The Grasshopper and the Grasshopper: Credit Rating Agencies incentives, Regulatory use of Ratings and the Subprime Crisis

Transcription

1 The Grasshopper and the Grasshopper: Credit Rating Agencies incentives, Regulatory use of Ratings and the Subprime Crisis Stefano Lugo Utrecht University School of Economics. Kriekenpitplein 21-22, 3584 EC Utrecht, Netherlands. Phone: ; E- mail address: January 14 th 2013 Abstract This paper investigates how distorted incentives and the regulatory use of ratings have brought toward inflated evaluation by Credit Rating Agencies (CRAs) in the market for subprime mortgage-backed securities. We use a sample of 9,427 Home Equity Loan tranches issued between 1992 and 2007 and compute for each one a measure of fully predictable ex-ante relative rating bias. Rating inflation translated partially (for Moody s) or even 1-to-1 (for S&P) into subsequent downgrades and was more likely for securities issued by bigger firms and for those rated by only one CRA (but only for S&P), a result coherent with rating shopping phenomena and distorted incentives due to the issuer-paid mechanism. However, despite relative biases were fully predictable ex-ante, markets only partially discounted them on Investment Grade (IG) securities pricing, while correctly priced them for Speculative Grade (SG) tranches. This difference can be explained by distorted incentives for the market due to the regulatory use of ratings. Moreover, during booming years more favorable ratings drove the prices for tranches with split evaluations. All in all, this evidence suggests that the buy-side too has played a relevant role in the disruption of ratings quality in the years leading to the 2007 crisis. JEL classification: G01, G14, G24, G38 Keywords: credit rating agencies, subprime, regulation, conflicts of interest, rating shopping

2 2 1. Introduction Credit Rating Agencies (CRAs hereafter) are accused to have played an active role in the disruption of the market of structured finance products backed by mortgages. By assigning inflated evaluations, the critique goes, they favored the bubble that resulted in the crisis started in This view implicitly consider markets as victims of misleading CRAs evaluations; however, buy-side could have also played an active role in allowing the disruption of ratings quality. In this paper, we use a sample of more than 9,400 Home Equity Loans tranches which have been rated and subsequently downgraded between June the 1 st 2007 and August the 1 st 2011 by either Moody s and/or S&P to investigate: a) whether rating inflation, at least on a relative basis, was detectable by markets ex-ante; b) which securities where more likely to receive an inflated rating before the crisis started; and c) whether and to what extent predictable rating inflation was priced by the market, and the role played by the latter in allowing rating quality disruption. While empirical studies have tried so far to detect the effect of rating shopping and distorted incentives on CRAs evaluations by focusing on market prices or realized default probabilities, we directly model the determinants of rating at issuance and compute an expected rating for each tranche. This approach allows to consider rating inflation only on a relative basis, but permits us to build a measure of predictable exante relative rating bias as the difference between expected and realized rating for each tranche. We find that expected relative rating inflation has partially (for Moody s) or even fully translated (for S&P) into subsequent downgrades once the crisis occurred; for example, a tranche having a predicted rating inflation of 1 notch (e.g. being rated Aa1/AA+ while a Aa2/AA rating was expected) has experienced on average a 1 notch harsher downgrade, ceteris paribus. Coherently with previous studies we find that, due to CRAs distorted incentives and rating shopping phenomena, securities issued by bigger firms and those that were not jointly rated (but only for S&P) were more likely to receive inflated evaluations. Again, those securities received harsher

3 3 downgrades once the crisis occurred, confirming ex-post that differences in evaluations of otherwise similar tranches were hardly imputable to unobservable characteristics. We then turn our attention on how the market was pricing ratings on these securities before the advent of the crisis. We find that rating biases were correctly (on relative terms) priced by the market for Speculative Grade (SG) tranches, but not for Investment Grade (IG) tranches. Even accounting for the effect on price of a predictable rating bias, tranches receiving a 1-letter better evaluation than otherwise identical issues still enjoyed 27 to 112 bps lower spreads at issuance when rated Baa/BB or above. The cumulated effect of predictable rating inflation on price is instead non statistically different from zero for SG securities. We interpret this result as an effect of the regulatory use of ratings: as modeled theoretically by Opp, Opp and Harris (2012), investors subjected to rating-based restrictions on their investment universe might favor rating inflation as they actually benefit from it. Restricted investors might in fact increase their expected absolute returns, at the cost of risk-adjusted expected returns, by consciously buying inflated-rating securities. As the regulatory benefit of rating inflation is much lower as long as securities are classified SG, this can explain why for those securities a predictable rating bias was instead correctly priced. Looking at issues rated by both Moody s and S&P, we find that, in case of split evaluation by the two agencies, prices were driven by the more favorable rating during the booming years Again, these empirical results suggest buy-side behavior has played a significant role in allowing rating quality disruption. Few recent papers have studied empirically the behavior of CRAs in the context of the subprime crisis (e.g. Ashcraft, Goldsmith-Pinkham and Vickery, 2010; He, Qian and Strahan, 2012; Griffin and Tang, 2012) and the role played by their evaluations in the pricing of these securities (e.g. Ashcraft, Goldsmith-Pinkham, Hull and Vickery, 2011; Mählmann, 2012). We add to this literature in several ways.

