credit worthiness falls into the lower end of the distribution of all issuers credit quality within the BB category. 7

3 Sample Selection and Empirical Methodology To construct our rating sample, we first merge two rating databases from EJR and S&P. EJR s issuer credit ratings are collected from Bloomberg and EJR s database via the company s website. EJR keeps its historical rating records back to July 1999. This database contains EJR s issuer ratings in a time series. Each observation is a credit rating corresponding to a certain rating action, including new rating assignments, affirmation, upgrades and downgrades. For each observation, we also have related identification information including company names and rating action dates, which we later use to merge EJR s database to others. EJR s original database covers the period from July 1999 to July 2009, with 23,223 observations representing 2,033 issuers. We eliminate issuers that only obtained a newly assigned rating but had not been followed since then. We also delete observations with a rating of NR, which indicates a rating withdrawal. These two steps reduce the EJR rating sample to 22,816 observations with 1,642 issuers. We obtain S&P s issuer credit ratings from S&P s rating Xpress data services. This database contains detailed information on S&P s credit ratings back to 1920s, including issuers long-term credit ratings and rating Watchlist and Outlook provision. Similar to EJR s rating database, each observation in S&P s rating database is also a credit rating corresponding to a rating action. In the initial database, there are 127,849 observations representing 17,298 private and public issuers globally. We restrict our analyses to U.S. issuers, which leaves us with 72,641 observations from 9,100 private and public issuers. We construct two quarterly panel datasets based on the two rating databases respectively, starting from the third quarter of 1999 to the third quarter of 2009. Following existing literature, we assign numerical values to S&P s and EJR s ratings on notch basis: AAA=1, AA+=2, AA=3, AA-=4, A+=5, A=6, A-=7, BBB+=8, BBB=9, BBB-=10, BB+=11, BB=12, BB-=13, B+=14, B=15, B-=16, CCC+=17, CCC=18, CCC-=19, CC=20, C=21, and D=22. Since both rating databases treat a credit rating with an rating action (rather 8

than a credit rating itself) as an observation, we assign a rating in the current quarter equal to the issuer s rating in the past quarter if no rating action happens. In addition, if two rating actions happen in the same quarter (which means there are two observations in the same quarter), we take the mean of the ratings based on the above numerical conversion. We then merge these two panel datasets by manually matching company names and year-quarter information. We successfully merged 1,574 out of 1,642 issuers from EJR s rating database. Since we are interested in issuers financial activities in our analyses, we restrict our sample to non-financial and non-utility issuers. Issuers financial information are obtained from COMPUSTAT quarterly database. Since our tests are based on the comparison of S&P s ratings to EJR s, we delete any issuer-quarter observations if either of the two ratings is not available. We further delete any observations if the issuer is rated as default by either S&P or EJR due to these issuers potential abnormal financing activities and ongoing restructuring. 6 Following these criteria, our primary sample consists of 25,713 issuer-quarter observations representing 964 issuers. Panel A of Table 1 presents descriptive statistics of the primary sample consisting of issuers rated by both S&P and EJR in Column (2). In comparison, Column (1) presents summary statistics for all non-financial and non-utility public U.S. issuers that are rated by S&P in the sample period. From Panel A, we can see that issuers that are rated by both rating agencies are, on average, larger than all the issuers rated by S&P. For example, issuers in our rating sample have a mean capitalization of 10,591 million U.S. dollars (median 2,798), compared to the mean of 7,541 (median 1,516) million for the other group. Similarly, the mean total assets of issuers rated by both agencies are 11,172 million (median 3,701), while the average assets of firms in the other group are 8,873 (median 2,013) million. These differences between the two groups of firms are highly significant at 1% level. In addition, issuers rated by both rating agencies have lower leverage, higher Altman s Z-Score and higher 6 This criterion is not crucial. All our results still hold when we include issuers rated as default by S&P or by EJR. 9

