Rating the raters: Are reputation concerns powerful enough to discipline rating agencies?

Abstract

Credit rating agencies (CRAs) are accused of bearing a strong responsibility for contributing to the subprime crisis by having been deliberately too lax in the ratings of some structured products. In response to this accusation, CRAs argue that such an attitude would be too dangerous for them, since their reputation is at stake. The objective of this article is to examine the validity of this argument within a formal model: Are reputation concerns sufficient to discipline rating agencies?

We show that the reputation argument only works when a sufficiency large fraction of the CRA income comes from other sources than rating complex products. By contrast when rating complex products becomes a major source of income for the CRA, we show that it is always too lax with a positive probability and inflates ratings with probability one when its reputation is good enough.

We provide some empirical support for this prediction, by showing that ceteris paribus, the proportion of subprime residential mortgage-backed securities (RMBS) that were rated AAA by the three main CRAs indeed increased over the last eight years.

We analyze the policy implications of our findings and advocate for a new business model of CRAs that we call the platform-pays model.

Introduction

Credit rating agencies (CRAs) are accused of bearing a strong responsibility in the subprime crisis, by having been too lax in the ratings of some structured products. For example the title of the article by Mason and Rosner (2007) is self-explanatory: “How misapplied bond ratings cause mortgage-backed securities and C.D.O. market disruptions”. The authors of that paper argue in particular that CRAs under-evaluated some parameters that were crucial in the assessment of risks such as those measuring correlation between defaults of large obligors. Many commentators explain this behavior by a fundamental conflict of interest generated by the new business model of CRAs, which collect most of their income from the issuers rather than from the investors, as was the case when John Moody started his company (and the ratings industry) in 1905. Since the development of photocopying machines in the early seventies, the cost of reproducing the books of ratings sold to investors has decreased dramatically, threatening the viability of the investors-pay model, where fees are collected from investors. With exceptions like Egan Jones, which does not have a significant market share of the ratings industry, most CRAs now get the bulk of their revenue from issuers. As some commentators have put it: “It is as if the referee was paid by one of the teams”. Another change is due to regulation: since the creation, in 1975, of the status of Nationally Recognized Statistical Ratings Organizations (NRSROs) the stakes associated with obtaining a good rating for issuers have increased considerably. This has been reinforced recently by the official recognition of the ratings provided by NRSROs in the computation of regulatory capital requirements for commercial banks in the Basel II accord. Finally, it is clear that CRAs have played a crucial role in the extraordinarily rapid development of structured finance products in recent years, and that they have benefited a lot from this development. For example Moody's net income has grown from $159 m in 2000 to $705 m in 2006 (with 44% of its revenue in 2006 coming from its structured finance activities).

In response to the accusation of having been deliberately too lax, CRAs essentially argue that such an attitude would be too dangerous for them, since their reputation is at stake.¹ Although there are important differences between the ratings industry and the accounting industry, a parallel is often made with the rapid fall of the accounting firm Arthur Andersen following its implication in the Enron scandal. CRAs argue that they cannot afford to be the next Arthur Andersen. The objective of this article is to examine rigorously the validity of this argument: Are reputation concerns sufficient to discipline rating agencies?

There is a very large literature on reputation models in game theory, but, somewhat surprisingly, there are few applications of these models to financial markets. Our model builds on one of the few exceptions, namely Benabou and Laroque (1992), who study how a financial guru can build a reputation and ultimately cash on it by manipulating market prices in one direction and trading in the opposite direction. We adapt this model to the rating industry by considering a financial market where, at discrete dates, new firms want to issue a security for financing some investment project with a cost normalized to unity. The project quality is a priori unknown (even to issuers). It is “good” with probability $\lambda$, or “bad” with probability $(1-\lambda )$. We assume that good firms should be financed but that without prior knowledge on the quality of the project, no financing should take place. A (monopoly) CRA observes the project quality and communicates a rating to the market. No issue takes place if rating is bad or denied. The CRA can be of two types: either it is fully committed to always tell the truth or it is opportunistic (i.e., chooses the rating that maximizes the expected present value of its profits). The reputation of the CRA is measured by the probability $q$ that investors assess to the CRA being committed. The strategy of an opportunistic CRA is described by the probability $x(q)$ that a CRA of reputation $q$ will be too lax (i.e., will give a good rating to a bad security). This influences the accuracy $a(q)$ of ratings (i.e., the probability that investors believe a good rating to mean a good project, given the reputation of the CRA). A Markov perfect equilibrium (MPE) is a pair of functions $(x,a)$ that are simultaneously rational from the respective viewpoints of CRAs and investors.

