Do Multiple Credit Ratings Signal Complexity? Evidence from the European Triple-A Structured Finance Securities

Transcription

1 Do Multiple Credit Ratings Signal Complexity? Evidence from the European Triple-A Structured Finance Securities March 2014 Frank J. Fabozzi EDHEC-Risk Institute Mike E. Nawas Bishopsfield Capital Partners Dennis Vink Nyenrode Business Universiteit

2 Abstract In much of the current research on market practices with respect to the use of credit ratings, the rating shopping hypothesis and the information production hypothesis feature prominently. Both of these hypotheses predict an inverse relationship between the number of ratings and a security s funding cost; that is, more ratings will reduce funding costs and, conversely, fewer ratings will increase funding costs. Our study finds precisely the opposite to have been the case for the mainstay of the structured finance securities market in Europe prior to 2007, namely the triple-a tranches of European residential mortgage-backed securities. Our findings suggest that structured finance markets may behave differently than what would be predicted by two hypotheses traditionally used to explain the number of ratings and funding costs: the rating shopping and information production hypotheses. Obtaining multiple credit ratings may be a signal for complexity, for which investors demand a risk premium. JEL Classifications: G12; G24; L11. Keywords: Credit ratings, regulation, mortgage-backed securities. EDHEC is one of the top five business schools in France. Its reputation is built on the high quality of its faculty and the privileged relationship with professionals that the school has cultivated since its establishment in EDHEC Business School has decided to draw on its extensive knowledge of the professional environment and has therefore focused its research on themes that satisfy the needs of professionals. 2 EDHEC pursues an active research policy in the field of finance. EDHEC-Risk Institute carries out numerous research programmes in the areas of asset allocation and risk management in both the traditional and alternative investment universes. Copyright 2015 EDHEC

3 1. Introduction The recent global financial crisis that began with the subprime mortgage problem in the United States in the summer of 2007 has called for a closer examination of market practices with respect to the use of ratings assigned to structured products assigned by credit rating agencies (CRAs). This has led to a well-established body of research in which two hypotheses feature prominently. The first is the rating shopping hypothesis, which asserts that underwriters of securities may shop for the most favourable (combination of) ratings in order to minimise funding cost. The second body of research involves the information production hypothesis, which puts forward the notion that funding costs can be minimised by increasing the number of credit ratings on a given security because this adds to the information available to potential investors in the security. These two hypotheses both predict an inverse relationship between the number of ratings and a security s funding cost; that is, more ratings will reduce funding cost and, conversely, fewer ratings will increase funding cost. Regulatory and supervisory bodies responsible for overseeing financial markets have picked up on these research findings. In Europe, where public concern about the social utility and risks of structured products in particular has been intense and widespread, European Union (EU) policymakers have adopted new regulations aimed at increasing the information content of structured finance securities. 1 Under these new rules, issuers are required to engage at least two CRAs to rate structured finance securities. The reason that authorities singled out structured finance securities for requiring more than one rating is that due to their complexity, they see a need to provide investors with more information on these securities. 2 In this paper, we first test whether prior to the 2007 crisis the inverse relationship between the number of ratings and funding cost that would be predicted by the rating shopping and information production hypotheses was empirically valid for the mainstay of the European structured finance market: the triple-a rated portion of residential mortgaged-backed securities (RMBS). We find that the predicted relationship by these two hypotheses is not supported. In fact, we find the opposite: for our sample we find evidence that the greater the number of ratings, the higher the funding costs. We suggest that our empirical findings may be explained by the complexity of the securities: in a market such as the European RMBS market there is a mix of relatively complex and relatively straightforward securities, and issuers may only find it necessary to obtain more than one triple-a rating for the more complex securities in order to issue securities at the lowest possible interest cost. In this view, the number of triple-a ratings obtained for a structured finance security is a signal for the complexity of the security because only securities with relative complexity warrant incurring the costs of engaging multiple CRAs. This is because with a fewer number of ratings the underwriters may struggle to find an adequate universe of investors to purchase the securities. In this view, investors perceive the number of triple-a ratings as a signal for complexity, for which they demand a risk premium. We test our complexity hypothesis in this paper in a number of ways. First, we test whether for structured finance securities funding costs are higher when they are rated triple-a by two or three CRAs rather than by one. Second, we test whether structured finance securities with complex features have more ratings than those without the complex features. Consequently, this paper provides two main contributions to the research on market practices with respect to the use of credit ratings. The first one is that in the European structured finance market, more ratings are associated with a higher funding cost. This contribution has global implications as it challenges the commonly held view that multiple ratings are obtained to decrease the funding cost (e.g. Skreta and Veldkamp (2009), Bongaerts, Cremers, and Goetzmann, (2012), and He, Qian, and Strahan (2012)). Our second contribution is that our results provide evidence that 1 - Regulation (EU) No 462/2013 of the European Parliament and of the Council of 21 May 2013 amending Regulation (EC) No 1060/2009 on credit rating agencies (CRA III). 2 - EU Press Release MEMO/13/571, 18 June