ROA. This evidence suggests that these issuers appear to be less risky and more productive than their counterparts. However, their higher Market-to-Book, lower Tangibility and higher R&D/Sales indicate that these issuers tend to invest more heavily on R&D to accommodate their higher growth opportunity and are possibly more difficult to evaluate due to their low proportion of fixed assets. As expected, these issuers ratings are more likely to be requested by EJR s client base. To associate the difference between S&P s and EJR s ratings with the severity of S&P s conflict of interest at the issuer level, we generate our left-hand-side variable to capture the rating difference. Following Jiang et al. (2011), we calculate the cardinal different between the two rating agencies ratings (EJR s rating minus S&P s rating) based on previous numerical conversion. For example, the rating difference equals 2 for a difference between AA+ and AA-, and equals 3 for a difference between AA+ and A+. A higher value of the rating difference variable indicates that compared to EJR, S&P issues a more favorable rating to the issuer. As a robustness check, we also employ alternative approaches to define rating difference. For example, we convert each S&P s and EJR s rating to the corresponding five-year default rates calculated using EJR s ratings in our sample. We then generate the rating difference variable as the difference between the default rates corresponding to the two agencies ratings. Alternatively, we also define a dummy variable that equals 1 if S&P s rating is higher than EJR s for a given issuer at a certain point of time, and equals 0 otherwise. More discussions and the results using these alternative definitions are presented in the Robustness Section. Panel B summarizes the difference between S&P s and EJR s ratings. It is worth noting that this difference is significantly different from 0, lending support to the notion that on average, S&P assigns more favorable ratings than EJR. The magnitude of the difference is also economically significant. For example, the mean of the difference is 0.195, suggesting that for an average issuer, EJR s ratings is about 0.2 notch lower than S&P s. Equivalently, S&P issues a rating that is one notch step-up from EJR s for approximately every one of out 10

five firms. 4 Empirical Results In this section, we discuss our main tests on the link between rating inflation and the issuerpays compensation model. We present our test results based on three proxies for the severity of S&P s conflict of interest: issuers short-term liquidity needs, S&P s revenue share and issuers management turnover. 4.1 Issuer s Short-term Liquidity Needs We start by examining the association between rating inflation and issuers amount of shortterm debt. Issuers that are exposed to a large amount of short-term debt are likely to issue new debt in the future to replace their expiring debt. Hence, they are more likely bring new rating business to the agency. In this case, the rating agency will be tempted to issue a favorable rating to obtain these clients lucrative future business. Table 2 presents the results of OLS regression models. We first examine Specification (1) where we include the logarithm of the issuer s total short-term debt amount as the independent variable. To capture any changes in rating standards over time as suggested in Blume, Lim, and MacKinlay (1998) and Baghai, Servaes, and Tamayo (2011), we also include year dummies in this specification. The positive coefficient on Ln(Short-term debt) confirms that S&P is more likely to issue favorable ratings when issuers have higher short-term liquidity needs. Issuers short-term debt volume may be correlated with the amount of their long-term debt, which in turn, captures issuers past relationship with the rating agency (Covitz and Harrison (2003)). To isolate the effect of this past relationship from the short-term debt variable, in Specification (2) we include the logarithm of the issuer s long-term debt amount. We can see that while issuers past relationship with the agency also plays a significant role in determining rating inflation, the effect of future business opportunities captured by the 11

amount of short-term debt remains significant at the 1% level. To further check the robustness of our results, we include additional issuer characteristics as controls in Specification (3), including the logarithm of Sales, Tangibility, R&D Expense/Sales (and R&D Missing Dummy), and Market-to-Book. Further more, one concern over our dependent variable is that by construction, the difference between S&P s and EJR s ratings will be larger when S&P s (EJR s) ratings are closer to (further from) the higher end (AAA) of the rating spectrum. In this case, if the right-hand-side variables we have included in the models happen to capture the relative positions of issuers ratings along the rating spectrum rather than the true factors that affect rating inflation, our results are biased. To address this concern, we generate dummy variables for S&P s rating categories on letter basis (AAA, AA, A, BBB, BB, B, CCC and CC/C) and include them as additional controls. The results in Specification (3) are consistent with previous specifications, confirming the positive relation between issuers short-term debt and rating inflation. The economic significance of this effect is sizable. One standard deviation increase of Ln(Short-term Debt) (2.44) is associated with 0.16 notch step-up in S&P s ratings compared to EJR s, amounting to 84% of the average rating difference in our sample (0.195 notch). In a study by Ederington and Goh (1998) that examines the value of information provided by stock analysts and rating agencies, they find that both agents provide new information to the market and that Granger causality of this information flows both ways. Inspired by this finding, we examine the relation between stock analysts information and rating inflation. More specifically, we include two variables, Number of Analysts and Standard Deviation of Analysts Reports (on EPS) in our model. We obtain this information from I/B/E/S monthly summary database and use the values of the two variables in the last month of each quarter for our issuer-quarter sample. The estimation is presented in Specification (4). S&P tends to issue less inflated ratings if an issuer is followed by more stock analysts, but is more likely to do so if their opinions are more dispersed. This finding indicates that rating agencies tendency to issue inflated rating may be constrained by other information providers. It 12