4 4 First, our approach allows to quantify the (relative) ex-ante rating bias; this in turns permits to study how rating bias translated into subsequent downgrades and, even more importantly, to what extent it was priced by the market. He, Qian and Strahan (2012) look for empirical evidence of CRAs misbehavior by studying whether markets discounted at issuance the number of CRAs rating the deal and the issuer s dimension; this approach implicitly assume markets efficient pricing and gives only an indirect measure of potential CRAs misbehavior. On the contrary, our approach gives us relevant insights on the potential active role played by the buy-side on the deterioration of the market. Results suggest not only that rating inflation was in place, particularly for not jointly rated tranches and securities issued by big firms, but that it was to a certain extent favored by the buyside itself. Second, we assess how buy-side caution in the structured finance market as evolved over time by looking at how tranches exhibiting split evaluations were priced. Mählmann (2012) provides evidence that Collateralized Debt Obligations (CDOs) prices were increasingly dependent on ratings during the booming years; we add to his findings by showing that not only ratings had an increasingly material impact, but during the booming years leading to the financial crisis the most benign rating was also the most influential on prices. Again, our results suggest markets were not only relying excessively on ratings, but were in this way actively favoring rating inflation. Finally, empirical literature so far has not assessed potential differences in the impact of distorted incentives for different CRAs or different tranches of the same deal (e.g. Calomiris, 2009). We set our analyses at tranche level and show that the two main CRAs, Moody s and S&P, were exposed to rating shopping and distorted incentives to different extents. Ex-ante rating inflation fully translated in downgrades for S&P but only partially for Moody s; moreover, tranches rated solely by S&P appeared to receive ceteris paribus higher ratings, a result that does not hold for Moody s. Both agencies however appear to have assigned ceteris paribus higher ratings to their bigger clients, a result coherent with those by He, Qian and Strahan (2012).

5 5 Our paper proceeds as follows. Section 2 contains the motivation of this work in light of existing literature. In Section 3 we present the dataset, while in Section 4 empirical strategies and results are illustrated. Section 5 closes with some final remarks and conclusions. 2. Motivation 2.1. Credit Ratings Agencies and the subprime crisis Credit Rating Agencies have been considered by many observers as one of the main culpable for the subprime crisis started in 2007, up to the point that for the first time in their more than secular history they have been sued for the consequences brought by their misleading evaluations. 1 CRAs have being accused of consciously assigning too high ratings in order to favor issuers; the rationale for this speculation lies on the fact that CRAs are paid for their evaluations of fixed income products by the same firms who originate and distribute them, creating a clear conflict of interest. In the view of CRAs, this distorted incentive should be more than counterbalanced by their concerns about their reputation. Few recent theoretical model (e.g. Bolton, Freixas and Shapiro, 2012; Mathis, McAndrews and Rochet, 2008; and Camanho, Deb and Liu, 2012) explain how the issuerpaying scheme might have led to short-term maximization profits targets to overcome long-term reputational concerns for CRAs, especially for structured finance products; as stressed by Stolper (2009) and Bar-Isaac and Shapiro (2012), given the complexity of these securities, ratings alignment can make it difficult to identify misbehavior by a specific CRA. Short-term profits maximization is pursued by assigning inflated ratings; using data from an anonymous top CRA, Griffin and Tang (2012) show that the practice to qualitatively adjust upward Collateralized Debt Obligations (CDOs) ratings behind quantitative model predictions was indeed fairly common. Ashcraft, Goldsmith-Pinkham and Vickery (2010) show that, at the very least, ratings were not incorporating all available information. Benign evaluations are expected especially for securities 1 ONTAINS+'ANV'

6 6 issued by big firms, both because they are CRAs best clients (He, Qian and Strahan, 2012) and because they are more likely to know how to gamble their evaluations methodologies (Mählmann, 2011). Rating inflation is also exacerbated by the so called rating shopping phenomenon - i.e. issuers asking different CRAs for ratings and making public only the more favorable ones (e.g. Skreda and Veldkamp, 2009; Sangiorgi, Sokobin and Spatt, 2009). Indeed, Ashcraft, Goldsmith-Pinkham and Vickery (2010) and Benmelech and Dlugosz (2010) find that the number of CRAs rating a structured finance product is inversely related to its likelihood of default during the crisis. Empirical evidence produced so far on potential CRAs misbehavior in the structured finance market relies, with the notable exception of Griffin and Tang (2012), on either market prices (e.g. He, Qian and Strahan, 2012) or ex-post information on defaults (e.g. Ashcraft, Goldsmith-Pinkham and Vickery, 2010). Market prices allow to verify the presence of a premium associated with certain features, but requires the implicit assumption that markets themselves are not biased in their evaluations. Studies focusing on ex-post default to assess ex-ante misevaluations while directly assessing ratings evaluations might be instead affected by backward looking prediction bias. Previous empirical studies always analyze issues at deal level but, as stated by Calomiris (2009) among others, rating inflation and rating shopping characterize different tranches of a certain deal to different extents. Moreover, rating shopping phenomena are often accounted for by simply including the number of rating agencies among explanatory variables; however, different CRAs might give a different weight to short-term profits maximization versus long-term reputational concerns depending on their market share and reputational capital, as argued by Mathis, McAndrews and Rochet (2009). Studies on corporate bonds ratings have indeed shown that potential differences among the main CRAs (Moody s and S&P) can exist with this respect (e.g. Covitz and Harrison, 2003). Calomiris (2009) reports Commercial Mortgage-backed Securities (CMBS) non-senior tranches were more likely to receive a rating by S&P and/or Fitch but not by

