VIEWS: 14 PAGES: 26 POSTED ON: 10/19/2012 Public Domain
The Journal of Credit Risk (95–119) Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies Andre Güttler European Business School, International University Schloss Reichartshausen, Department of Finance, Accounting and Real Estate, 65375 Oestrich-Winkel, Germany; email: andre.guettler@ebs.edu Using a sample of corporate issuers rated by Moody’s and Standard & Poor’s (S&P) for the years 1994–2005, we ﬁnd evidence that Moody’s rating migration rates are higher given a rating change by S&P. This seems to be tentative evidence that S&P assigns ratings in a timelier manner than Moody’s. Furthermore, we ﬁnd that the tendency toward rating convergence is stronger for Moody’s than for S&P. Our ﬁndings are important given the concerns regarding the agencies’ inherent incentives and their dominant market position. 1 INTRODUCTION The international market for debt securities has experienced tremendous growth. At the end of 2007 the amount outstanding – for ﬁnancial institutions and corporate issuers only – exceeded US$19 trillion (BIS (2007)). Because credit analysis is costly to dispersed investors, it is more efﬁcient to centralize credit analysis in a single external credit rating agency (CRA), or at least in a small number of CRAs. Thus, in virtually all cases, the creditworthiness of debt securities is assessed and made public by external CRAs. In the international market for debt securities, Moody’s and Standard & Poor’s (S&P)1 together command a market share of 80%, on average (Löfﬂer (2007)), while A previous version of this paper was entitled “Conditional Rating Transitions: The Case of S&P and Moody’s”. We thank two anonymous referees, Patrick Behr, Albert L. Chun, Axel Eisenkopf, Michael Kunisch, Peter Raupach, Sascha Steffen and seminar participants at the Deutsche Bun- desbank, the annual meeting of the Swiss Society for Financial Market Research and the annual meeting of the German Finance Association for their helpful comments. All errors and opinions expressed in this paper are, of course, our own. Financial support by the E-Finance Lab is gratefully acknowledged. 1 Throughout the paper we refer to the two CRAs in alphabetical order. The chosen order has no implications with regard to the timeliness of the respective CRA’s rating assessments or the level of rating convergence. 95 96 A. Güttler their combined speciﬁc coverage differs across regions and sectors. Obtaining a rating from both of the leading CRAs is standard in most rating markets, particularly in the US. Other CRAs such as Fitch Ratings or Dominion Bond Rating Service (DBRS) concentrate on particular business sectors or certain regions. Thus, Moody’s and S&P are the main providers of external credit risk assessments in most markets, and therefore the timeliness of the two dominant CRAs’ ratings and their tendency toward rating convergence are particularly important to market participants. Given the outstanding importance of credit ratings and the oligopolistic market structure, the CRAs’ business model has been viewed with increasing skepticism in recent years. Most CRAs charge companies that request credit ratings signiﬁcant rating fees. Although the CRAs also offer paid research services, rating fees constitute the main part of their revenues.2 This could give rise to a conﬂict of interest, since CRAs might have a ﬁnancial incentive to assign ratings that are favorable to issuers.3 The outcome might therefore diverge from the CRAs’ proper function of assigning objective risk assessments to market participants. Yet, the CRAs’ potential conﬂict of interest would seem to be resolved by reputa- tional arguments (Ramakrishnan and Thakor (1984) and Millon and Thakor (1985)). In addition, a survey among rated issuers reveals that only 2.7% of them would agree with the assertion that the CRAs’ current business model encourages agencies to assign high ratings in order to satisfy issuers (Baker and Mansi (2002)). Thus, users trust the credit ratings because of the CRAs’ longstanding reputation in assessing a huge number of debt securities over time. CRAs also promote their reputation by releasing reports detailing their rating performance. These usually demonstrate that historical default frequencies rise with worsening rating levels. More recently, rating agencies also provide more information regarding rating accuracy and rating stability. For example, Moody’s (2010) reports a one-year (ﬁve-year) accuracy ratio of 83% (65%), an average rating prior to default of B2/B3, and a (large) rating action rate of 31.1% (8.5%) for corporate ratings. In addition, empirical results brought forward by Covitz and Harrison (2003) and Löfﬂer (2007) seem to conﬁrm that reputation predominates over conﬂicting ﬁnancial interests at leading CRAs. However, despite the theoretical and empirical support for the CRAs, in the wake of the ﬁnancial crisis that started in 2007, criticism is on the rise. Outside structured ﬁnance, their failure to foresee the crises of companies such as Enron, WorldCom or 2 Historically, the main objective of the rating agencies has been to provide rating information to investors who paid for these services. 3 Cantor and Packer (1997) ﬁnd that this incentive seems to be particularly strong for smaller rating agencies. However, Jewell and Livingston (1999) and Feinberg et al (2004) show that those rating agencies (ie, other than Moody’s and S&P) provide incremental information by their ratings. The Journal of Credit Risk Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies 97 Parmalat raised concerns about the timeliness of credit rating adjustments (Economist (2005)). Although one has to admit that credit risk events related to fraud are very difﬁcult to assess, Enron’s default, involving an outstanding nominal amount of over US$20 billion, is often held against the leading CRAs. Regardless of the fact that most of the negative information was already known – as is best reﬂected in the stock price collapse from over US$100 to US$3 – Moody’s and S&P’s downgrading to junk status came very late. Two additional arguments might help to explain the mixed evidence regarding the reputational incentive in the rating business. First, outcomes relating to the perfor- mance of stock analysts are stressed. Among others, Trueman (1994) demonstrates that “the likelihood that the analyst releases a forecast similar to those previously announced by other analysts is greater than could be justiﬁed by his own informa- tion.” The results indicate that analysts try to favorably affect investors’ assessments of their own forecasting ability.4 Second, a CRA faces a trade-off between reputa- tional and ﬁnancial capital,5 as described by Boot et al (1993). In order to maximize its reputational capital, the CRA has to make the most of the rating’s value for both issuers and investors by minimizing the difference between the assigned rating and the ﬁrm’s true quality. However, a CRA also has to stand up to competition from other agencies (Bannier and Tyrell (2006)). Thus, to maximize its ﬁnancial capital, a CRA may announce a rating that takes into account the potential effects of competition, ie, other CRAs’ credit ratings, while not necessarily reﬂecting the ﬁrm’s default risk.6 Following these theoretical arguments, CRAs might be inclined to take a very con- servative approach toward the risk of announcing incorrect rating assessments, and can therefore be expected to try to match as closely as possible a competitor’s rating (or the competitors’ average rating). This conservative line minimizes the likelihood that outside observers will be able to recognize inaccurate ratings. In this study we use rating information from Moody’s and S&P for the observa- tion period January 1994–December 2005. This includes rating changes as well as Watchlist additions. We use survival analysis to estimate conditional rating change intensities,7 ie, we analyze how promptly rating changes by one CRA occur given the other CRA’s rating action or rating level. First, we assess differences in the time- liness of the two CRAs’ rating changes. This is done by estimating rating intensities 4 However, conﬂicting empirical evidence against herding among professional ﬁnancial analysts is presented by Bernhardt et al (2006). 