4 the complexity of structured finance securities is a driver for issuers to obtain a higher number of credit ratings, and, vice versa, that tranches with less complexity are associated with a lower number of obtained ratings. Both contributions have ramifications for the rating shopping view in that the rating shopping hypothesis would predict that securities with one or two triple-a ratings would have higher funding costs than securities with three triple-a ratings, to reflect the risk that in such cases an issuer may have shopped for the one or two particular CRAs who were prepared to assign a triple-a rating to the security and being unable to convince all three CRAs to do so. The results also relate and contribute to a several issues in the literature regarding the effects of complexity on the pricing of structured finance securities. Furfine (2014) finds that the underwriter effectively determines a deal s complexity and shows that loan performance is worse for loans packaged in more complex securitisations compared to less complex securitisations. In this respect, our results suggest that the issuer chooses more credit ratings for more complex deals. Interestingly, Bas-Isaac and Shapiro (2013) suggest that more complex deals imply a reduced accuracy of credit ratings on structured finance products in the period just preceding the financial crisis. Focusing on the actual credit performance of tranches, Griffin and Tang (2012) call into question the traditional rating shopping view and argue that tranches rated by multiple agencies actually perform worse than deals that are rated by one rating agency. Our empirical findings shed some light on the discussion, as they suggest that investors are aware of additional risks associated with multiple credit ratings. Our results imply that investors take into account the risk associated with more complex deals and perceive the number of ratings obtained by the issuer as a signal for complexity and demand a risk premium. For policymakers, our empirical results open up the possibility that the EU requirement of two credit ratings for structured finance tranches, which came into force on 20 June 2013, may lack effectiveness, as it does not focus on structured finance securities with complex features. Policymakers should be cognizant of the risk that a requirement of a minimum number of ratings may reduce the information content on the complexity of a security that could otherwise (i.e. without the requirement) be signalled by the number of ratings. The European Commission will revisit the mandate by 1 January That review will be followed by a report to the European Parliament and to the Council and may be accompanied by a further legislative proposal. 3 Our paper will help the European Commission in assessing whether or not its current position regarding the requirement of number of ratings is prudent and effective. The rest of the paper is organised as follows. In Section 2 we review the hypotheses on the impact of the number of credit ratings on the funding cost. Section 3 describes the data. Section 4 examines the impact of the number of triple-a credit ratings on the funding cost of an RMBS. Section 5 examines the relation between complexity and the number of triple-a ratings. We conclude in Section Hypotheses on the Impact of the Number of Credit Ratings on Funding Cost To understand the impact that the number of ratings could have on funding cost, it is important to understand the process by which credit ratings are obtained for structured finance securities. The originator of the security will typically request one or more of the CRAs to rate a planned transaction. Throughout the structuring phase of the security it remains at the originator s discretion whether to withdraw or maintain such a rating request. Prior to the issuance of the security, the originator would have obtained detailed feedback from a CRA it has approached for a rating, as to the rating that is likely be assigned given the characteristics of the collateral and the structure of the transaction Source: //

5 There are various hypotheses identified in the literature that offer an explanation for the selection of the CRA or CRAs to rate the tranches of a structured finance transaction and the decision as to the number of agencies to obtain a rating from. These hypotheses include the rating shopping hypothesis and information production hypothesis mentioned earlier. The rating shopping hypothesis asserts that an issuer will search (i.e., shop) for an additional credit rating for a debt obligation in order to decrease its cost of capital. Examples of research in this area include Sangiorgi, Sokobin, and Spatt (2009), Farhi, Lerner, and Tirole (2013), Bolton, Freixas, and Shapiro (2012), and Opp, Opp, and Harris (2013). One form of the rating shopping hypothesis relates to shopping for a higher rating, and some forms focus on shopping for as many ratings as possible even if at the same rating level. An important example of the first form, applied in the structured finance market, is the research by Skreta and Veldkamp (2009). The second form can be found in He, Qian, and Strahan (2012), who state that having a smaller number of ratings on a security will increase funding cost because investors price the risk that issuers shopped for the best rating when tranches have fewer than three ratings. By shopping, an issuer could censor out pessimistic ratings, thus reducing the number of ratings observed by investors. Since the focus in this paper is on triple-a rated tranches, we only test the second form of the rating shopping hypothesis, that is, where the effect of the number of ratings on funding cost is measured. This is because an issuer cannot shop for a better rating (the other form of the rating shopping hypothesis) if the CRA it obtains a rating from awards the highest (i.e. triple-a) rating to a tranche in the first place. Another body of research examines the information production of credit ratings. The information production hypothesis, suggested by Bongaerts, Cremers and Goetzmann (2012), asserts that multiple credit ratings may provide investors with more information regarding the credit risk of a debt obligation. More specifically, obtaining additional credit ratings reduces the uncertainty about the creditworthiness of a debt obligation: an extra rating in agreement with existing ratings would reduce credit quality uncertainty and thereby lower credit spreads. They do not find evidence to support this hypothesis using a sample of U.S. corporate bond issues from 2000 to The relevant rating shopping hypothesis (i.e. the hypothesis dealing with triple-a rated securities) and the information production hypothesis both suggest the same about the impact of more than one triple-a rating for a tranche on that tranche s funding cost: an inverse relationship. The arguments for the inverse relationship, however, are different. The rating shopping hypothesis suggests that investors will penalise securities that have one or two triple-a ratings rather than three triple-a ratings, 4 because the investors will price in the risk that in those cases the issuer may have shopped for the one or two CRAs that indeed were prepared to assign a triple-a rating to the security. The CRA whose rating is missing may have been unwilling to assign a triple-a rating and the issuer therefore may have chosen not to enter into (or to terminate an existing) contract with this reluctant CRA. So securities with one or two triple-a ratings will have higher funding costs than securities with three triple-a ratings. The argument of the information production hypothesis for the inverse relationship is that with every additional rating, even if at the same (triple-a) level, additional information about the security is produced, which reduces the risk associated with investing in the security, which then can be rewarded by a lower funding cost. An alternative to the rating shopping and information production hypotheses is that the decision by an issuer to use more than one triple-a rating is due to the complexity of the structure. We refer to this as the complexity hypothesis and suggest that multiple ratings, even when they are the same rating, may reflect the issuer s view of how the market would perceive the transaction in terms of the risk associated with its complexity. In this framework, a rational issuer would only 4 - Three rating are mentioned here because there are three major credit rating agencies which typically rate structured finance securities: Standard & Poor s, Moody s, and Fitch. 5