7 7 Moody s; he interprets this piece of empirical evidence as proving that the latter agency was less favored by issuers seeking inflated evaluations via rating shopping. In this work we thus want to provide further empirical evidence of rating inflation and rating shopping using a slightly different empirical approach described in Section than previous literature. By estimating ex-ante relative rating biases at tranche level for each of the two main CRAs, we can study whether: a) ex-ante expected biases translated into subsequent downgrades; b) whether tranche rated by only one agency and those issued by bigger firms appeared to receive more inflated ratings, as rating shopping and issuer pay scheme arguments respectively suggest; and c) whether any difference between Moody s and S&P exists with this respect Security pricing and the regulatory role of ratings Giving CRAs inflated evaluations a central role in explaining the subprime crisis implicitly implies markets are not able to independently assess the credit risk of a security. Papers assessing the relationship between assigned ratings and prices for structured finance products (Ashcraft, Goldsmith-Pinkham, Hull, and Vickery, 2011, He, Qian and Strahan, 2011 and Mählmann, 2012) show how both tranches characteristics and ratings concurred in explaining spreads at issuance during the years preceding the subprime crisis. Two specular conclusions can be drawn by their empirical results: on the one hand, ratings appeared to contain relevant information for investors, thus driving securities prices even when controlling for other observable characteristics. On the other end, however, observable characteristics still influenced prices, meaning markets did not blindly rely only on CRAs evaluations alone but kept discounting observable characteristics. An interesting question thus emerge: to what extent rating inflation was actually beneficial for security issuers in terms of spread reduction if observable characteristics allow to spot a rating bias? By quantifying the ex-ante expected relative rating bias, we are able to test whether the joint effect of an higher rating and an identifiable rating upward adjustment actually resulted in an advantage for the issuer. If markets efficiently price rating inflation, we would expect the (predictable) rating

8 8 inflation negative effect on price to fully counterbalance the positive effect of an higher assigned rating. Viceversa, if markets did not fully discount predictable rating bias, the joint effect with an higher rating would still result in a lower spread. 2 Mispricing of ratings bias could happen not only because of less than fully efficient financial markets, but also as a consequence of the regulatory use of ratings. Most of theoretical works on CRAs behavior listed in Section 2.1 model the buy-side for structured finance products as noise, uninformed or at least less than fully rational investors. On the contrary, Opp, Opp and Harris (2012) model a situation where - despite fully rational agents - rating inflations still happens because it benefits investors. A rating over a certain threshold can in fact allow investors subjected to rating-based regulations to buy a product that would have been precluded to them otherwise. An higher rating means lower capital requirements for banks and insurance companies, thus increasing the attractiveness of the security. All in all, due to regulatory use of ratings, a security receiving a more benign evaluation is more valuable for a regulated investor than an otherwise identical one with a lower rating, even if markets are aware that the two are characterized by the same level of credit risk. As such, it is fully rational for financial markets to consider an higher evaluation by a CRA as a valuable feature by itself. This is particularly relevant as institutional investors subjected to rating-based regulations are a big stake of the market: for example, Campbell and Taksler (2003) report regulated investors hold around half of all corporate bonds. Indeed, few recent papers (e.g. Bongaerts, Cremers and Goetzmann, 2010; Kisgen and Strahan, 2010; and Becker and Milbourn, 2011) show the relevance of regulatory use of ratings in determining their effect on price for the corporate bond market. The relevance for the structured finance market during the 200s might have been even higher: IMF (2008) reports hedge funds accounted for only 19% of the total exposure to U.S. subprime by the end of 2006, while banks (including investment banks) alone accounted for 51.1%. With the bulk of 2 Again, it is worth stressing that we assess relative rating inflation; of course, on an ex-post basis basically all evaluations of mortgage-backed products were inflated before 2007.

9 9 the buy-side subjected to rating-based regulation, it is thus fair to assume that ratings can have a strong material impact on prices that goes beyond their role as signals of credit quality. Moreover, low interest rates makes investors keen to take additional risk to keep a desired level of expected absolute rate of return (e.g. Bekaert, Hoerova and Lo Duca, 2012); inflated ratings allow to increase risk profile behind what officially permitted, or to reduce capital requirements for a given level of credit risk in exchange for accept lower risk-adjusted returns. The regulatory use of ratings could thus explain a only partial pricing of easy-to-spot rating inflation: on the one hand, a security with the same level of credit risk of another one but a better rating will pay a lower spread just because regulated investors seeking higher absolute expected returns are allowed to buy it, creating an higher demand pressure; on the other hand, there would be no point for them in preferring it to a security with the same rating but a lower credit risk if they were to pay the same spread. For the same reason, we should expect a regulatory effect in the mispricing of rating bias only when the tranche is rated Investment Grade. Speculative securities are usually bought by non or poorly regulated investors allowed to take any level of risk: IMF (2008) reports that by mid-2007 only around 20% of structured finance products held by hedge funds were senior tranches; for banks, the proportion was about 60%. With no regulatory-based advantage of getting an higher ratings, we would thus expect fully predictable relative rating bias to be correctly priced for SG securities. 3. Dataset Given the aim of this study, we build our dataset by searching for all US tranches available in Bloomberg as on August 6 th 2011 labeled as ABS/CMO HOMEEQ which have been issued between 1992 and 2007 and that experienced a downgrade or have been placed on a watchlist by Moody s and/or S&P since June 1 st We choose to set the start of the crisis on June 2007 because on the 22 nd of that month Bear Sterns pledged up a USD 3.2 billion loan in a desperate attempt to save one of its hedge funds invested in structured finance products; the biggest attempt to rescue an hedge fund since Long Term Capital Management in 1998 unveiled once for all the