5 Financial capital signiﬁes (short-term) increases in proﬁts that might arise due to increased rating fees or to decreased (personnel) costs. 6 In practice, rating agencies aim to minimize these incentives by not discussing competitors’ rating in rating committee meetings. 7 Rating change intensities can also be termed rating migration rates. Technical Report www.thejournalofcreditrisk.com 98 A. Güttler for downgrades (upgrades) by each CRA conditional on whether there was a down- grade or negative Watchlist addition (upgrade or positive Watchlist addition) by the other CRA. Second, we examine whether the rating intensity is higher in the case of downgrades (upgrades) by one CRA if the rating level by the other CRA was lower (higher). This is done to provide evidence for a rating convergence to the other CRA’s rating level. We accomplish a relative comparison of Moody’s and S&P’s rating intensities with our approach. It allows us to judge which of the two CRAs reacts in a timelier manner, and which CRA is less likely to engage in rating convergence. This restriction to a relative comparison is a consequence of the fact that rating changes are not random events. The ﬁrst CRA made a rating change presumably because it received public or private information that changed its risk assessment sufﬁciently to justify the observed rating change. Assuming the second CRA receives the same information8 and that it employs a comparable rating process,9 one would expect it, on average, to announce a similar rating change.10 Therefore, we do not provide an absolute comparison of the two CRAs’ ratings against an external benchmark such as a market-based credit risk measure. Rather, we assess relative differences between Moody’s and S&P regarding timeliness and rating convergence. In this paper we deliver two main results: First, we ﬁnd evidence that Moody’s rating change intensities are higher given a rating change by S&P. This effect is somewhat stronger in the case of downgrades than for upgrades and seems to be tentative evi- dence that S&P assigns ratings in a timelier manner than Moody’s. Second, we ﬁnd that the tendency toward rating convergence is stronger for Moody’s than for S&P. The respective increase in upgrade intensities is on average 38% larger for Moody’s when we observe a better rating by S&P than vice versa. Both results are novel. In this paper we complement two streams of literature. First, our results extend the evidence regarding the relative informational value of Watchlist additions versus rating changes. Norden and Weber (2004) show that negative Watchlist additions by Moody’s and S&P exhibit signiﬁcant abnormal performance in stock and CDS markets, whereas actual downgrades do not. Using a market price expectations model, Hand et al (1992) ﬁnd that negative and positive Watchlist additions by both CRAs are associated with stronger abnormal bond and stock price effects than in the case 8 Even if CRAs receive the same information about an issuer they may not necessarily assign the same ratings since the interpretation of information could be different. 9 Both CRAs employ a long-term rating approach that is meant to look through the business cycle. However, Moody’s uses an expected loss approach that incorporates the probability of default and the loss severity, whereas S&P only takes account of the probability of default. 10 Analysis of split ratings reveals that most of Moody’s and S&P’s respective ratings are the same or just one notch apart (Cantor and Packer (1997)). The Journal of Credit Risk Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies 99 of actual rating changes. In contrast, we ﬁnd no evidence that Watchlist additions by one CRA increase the rating intensities of the other CRA. Given the close relationship between rating intensities and rating transition prob- abilities, our ﬁndings also relate to research into the conditional factors of transition matrices. In this line of literature it has been shown that rating transition probabilities are inﬂuenced by several factors that are not included in the standard Markov setup. These factors include previous rating changes by the analyzed CRA (Altman and Kao (1992); Lando and Skødeberg (2002); Christensen et al (2004); Güttler and Raupach (2010); and Hamilton and Cantor (2004)), the business cycle (Nickell et al (2000) and Bangia et al (2002)), latent systematic factors (Wendin and McNeil (2006)), and a bond’s age (Altman and Kao (1992); Kavvthas (2001); and Figlewski et al (2008)). We contribute to these works by identifying additional rating information, ie, rating changes by another CRA and rating level differences between the two CRAs, as being predictors of conditional rating transition probabilities of a CRA. The remainder of the paper is organized as follows. Section 2 provides a description of our dataset. Section 3 presents a short overview of the survival analysis methodol- ogy. Section 4 is devoted to the empirical results, while the last section concludes. 2 DATASET DESCRIPTION This study covers the period January 1994–December 2005. We use rating changes as well as positive and negative Watchlist entries by Moody’s and S&P from Bloomberg.11 Watchlist entries are introduced by the two CRAs to provide investors with timelier information about the issuers’ credit risk. For instance, Steiner and Heinke (2001) provide evidence that Watchlist additions are anticipated to a lesser extent than rating changes. We therefore also take account of Watchlist entries in addition to rating changes. Furthermore, we use withdrawn ratings as a signal that the rating relationship between CRA and issuer was terminated. Given the broad range of different ratings for a given obligor, we construct a single rating history for the long-term, foreign currency, senior unsecured ratings of each issuer. We make use of a mapped numerical rating scale ranging from 1 (AAA/Aaa) to 19 (CC/Ca) throughout, ie, we assume that S&P’s AAA corresponds to Moody’s Aaa, etc. Despite existing differences between the CRAs’ rating method- ologies, this mapping procedure is widely used in academic practice (Morgan (2002)), and by regulatory authorities. The SEC’s important investment/noninvestment bound- ary BBB /BBC for S&P and Baa3/Ba1 for Moody’s is one example among others. 