also implies that stock analysts in the equity market may have a disciplinary role on rating agencies behavior in the credit market. One limitation of the models so far is that they do not control for unobservable characteristics of issuers that may be correlated with S&P s incentives to issue high ratings. To address this concern, we rectify our model by including issuer fixed effects. This model is estimated in Specification (5). We can see that the effects of the issuer s short-term debt remians highly significant at the 1% level. The economic significance is also comparable to previous specifications. This result further confirms the positive relationship between rating inflation and issuers short-term liquidity needs. 4.2 The Rating Agency s Revenue Share We now turn to our tests using S&P s revenue share as the second proxy for the severity of its conflict of interests. While many issuers obtain more than one credit rating from major rating agencies, only fewer than ten percent of investors are required to hold securities from issuers with two or more ratings (Baker and Mansi (2001)). Therefore, major rating agencies face competition from each other in their rating business. Becker and Milbourn (2010) find that as S&P and Moody s face more competition from Fitch, measured by Fitch s market share, they produce more issuer-friendly and less informative ratings. In similar spirit of this study, we expect that S&P will be tempted to provide favorable ratings to an issuer when it senses the threat of losing it as its future client. The extent of this threat is positively correlated with the competition S&P faces, measured by S&P s revenue share on a per-client basis. More specifically, we define Fraction of Bond Issue Volume Rated by S&P in the Past n Quarters as the amount of bonds issued by an issuer during the past n quarters that are rated by S&P as a fraction of those that are rated by the three major rating agencies (S&P, Moody s and Fitch) in total. A lower fraction then indicates the issuer s intention to seek alternative ratings from S&P s competitors, and hence, stronger competition for S&P. We trace each issuer s debt-issuance activity on quarterly basis back to the past two years 13

at each time point and calculate revenue share based on past 2, 4, 6, and 8 quarters, respectively. Issuers debt issuance information is obtained from The Fixed Investment Securities Database (FISD). This database provides key characteristics on almost all publicly traded bond issuance. We merged this database to our primary rating sample using issuers 6-digit CUSIPs. Since firms only issue debt periodically, we omit an issuer-quarter observation if the issuer has not had any bond issues during the past n quarters, where n corresponds to different revenue share windows. Figure 2 plots S&P s revenue share in the past 4 and 8 quarters on quarterly basis from 1999 to 2007. We exclude 2008 and 2009 because of the abnormally small amount of bond issue during the financial crisis. In comparison, we also include the revenue share measure similar to that in Becker and Milbourn (2010), which is defined as the number of bonds issued by an issuer during the past n quarters that are rated by S&P as a fraction of those that are rated by the three major rating agencies (S&P, Moody s and Fitch) in total. A few observations are worth noting. First, the two measures of revenue share move closely with each other and are close to 50% between 1999 and 2003. Second, consistent with the finding in Becker and Milbourn (2010), there is an apparent decline in S&P s revenue share starting from the second half of 2003. These features arise from the fact that many issuers obtain two ratings from both S&P and Moody s before 2003 when competition in the rating industry was mostly limited to the two major rating agencies and Fitch s market share was relatively small. Starting from 2005, however, Fitch has been playing an growing role in the rating industry because of its inclusion as a rater to Lehman Brothers Aggregate U.S. Bond index (now Barclays Capital Aggregate Bond Index). This change shifts S&P s revenue share from close to 50% to around 33% on the average level. Table 3 presents the issuer-fixed-effect OLS regression analyses. In all specifications, we include year dummies to control for the increase in Fitch s market share over time as well as other time-related unobservable variables as in Table 2. 7 Consistent with our hypothesis, 7 We also check the robustness of our results by explicitly controlling for Fitch s market share in our models. The results are qualitatively similar. 14