6 obtain multiple ratings if required, in order to reasure investors that the complexity of the security has been reviewed by more than one CRA. In this view, investors perceive the number of triple-a ratings as a signal for complexity, for which they demand a risk premium. There has been some research aimed at examining the potential impact of complexity of a security on the process of obtaining credit ratings. In particular Skreta and Veltkamp (2009) focused on the topic, suggesting that the incentive to shop for ratings may be greater with complex securities than with simple securities. They put forward a theory where for simple assets, agencies issue nearly identical forecasts. Asset issuers then disclose all ratings because more information reduces investors uncertainty and increases the price they are willing to pay for the asset. For complex assets, ratings may differ, creating an incentive to shop for the best rating. Their theory does not consider, though, whether or not issuers seek to obtain more credit ratings for complex securities than for simple securities, nor did the authors test their theory on complexity empirically. In addition to Skreta and Veldkamp (2009), several other papers analyze the effects of complexity on the pricing of securities. Carlin, Kogan and Lowery (2013) show that complexity affects both the liquidity and price volatility of assets, arguing that this might be explained by the likelihood that more computational errors are made when valuing complex assets. Arora et al. (2009) show that complex assets are difficult to price. By empirically testing our complexity hypothesis, which suggests a higher funding cost when securities have additional triple-a credit ratings, we automatically test the rating shopping and information production hypotheses that both suggest the opposite (a lower funding cost when there is more than one triple-a rating). Furthermore, our empirical findings build on and further develop an understanding of the impact of complexity on asset pricing as we investigate our complexity hypothesis, in detail, as follows. First, we examine in fact whether or not, and to what extent, multiple triple-a ratings lead to higher funding costs. We delve deeper into the relationship between multiple triple-a ratings and funding cost by testing whether the difference in funding costs of securities with two versus three triple-a ratings depends on any particular combination of the two CRAs that are retained. Second, we examine whether or not securities that have certain complex features are likely to have more triple-a ratings than securities without the complex features. 3. Data and Descriptive Statistics The only structured finance asset class within the European structured finance market that has had a sufficiently large number of issuances to conduct meaningful empirical analysis is RMBS. We collected the entire set of triple-a rated Euro-denominated RMBS issued at par between 1999 and 2006 as reported in Structured Finance International (SFI), a publication of Euromoney Institutional Investor Plc. The cut-off of 2007 was intentional for two reasons. First, regulatory bodies designed the requirement to have more than one credit rating on structured finance securities to address concerns with respect to market practices that led to the crisis in 2007, so the relevant dataset has to be up to Second, since the crisis, the number of new structured finance securities that are sold to investors 5 has dropped so dramatically that it impedes empirical analysis. We focus our analysis on triple-a tranches as this represents by far the largest part of the RMBS market and because in this market the lower rated tranches typically only exist by virtue of the need to provide credit support to a related triple-a tranche. That is, issuers in the RMBS market almost always create, for each transaction, more than one tranche of securities, ranked in order of seniority. Investing in the most senior tranche of a transaction carries less risk for an investor than investing in one of the junior tranches of the same transaction because the senior tranche is protected from credit losses by the junior tranches. Issuers typically evidence this lower risk Since the crisis of 2007 it has become common for issuers of structured finance securities not to sell its securities to third parties but to retain them as collateral for loans from the European Central Bank.