11 Wedisregard neutral Watchlist entries. First, the direction and therefore the meaning of these Watchlist additions are unclear. Second, these Watchlist entries are rarely made by both CRAs. Technical Report www.thejournalofcreditrisk.com 100 A. Güttler TABLE 1 Regional and sectoral distribution. (a) Regional distribution Rating changes ‚ …„ ƒ All companies Moody’s S&P ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ North America 3282 0.701 4754 0.703 4716 0.681 Europe 695 0.149 902 0.133 901 0.130 Asia 415 0.089 605 0.089 717 0.103 Latin/Central America 174 0.037 368 0.054 405 0.058 Other 114 0.024 134 0.020 190 0.027 (b) Sectoral distribution Rating changes ‚ …„ ƒ All companies Moody’s S&P ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ Other nonﬁnancial 3544 0.757 5286 0.782 5359 0.773 Utility 63 0.013 65 0.010 47 0.007 Bank 667 0.143 889 0.131 961 0.139 Other ﬁnancial 272 0.058 319 0.047 345 0.050 Insurance 134 0.029 204 0.030 217 0.031 The table provides an overview of companies that held long-term senior unsecured credit ratings by Moody’s and S&P in the observation period January 1994–December 2005. Rating data is extracted from Bloomberg. Columns 1 and 2 show the regional and sectoral distribution of 4680 sample companies which were assigned a credit rating by both CRAs for at least one day during the observation period 1994–2005. Columns 3 and 4 (5 and 6) exhibit the respective distribution of rating changes that have been assigned by Moody’s (S&P). We constrain the sample to those rating changes which have also been rated by the second CRA during that time. In part (a), “Other” comprises Africa, the Arabian Peninsula and the Paciﬁc region. In part (b), the category “Other ﬁnancial” includes asset managers, brokerage ﬁrms and the like. To investigate our two research questions we start with 4680 issuers which were rated by both Moody’s and S&P in the observation period January 1994–December 2005. Our criterion for including companies in our analysis is that both CRAs rated a company for at least one day in our observation period. In addition to the rating history of January 1994–December 2005, we also use rating information from January 1990–December 1993 to condition for previous rating actions. The starting point of our observation period is determined by our data source, Bloomberg, which begins at this point in time. Our results in columns 1 and 2 of part (a) of Table 1 provide the regional distribution at the company level. 70% of our sample companies are from North America, 15% The Journal of Credit Risk Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies 101 from Europe, 9% from Asia and 4% from Latin and Central America. The remainder of the sample is distributed over Africa, the Arabian Peninsula and the Paciﬁc region. Columns 3 and 4 (5 and 6) provide results on the rating level for Moody’s (S&P), which are very similar to the company level. For rating changes by either CRA we require valid rating observations by the other CRA for the time from the last rating change to the point in time of the rating action of interest. This enables us to use time-varying covariates, ie, rating information, from the other CRA in the survival analysis. By using an international sample of bond rating changes we are able to increase our sample size considerably. Additionally, our results should be relevant not only for single markets, such as the US, but also for all other rating markets. Potential differences in rating methodologies and the meaning of ratings across countries are controlled for by our econometric setup (see Section 3). The sample’s sectoral distribution is shown in columns 1 and 2 in part (b) of Table 1 on the facing page on the company level. With a share of 76%, the nonﬁnancial and nonutility companies represent the vast majority. Banks account for 14% and insurance companies for an additional 3%. The remainder is distributed among utilities (1%) and other ﬁnancial ﬁrms (6%). Columns 3 and 4 (5 and 6) provide results on the rating level for Moody’s (S&P) which are very similar to the company level. To enlarge the sample and ensure that it is representative of the population of rated ﬁrms, our analysis is not restricted to a particular business sector. Again, any remaining differences are controlled for by our econometric setup (cf, Section 3). Table 2 on the next page provides information on the magnitude of the rating changes. Overall, the number of rating changes is 6763 for Moody’s and 6929 for S&P. The share of upgrades is higher in the case of Moody’s: 34% of its rating changes were upward, compared with only 29% of S&P’s. The higher number of downgrades for both CRAs is in line with the ﬁndings of Blume et al (1998), among others. In general, it is explained by decreasing credit quality coupled with more stringent rating standards. In addition, for both CRAs, the share of downgrades that exceed one rating notch is higher than the share of large upgrades. In the case of Moody’s (S&P), large changes account for 41% (34%) of all downgrades versus 29% (24%) of all upgrades. Thus, downgrades are on average of a greater magnitude than upgrades. Besides, Moody’s adjusts its ratings by more than one rating notch more often than S&P does. The differences between the two CRAs’ rating change magnitudes are signiﬁcant at least at the 5% level using standard contingency tests. The number of Watchlist additions shows differences between Moody’s and S&P, too. On the negative side, S&P assigned 4710 additions whereas Moody’s accounts for only 3577 entries. On the other hand, Moody’s published 1504 positive Watchlist additions for our selected companies; S&P released only 1273 positive additions. Technical Report www.thejournalofcreditrisk.com 102 A. Güttler TABLE 2 Distribution of the magnitude of the rating changes. (a) Moody’s Downgrades Upgrades ‚ …„ ƒ ‚ …„ ƒ One Two > Two One Two > Two All notch notches notches All notch notches notches Observations 4437 2629 1226 582 2326 1645 469 212 In % 0.594 0.276 0.131 0.708 0.201 0.091 (b) S&P Downgrades Upgrades ‚ …„ ƒ ‚ …„ ƒ One Two > Two One Two > Two All notch notches notches All notch notches notches Observations 4897 3223 1045 629 2032 1540 269 223 In % 0.660 0.212 0.128 0.759 0.132 0.109 The table provides an overview of the magnitude of the rating changes. Columns 1–4 include only companies with two long-term senior unsecured credit ratings by both Moody’s and S&P in the observation period January 1994– December 2005. In columns 5–8 we constrain the sample to those rating changes which have also been rated by the second CRA during that time. Rating data is extracted from Bloomberg. 3 METHODOLOGY Close parallels exist between the time series of ratings and other time-to-event data such as clinical trials involving treatment and response. In the so-called survival analysis, patients are followed over time as they progress from one state (survival) to another (death). The patients’ survivals and deaths are related by the concept of intensity.12 It predicts the rate at which the patient will die by a speciﬁc point in time given that he or she has survived up to this time. One of the main advantages of survival analysis is the inclusion of censored observations, as one is not forced to exclude parts of the sample that could otherwise cause a selection bias. As in the case of conventional regression approaches, relevant independent vari- ables, which are denoted covariates in survival analysis, are employed to assess differ- ences in failure intensity among different patients. In the following the rating change intensity of group i at a given time t , ˛i (t ), is shown as: ˛i .t/ D ˛0i .t / exp.ˇ1 z1 .t / C ˇ2 z2 / (3.1) 12 “Hazard rate” is an alternative term often used to denote this concept. The Journal of Credit Risk Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies 103 In this proportional intensity regression model, ˛0i .t / is an unspeciﬁed baseline intensity, z1 .t/ is a vector of time-varying covariates, z2 is a vector of covariates which are constant over t , and ˇ1 and ˇ2 are the regression parameters associated with the covariates (Cox (1972)). We follow the approach of Lando and Skødeberg (2002) and concentrate on rating changes by one notch to the neighboring rating class. In speciﬁcation (1), time t is deﬁned in three different ways. First, t is the time period between the reported rating change and the last rating change or, as the case may be, the initial rating. It thus expresses the time during which the company stayed within the “old” rating class. Second, in the case of a rating that is subsequently withdrawn, t is the time the company stayed within the rating class until it was withdrawn. Third, we handle right-censoring by deﬁning t as the time between the company’s entry into a rating class and the end of the observation period (December 31, 2005). As we use time-varying covariates, z1 .t /, with rating information on the second