10 10 gravity of the situation. 3 The focus on Home Equity Loan issues is justified by their relevance among non-prime residential mortgage-backed securities and asset-backed securities in general before the crisis started; Benmelech and Dlugosz (2010) for example estimate 54% of structured finance products downgrades in 2007 were on HEL issues. The focus on a specific category of issues backed by residential mortgages allows to work on a fairly homogenous sample of products, while controlling for several characteristics likely to influence the phenomena under study. As shown in Panel A of Table 1, we identify 9,427 tranches representing 1,782 deals, for a total notional amount of almost USD 584 billion. [Insert Table 1 about here] Looking at the temporal distribution of our sample, few interesting patterns emerge; first, it is easy to notice the tremendous growth in both the number of deals and their notional values in the years leading to the start of the crisis; issues from 2005 onward represent almost 65% of the sample in USD amount (378,320 millions) and more than 60% in terms of number of tranches. Second, the market share of the two main CRAs has evolved over time; up to 2000, less than 58% of the tranches were rated by both Moody s and S&P. When only one CRAs evaluation was present, issuer expressed a clear preference for Moody s, who rated almost 70% of Non Jointy Rated (NJR) tranches. As the market developed, the percentage of Jointly rated (JR) securities increased to more than 80%, with a peak of 95% in 2003; at the same time S&P became the main agency for NJR securities, rating almost 83% of them in 2005 (15.17% of total 2005 vintage tranches). Benmelech and Dlugosz (2010) notes that tranches rated solely by S&P were, ceteris paribus, more likely to default by 2008; this piece of evidence jointly with this shift in market shares suggest that S&P might have been more exposed to rating shopping than its main competitor. Panel B of Table 1 shows the dataset distribution by depositor. As already underlined by SEC (2008) and He et al. (2012) among others, the market showed an high degree of concentration: we find that, out of the 3 $3.2 Billion Move by Bear Stearns to Rescue Fund. The New York Times, June 23, 2007.

11 depositors in our dataset, the first 10 in terms of issued tranches account for more than 52% of the notional amount and more than 46% in terms of number of deals. Table 2 reports the rating distribution at the beginning and at the end of our sampling period. Data draw a clear picture of the effect of the crisis on issues evaluations; while on June 2007 more than 35% of tranches were rated triple A by Moody s and/or S&P and more than 95% were considered Investment Grade (i.e. Baa/BBB or higher), by the end of July % (49%) of tranches rated by Moody s (S&P) were assigned a triple C or lower rating. [Insert Table 2 about here] In order to analyze the determinants of ratings deterioration of these securities as well as of their pricing at issuance, we use several customary tranche and deals characteristics collected from Bloomberg, whose descriptive statistics are reported in Table 3. [Insert Table 3 about here] FICO Score is the value-weighted average FICO Score of borrowers on underlying mortgages (FICO Score); the median value for all securities in our sample is 624, which is in line with the median value for all securitized subprime mortgages (626) as reported by Ashcraft et al. (2010). Since only 4,799 out 9,427 tranches in our sample have a reported value for FICO Score, we follow in our analyses Ashcraft et al. (2010) by including an indicator equal to 1 when FICO Score is missing and 0 otherwise. Amount is the tranche notional amount expressed in USD millions; while the average face value is around 60 millions, the biggest tranche has a reported value of more than 2.5 billions. Credit Support is the level of tranche overcollateralization expressed as a percentage of its notional value measured at issuance. WAL is the value-weighted average loans life by the time the tranche is issued expressed in months. To account for other credit quality characteristics at mortgage level we do not observe, we use the weighted average of annual interest rates, expressed as a percentage of face value, paid on underlying mortgages at origination (WAC). Following He et

12 12 al. (2012), we also include an Herfindahl-Hirschman index to measure geographical concentration on the collateral pool (Geo HH); the index ranges from 0 to 1 and it is computed summing the squared percentage (expressed in decimals) for each of the three mean geographical area represented in each pool. Table 3 reports statistics for both the whole sample and the subsamples of jointly and not-jointly rated securities; it is worth noticing how NJR tend to show a significantly (at 1% level) lower credit quality of the collateral at origination than JR counterparts, as proxied by a lower mean FICO Score and Credit Support and an higher mean WAC. The price at issuance is measured using the spread in basis points paid over the reference index (Spread); to keep pricing analyses homogenous, we consider only tranches paying a spread over 1 month LIBOR, which represent 74% of our sample. While the mean spread was around 1%, the median value in our sample is only 59 bps, indicating bulks of tranches were paying low interests while few were considered as highly speculative. Moody s (S&P) downgrade is the difference between rating on August 1 st 2011 and June 1 st 2007 assigned by Moody s (S&P). As customary, we measure downgrades by converting the alphanumeric rating scale in notches (i.e. Aaa/AAA =1, Aa1/AA+ = 2 and so on) and computing the difference in ratings expressed numerically. As seen in Table 2 tranches have experienced harsh downgrades, with Moody s (S&P) evaluations going down on average by 12 (9) notches. The two CRAs downgrades appears quite different when we look at differences between JR and NJR tranches; while S&P applied significantly (at 1% level) harsher downgrades on NJR securities, for Moody s the mean downgrade in the JR subsample is more than 3 notches higher than for the NJR subsample, also significant at 1% level. We will investigate more thoughtfully this difference in the next section controlling for tranches characteristics, and limit here to underlying that this preliminary piece of evidence again points toward S&P appearing more exposed to rating shopping than Moody s. Definitions of all variables are summarized in Table I of Appendix A.