7 by contracting a CRA that is prepared to assign the highest possible credit rating to the senior tranche: triple-a. As a result of the lower risk to the investor, the funding cost to the issuer of the (triple-a) senior tranche is lower than the funding cost to the issuer of the (lower rated) junior tranches. For each transaction, the issuer tries to maximise the relative size of the (low cost) senior tranche compared to the (high cost) junior tranches. However, there is a limit to this maximisation. If this issuer reduces the relative size of the junior tranches too much, the CRA may become unwilling to assign a triple-a rating to the senior tranche, as they may decide that the senior tranche has lost too much of the junior tranche loss-protection. So the mix of senior and junior securities per transaction is a result of careful structuring and negotiation between the issuer (and its underwriters) with the CRAs. This way of structuring to a certain rating level is an important reason why the securities are called structured finance securities. In practice, in the European RMBS market almost all transactions are structured such that the senior tranche is by far the largest and is indeed rated triple-a by at least one CRA. SFI reports the spread at which each tranche is issued. There are two reasons why we use this spread measure rather than secondary market spreads. The first is that secondary market spreads vary continuously throughout a security s life and will be impacted by not only the rating but also by the actual performance of the collateral underlying the security (defaults and recoveries). A secondary market spread will reflect more information on the security s (actual and expected) performance than the credit rating, which in practice is not changed continuously: credit ratings are reviewed by CRAs periodically (under normal circumstances annually or semi-annually), not continuously. This problem does not exist in the case of new issuance spreads. The second reason for using the new issuance spread is the difficulty of obtaining reliable secondary market spreads since such spreads are typically derived from pricing matrices or dealer indicative quotes. So even though secondary market trades could theoretically be preferable in that they provide a cross-sectional snapshot of where tranches trade in the market at a given point in time, no reliable secondary market spread data are available due to the lack of active trading in the structured finance sector. There are tranches that are both fixed-rate and floating-rate. For our analysis we want to have a consistent benchmark for assessing the primary market spread. If fixed-rate tranches were to be included in our study, then it would be necessary to determine the appropriate benchmark yield curve for each tranche in the sample. By restricting the tranches in our sample to floating-rate tranches where the reference rate is the same interest rate benchmark, we avoid this problem. The coupon reset formula for a floating-rate security is the reference rate plus the quoted margin. In our study, we use only floating-rate tranches benchmarked off the European interbank offered rate (EURIBOR) and trading at par. 6 The quoted margin, or spread, over the prevailing EURIBOR at issue is the tranche s funding cost. The quoted margin for a floating-rate tranche issued at par is the additional per annum compensation for the risk to investors of purchasing that particular tranche. For securities issued above or below par, the quoted margin reflects, in addition to credit risk, a yield adjustment not related to risk; for that reason only EURIBOR referenced floating-rate tranches issued at par are included in our sample. Our final sample consists of 441 Euro-denominated triple-a RMBS tranches. The sample represents 83% of the entire set of Euro-denominated triple-a RMBS issued in the period 1999 to 2006 with a total par value of billion (87% of the entire set). Table 1 provides summary statistics for our sample, showing relevant statistics regarding the dummy variables and the continuous variables that we measured. 6 - EURIBOR reflects the interest rate at which highly credit rated banks can borrow, in euros, from other banks on an unsecured basis. EURIBOR is determined and communicated on a daily basis for a variety of maturities. 7