CRA, we can make use of more than one observation per rating change, ie, a change in the respective covariate at a given time. The second covariate, z2 , is not time-varying as, for instance, either the CRA’s last rating change was a downgrade or not. Including z2 in speciﬁcation (1) controls for the effect that consecutive rating changes in the same direction are more frequent than in the opposite direction (see, for example, Lando and Skødeberg (2002)). The model is semi-parametric since the rating change intensity has two components: a nonparametric baseline intensity (corresponding to the intercept in conventional regressions) and a parametric element which is determined by the covariates. Thus, by using this approach we do not have to specify the form of the underlying distribution of survival times. We instead assume a constant and proportional relationship between the covariates and the dependent variable for all observations. As it is known that rating intensities might also be driven by other factors such as the state of the business cycle (see, for example, Bangia et al (2000)) or varying rating standards (Blume et al (1998)) we absorb such time-varying, ﬁrm-invariant effects through the baseline intensity.13 We further control for ﬁrm-speciﬁc effects by employing different baseline intensities, ˛0i .t/, for different groups i . Speciﬁcally, we form groups according to line of business, the company’s region and the credit cycle.14 The credit cycle is approximated by a dummy which equals 1 if the downgrade ratio (number of Moody’s and S&P’s downgrades over upgrades in our sample) is larger than the median ratio and 0 otherwise. Thus we obtain stratiﬁed estimates that yield equal coefﬁcients ˇ1 and ˇ2 across groups but with baseline intensities unique to each group. 13See, for instance, the discussion in Lando and Skødeberg (2002) 14 The time since issuance of the bond (see, for example, Altman and Kao (1992)) also inﬂuences rating intensities. In our setup, however, this effect would seem to be of only minor relevance since we concentrate on companies and not on single bonds. Technical Report www.thejournalofcreditrisk.com 104 A. Güttler 4 EMPIRICAL RESULTS Throughout we use the Cox regression of Equation (3.1) where only the employed covariates z1 .t/ and z2 are altered. The variation of the intensity of a certain rating change, ie, from Aa1 to Aa2, that is due to a speciﬁc covariate can be calculated by exp.ˇ1 /. To compare the results of the two CRAs for differing covariates, we calculate the mean variation of the rating intensities over all eighteen one-notch rating transitions, weighted by the number of actual rating changes out of the starting rating class. The following list provides an overview of the time-dependent covariates employed. Table 3 on page 106: downgrades by Moody’s. – Columns 5/6: downgrade by S&P. – Columns 7/8: Watchlist negative by S&P. – Columns 9/10: riskier rating by S&P. Table 4 on page 108: downgrades by S&P. – Columns 5/6: downgrade by Moody’s. – Columns 7/8: Watchlist negative by Moody’s. – Columns 9/10: riskier rating by Moody’s. Table 5 on page 110: upgrades by Moody’s. – Columns 5/6: upgrade by Moody’s. – Columns 7/8: Watchlist positive by Moody’s. – Columns 9/10: less risky rating by Moody’s. Table 6 on page 112: upgrades by S&P. – Columns 5/6: upgrade by S&P. – Columns 7/8: Watchlist positive by S&P. – Columns 9/10: less risky rating by S&P. For all these covariates we expect an increase in the rating change intensity. For instance, the downgrade intensity for companies rated by Moody’s should be higher for ﬁrms that experience a downgrade by S&P. Nevertheless, we have no a priori expectation as to which of these covariates causes the strongest increase in rating change intensities. In addition, we include the second, time-independent covariate z2 in all speciﬁcations but omit these results in the tables. The Journal of Credit Risk Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies 105 4.1 Downgrades Table 3 on the next page shows results for all downgrades by one rating notch by Moody’s and rating actions by S&P as covariates of interest. In all three different speciﬁcations of the table, z2 equals 1 if the issuer was previously downgraded by Moody’s. Column 5 shows coefﬁcients where z1 .t / equals 1 if the respective issuer was downgraded by S&P at t . Fifteen out of the eighteen regression coefﬁcients are signiﬁcantly different from zero on at least the 10% signiﬁcance level.15 The weighted mean variation of the downgrade intensity is 2.27, ie, a downgrade by Moody’s is 127% more likely if there was a downgrade by S&P. Column 7 shows coefﬁcients where z1 .t/ equals 1 if the respective issuer was put on Watchlist negative by S&P at t . Here we ﬁnd a somewhat weaker effect, as the weighted mean variation of the downgrade intensity is 80% higher in the case of a negative Watchlist entry by S&P. Column 9 shows coefﬁcients where z1 .t / equals 1 if the respective ﬁrm had a lower, ie, riskier, rating by S&P at t. The results are very similar to the case of a downgrade by S&P, as the downgrade intensity is 134% higher. Table 4 on page 108 displays corresponding results for all downgrades by one rat- ing notch by S&P and rating actions by Moody’s as covariates of interest. In all three different speciﬁcations of the table, z2 equals 1 if the issuer was previously down- graded by S&P. Column 5 shows coefﬁcients where z1 .t / equals 1 if the respective issuer was downgraded by Moody’s at t . Thirteen out of the eighteen regression coef- ﬁcients are signiﬁcantly different from zero. The downgrade intensity is 82% higher if there was a downgrade by Moody’s. Comparing these results with those shown in column 5 of Table 3 on the next page, we ﬁnd that Moody’s downgrade intensity is 25% larger (2.27 versus 1.82) in the case of a downgrade by S&P than vice versa. Column 7 shows coefﬁcients where z1 .t / equals 1 if the respective issuer was put on Watchlist negative by Moody’s at t . The downgrade intensity is 72% higher in the case of a negative Watchlist entry by Moody’s. This result is almost on a par with the Table 3 on the next page column 7 results. Column 9 exhibits coefﬁcients where z1 .t/ equals 1 if the respective ﬁrm had a lower, ie, riskier, rating by Moody’s at t. These results are rather similar to those for the two other speciﬁcations of the table, as S&P’s downgrade intensity is 75% higher. Comparing these results with those shown in column 9 of Table 3 on the next page, we ﬁnd that Moody’s downgrade intensity is 34% larger (2.34 versus 1.75) in the case of a lower rating by S&P than vice versa. All in all, we ﬁnd that rating actions by one CRA cause signiﬁcantly higher rating intensities for most one-notch downgrades by the other CRA. In line with ﬁndings by Norden and Weber (2004) and Hand et al (1992) we observe that negative Watchlist 15 We employ the 10% signiﬁcance level throughout. Technical Report www.thejournalofcreditrisk.com 106 The Journal of Credit Risk A. Güttler TABLE 3 Rating downgrade intensities for issuers rated by Moody’s conditional on S&P’s rating information. [Table continues on next page.] Downgrade Watchlist Rating Ratings by S&P negative by S&P lower by S&P ‚ …„ ƒ Potential Actual ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ From To changes changes Coefﬁcient p value Coefﬁcient p value Coefﬁcient p value Aa1 Aa2 92 38 0.696 0.177 2.101 0.002 0.412 0.621 Aa2 Aa3 170 57 1.528 0.000 0.154 0.681 0.689 0.112 Aa3 A1 220 107 1.064 0.000 1.009 0.000 0.542 0.004 A1 A2 329 152 0.560 0.001 0.739 0.000 0.562 0.003 A2 A3 450 216 0.772 0.000 0.925 0.000 0.635 0.000 A3 Baa1 472 196 1.140 0.000 0.665 0.000 0.834 0.000 Baa1 Baa2 500 234 0.867 0.000 0.585 0.000 0.596 0.000 Baa2 Baa3 535 261 0.825 0.000 0.304 0.056 0.743 0.000 Baa3 Ba1 411 149 0.819 0.000 0.737 0.000 0.763 0.000 