13 13 Together with characteristics at tranche or collateral level presented above, we use in our analyses a set of indicators for the deal structure and tranche typology presented in Table 4. [Insert Table 4 about here] As predicted by Skreta and Veldkamp (2009) theoretical model, NJR tranches tend to be associated with more complex structures, such as Principal-Only/Interest-Only and Sequential Payers, while plain Pass-through securities are more commonly rated by both CRAs. Also, coherently with Calomiris (2009) empirical evidence on Commercial Mortgage-Backed Securities, Subordinated tranches are more commonly rated by only one CRA (Difference in frequency among the two subsamples is 11%, significant at 1% level), confirming the importance of studying rating shopping phenomena and their consequences at tranche level. 4. Empirical evidence 4.1. Predictable rating bias and subsequent downgrades One of the main problems in analyzing CRAs behavior in this context is that, as underlined by Stolper (2009), it is not easy to disentangle the effect of misapplied ex-ante high ratings from that of an exogenous unpredictable shock to explain why many securities rated as investment grade have experienced such poor performances. To address this problem, we follow Ashcraft, Goldsmith- Pinkham and Vickery (2010) and combine ex-ante (i.e. before the crisis) and ex-post (i.e. after the crisis) empirical evidence. By showing that tranches more likely to have received an inflated rating at issuance tend to experience, ceteris paribus, harsher downgrades ex-post we can provide empirical evidence of CRAs having inflated some of their evaluations with no grounding for that. In order to directly assess ex-ante rating inflation at tranche level, we estimate an ordered LOGIT model for the rating assigned at issuance by each CRA, using as explanatory variables the tranche characteristics at issuance presented in Section 3. A similar approach has been used for corporate

14 14 bonds rating by Bhojraj and Segupta (2003) and Cremers, Nair and Wei (2007) among others. Coefficients estimates are reported in Table 5. [Insert Table 5 about here] From these models we are able to get estimated probabilities of each rating level for each tranche given its observable ex-ante characteristics; we thus compute an ex-ante measure of Relative Expected Rating Bias as in Equation 1: = ( ) (1) Where r is a rating expressed in notches (for the Alphanumeric rating scale, where r ranges from 1 to 21), or in a numeric scale from 1 (Aaa/AAA) to 7 (Caa/CCC or below) using Alphabetic ratings as in Table 2; R a is the maximum numerical value assigned in either the Alphanumeric rating scale (21) or in the Alphabetic rating scale (7); p r is the estimated expected probability of receiving a rating by Moody s or S&P given tranche characteristics. Expected rating probabilities for each CRA and rating scale are computed using Models (1) to (4) presented in Table 5. Rating is the actual evaluation assigned by either Moody s or S&P. Given this definition, Rating Bias ranges from -7 (-21) to +7 (+21) using the Alphabetic (Alphanumeric scale). It is important to stress that, by construction, Rating Bias is a measure of relative rating inflation; it does not capture the average upward rating bias that all tranches might have experienced during the booming years, but can quantify differences in assigned ratings to single tranches that cannot be explained by observables. Once we have a measure of fully predictable ex-ante relative rating bias, it is possible to see how this measure relates to ex-post downgrades; we thus estimate models for the determinants of rating revision including Rating Bias among the covariates. To control for the initial rating class and for the age of tranches by the time the crisis occurred, we include ratings (at Alphabetic level) and vintage indicators. Since the analysis focus on downgrades, we exclude from all analyses tranches

15 15 that were already assigned a rating of Caa/CCC or below before the advent of the crisis. 4 Coefficients estimates are reported in Table 6. [Insert Table 6 about here] For both Moody s and S&P, an ex-ante predictable relative rating inflation resulted in a significantly (at 1% confidence level) higher subsequent downgrade. For Moody s, a predicted rating bias by 1 level in the alphabetic scale (e.g. a tranche predicted to be a Baa but having a single A rating) has on average translated into a 0.8 notches harsher downgrade, ceteris paribus; using the alphanumeric scale, a one notch expected rating inflation ex-ante resulted in a 0.3 notches bigger downgrade ex-post. The effect is even stronger for S&P, where a 1 level of alphabetic (alphanumeric) class rating inflation resulted in a 2.8 (0.9) notches harsher downgrade, as Models (3) and (4) estimates show. We perform a number of robustness checks for these results, reported in Table I of Appendix B; first, we include controls for tranche characteristics and deal structure presented in Tables 4 and 5. Second, we use as an alternative proxy for rating bias the Probability of Inflated Rating, computed as in Equation 2. =!"#$ (2) Where R a Rating and p r are as previously defined. Probability of Inflated Rating is thus the estimated probability that, given its characteristics, a tranche has received an upward rating adjustment. Again, this measure only captures relative differences in the evaluation of otherwise similar issues, and does not account for the potential average bias of all tranches. The main result 4 In unreported analyses we repeated all the estimations including also Caa/CCC or below rated securities; all results are qualitatively similar.