8 4. The Impact of the Number of Triple-A Ratings on the Funding Cost of a Security In this section, we first describe the regression model that we use for measuring the relationship between (combinations of) triple-a ratings and the funding costs of securities, after which we present the empirical results Empirical Model We estimate the following two regression models to test the impact of the number of credit ratings on the funding cost for our RMBS sample: where Spread it represents the new issuance spread of tranche i at time t. Two Raters it stands for a dummy variable and corresponds to a tranche that is rated by two CRAs. Moody s-fitch Rating it stands for a dummy variable and corresponds to a tranche that is rated exclusively by Moody s and Fitch. Moody s-s&p Rating it stands for a dummy variable and corresponds to a tranche that is exclusively rated by Moody s and S&P. S&P-Fitch Rating it stands for a dummy variable and corresponds to a tranche that is rated exclusively by S&P and Fitch. Three Raters it stands for a dummy variable and corresponds to a tranche that is rated by all three CRAs. 7 In Tables 2 and 3 we report the impact on the spread of the number of credit ratings and other commonly used control variables in a pooled time-series and cross-sectional panel dataset. Given the nature of our data, we had to deal with three potential econometric issues. First, to remove systematic heterogeneity from the error term, we used a heteroskedasticity-consistent variancecovariance matrix as suggested by White (1980). Second, observations in the aggregate may be affected by the same macroeconomic conditions; therefore, it is necessary to control for the time effect. To deal with the potential error-dependence problem, we follow Petersen (2009) and use dummy variables that correspond to different quarters. Each dummy variable is equal to one if the securitisation from which the tranche is included was issued during the corresponding quarter, and zero otherwise. Because of the use of time dummies, we do not include any other macroeconomic variables in our analysis. Third, when bond metrics such as spreads are the unit of observation, a problem arises when there are multiple observations for the same issuer. As a result, the observations cannot be treated as independent of each other. For this reason, we follow Petersen s suggestion and take into account issuer fixed effects in our analysis: we cluster for all tranches issued by the same RMBS originator. Finally, we control for size (log of tranche amount) and for the subordination level of each tranche in the sample to obtain robust cross-sectional results. The need to control for a tranche s subordination level is due to the triple-a nature of our sample (as explained in Section 3). The subordination level as a structuring feature is used in securitisation transactions to support each of the triple-a tranches in our sample. We measure the level of subordination below each triple-a tranche by the tranche s attachment point (i.e. the point at which credit losses can no longer be absorbed by tranches subordinated to that tranche). To compute the subordination levels for all tranches in our sample, we first divided the par value of each tranche by the total amount of the transaction s liabilities. We then calculated a tranche s attachment point as the percentage of the total liabilities subordinate to that tranche. Thus, the tranche will not suffer any losses until after that percentage of the liabilities has been lost Empirical Results Table 2 shows the empirical results for various regressions of regression model (1); in each column we display the results of a particular regression. In the first two regressions, one rater is the 7 - Contrary to the dummy variables applied for analysing specific combinations of two triple-a ratings, we do not introduce dummy variables related to analysing the effect of which particular single CRA rates a tranche triple-a in the cases where the tranche is rated only by one CRA. There are too few cases to generate meaningful results. As can be seen in Table 1, Panel A, only 45 of our triple-a securities were rated by one of the three CRAs.

9 omitted class. Regressions (1) and (2) show the effect on the spread at issuance of two and three triple-a ratings compared to tranches where there was only one triple-a rating. Regressions (3) and (4) list the empirical results when three raters is the omitted class. In Regressions (5) and (6), two raters is the omitted class. In regressions (1), (3) and (5) we have controlled for issuer fixed effects and in regressions (2), (4) and (6) we have not. It can be seen in regressions (1) and (2) that all the coefficients of two and three raters are positive and highly significant with computed t-statistics between 3.20 and This means that Eurodenominated RMBS securities rated triple-a by two or three CRAs between 1999 and 2006 had on average a higher spread at issuance than securities with only one triple-a rating. For example, in regression (2), two triple-a ratings show a spread increase on average of almost 4 basis points (t-statistic of 3.20) compared to one triple-a rating. We observe a spread increase of about 6 basis points (t-statistic of 5.15) for three ratings compared to one rating. We also observe that on average securities with three triple-a ratings have a higher spread than securities with two triple-a ratings. For example, in regression (6) we see that three triple-a ratings give a spread increase of more than 2 basis points (t-statistic of 3.45) compared to two triple-a ratings. By using regression model (2), we analyse whether the results described above are sensitive to any particular combination of CRAs. That is, when comparing two triple-a ratings with one or three raters, does the positive impact between the number of triple-a ratings and spread at issuance hold irrespective of which combination of CRAs is retained for providing two triple-a ratings? The results can be found in Table 3. In regressions (1) and (2) in Table 3, we can see that for all combinations of two triple-a ratings, on average the spread at issuance is higher than when only one triple-a rating is obtained: the spread increase is about 3 to 4 basis points with t-statistics ranging from 2.55 to The results presented in Tables 2 and 3 are consistent. We see that both prior to and after controlling for issuer fixed effects, the coefficients are positive and highly significant and robust for every combination of CRAs. So, we can conclude that for the European RMBS market pre-crisis, there was on average a positive relationship between the number of triple-a ratings assigned to a security and the spread at issuance. Apparently one must reject the rating shopping and information production hypotheses for this market. 5. Could the Number of Triple-A Ratings Be a Signal of Complexity? In this section, we explain which indicators we use as a proxy for complexity in structured finance securities. We then describe the empirical model that we use to measure complexity. Finally, we present our empirical results Complexity in Structured Finance As set out in Section 3, the most commonly applied structured finance technique to obtain a triple-a rating is by applying subordination to enhance the creditworthiness of the structure s most senior tranche. The higher the risks associated with the pool of assets, the greater the amount of subordination in the capital structure needed to protect the most senior tranche from investment losses to such an extent that a CRA will assign the highest triple-a rating to this senior tranche. So, other factors being equal, a high level of subordination can be seen as a proxy for complexity of a triple-a tranche, be it the quality of the underlying pool, the predictability of losses, the rigour of the legal structure or any other way causing complexity. In addition, for European RMBS, slicing-off the bottom portion of a triple-a rated senior tranche to create a so-called super-senior tranche rated triple-a and a subordinated tranche also rated triple-a, has been a common feature to cater to different investor requirements: some investors 9