Volume 7/Number 1, Spring 2011 Ba1 Ba2 249 93 0.775 0.001 0.437 0.109 0.971 0.000 Ba2 Ba3 244 92 0.624 0.010 0.091 0.716 0.576 0.058 Ba3 B1 340 139 0.988 0.000 0.661 0.002 0.651 0.016 Technical Report TABLE 3 Continued. Downgrade Watchlist Rating Lead–lag relationships and rating convergence among credit rating agencies Ratings by S&P negative by S&P lower by S&P ‚ …„ ƒ Potential Actual ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ From To changes changes Coefﬁcient p value Coefﬁcient p value Coefﬁcient p value B1 B2 472 237 0.623 0.000 0.464 0.003 0.689 0.002 B2 B3 544 252 0.723 0.000 0.371 0.012 1.440 0.000 B3 Caa1 413 194 0.815 0.000 0.183 0.303 1.077 0.000 Caa1 Caa2 244 103 0.079 0.706 0.137 0.597 0.987 0.009 Caa2 Caa3 138 46 0.694 0.040 0.592 0.079 1.511 0.002 Caa3 Ca 114 63 0.097 0.760 0.362 0.182 0.670 0.220 Weighted mean variation of rating intensity 2.267 1.796 2.341 The table shows rating-downgrade intensities for companies that held a rating by Moody’s in the observation period January 1994–December 2005. We include only companies with long-term senior unsecured credit ratings by both Moody’s and S&P in the observation period. We further employ rating data for the period January 1990–December 1993 to condition upon previous rating actions. Rating data is extracted from Bloomberg. The ﬁrst two columns report the type of rating transition studied, eg, the ﬁrst row exhibits rating changes from “Aa1” to “Aa2”. We concentrate on rating downgrades by one rating notch. The third column shows the number of all potential rating transitions out of the “From” rating www.thejournalofcreditrisk.com class, ie, censored and uncensored observations, whereas the fourth column provides all observed rating transitions out of the “From” into the “To” rating class. Columns 5 and 6 provide results for the speciﬁcation with downgrades by S&P as time-varying covariates. Columns 7 and 8 show results for the speciﬁcation with negative Watchlist entries by S&P as time-varying covariates. Columns 9 and 10 provide results for the speciﬁcation with lower, ie, riskier, ratings by S&P as time-varying covariates. In all three speciﬁcations we omit results for the covariate of previous downgrades by Moody’s. We use a proportional intensity regression model by maximizing the partial likelihood of every rating downgrade by one full notch to the neighboring rating class. Each line thus comprises three separate regressions with discrete, time-varying covariates by S&P and covariates by Moody’s which are constant over time. We exclude the rating sequence “Aaa” to “Aa1” to allow for previous downgrades by Moody’s. A positive (negative) coefﬁcient implies that the downgrade intensity is higher (lower) compared with the case of no downgrade (columns 5 and 6), no negative Watchlist entry (columns 7 and 8) or no lower rating (columns 9 and 10) by S&P. The last line shows the mean variation of the rating intensity, exp.ˇ1 /, over all one-notch rating transitions, weighted by the number of potential rating changes out of the starting rating class. We furthermore employ strata for regions, sectors and whether the rating change occurred in a month with an above-average rating downgrade ratio. We employ the robust sandwich estimate of Lin and Wei (1989) for the covariance matrix. 107 108 The Journal of Credit Risk A. Güttler TABLE 4 Rating downgrade intensities for issuers rated by S&P conditional on Moody’s rating information. [Table continues on next page.] Downgrade Watchlist negative Rating lower Ratings by Moody’s by Moody’s by Moody’s ‚ …„ ƒ Potential Actual ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ From To changes changes Coefﬁcient p value Coefﬁcient p value Coefﬁcient p value AAC AA 63 42 0.105 0.790 0.584 0.193 1.818 0.012 AA AA 178 120 0.444 0.052 1.151 0.000 0.567 0.030 AA AC 325 209 0.695 0.000 1.108 0.000 0.384 0.013 AC A 440 217 1.035 0.000 0.535 0.002 0.866 0.000 A A 500 247 0.615 0.000 0.454 0.002 0.184 0.178 A BBBC 558 288 0.414 0.002 0.086 0.530 0.513 0.000 BBBC BBB 586 278 0.780 0.000 0.642 0.000 0.493 0.000 BBB BBB 593 258 0.820 0.000 0.587 0.000 0.801 0.000 BBB BBC 391 186 0.617 0.000 0.499 0.002 0.524 0.001 Volume 7/Number 1, Spring 2011 BBC BB 306 141 1.087 0.000 0.417 0.013 0.476 0.013 BB BB 389 203 0.587 0.001 0.531 0.001 0.356 0.048 BB BC 528 272 0.488 0.001 0.467 0.002 0.675 0.000 Technical Report TABLE 4 Continued. Downgrade Watchlist negative Rating lower Lead–lag relationships and rating convergence among credit rating agencies Ratings by Moody’s by Moody’s by Moody’s ‚ …„ ƒ Potential Actual ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ From To changes changes Coefﬁcient p value Coefﬁcient p value Coefﬁcient p value BC B 610 317 0.238 0.059 0.245 0.106 0.462 0.003 B B 380 204 0.051 0.737 0.086 0.621 0.108 0.512 B CCCC 251 137 0.164 0.377 0.710 0.003 0.566 0.005 CCCC CCC 122 47 0.229 0.498 0.987 0.005 0.325 0.289 CCC CCC 54 18 0.319 0.591 1.275 0.005 1.849 0.007 CCC CC 48 39 0.902 0.037 0.526 0.126 0.554 0.095 Weighted mean variation of rating intensity 1.817 1.725 1.748 The table shows rating downgrade intensities for companies that held a rating by S&P in the observation period January 1994–December 2005. We include only companies with long-term senior unsecured credit ratings by both Moody’s and S&P in the observation period. We further employ rating data for the period January 1990–December 1993 to condition upon previous rating actions. Rating data is extracted from Bloomberg. The ﬁrst two columns report the type of rating transition studied, eg, the ﬁrst row exhibits rating changes from “AAC” to “AA”. We concentrate on rating downgrades by one rating notch. The third column shows the number of all potential rating transitions out of the “From” www.thejournalofcreditrisk.com rating class, ie, censored and uncensored observations, whereas the fourth column provides all observed rating transitions out of the “From” into the “To” rating class. Columns 5 and 6 provide results for the speciﬁcation with downgrades by Moody’s as time-varying covariates. Columns 7 and 8 show results for the speciﬁcation with negative Watchlist entries by Moody’s as time-varying covariates. Columns 9 and 10 provide results for the speciﬁcation with lower, ie, riskier, ratings by Moody’s as time-varying covariates. We omit in all three speciﬁcations results for the covariate of previous downgrades by S&P. We use a proportional intensity regression model by maximizing the partial likelihood of every rating downgrade by one full notch to the neighboring rating class. Each line thus comprises three separate regressions with discrete, time-varying covariates by Moody’s and covariates by S&P which are constant over time. We exclude the rating sequence “AAA” to “AAC” to allow for previous downgrades by S&P. A positive (negative) coefﬁcient implies that the downgrade intensity is higher (lower) compared with the case of no downgrade (columns 5 and 6), no negative Watchlist entry (columns 7 and 8), or no lower rating (columns 9 and 10) by Moody’s. The last line shows the mean variation of the rating intensity, exp.ˇ1 /, over all one-notch rating transitions, weighted by the number of potential rating changes out of the starting rating class. We furthermore employ strata for regions, sectors and whether the rating change occurred in a month with an above-average rating downgrade ratio. We employ the robust sandwich estimate of Lin and Wei (1989) for the covariance matrix. 