We first show (Proposition 1) that, when the fraction of the CRA income that comes from other sources than rating complex products and/or the proportion of successful projects are large enough, there is a unique MPE where an opportunistic CRA always tells the truth. In this case, reputation is a good disciplining device for CRAs. By contrast when the fraction of the CRA income that comes from rating complex products² and/or the proportion of unsuccessful projects becomes large, we show (Proposition 2) that there is a unique MPE where the CRA is always too lax with some probability and lies with probability one when its reputation is good enough. This implies the possibility of what we call confidence cycles, consisting of several phases. First, starting from a situation where investors are not very trustful, issuing volumes are low, and credit spreads are high, the CRA tries to increase its reputation by being very strict. Then investors become more optimistic, the reputation of the CRA increases, spreads decrease, and issuing volume increases. But this is precisely when CRAs become more lax and risk of default increases. Ultimately there is a default, which provokes a crisis of confidence: the opportunistic CRA is detected, its reputation brutally falls down, spreads become high again and issuing volumes decrease dramatically. Thus in this case, it may take a long time to detect an opportunistic CRA and the conflict of interest is not solved by reputation concerns. Instead, reputation building strategies by CRAs generate inefficient confidence cycles.

We also extend our model to the case where the screening technology of the CRA is imperfect. That is, we assume that the CRA imperfectly observes the project quality: a good project may fail and a bad project may succeed. In the particular case where the CRA only has uncertainty on the bad project outcome, we show (Proposition 3) that the possibility of success for a bad project has a negative impact on the CRA's discipline. In the general case where the CRA imperfectly observes the quality of each project, we show (Proposition 4) that things are even worse: the opportunistic CRA is never disciplined at equilibrium.

We provide some empirical support for the opportunistic behavior of CRAs by using data on subprime residential mortgage-backed securities (RMBS) tranches that were rated by Fitch, Moody's, and S&P (the three main CRAs) over the period 2000–2008. After accounting for all characteristics of these tranches that were observable by CRAs at the time of issuance, we show that the fraction of these tranches that were rated AAA has indeed increased over the last eight years.

We analyze the policy implications of our findings and advocate for a change in the business model of CRAs. Rather than closely regulating ratings, a formidable task for regulators, we argue in favor of a new business model, what we call the platform-pays model, that would be applicable to all securities that use a trading platform such as an exchange, a clearing house, or even a central depository.

Related literature: With the exception of Kuhner (2001), who only considers a one-period model, the theoretical literature on rating agencies has not yet considered the question of reputation building by a strategic CRA. Indeed, most of the literature on CRAs considers non-strategic agencies. For example Skreta and Veldkamp (2008) assume that each CRA observes a noisy signal (rating) on the asset quality. They show that if issuers can choose one rating from among multiple ratings to disclose to the market (what is called “shopping for ratings”) then an increase in the complexity of securities (noise in signal) produces ratings inflation. Increasing competition among agencies does not solve this problem since it enlarges the range of ratings. Boot et al. (2006) show that credit ratings provide a “focal point” for firms and investors that can serve as a coordinating mechanism in situations where multiple equilibria are possible. On the contrary, Carlson and Hale (2006) establish that introducing a CRA can bring multiple equilibria to a market that otherwise would have possessed a unique equilibrium.