10 sought only to purchase triple-a securities that were protected by a subordinated tranche that itself was of such a high credit quality that a CRA was prepared to assign it a triple-a rating, even though this tranche was not the most senior in the capital structure. Consequently, the investors in the super-senior tranche were protected by a tranche that was of triple-a quality itself and even further removed from the risk of investment losses. The creation of two triple-a tranches in one transaction is another indicator of complexity as they are clearly engineered for reasons beyond achieving a triple-a rating. Based on these two measures of complexity, in our complexity hypothesis framework we expect that issuers will seek more triple-a ratings on tranches with higher subordination levels; and we also expect that issuers will seek more triple-a ratings in structures that have applied the supersenior/subordinate triple-a splitting-off feature. 5.2 Empirical Model Empirically testing our complexity hypothesis requires a different model than the one we used in Section 4, where we ran ordinary least squares regressions to estimate the relationship between funding cost and number of triple-a ratings. To examine the impact of our indicators of complexity on the number of ratings, we do not measure the direct relationship between those indicators and the funding cost. This is because the indicators of complexity are structuring techniques that precisely aim to reduce the funding cost for the issuer. For example, modelling the simple relationship between the funding cost of the triple-a tranche and the level of subordination would only display the obviously inverse relationship between the two but indicate nothing about the impact of the complexity of the pool (as measured by the subordination level) on the number of ratings. We therefore use a different empirical model. We want to investigate whether complexity of the triple-a tranche has an impact on the number of credit ratings. The issuer can choose to obtain one, two or three credit ratings. We consider three unique situations in which the issuer chooses the number of ratings given a set of tranche complexity characteristics. The first situation is when the issuer chooses three triple-a ratings instead of two. Secondly, the issuer chooses to obtain three triple-a ratings instead of one. Third, the issuer chooses two triple-a ratings over one. To investigate the impact of complexity characteristics on the probability on the number of obtained credit ratings, we estimate the following probit model: (3) where Y i is an indicator that takes the value of 1 if tranche i has either two or three credit ratings, x i is a vector of complexity factors, β is a vector of parameters to be estimated, and Φ is a standardised normal cumulative distribution function. We have also applied the logit approach. The results obtained from this approach are consistent with our probit model. Therefore, we only report the results based on the probit model. Table 4 Panel A presents the results of probit regressions where Y i is an indicator that takes the value of 1 if tranche i has three credit ratings and zero two credit ratings. In Panel B we repeat the analysis, however then Y i is an indicator that takes the value of 1 if tranche i has three credit ratings and zero two credit ratings. Finally, in Panel C the same analysis is applied for when Y i is an indicator that takes the value of 1 if tranche i has two credit ratings and zero one credit rating. Our right hand side variables that are of interest to the analysis at hand are our two indicators of complexity: the level of the subordination for each triple-a tranche and whether the issuer has applied the super-senior/subordinate triple-a splitting-off feature. Super Senior it represents a dummy variable of 1 when a tranche is a super senior tranche (i.e. supported by a subordinated triple-a tranche), zero otherwise. Subordinated it represents a dummy variable of 1 when a tranche is a subordinated triple-a tranche, zero otherwise. Senior it represents triple-a securities that have 10

11 not been split into super senior and subordinated triple-a tranches, and is the omitted class. Furthermore, in our baseline specification we apply controls for size (because we expect tranches with a larger size to have more credit ratings) and for time (as it is well known that over time RMBS originators have increased the number of obtained credit ratings per tranche, we include year fixed effects). 5.3 Empirical Results In Table 4 we report the results of probit regressions. The table shows the regression coefficients of the complexity factors on the probability of having multiple ratings. We report robust z-statistics within brackets underneath the estimated coefficients. Panel A displays the first of the three unique situations that we analyse, namely when the issuer chooses to obtain three triple-a ratings compared to two. In column (2) we present the result of a probit regression, in which super-senior and subordinated variable are the only variables on the right-hand side. We find that tranches that are super-senior and those that are subordinated are 65.9% and 81.9% more likely to have three ratings compared to two. In column (3), subordination level is the only variable on the right-hand side. The results are significant at the 1% level and show that tranches with a higher subordination level are more likely to have three ratings. In column (4) we include all variables: size, senior-subordinate, subordination level and year fixed effects. For subordinated tranches the effect is virtually unchanged: subordinated triple-a tranches are 84.1% more likely to have three ratings compared to two. For super-senior tranches, however, the sign is now negative instead of positive and significant only at the 10% level. The effect of subordination level, our other proxy for complexity, is now much stronger and remains highly significant. The results displayed in column (4) can be interpreted as follows. In column (4) all variables are included simultaneously. The variable super-senior tranche is, by its nature, strongly correlated with the variable subordination level, and we cannot rule out the possibility that the positive effect of the senior tranche on the number of credit ratings is simply a consequence of senior tranches have higher subordination levels. This interpretation is strengthened by the fact that the inclusion of subordination level reduces the statistical precision of the super-senior tranche in the analysis. In Panel B we repeat the analysis displayed in Panel A, but now the object of analysis is the choice for three triple-a ratings compared to one, whereas in Panel A the object of analysis was the choice for three triple-a ratings compared to two. For two of our proxies for complexity, the results are similar to our earlier findings in Panel A: where we found that the presence of a subordinated tranche and higher subordination levels are associated with the choice for three credit ratings. As regards the super-senior tranches, no significant impact on the choice for three versus one triple-a rating is found, either when considered independently in column (2) or after adding all the factors in column (4). In Panel C the object of analysis is the choice for two triple-a ratings compared to one. In this specific situation our factors for complexity show no significant impact on the choice between two and one triple-a ratings. In columns (2) and (3) we can see that the factors all have the expected positive signs, but the measurements do not pass our significance tests. In sum, the findings described above suggest that the higher the subordination level (an indicator for complexity) the greater the probability that tranches will have three triple-a ratings when compared to either two or one triple-a ratings. The impact of having the super-senior/subordinated triple-a splitting off feature on the choice for multiple credit ratings is less straightforward. For the super-senior portions of the transactions where the splitting off feature has been applied, we virtually find no significant results (and sometimes with the wrong sign, albeit insignificant). For the subordinated triple-a tranches, we do find a highly significant impact on the choice between three and two, and between three and one triple-a rating. As stated before, for the situations 11