109 110 The Journal of Credit Risk A. Güttler TABLE 5 Rating upgrade intensities for issuers rated by Moody’s conditional on S&P’s rating information. [Table continues on next page.] Upgrade Watchlist Rating Ratings by S&P positive by S&P higher by S&P ‚ …„ ƒ Potential Actual ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ From To changes changes Coefﬁcient p value Coefﬁcient p value Coefﬁcient p value Aa1 Aaa 64 10 1.962 0.118 2.077 0.087 0.410 0.593 Aa2 Aa1 147 34 1.822 0.000 0.856 0.009 N/A N/A Aa3 Aa2 184 71 0.166 0.527 0.494 0.086 0.611 0.106 A1 Aa3 259 82 1.197 0.000 0.822 0.032 0.757 0.002 A2 A1 332 98 0.816 0.000 1.321 0.000 1.072 0.000 A3 A2 398 122 1.319 0.000 1.128 0.002 0.471 0.019 Baa1 A3 388 122 0.544 0.008 0.377 0.371 0.656 0.001 Volume 7/Number 1, Spring 2011 Baa2 Baa1 395 121 0.941 0.000 1.679 0.000 1.071 0.000 Baa3 Baa2 414 152 0.561 0.002 1.002 0.000 0.901 0.000 Technical Report TABLE 5 Continued. Upgrade Watchlist Rating Ratings by S&P positive by S&P higher by S&P ‚ …„ ƒ Potential Actual ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ Lead–lag relationships and rating convergence among credit rating agencies From To changes changes Coefﬁcient p value Coefﬁcient p value Coefﬁcient p value Ba1 Baa3 276 120 1.176 0.000 0.780 0.017 0.554 0.009 Ba2 Ba1 262 110 0.591 0.051 0.412 0.364 0.737 0.005 Ba3 Ba2 303 102 0.518 0.015 0.880 0.002 0.923 0.000 B1 Ba3 363 128 1.183 0.000 0.877 0.001 0.855 0.003 B2 B1 443 151 1.058 0.000 1.067 0.000 0.786 0.002 B3 B2 322 103 1.491 0.000 0.994 0.004 1.604 0.006 Caa1 B3 200 59 0.628 0.087 0.977 0.011 1.636 0.029 Caa2 Caa1 134 42 1.176 0.009 0.561 0.290 2.046 0.028 Caa3 Caa2 69 18 2.657 0.000 2.066 0.119 0.378 0.613 Weighted mean variation of rating intensity 2.957 2.847 2.593 The table shows rating upgrade intensities for companies that held a rating by Moody’s in the observation period January 1994–December 2005. We include only companies with long-term senior unsecured credit ratings by both Moody’s and S&P in the observation period. We further employ rating data for the period January 1990–December 1993 to condition upon previous rating actions. Rating data is extracted from Bloomberg. The ﬁrst two columns report the type of rating transition studied, eg, the ﬁrst row exhibits rating www.thejournalofcreditrisk.com changes from “Aa1” to “Aaa”. We concentrate on rating upgrades by one rating notch. The third column shows the number of all potential rating transitions out of the “From” rating class, ie, censored and uncensored observations, whereas the fourth column provides all observed rating transitions out of the “From” into the “To” rating class. Columns 5 and 6 provide results for the speciﬁcation with upgrades by S&P as time-varying covariates. Columns 7 and 8 show results for the speciﬁcation with positive Watchlist entries by S&P as time-varying covariates. Columns 9 and 10 provide results for the speciﬁcation with higher, ie, less risky, ratings by S&P as time-varying covariates. In all three speciﬁcations we omit results for the covariate of previous upgrades by Moody’s. We use a proportional intensity regression model by maximizing the partial likelihood of every rating upgrade by one full notch to the neighboring rating class. Each line thus comprises three separate regressions with discrete, time-varying covariates by S&P and covariates by Moody’s which are constant over time. We exclude the rating sequence “Ca” to “Caa3” to allow for previous upgrades by Moody’s. A positive (negative) coefﬁcient implies that the upgrade intensity is higher (lower) compared with the case of no upgrade (columns 5 and 6), no positive Watchlist entry (columns 7 and 8), or no higher rating (columns 9 and 10) by S&P. The last line shows the mean variation of the rating intensity, exp.ˇ1 /, over all one-notch rating transitions, weighted by the number of potential rating changes out of the starting rating class. We furthermore employ strata for regions, sectors and whether the rating change occurred in a month with a below-average rating downgrade ratio. We omit estimation results if we have too few observations of the respective covariate and mark these cases “N/A”. We employ the robust sandwich estimate of Lin and Wei (1989) for the covariance matrix. 111 112 The Journal of Credit Risk A. Güttler TABLE 6 Rating upgrade intensities for issuers rated by S&P conditional on Moody’s rating information. [Table continues on next page.] Upgrade Watchlist positive Rating higher Ratings by Moody’s by Moody’s by Moody’s ‚ …„ ƒ Potential Actual ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ From To changes changes Coefﬁcient p value Coefﬁcient p value Coefﬁcient p value AAC AAA 25 4 N/A N/A N/A N/A N/A N/A AA AAC 65 7 N/A N/A N/A N/A N/A N/A AA AA 151 35 0.431 0.301 1.329 0.002 1.059 0.038 AC AA 282 60 0.692 0.027 0.881 0.011 0.969 0.002 A AC 335 88 0.812 0.001 0.443 0.282 0.137 0.624 A A 392 128 0.310 0.211 1.019 0.000 0.265 0.214 BBBC A 441 139 0.872 0.000 0.769 0.001 1.122 0.000 BBB BBBC 500 170 0.948 0.000 0.929 0.000 0.541 0.002 BBB BBB 366 162 1.144 0.000 1.349 0.000 0.476 0.010 Volume 7/Number 1, Spring 2011 BBC BBB 293 128 0.515 0.007 1.253 0.000 0.485 0.034 BB BBC 313 127 1.062 0.000 1.320 0.000 0.762 0.001 BB BB 380 124 0.962 0.000 0.044 0.920 0.978 0.001 Technical Report TABLE 6 Continued. Upgrade Watchlist positive Rating higher Lead–lag relationships and rating convergence among credit rating agencies Ratings by Moody’s by Moody’s by Moody’s ‚ …„ ƒ Potential Actual ‚ …„ ƒ ‚ …„ ƒ ‚ …„ ƒ From To changes changes Coefﬁcient p value Coefﬁcient p value Coefﬁcient p value BC BB 431 138 1.199 0.000 1.429 0.000 0.503 0.035 B BC 295 119 1.039 0.000 1.603 0.000 0.703 0.003 B B 172 58 1.623 0.002 0.152 0.817 0.019 0.952 CCCC B 106 31 1.176 0.122 0.122 0.781 0.295 0.679 CCC CCCC 52 16 1.038 0.149 1.238 0.313 1.577 0.105 CCC CCC 15 6 N/A N/A N/A N/A N/A N/A Weighted mean variation of rating intensity 2.513 2.777 1.878 The table shows rating upgrade intensities for companies that held a rating by S&P in the observation period January 1994–December 2005. We include only companies with long-term senior unsecured credit ratings by both Moody’s and S&P in the observation period. We further employ rating data for the period January 1990–December 1993 to condition upon previous rating actions. Rating data is extracted from Bloomberg. The ﬁrst two columns report the type of rating transition studied, eg, the ﬁrst row exhibits rating changes from “AAC” to “AAA”. We concentrate on rating upgrades by one rating notch. The third column shows the number of all potential rating transitions out of the “From” rating class, ie, censored and uncensored observations, whereas the fourth column provides all observed rating transitions out of the “From” into the “To” rating class. Columns 5 and www.thejournalofcreditrisk.com 6 provide results for the speciﬁcation with upgrades by Moody’s as time-varying covariates. Columns 7 and 8 show results for the speciﬁcation with positive Watchlist entries by Moody’s as time-varying covariates. Columns 9 and 10 provide results for the speciﬁcation with higher, ie, less risky, ratings by Moody’s as time-varying covariates. In all three speciﬁcations we omit results for the covariate of previous upgrades by S&P. We use a proportional intensity regression model by maximizing the partial likelihood of every rating upgrade by one full notch to the neighboring rating class. Each line thus comprises three separate regressions with discrete, time-varying covariates by S&P and covariates by Moody’s which are constant over time. We exclude the rating sequence “CC” to “CCC ” to allow for previous upgrades by S&P. A positive (negative) coefﬁcient implies that the upgrade intensity is higher (lower) compared with the case of no upgrade (columns 5 and 6), no positive Watchlist entry (columns 7 and 8) or no higher rating (columns 9 and 10) by Moody’s. The last line shows the mean variation of the rating intensity, exp.ˇ1 /, over all one-notch rating transitions which is weighted by the number of potential rating changes out of the starting rating class. We furthermore employ strata for regions, sectors and whether the rating change occurred in a month with a below-average rating downgrade ratio. We omit estimation results if we have too few observations of the respective covariate and mark these cases “N/A”. We employ the robust sandwich estimate of Lin and Wei (1989) for the covariance matrix. 