Table 3 shows that issuers are more likely to receive a higher rating from S&P when its revenue share is lower. For example, using the past-4-quarter window, the coefficient on Fraction of Bond Issues Volume Rated by S&P is -0.318 and is significantly at 5% level. This means that one standard deviation decrease in S&P s revenue share in the past 4 quatres (0.27) will lead S&P to issue a rating about 0.09 notch higher than EJR s, amounting to 44% change of the sample mean of rating difference. Similar magnitude is found in specifications using revenue shares based on the past 2 and past 6 quatres. 8 In Specification (4), the coefficient of the past-8-quarter revenue share is not significant at the 10% level, indicating that the effect of revenue share on the agency s rating strategy may be more concentrated in shorter terms. Overall, the negative relation between S&P s revenue share and its tendency to issue inflated ratings again reveals S&P s conflict of interest due to the issuer-pays business model. 4.3 Issuers Management Turnover As our final measure of the severity of S&P s conflict of interests, we explore issuers management turnover. A firm s CEO and CFO play an important role in the rating process (Graham and Harvey (2001), Kisgen (2006), Kisgen (2007), Kisgen (2009) and Norris (2009)). CFOs and CEOs usually determines which rating agency to request a rating from, and are also actively involved in the credit rating process (Fight (2001)). Since newly appointed corporate leaders are more inclined to change the firm s operational and financial strategies (Giambatista, Rowe, and Riaz (2005)), we hypothesize that issuers new management is more likely to switch rating agencies. Hence, S&P is tempted to issue an inflated rating following the appointment of a new CFO or CEO in order to build a good relationship with the new management and generate more future business. To examine this hypothesis, we obtain CEO and CFO information from COMPUSTAT 8 The coefficients of other variables are similar to those in Table 2. One difference is that the coefficients of Ln(Short-term Debt) now become insignificant. This is because by construction, our sample in Table 3 only consists of firms that have recently issued bonds in the past few quarters. This, in turn, may have diluted the liquidity needs effect Ln(Short-term Debt) is supposed to capture, and hence, reduces the magnitude of its coefficient. 15

EXECUCOMP annual database. We identify issuers CEOs using data item CEOANN, where a CEO is identified if CEOANN=CEO. Following Gopalan, Song, and Yerramilli (2010), we identify CFOs based on managers titles from data item TITLEANN. A CFO is identified if a manager s title contains: CFO, chief financial officer, finance, treasurer, VP-finance or a combination of two or more of them. We identify that a new CFO (CEO) is assigned if an issuer s current CFO (CEO) is different from the one in the last fiscal year. To be consistent with EXECUCOMP annually based data, we aggregate the difference between S&P s and EJR s ratings to annual level by taking the mean of its values in the four quarters during each fiscal year. We further restrict our analysis to issuer-year observations where information on both CEO and CFO is available. Table 4 presents the results of the issuer-fixed-effect OLS regression models. We see that in Specification (1), there is a significant boost in rating inflation in the year when a new CFO is appointed (new CFO (t)) and in the following year (new CFO (t-1)). The appointment of a new CFO is associated with a total of 0.29 notch (0.18+0.11) increase in the rating difference during these two years. In Specification (2), while the coefficients of new CEO (t) and new CEO (t-1) are also positive, their significance levels are lower. This result indicates that CFOs seem to have a bigger impact than CEOs in affecting the rating agencies strategies. This evidence is consistent with prior studies that find CFOs are more influential in certain areas related to the management of an issuer s financial system because of their ultimate responsibility in those areas (Mian (2001), Geiger and North (2006) and Jiang, Petroni, and Wang (2008)). Specification (3) includes both CFO and CEO appointment dummies in the regression. The coefficients on the CFO dummies are comparable to those in Specification (1), suggesting that the effects of a new CFO are not likely to be driven by concurrent CEO changes. This result lends further support to the notion that issuers new management generates incentives for the rating agency to build a good relationship through an inflated rating. 16