16 16 that ex-ante predictable rating bias resulted in higher downgrades once the crisis occurred is confirmed for both CRAs. 5 Table 6 also reports in square brackets the t-statistics for a test on Rating Bias coefficients in Models (2) and (4) being different from 1; while for Moody s the coefficient is significantly (at 1% confidence level) lower than 1, for S&P we cannot reject the hypothesis that a 1 notch of ex-ante rating inflation has in fact fully translated in 1 notch of ex-post rating downgrade; this means that non observable characteristics on average could not even partially justify differences in evaluations assigned to otherwise similar issues. Our measure of predictable rating inflation is useful also to assess to what extent rating biases were driven by CRAs distorted incentives, as previous literature suggests. Ashcraft, Goldsmith-Pinkham and Vickery (2010) and Benmelech and Dlugosz (2010) show for example that tranches rated by fewer CRAs were more likely to default, an indirect evidence of rating shopping; our approach allows to directly look also at whether the presence of only one agency resulted in a (relatively) inflated rating at issuance, and to consider potential differences among CRAs with this respect. He, Qian and Strahan (2012) find that markets discounted the issuer dimension in the security price, suggesting their securities evaluation were expected to be more likely inflated. Again, our approach allow us to see whether this was indeed the case and directly assess whether the issuer dimension was in fact positively correlated with a stronger relative rating inflation. In Panel A of Table 7 are reported the estimated coefficients for ordered LOGIT models, where the dependent variable represents the rating at Alphabetic level assigned by either Moody s (Models (1) to (3)) or S&P (Models (4) to (6)). Along with all the determinants included in Table 5, we add an indicator (NJR) equal to 1 when either Moody s or S&P (but not both) rate the issue and 0 otherwise; to proxy for the issuer dimension, we use the natural logarithm of 1 + the notional 5 Analyses reported in Table I of Appendix B use Rating Bias and Probability of Inflated Rating computed at Alphanumerical level, estimating rating probabilities using coefficients of Models (2) and (4) of Table 6. Results using rating inflation proxies computed at Alphabetic level are qualitatively similar.

17 17 amount (in USD millions) of tranches by the same depositor in our sample (Ln (1+ Amount Depositor)). Panel B reports marginal effects for the two variables of interest, computed keeping all other variables equal to their sample median. Two interesting results emerge: first, tranches issued by a bigger depositor are more likely to receive an higher rating, ceteris paribus. An increase of Ln(1 + Amount Depositor) from 9.07 (it s 25 th sample percentile) to (it s 75 th sample percentile) augments by more than 5% the likelihood of receiving a triple or double-a rating by Moody s, and by more than 4% for S&P. The second - and more interesting - result is that NJR also increase the likelihood of an higher rating but only for S&P; beside statistical significance, the result is also economically relevant: for all other variables equal to their sample median values, NJR tranches have a 23% higher probability to receive a AAA or AA rating by S&P than JR tranches, while for Moody s no appreciable difference between JR and NJR tranche is present. We thus find empirical evidence that, as suggested by Calomiris (2009), not all CRAs were exposed to rating shopping to the same extent. [Insert Table 7 about here] Client Dimension and NJR could in principle proxy for unobservable characteristics actually justifying differences in ratings; again, we assess this potential concern by providing ex-post evidence of the correlation between expected rating inflation and downgrades. We use Rating Bias computed using rating probabilities estimated from Models (1) and (2) of Table 5. It is worth pointing out that, since nor NJR nor Ln(1 + Amount Depositor) either are included in those two models, Rating Bias reflects all the rating error relating to the number of rating agencies or the depositor dimension. We thus estimate a model for the determinants of downgrades including, along with Rating Bias, the cross product of the latter with NJR: if NJR tranches higher rating was actually justified ex-ante, we should expect the cross product to have a negative impact on the expost rating downgrade. The same empirical strategy applies for the dimension of the depositor; for

18 18 the sake of making results more easily interpretable, we use a dummy (Big Client) equal to 1 when Ln(1 + Amount Depositor) is higher or equal to its sample median value and 0 otherwise. Table 8 reports estimation results: the correlation between ex-ante rating bias and ex-post rating downgrades by either Moody s or S&P is not diminished for NJR tranches or tranches issued by a bigger depositor. We thus have evidence that rating inflation was particularly an issue for tranches rated only by S&P and/or issued by bigger firms, confirming that distorted incentives played a significant role in the misrepresentation of structured finance products credit quality by CRA before the advent of the crisis. [Insert Table 8 about here] 4.2. Ratings and market pricing To study whether markets correctly discounted rating inflation at least in relative terms, we build a pricing model for securities including as determinants both indicators of assigned rating class at issuance and the ex-ante rating bias at Alphabetic level, computed as in Equation 1 using coefficients estimates of Models (1) and (2) of Table 5. We also include market controls for the general level of prices by the time of issuance: to account for the term-structure of interest rates, we include the values of the three principal components of the US Government yield curve, usually thought of as representing the Level, Slope and Curvature of the yield curve (Litterman and Scheinkman, 1991), as in the settlement day; we also include the difference between average yields on corporate bonds rated Aaa and Baa by Moody s as a proxy for the Default Premium asked by the market (e.g. Chen, Collin-Dufresne and Goldstein, 2009). All yields data are from the Federal Reserve website. 6 OLS coefficients estimates for our pricing models are reported in Table 9. [Insert Table 9 about here] 6