113 114 A. Güttler additions reveal vital information to market participants. Speciﬁcally, our relative comparison of the downgrade intensities provides evidence that Moody’s downgrade intensities seem to be more strongly affected by both downgrades and lower rating levels by S&P than vice versa. The latter results appear to conﬁrm those presented in a paper by Welch (1990) in which he shows that stock recommendations by securities analysts have a signiﬁcant inﬂuence on the stock recommendations of the subsequent two analysts. However, we ﬁnd no differences in downgrade intensities in the case of negative Watchlist additions. 4.2 Upgrades Table 5 on page 110 shows results for all upgrades by one rating notch by Moody’s and rating actions by S&P as covariates of interest. In all three different speciﬁcations of the table, z2 equals 1 if the issuer was previously upgraded by Moody’s. Column 5 shows coefﬁcients where z1 .t/ equals 1 if the respective issuer was upgraded by S&P at t . Sixteen out of the seventeen feasible regression coefﬁcients are signiﬁcantly different from zero. Moody’s upgrade intensity is 196% higher if there was an upgrade by S&P. Column 7 shows coefﬁcients where z1 .t / equals 1 if the respective issuer was put on Watchlist positive by S&P at t . We ﬁnd that the weighted mean variation of the upgrade intensity is 185% higher in the case of a positive Watchlist entry by S&P. Column 9 shows coefﬁcients where z1 .t / equals 1 if the respective ﬁrm had a higher, ie, less risky, rating by S&P at t . One out of the eighteen separate regressions did not converge due to an insufﬁcient number of observations. These results are slightly weaker, as the upgrade intensity is only 159% higher. Table 6 on page 112 displays corresponding results for all upgrades by one rating notch by S&P and rating actions by Moody’s as covariates of interest. In all three different speciﬁcations of the table, z2 equals 1 if the issuer was previously upgraded by S&P. Three out of the eighteen separate regressions did not converge due to an insufﬁcient number of observations. Column 5 exhibits coefﬁcients where z1 .t / equals 1 if the respective issuer was upgraded by Moody’s at t . Eleven out of the ﬁfteen feasible regression coefﬁcients are signiﬁcantly different from zero. The upgrade intensity is 151% higher if there was an upgrade by Moody’s. Comparing these results to those in column 5 of Table 5 on page 110, we ﬁnd that Moody’s upgrade intensity is 18% larger (1.96 versus 2.51) in the case of an upgrade by S&P than vice versa. Interestingly, these upgrade intensity increases are much larger than the results shown in column 7 of Table 3 on page 106 and Table 4 on page 108 for negative Watchlist additions. Column 7 shows coefﬁcients where z1 .t / equals 1 if the respective issuer was put on Watchlist positive by Moody’s at t . The weighted mean variation of the upgrade intensity is 178% higher in the case of a positive Watchlist entry by Moody’s. This result is almost on a par with the results in column 7 of Table 5 on page 110. The Journal of Credit Risk Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies 115 Column 9 exhibits coefﬁcients where z1 .t / equals 1 if the respective ﬁrm had a higher, ie, less risky, rating by Moody’s at t . These results are much weaker compared with the two other speciﬁcations of the table, as S&P’s upgrade intensity is only 88% higher. Comparing these results to those in column 9 of Table 5 on page 110, we ﬁnd that Moody’s upgrade intensity is 38% larger (2.59 versus 1.88) in the case of a higher rating by S&P than vice versa. To summarize, we ﬁnd that rating actions by one CRA cause signiﬁcantly higher rating intensities for most one-notch upgrades by the other CRA. We thus add evi- dence to Hand et al (1992), who ﬁnd weak results for positive Watchlist additions, as we observe that positive Watchlist additions by the other CRA increase upgrade intensities even more sharply than negative Watchlist additions increase downgrades. Speciﬁcally, our comparison of upgrade intensities provides evidence that Moody’s upgrade intensities seem to be more strongly affected by both upgrades and higher rating levels by S&P. The latter results again appear to conﬁrm those presented in a paper by Welch (1990). However, we ﬁnd no different upgrade intensities in the case of positive Watchlist additions. 4.3 Robustness checks The central assumption of the Cox regression is that of proportionality, ie, that a unit variation in a covariate causes the baseline intensity to be multiplied by the exponential of the covariate’s coefﬁcient. Thus, enlargements and reductions of each covariate are proportional. As a robustness check we test this proportionality assumption (Gramb- sch and Therneau (1994)). For the 206 valid regression coefﬁcients of interest in Tables 3–6, the employed speciﬁcation test rejects the proportionality hypothesis in twenty-ﬁve cases (at the 5% signiﬁcance level).16 For these exceptions we apply the graphical Kaplan–Meier approach, which allows the comparison of survival curves for different covariate values. In our case of binary covariates we obtain two survival curves for each case where the survival probability should be lower for covariates that equal one, ie, where we observe a positive rating intensity for the respective covariate. The graphical comparison should also clarify whether the survivor curves cross. We ﬁnd in only eight of the twenty-ﬁve cases an intersection of the two survivor curves. Otherwise we ﬁnd no rejection of the proportionality hypothesis through the use of the Kaplan–Meier approach. Given that the proportionality hypothesis is hardly rejected in either of the two tests, our Cox regression model seems to be correctly speciﬁed. 16 Results are not shown here but are available upon request. Technical Report www.thejournalofcreditrisk.com 116 A. Güttler 5 DISCUSSION AND CONCLUSION Using a sample of issuers rated by both Moody’s and S&P, in this paper we have analyzed whether one of the two dominant CRAs adjusts its ratings in a timelier manner than the other. In addition, we have examined whether there is a tendency by one (or both) of the CRAs toward the other CRA’s rating. We ﬁnd evidence that Moody’s rating change intensities are higher given a rating change by S&P. This effect is somewhat stronger in the case of downgrades than for upgrades and seems to be tentative evidence that S&P assigns ratings in a timelier manner than Moody’s. Second, we ﬁnd that the tendency toward rating convergence is stronger for Moody’s than for S&P. Despite delivering novel results, we admit the following limitations: ﬁrst, due to data restrictions, the period covered in the study is too short to determine any credit- cycle effects. Research done with credit ratings of one single CRA typically covers a longer horizon. We nevertheless try to control for credit cycle effects by incorporating the state of the credit cycle as a control variable in our speciﬁcations. Second, our results might – at least partly – be driven by different rating methodologies. Whereas Moody’s uses an expected loss approach that incorporates the probability of default and the loss severity, S&P only takes account of the probability of default. With particular regard to rating convergence, however, practitioners and regulators do not seem to differentiate much between Moody’s and S&P. For example, the often used junk rating threshold is deﬁned in the same way as our rating matching. Third, the reﬁned rating scale Caa1, Caa2, Caa3 was introduced by Moody’s around 1997. This may lead to an over-representation of the transitions to rating