12 analysing our set of complexity factors that influence the choice for two versus one triple-a ratings, we do see the expected signs, but the results do not pass our significance tests. 6. Conclusion and Policy Implications In much of the current research on market practices with respect to the use of credit ratings, the rating shopping hypothesis and the information production hypothesis feature prominently. Both of these hypotheses predict an inverse relationship between the number of ratings and a security s funding costs; that is, more ratings will reduce funding costs and, conversely, less ratings will increase funding costs. We find precisely the opposite to have been the case for the mainstay of the structured finance securities market in Europe prior to 2007, namely the triple-a tranches of European RMBS. Looking at the relationship between the number of CRAs and funding costs we find that on average, tranches with three triple-a ratings had higher a funding cost than tranches with one or two triple-a ratings, and tranches with two triple-a ratings had a higher funding cost than tranches with only one triple-a rating. We develop a complexity hypothesis to help explain the observed positive relationship between the number of ratings and the funding spread. We suggest that the number of ratings may be a signal for complexity, for which investors demand a risk premium. We test an empirical model that measures the relationship between the number of triple-a ratings and certain complexity features that are key to the structured finance market: the application of subordination in order to achieve a triple-a rating and the splitting of a triple-a tranche into a super-senior and a subordinated triple-a tranche. Our complexity hypothesis is that these complexity features are likely to cause issuers to obtain a greater number of triple-a ratings in order to secure a successful placement of these tranches with investors, notwithstanding the complexity of the securities in question. Indeed we find, for the same data set of European RMBS, that the higher the subordination level, the greater the probability that tranches will have three triple-a ratings rather than either two triple-a ratings, or one. The impact of our other complexity indicator (i.e. having the super-senior/ subordinated triple-a splitting off feature) is less clear. For the resultant subordinated triple-a tranches, we do find a highly significant impact on the choice between three and two ratings, and between three ratings and one triple-a rating, but for the super-senior tranches we find less convincing results. For the choice between two versus one triple-a rating our results do not pass our significance tests. Academically, these results are relevant globally for our understanding of how credit ratings may impact market behaviour. Our findings suggest that the structured finance market may behave differently than what would be predicted by the rating shopping and information production hypotheses. Obtaining multiple credit ratings may be a signal for complexity, as suggested by a number of our findings. Hence further analysis on the role of complexity features of a security on market practices with respect to credit rating agencies is called for. For European policymakers the relevance lies in the review, scheduled to take place before 1 January 2016, of the efficacy of the EU Regulation that requires issuers to engage at least two CRAs for the rating of structured finance securities. The current regulation, which came into force on 20 June 2013, may lack effectiveness, as it does not focus on structured finance securities with complex features. Policymakers should be cognizant of the risk that a requirement of a minimum number of ratings may reduce the information content on the complexity of a security that could otherwise (i.e. without the requirement) be signalled by the number of ratings. Our paper will help the European Commission in assessing whether or not its current position regarding the requirement of number of ratings is prudent and effective. 12