19 19 The dependent variable is the spread in basis points over Libor. Models (1) and (3) include for Moody s and S&P respectively, as previously defined. Markets appears indeed to discount relative rating inflation associated to observable characteristics: ceteris paribus, a 1 class expected rating bias result in a 19 bps higher spread using Moody s ratings and 27 bps higher using S&P evaluations. In Models (2) and (4) we include the cross product of Rating Bias with rating class indicators to see how this is priced for different grades issues. We then compute for each rating class the marginal overall effect of a fully predictable 1 class rating bias as in Equation 3 % &'! = ( )_!"#+!"# -!. + ( )_!"#+ ( )_!"#+01 (3) Where ( are Models (2) and (4) coefficients estimates reported in Table 9 and _! (!$) is an indicator for the tranche being given an Alphabetic rating ( + 1, where + 1 is a lower rating class than ). The marginal effects thus measure the change in price driven by a 1 class rating increase when this can be fully accounted for as an upward bias. The computed marginal effects for each rating class for Moody s and S&P are reported in Columns (5) and (6) of Table 9 respectively. For both CRAs and all Investment Grade rating levels, the positive effect of an higher rating more than counterbalance the negative effect associated with an evident upward adjustment; for example, a tranche rated triple A by Moody s pays on average a spread 27 bps lower than an otherwise identical tranche rated double A, even if this relative rating bias is fully evident. For S&P the effect is somehow smaller but still economically and statistically significant: a one class inflated triple A by S&P still pays at issuance a 13.5 bps lower spread than an identical double A security. On the contrary, Speculative Grade (SG, i.e. Ba/BB and below) tranches prices exhibit a null marginal overall effect of rating bias, meaning the predictable rating relative inflation was fully discounted for by the market. As treated in Section 2.2, this result is highly suggestive of mispricing being partially explained by the regulatory use of ratings. To further investigate whether the buy-side was simply mislead by rating evaluations or played a conniving role in the market bubble, we study how ratings influenced price at issuance when

20 20 discordant evaluations were present; we would expect a conscious market to err on the side of caution in case of split evaluations, giving an higher weight to the more conservative rating (e.g. Liu and Moore, 1987); on the contrary, a conniving market would rely more on the highest evaluation, giving further incentives to issuers to try to fool the rating system and get an inflated evaluation by at least one agency. To study how markets priced split ratings, we follow the approach proposed by Cantor, Packer and Cole (1997); first, we estimate a pricing model for HEL securities using only tranches jointly rated by Moody s and S&P receiving exactly the same evaluation by the two agencies. The pricing model includes as explanatory variables: a) rating indicators (at Alphanumeric level); b) all issue characteristics and structure indicators presented in Tables 3 and 4; c) geographical controls on the collateral, representing the percentage of loans originated in each state; d) the 4 controls for market interest rates and prices in the settlement day (Level, Slope, Curvature and Default Premium). Estimated coefficients for the pricing model are reported in Table II of Appendix A. Second, we use the previously estimated model to compute out-of-sample pricing errors (3) for JR tranches with split evaluations as in Equation 4, 3! = 4! 46 5 (4) Where 4! is the tranche i spread over 1-month Libor and 46 5 is the predicted spread for tranche i using rating rule r, where r is either equal to h (i.e. using the higher assigned rating) or l (i.e. using the lower rating). To compare the performance of the high rating and low rating pricing rules, we then construct three statistics - Mean Bias (7), Mean Absolute Error (MAE) and Mean Absolute Error Adjusted (MAEA) defined as in Equation (5) to (7) 9 7 = 3!! 8 (5)

21 21 9 %:& = 3! 8! (6) 9 %:&: = 3! 7 8! (7) Where N is the number of split evaluation tranches. If markets act with caution we would expect %:& < < %:& >, i.e. in case of split ratings prices are more driven by the more conservative rating. Following Cantor, Packer and Cole (1997) we also confront the two pricing rule adjusting for their known mean bias (7 ) using the Mean Absolute Error Adjusted statistics; again, we expect %:&: < < %:&: > if markets behave in a (relatively) prudent fashion. Table 10 reports the computed statistics for the two pricing rules (Low and High rating) for the whole sample of split rating JR tranches, as well as for two subsamples of tranches issued a) before 2005 and b) in the period; the latter represents the booming years of the market for subprime mortgages securitization leading to the 2007 crisis (e.g. He, Qian and Strahan, 2012). Over the whole sampling period, more conservative ratings were indeed driving prices at issuance: MAE l is bps while MAE h is 37.02, the difference being statistically significant at 1% confidence level. The difference is somehow smaller when adjusting for the known pricing error (30.55 bps against 33,58 bps), but still statistically significant at 1% level. However, in the booming years results are overturned: the pricing rule based on the higher ratings leads to a Mean Absolute Error 5.16 bps lower, a difference statistically significant at 1% confidence level. Again, results are somehow weaker in economic significance when adjusting for 7, but still statistically significant at 1% confidence level. [Insert Table 10 about here] We thus have strong empirical evidence that markets played a crucial role in allowing for inflated evaluations by CRAs; first, spread at issuance did not fully adjusted for fully predictable ex-ante