Caa (Caa2) in the period 1994–97. The scale CCCC, CCC, CCC was introduced earlier by S&P, which may bias our results for the ﬁrst years. A further potential criticism might be that we only include Moody’s and S&P in our study and disregard all other CRAs. Of course, the inclusion of additional rat- ing actions by the other CRAs might provide interesting insights into the lead–lag relationships and the rating convergence between large and smaller CRAs. However, since Moody’s and S&P are by far the largest rating agencies for securities in capital markets in the world – as is demonstrated by their huge share of the market – we are able to cover a large dataset solely with those issuers rated by both. The inclusion of further, much smaller rating agencies would decrease the available dataset substan- tially. Even the number three in the market for credit ratings, Fitch Ratings, is not large enough to possess a long and broad history of corporate ratings. Other rating agencies do not even operate worldwide, such as the Japanese rating agencies JCR and R&I, or are limited to a speciﬁc sector, such as AM Best, which specializes in the insurance and banking sector. The Journal of Credit Risk Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies 117 Further research might enlarge the scope of our study to include structured ﬁnance ratings. The ﬁnancial crisis that began in 2007 has raised serious questions as to whether the leading CRAs reacted quickly enough to warn investors about the risks of investing in securities backed by US subprime mortgages. Including Fitch Ratings in this area as a third CRA might be possible since it commands a much larger market share in this particular segment of the rating business. REFERENCES Abad-Romero, P., and Robles-Fernandez, D. (2006). Risk and return around bond rating changes: new evidence from the Spanish stock market. Journal of Business Finance and Accounting 33(5), 885–908. Altman, E. I., and Kao, D. L. (1992).The implications of corporate bond ratings drift. Financial Analysts Journal 48, 64–67. Baker, H. K., and Mansi, S. A. (2002). Assessing credit rating agencies by bond issuers and institutional investors. Journal of Business Finance and Accounting 29(9), 1367–1398. Bangia, A., Diebold, F., Kronimus, A., Schagen, C., and Schuermann, T. (2002). Ratings migration and the business cycle, with applications to credit portfolio stress testing. Jour- nal of Banking and Finance 26(2), 445–474. Bannier, C. E., and Tyrell, M. (2006). Modelling the role of credit rating agencies. Do they spark off a virtuous circle? Working Paper, University of Frankfurt. Barron, M. J., Clare, A. D., and Thomas, S. H. (1997). The effect of bond rating changes and new ratings on UK stock returns. Journal of Business Finance and Accounting 24(3), 497–509. BBA (2006). BBA Credit Derivatives Report 2006. Bernhardt, D., Campello, M., and Kutsoati, E. (2006). Who herds? Journal of Financial Economics 80(3), 657–675. BIS (2007). BIS Quarterly Review. Blume, M. E., Lim, F., and MacKinlay, A. C. (1998). The declining credit quality of US corporate debt: myth or reality? Journal of Finance 53(4), 1389–1414. Boot, A., Greenbaum, S., and Thakor, A. (1993). Reputation and discretion in ﬁnancial contracting. American Economic Review 83(5), 1165–1183. Cantor, R., and Packer, F. (1997). Differences of opinion and selection bias in the credit rating industry. Journal of Banking and Finance 21(10), 1395–1417. Christensen, J. H. E., Hansen, E., and Lando, D. (2004). Conﬁdence sets for continuous- time rating transition probabilities. Journal of Banking and Finance 28(11), 2575–2602. Covitz, D. M., and Harrison, P. (2003). Testing conﬂicts of interest at bond ratings agencies with market anticipation: evidence that reputation incentives dominate. Working Paper, Finance and Economics Discussion Series no. 2003-68, Board of Governors of the US Federal Reserve System. Cox, D. R. (1972). Regression models and life tables. Journal of the Royal Statistical Society B 34, 187–220. Economist, The (2005). Credit-rating agencies. Who rates the raters? (March 26.) Technical Report www.thejournalofcreditrisk.com 118 A. Güttler Feinberg, M., Shelor, R., and Jiang, J. (2004). The effect of solicitation and independence on corporate bond ratings. Journal of Business Finance and Accounting 31(9), 1327– 1353. Figlewski, S., Frydman, H., and Liang, W. (2008). Modeling the effect of macroeconomic factors on corporate default and credit rating transitions. Working Paper, NYU Stern School of Business. Grambsch P. M., and Therneau, T. M. (1994). Proportional hazards test and diagnostics based on weighted residuals. Biometrika 81(3), 515–526. Güttler, A., and Raupach, P. (2010). The impact of downward rating momentum. Journal of Financial Services Research 37(1), 1–23. Hamilton, D. T., and Cantor, R. (2004). Rating transitions and defaults conditional on watch- list, outlook and rating history. Moody’s Special Comment. Hamilton, D. T., Cantor, R., and Ou, S. (2002). Default and recovery rates of corporate bond issuers. Moody’s Special Comment. Hand, J. R. M., Holthausen, R. W., and Leftwich, R. W. (1992). The effect of bond rating agency announcements on bond and stock prices. Journal of Finance 47(2), 733–752. Jewell, J., and Livingston, M. (1999). A comparison of bond ratings from Moody’s, S&P and Fitch IBCA. Financial Markets, Institutions, and Instruments 8(4), 1–45. Kavvathas, D. (2001). Estimating credit rating transition probabilities for corporate bonds. Working Paper, University of Chicago. Lando, D., and Skødeberg, T. M. (2002). Analyzing rating transitions and rating drift with continuous observations. Journal of Banking and Finance 26(2), 423–444. Lin, D. Y., and Wei, L. J. (1989). The robust inference for the Cox proportional hazards model. Journal of the American Statistical Association 84, 1074–1078. Löfﬂer, G. (2007). Proﬁts ﬁrst, or clients ﬁrst? Some lessons from Moody’s stock price. Working Paper, University of Ulm. Millon, M. H., and Thakor, A. V. (1985). Moral hazard and information sharing: a model of ﬁnancial information gathering agencies. Journal of Finance 40(5), 1403–1422. Moody’s (2010).The performance of Moody’s corporate debt ratings. December 2009 Quar- terly Update. Moody’s Investors Service, Special Comment. Morgan, D. (2002). Rating banks: risk and uncertainty in an opaque industry. American Economic Review 92(4), 874–888. Nickell, P., Perraudin, W., and Varotto, S. (2000). Stability of rating transitions. Journal of Banking and Finance 24(1), 203–227. Norden, L., and Weber, M. (2004). Informational efﬁciency of credit default swap and stock markets: the impact of credit rating announcements. Journal of Banking and Finance 28(11), 2813–2843. Perry, L. G. (1985). The effect of bond rating agencies on bond rating models. Journal of Financial Research 8, 307–315. Poon, W. P. H., and Firth, M. (2005). Are unsolicited credit ratings lower? International evidence from bank ratings. Journal of Business Finance and Accounting 32(9), 1741– 1771. Ramakrishnan, R. T. S., and Thakor, A. V. (1984). Information reliability and a theory of ﬁnancial intermediation. Review of Economic Studies 51(3), 415–432. The Journal of Credit Risk Volume 7/Number 1, Spring 2011 Lead–lag relationships and rating convergence among credit rating agencies 119 Steiner, M., and Heinke, V. G. (2001). Event study concerning international bond price effects of credit rating actions. International Journal of Finance and Economics 6(2), 139–157. Trueman, B. (1994). Analyst forecasts and herding behavior. Review of Financial Studies 7(1), 97–124. Welch, I. (1990). Herding among security analysts. Journal of Financial Economics 58(3), 369–396. Wendin, J., and McNeil, A. J. (2006). Dependent credit migrations. Working Paper, ETH Zurich. Technical Report www.thejournalofcreditrisk.com