13 Table 1 Summary of Statistics Variables Used in the Analyses This table reports summary statistics of Euro-denominated RMBS issued and sold in the European market (83% of the entire set). Panel A reports the statistics for the dummy variables. Number of Ratings represent the number of ratings for each tranche. Tranches with dual ratings are tranches that obtained two ratings. Moody s-fitch Rating stands for a dummy variable of 1 when the tranche is rated exclusively by Moody s and Fitch, zero otherwise. Moody s-s&p Rating stands for a dummy variable of 1 when the tranche is exclusively rated by Moody s and S&P, zero otherwise. S&P-Fitch Rating stands for a dummy variable of 1 when the tranche is rated exclusively by S&P and Fitch, zero otherwise. Tranches with Single Rating represent the number of tranches rated by one CRA. Senior represents a dummy variable of 1 when the tranche has not been divided in a super senior and subordinated triple-a tranche, zero otherwise. Super Senior tranche represents a dummy variable of 1 when the tranche represents the super senior part of the senior triple-a tranche, zero otherwise. Subordinated triple-a tranche represents a dummy variable of 1 when the tranche is the subordinated part of the senior tranche, zero otherwise. Panel B reports the statistics for the continuous variables. Spread at Issue is the tranche s quoted spread over the EURIBOR Interbank offered rate for a tranche issued at par. Log Principal Amount is the log of the size of the tranche. Subordination Level is the cumulative subordination level of each tranche in a transaction. 13

14 Table 2 Regressions of Triple-A RMBS Spread to Number of Ratings This table shows the results of regressing a tranche s new issuance spread on the number of credit raters, internal credit enhancement, and the size of the tranche. Regression (1) (6) are based on a sample of triple-a Euro-denominated RMBS tranches issued (at par) between January 1, 1999 to December 31, 2006 (83% of the entire set). One Rater represent a dummy variable of 1 when the tranche is rated by one rater, zero otherwise. Two Raters represent a dummy variable of 1 when the tranche is rated by two raters, zero otherwise. Three Raters represent a dummy variable of 1 when the tranche is rated by three raters, zero otherwise. Subordination Level is the cumulative subordination level of each tranche in a securitisation structure. Log of Principal is the log of the tranche size in euro millions. Time effects are included in regression (1)-(6) as control variables but are not shown in the tables. Issuer effects are included in regressions (1), (3) and (5). The table shows the coefficient and White (1980) heteroskedasticity-adjusted t-statistic in brackets. Dashes, - denote not included in the analysis. The symbols ***, **, and * denote parameter estimates for which zero falls outside the 99%, 95% and 90% posterior confidence intervals, respectively. Table 3 Regressions of Triple-A RMBS Spread to Dual Ratings This table shows the results of regressing a tranche s new issuance spread on the number of credit raters, dummies for dual ratings by combinations of credit raters, internal credit enhancement, and the size of the tranche. Regression (1) (4) are based on a sample of triple-a Euro-denominated RMBS tranches issued (at par) between January 1, 1999 to December 31, 2006 (83% of the entire set). Dual Rating by Moody's and Fitch stands for a dummy variable of 1 when the tranche is rated exclusively by Moody s and Fitch, zero otherwise. Dual Rating by Moody's and S&P stands for a dummy variable of 1 when the tranche is exclusively rated by Moody s and S&P, zero otherwise. Dual Rating by S&P and Fitch stands for a dummy variable of 1 when the tranche is rated exclusively by S&P and Fitch, zero otherwise. Three Raters stands for a dummy variable of 1 when the tranche is rated by all three credit rating agencies, zero otherwise. Subordination Level is the cumulative subordination level of each tranche in a securitisation structure. Log of Principal is the log of the tranche size in euro millions. Time effects are included in regression (1)-(4) as control variables but are not shown in the tables. Issuer effects are included in regressions (1) and (3). The table shows the coefficient and White (1980) heteroskedasticity-adjusted t-statistic in brackets. Dashes, - denote not included in the analysis. The symbols ***, **, and * denote parameter estimates for which zero falls outside the 99%, 95% and 90% posterior confidence intervals, respectively. 14

15 Table 4 Multiple Triple-A Credit Ratings and Tranche Characteristics This table shows probit regression results. The sample includes floating-rate tranches triple-a Euro-denominated RMBS issued at par in the period January 1, 1999 to December 31, 2006 (83% of the entire set). In Panel A the dependent variable is a binary variable that takes the value of one if the tranche has three credit ratings, and zero if the tranche has two ratings. In Panel B the binary variable that takes the value of one if the tranche has three credit ratings, and zero if the tranche has one rating. in Panel C the binary variable that takes the value of one if the tranche has two credit ratings, and zero if the tranche has one rating. Log Principal Amount is the log of the size of the tranche. Subordination Level is the cumulative subordination level of each tranche in a transaction. Super Senior Triple-A tranche represents a dummy variable of 1 when the tranche represents the super senior part of the senior triple-a tranche, zero otherwise. Subordinated Triple-A tranche represents a dummy variable of 1 when the tranche is the subordinated part of the senior tranche, zero otherwise. Panel B reports the statistics for the continuous variables. Dashes, - denote not included in the analysis. Year are dummy variables that indicate the tranche was issued and sold. Reported are regression coefficients with robust z-statistics in brackets. The symbols ***, **, and * denote parameter estimates for which zero falls outside the 99%, 95% and 90% posterior confidence intervals, respectively. 15