22 22 rating bias for IG tranches, but did for SG tranches, a result highly consistent with buy-side favoring rating inflation due to the regulatory use of ratings. Moreover, we show that during the market booming years 2005 and 2006 prices were following the higher rating in case of split evaluations, again suggesting that rating inflation have been allowed by a too aggressive buy-side. 5. Conclusion In this paper we study the role played by regulatory use of ratings, distorted incentives, and rating shopping in the disruption of the mortgage-backed structured finance market that resulted in the crisis started in By modeling the determinants of ratings assigned before the crisis started, we are able to compute a purely ex-ante measure of relative rating inflation. Tranches exhibiting a stronger rating upward bias experienced harsher downgrades once the crisis occurred, confirming ex-post that non-observable characteristics could not explain observed differences in ex-ante evaluations of otherwise identical issues. Upward relative rating biases were more likely for securities issued by bigger firms and those rated only by S&P, but not those rated only by Moody s. Our results thus confirm those of previous literature focusing on ex-post default rates (e.g. Ashcraft., Goldsmith-Pinkham and Vickery, 2010) or securities pricing (e.g. He, Qian and Strahan, 2012); however, by directly studying the determinants of assigned ratings at tranche level, we are able to provide evidence that, as suggested by Calomiris (2009), rating shopping did not affect in the same way all CRAs. While we show relative rating biases among similar securities were identifiable ex-ante, we provide empirical evidence they were fully priced at issuance by the market only for Speculative Grade (SG) tranches, but not for Investment Grades (IG) tranches. As the classification as SG or IG has a material effect on the attractiveness and eligibility of securities for investors subjected to ratingbased regulations, we interpret this as empirical evidence of the role played by the regulatory use of rating in favoring rating inflation, as theorized by Opp, Opp and Harris (2012). We also find that,

23 23 in presence of split evaluations by Moody s and S&P, the more benign rating was driving prices in the booming years ; this confirms that the buy-side played an active role in allowing the disruption of ratings quality and wasn t just fooled by CRAs misevaluations. Our results have relevant implications for the ongoing debate on the regulation of the rating industry. First, we provide new empirical evidence of the role played by distorted incentives for CRAs in the years leading to the 2007 crisis, adding to existing literature calling for a reform of the rating industry regulation. One of the most debated feature of the CRAs industry is the issuer-paid scheme: Strobl and Xia (2012) for example provide empirical evidence in support of less inflated evaluations when an investor-paid scheme is in place, supporting the theoretical conclusions of Mathis, McAndrews and Rochet (2009). Our results however provide empirical support to Opp, Opp and Harris (2012) argument that rating inflation is also favored by distorted incentives for investors. In order to minimize the effect of distortions in the rating industry, future reforms should thus focus on the role played by incentives for the buy-side as much as on incentives for CRAs.

24 24 References Ashcraft, A., Goldsmith-Pinkham, P. and Vickery, J. (2010). MBS ratings and the mortgage credit boom Working paper, Federal Reserve Bank of New York. Ashcraft, A., Goldsmith-Pinkham, P., Hull, P. and Vickery, J. (2011). Credit ratings and security prices in the Subprime MBS market. American Economic Review, 101, Becker, B. and Milbourn, T. (2011). How did increased competition affect credit ratings? Journal of Financial Economics, 101, Bekaert, G., Hoerova, M. and Lo Duca, M. (2012). Risk, uncertainty and monetary policy. Working Paper. Benmelech, E. and Dlugosz, J. (2010). The credit rating crisis. NBER Macroeconomics annuals, 24, Bar-Isaac, H., and Shapiro, J. (2012). Ratings quality over the business cycle. Journal of Financial Economics, forthcoming. Bhojraj, S., and Sengupta, P. (2003). The effect of corporate governance mechanisms on bond ratings. Journal of Business, 76, Bolton, P., Freixas, X. and Shapiro, J. (2012). The credit ratings game. Journal of Finance, 67, Bongaerts, D., Cremers, K. and Goetzmann, W. (2012). Tiebreaker: certification and multiple credit ratings. Journal of Finance, 67, Calomiris, C. W. (2009). The debasement of ratings: what s wrong and how we can fix it. Working Paper. Camanho, N., Deb, P. and Liu, Z. (2012). Credit rating and competition. Working Paper. Campbell, J. Y. and Taksler, G. B. (2003). Equity volatility and corporate bond yields. Journal of Finance, 58, Cantor, R., Packer, F. and Cole, K. (1997). Split ratings and the pricing of credit risk. Federal Reserve Bank of New York Research Paper No Covitz, D. and Harrison, P. (2003). Testing conflicts of interest at bond ratings agencies with market anticipation: evidence that reputation incentives dominate. Board of Governors and the Federal Reserve System, Finance and Economics Discussion Series, Cremers, K. M., Nair, V. B. and Wei, C. (2007) Governance mechanisms and bond prices. Review of Financial Studies, 20(5), Griffin, J. M. and Tang, D. Y. (2012). Did subjectivity play a role in CDO credit ratings? Journal of Finance, forthcoming. He, J., Qian, J. and Strahan, P. (2012) Are all ratings created equal? The impact of issuer size on the pricing of mortgage-backed securities. Journal of Finance, forthcoming. Kisgen, D. and Strahan, P. (2010). Do regulations based on credit ratings affect a firm's cost of capital? Review of Financial Studies, 23,