The (meager) literature on strategic CRAs only considers one-period models. Lizzeri (1999) and Faure-Grimaud et al. (2007) take a mechanism design approach and consider a model where a CRA commits to a fee and a disclosure rule and choose them so as to maximize expected profits. Lizzeri (1999) models non-contingent fees and ratings concerning the (possibly negative) value of an object. He shows that at equilibrium a monopoly CRA will only reveal whether the object has positive or negative value (if all objects have positive values then the CRA discloses no information). Doing so, the CRA can capture a large share of the surplus. When introducing competition à la Bertrand, CRAs make zero profit and fully disclose information. Faure-Grimaud et al. (2007) consider contingent fee and ex post efficient (renegotiation-proof) contracts. They show that at equilibrium a monopoly CRA will fully disclose information. When introducing competition à la Bertrand CRAs make zero profit and fully disclose information on firms that have values higher than the CRAs’ marginal observation cost (firms with lower values reject the contract since they would be worth less than the fee paid to obtain rating). They also show that full disclosure is not robust when introducing the possibility for ownership (disclosure rights) to the firms. According to Kuhner (2001), rating agencies are more likely to reveal their private information if their rating cannot become self-fulfilling from an ex post point of view. This is sustained by experimental studies (Sanela and Niessen, 2007).

We are thus the first to tackle explicitly the issue of reputation building by a strategic CRA in a dynamic model: Are reputation concerns sufficient to guarantee the truthful behavior of a strategic CRA? We build on two important papers that have modeled reputation building on financial markets: Benabou and Laroque (1992) study the behavior of a guru who can influence the formation of beliefs by uninformed investors, while Diamond (1989) explains why new firms with insufficient reputational capital have to borrow from banks before they can issue direct debt on financial markets.

Our paper is also related to the literature on reputation building through communication. Following the seminal papers of Kreps and Wilson, 1982 and Milgrom and Roberts, 1982 it is standard to model reputation through the introduction of a committed type (for an overview, see, e.g., Mailath and Samuelson, 2006). Some papers consider informed players that care about their reputation not because others players will treat them differently, but simply because the informed players want their advice to be believed in the future. Sobel (1985) considers a sender–receiver finitely repeated game with perfect monitoring where both players are long-run. The sender can be of two types: he either has identical preferences to the receiver (he is a “friend”) or has completely opposed preferences (he is an “enemy”). Instead of considering a discount factor as usual, the author considers that the importance of the date's play varies stochastically over time. Exploiting what happens in the last period by backward induction he shows that the enemy's equilibrium strategy is such that the probability of lying increases in the importance of the date's play. Benabou and Laroque (1992) consider a sender–receiver infinitely repeated game with imperfect observation (sender with noisy private information) where the receiver is short-run. The sender can be of two types: either committed to tell the truth or has completely opposed preferences to the receiver (“enemy”). Benabou and Laroque show that the sender's ability to manipulate information is limited only in the long run. Moreover if different senders follow one another then learning about the sender's type remains incomplete even in the long run, leaving a constant scope for manipulation. Morris (2001) adapts Benabou and Laroque's model to other sender's preferences. The sender can be of two types: either he is a “friend” or he is a “bad” type that wants as high an action as possible. Just as the bad type sometimes has an incentive to tell the truth (despite a short-run incentive to lie) in order to enhance its reputation, the good type may also have an incentive to lie (despite a short-run incentive to tell the truth) in order to enhance its reputation. In the infinite-horizon model, he shows that the good type does not necessarily have an incentive to tell the truth and that for at least some discount rates for the bad type and utility functions for the decision maker, there is not a truth-telling equilibrium. In particular, even if the good type is arbitrarily patient and the bad type is arbitrarily impatient, an informative equilibrium may not exist.

The rest of the paper is organized as follows. Section 2 presents the basic model and provides the main results. Section 3 illustrates the properties of the non-truthful equilibrium previously obtained. Section 3 extends the basic model to the case where CRA's observations are imperfect. Section 4 offers the empirical analysis. Finally, Section 5 discusses some policy implications and a possible extension of our model.