4.4 The Information Value of Credit Ratings The results so far indicate that the issuer-pays rating model leads to a significant amount of rating inflation. However, it is not clear whether inflated ratings have any welfare implications. If investors anticipate bias in S&P s ratings and correctly adjust for such bias, rating inflation is not likely to lead to misallocation of resources. To explore investors knowledge about rating bias, we examine the association between issuers bond yield spreads and the proxies for the severity of S&P s conflict of interest. If investors adjust for rating inflation, we expect to see a significant relation between yield spreads and these proxies, after controlling for S&P s contemporary credit ratings. Following Beaver et al. (2006), we obtain information on Treasury Spread for new bond issuance, defined as the difference between the issue s offering yield and the yield on a benchmark treasury security (a U.S. treasury bond) with similar duration and maturity from the Fixed Investment Securities Database (FISD). Since major rating agencies define an issuer s credit rating based on the issuer s ability to pay back its senior unsecured bonds, we restrict our sample to the issuance of senior unsecured bonds during the sample period. In addition, we exclude bonds that are callable, puttable, convertible, exchangeable, with sinking fund or with refund protection. We also generate the following financial variables for each issue as control variables: Enhancement is a dummy variable that equals 1 if the issue has credit enhancements; Covenants is a dummy variable that equals 1 if the debt issue contains covenants in the contract; Ln(Bond Issue Amount) is the logarithm of the par value of the debt issue in millions of dollars; Maturity in Years is the number of years to maturity of the debt. We then regress Treasury Spread on our three proxies for the severity of S&P s conflict of interest, controlling for S&P s issuer credit ratings. The results are presented in Table 5. Notice that most variables that can predict rating inflation show up insignificantly (certain variables show opposite signs as expected). One exception is Ln(Short-term Debt). In 17

Specification (1), its coefficient is positively significant at 5% level. 9 However, its economic significance is relatively small. One standard deviation increase in Ln(Short-term Debt) is only associated with 7.6% increase in the yield spread for an average firm. This change is marginal compared to the impact of the issuer s short-term debt on rating inflation. Furthermore, the effects of short-term debt becomes insignificant in other specifications where we control for additional proxies. Overall, we can not reject the null hypothesis that investors do not adequately adjust for the potential rating bias. These results suggest that investors may not well understand the information value of credit ratings, and hence, rating inflation can lead to misallocation of resources. These results further imply that regulators intervention in the credit rating industry would be beneficial to investors who use credit ratings to guide their investment decisions, and can potentially increase social welfare. 5 Robustness 5.1 Adjusted and Broader Rating Categories One concern over the tests so far is that S&P s ratings are usually through-the-cycle and may not necessarily reflect the agency s opinions on issuers short-term credit worthiness. This implies that compared to EJR s ratings, S&P s ratings tend to be more forward-looking and more stable. Our previous results may capture the difference in the nature of the two rating agencies rating technology rather than rating inflation. To resolve this concern, we take into account S&P s watchlist and outlook provision. These two rating actions, by definition, reflect information in a more timely manner and are usually treated as a refinement of long-term credit ratings. We adjust S&P s long-term ratings downwards (closer to the lower end of the rating spectrum) by one notch if S&P have put an issuer on negative watchlist or outlook, and upwards (closer to AAA ) by one notch if S&P have put an issuer on positive watchlist or 9 This result is consistent with Gopalan, Song, and Yerramilli (2010). In their paper, they find that bonds issued by firms with a higher proportion of short-term debt trade at higher yield spreads, even after controlling for issuers credit ratings. 18

outlook. We then recalculate the rating difference variable based on S&P s ratings adjusted for watchlist and outlook provision. Another concern over our results is that previous tests utilize rating categories on notch basis that takes into account rating modifiers ( + and - ). One may argue that this notch-based definition may capture rating differences that are too small in magnitude to be economically meaningful. As a robustness check, we suppress rating modifiers and re-define rating differences on letter basis. More specifically, a rating of AA+ is now considered the same as a rating of AA or AA-, and is more favorable than a rating of A+, A or A-, which belongs to the same A letter category. We then recalculate the difference between S&P s and EJR s ratings using a similar numerical conversion as before: AAA=1, AA=2, A=3, BBB=4, BB=5, B=6, CCC=7, CC=8, C=9 and D=10. Using these adjusted ratings, we re-estimated selected specifications from Table 2, Table 3, and Table 4. The estimation results are presented in Table 6. We can see that the coefficients of all of the key variables remain significant, and all of them are of the correct sign. This evidence confirms that our previous results are robust to watchlist/outlook-adjusted ratings and letter-based rating categories. 5.2 Endogeneity Issuers amount of debt is endogenously determined. Issuers who obtain an inflated rating may want to take advantage of the lower cost of capital and issue more debt. This raises the concern that our previous tests in Table 2 may be subject to a reverse causality problem. To resolve this endogeneity problem, we replace Ln(Short-term Debt) with a new variable Ln(Long-term Debt Due). Ln(Long-term Debt Due) is defined as the logarithm of the amount of long-term debt that is due within one year. Similar to Ln(Short-term Debt), this variable also measures how much future business an issuer is likely to bring to the rating agency. However, since the repayment schedule of long-term debt have already been determined many years in the past, it is not likely to be affected by the rating agency s current ratings. In 19

other words, Ln(Long-term Debt Due) is less likely to be subject to the endogeneity problem. We repeat the estimations in Table 2 using Ln(Long-term Debt Due) as the main independent variable. Table 7 reports the results. We can see that the coefficient on Ln(Long-term Debt Due) is positive and significant, confirming that the endogenous choice of debt is not likely to drive the results in Table 2. 5.3 Rating Shopping One may argue that some issuer characteristics used in our previous tests such as the amount of short-term debt may also capture issuers engagement in rating shopping. Rating shopping refers to the situation where an issuer approaches different rating agencies for preliminary ratings and then cherry-pick the most favorable one to disclose to the public, while conceal the lower ratings. Since rating agencies signals on issuers credit quality are noisy, the final observed rating tend to be higher if an issuer has shopped for ratings, even though none of the agencies has overstated the issuer s credit worthiness. In this case, if certain issuer characteristics happened to capture issuers involvement of rating shopping, which in turn, leads to a higher rating, our results are biased. To address this concern, we explicitly control for issuers rating shopping activities in our previous tests. Notice that in our second sets of tests where we use S&P s revenue share as a proxy for the severity of S&P s conflict of interest, rating shopping is not likely to explain our results. If the lower S&P s revenue share just captures firms tendency to shop with other rating agencies for better ratings, then we would expect to see S&P s ratings to be lower (relative to EJR s) when its revenue share is low, compared to the case when S&P s revenue share is high. This is because based on the rating shopping argument, a high revenue share suggests that S&P s ratings must be sufficiently high to make issuers less inclined to approach S&P s competitors for more favorable ratings. On the other hand, a low revenue share should be associated with lower ratings from S&P since this is the case when issuers are most likely to shop for better ratings. This intuition, however, is opposite to what we 20

found in Table 3. Similar argument can also be found in Becker and Milbourn (2011). Given this intuition, we focus our robustness checks on the tests based on issuers short-term debt volume and their management turnover. Following the definition of rating shopping, we define a Rating Shopping dummy equal to 0 if an issuer has three published ratings from S&P, Moody s and Fitch, and equal to 1 if it only has one published rating from S&P. This definition relies on the assumption that an issuer is not likely to have picked the highest rating if it discloses all the ratings from agencies they have approached (we only consider the three major rating agencies). Using the bond-rating information from FISD database, we identify an issuer to have a published issuer credit rating from Moody s (Fitch) if one of this issuer s outstanding senior unsecured bonds is rated by Moody s (Fitch) at that time. This approach is based on the fact major rating agencies provide an issuer credit rating for every borrower for which it rates any security. This approach generates comparable statistics as previous studies. For example, in our sample, over 95% of issuers obtain issuer s ratings from both S&P and Moody s, and about 60% of issuers obtain a third rating from Fitch, consistent with Bongaerts, Cremers, and Goetzmann (2010). The results controlling for rating shopping are presented in Table 8. As expected, the rating shopping dummy shows up positively, indicating that S&P indeed issues a more favorable rating if issuers have shopped for ratings. The coefficient of Rating Shopping, however, is not significant at the 10% level. More importantly, we can see from these specifications that all our previous results still hold with similar economic magnitude. These results confirms that the relation between rating inflation and the issuer-pays model we uncover throughout the paper is not driven by issuers endogenous selection choices. 5.4 Alternative Definitions of the Dependent Variable To further check the robustness of our results, we use alternative definitions for the dependent variable. First, we noticed that even though the differences between S&P s and EJR s ratings 21