Complex Securities and Underwriter Reputation: Do Reputable Underwriters Produce Better Securities?

Transcription

1 Complex Securities and Underwriter Reputation: Do Reputable Underwriters Produce Better Securities? John Griffin University of Texas, Austin Alessio Saretto University of Texas, Dallas June 30, 2012 Richard Lowery University of Texas, Austin Abstract Conventional wisdom and empirical evidence often suggests that banks value their reputation, and this gives them an incentive to work in the best interest of their clients. We develop a model of reputation and complexity, where strategic underwriters with high reputation will often produce complex assets that harm investors. The securities will perform well during good states but underperform during market downturns. We examine these predictions using a unique sample of $10 trillion dollars of CLO, MBS, ABS, and structured finance CDOs issued between 2000 and Contrary to the conventional view, we find that securities issued by more reputable banks did not outperform, but rather exhibited faster rating deterioration and default. The underperformance is present because high reputation underwriters issued more securities in the poorest performing parts of structured finance, and because within the MBS, ABS, and CDO markets they issued poorly performing securities. Further, among all high reputation underwriters, we fail to find a commitment-type underwriter that appears to have protected their customers. Overall, our theoretical and empirical evidence suggests that investors and regulators should reconsider the classical view of bank reputation. We thank Andres Almazan, Emre Carr, Jonathan Cohn, Joshua Coval, Jordan Nickerson, Stanislava Nikolova, Dragon Tang, Sheridan Titman, industry insiders and government officials who wish anonymity, and seminar participants at the University of Texas for useful comments. We especially thanks Zack Liu, Nathan Swem, and James Zhu for excellent research assistance. We are thankful to the McCombs School of Business for Research Grant Support. All errors are our own. john.griffin@mccombs.utexas.edu. Department of Finance, Red McCombs School of Business, 1 University Station B6600, Austin, TX Phone (512) richard.lowery@mccombs.utexas.edu. Department of Finance, Red McCombs School of Business, 1 University Station B6600, Austin, TX Phone (512) asaretto@utdallas.edu. Department of Finance, Naveen Jindal School of Management, SM 42, 800 West Campbell Road, Richardson, Texas Phone: (972)

2 1. Introduction Although there is no shortage of opinions, government inquiries, and news commentary, little academic research examines the role of the issuers in creating structured products which lost trillions of dollars for their clients. Much conventional wisdom, theory, and empirical evidence suggest that reputable banks work in the best interest of their clients because it is in their best interest to do so. In a Bloomberg 2010 interview regarding the perception of the firm, Lloyd Blankenfein, CEO of Goldman Sachs, articulates the conventional view: Loss of reputation can bring down a firm. It will not bring down Goldman Sachs because we enjoy, I believe the confidence and trust of our clients. We dont take it for granted, and want to work very, very hardto keep what we have... This intuition is also is supported in many standard models of reputation and product quality, including Booth and Smith (1986), Chemmanur and Fulghieri (1994), and Bolton, Freixas, and Shapiro (2007) in finance and in the game theoretic analyses of the product choice game, for example in Liu (2011) and Mailath and Samuelson (2006). More importantly, this intuition is also largely borne out in academic studies showing the value of investment bank reputation in the IPO (Beatty and Ritter (1986), Carter and Manaster (1990), and Lewellen (2006)), bond (Fang (2005)), loan (Ross (2010)), and acquisition market (Golubov, Petmezas, and Travlos (2012)). 1 We show theoretically and empirically that this intuition can break down with complex securities. We first model the incentives of underwriters to develop securitized products, with an emphasis on the interaction between the inherent complexity of these assets and the importance of reputation in such an opaque market. We define reputation in the sense of Milgrom and Roberts (1982) and Kreps, Milgrom, Roberts, and Wilson (1982), where there is uncertainty over whether a long-lived player maximizes his own utility or commits to a particular set of actions. In the context of the market we study, a securities underwriter has high reputation if there is a high prior probability that he is committed to working in the interest of his clients who purchase the securities. Perhaps surprisingly, we find that 1 Fernando, May, and Megginson (2012) use the collapse of Lehman Brothers to show that investment bank relationships are valuable to their clients. Our evidence suggests that investment bank relationships can also have substantial costs to clients.

3 high reputation underwriters have incentives to produce complex assets that perform poorly from the perspective of the customer. The model starts with the assumption that there might be gains from transferring risk exposure from banks and dividing such risk among those who are risk tolerant and those who are risk averse. Underwriters can accomplish this by investigating the correlation structure of a pool of assets and then constructing a security from those assets that are least correlated. However, an underwriter who chooses to investigate correlations may package correlated assets together to produce a security that has good returns in good states but worse returns in bad states; they do so to misrepresent the value of the security. The value destroying securities are costly to produce but far less costly than the securities that actually provide insurance. The collateral will be cheaper to source and provide higher yield for investors during good states but will underperform during bad states, as illustrated with CDOs by Coval, Jurek, and Stafford (2009b). Low reputation underwriters will not necessarily put in the effort to develop good securities, but, with no reputation to preserve, they do not face the same incentives to produce securities that look great in good states. Our model can thus explain why high reputation underwriters expended time, effort, and resources to develop securities that perform poorly in states where cash flow is most valuable, even while the clients willingly accept the complex, difficult to value securities. Complexity plays a role because we assume that the process of selecting and repackaging securities to create a structured finance product, while permitting precisely constructed payoff profiles, makes the resulting security difficult to understand and to value. Specifically, for an asset like a CDO, for example, investors would rarely have access to all of the detailed collateral-level data and thus would be unable to use the cash flows observed during boom states to back out how the security would have performed in a bad state (i.e., housing downturn). With simpler assets such as such as corporate debt, investors can use accounting information to determine if the underwriter misrepresented the nature of payoffs. High 2

4 reputation strategic underwriters may produce better non-complex securities simply for fear that any misrepresentation will be quickly uncovered. This can potentially explain why our model and findings contrast with much of the theoretical and empirical literature that focuses on less complex assets. Our findings also clarify the potentially detrimental aspects of financial complexity that have been raised by former industry participants such as Partnoy (1999). Having developed the model, we evaluate the core empirical predictions through a novel Bloomberg database consisting of a large sample of structure finance securities worth over $10.3 trillion USD in underlying collateral, and for which we can observe the identity of the underwriters and the security entire rating history. The 13,008 deals in our final sample cover collateralized loan obligations (CLO), non-agency mortgage-backed securities (MBS), asset-backed securities (ABS), and collateralized debt obligations (CDO) issued between January 2000 throughout December First, the most central empirical prediction of the model is that strategic underwriters with perceived high reputation may create securities than perform worse through a crisis than securities created by underwriters with low reputation. If most high reputation underwriters are strategic, then, this pattern will hold overall. Interestingly, we find that, overall, high reputation underwriters issue securities which do not outperform and, in fact, underperform securities issued by low reputation underwriters for the time period from June 2007 throughout December The underperformance occurs in the MBS, ABS, and CDO market. The results hold unconditionally and with controls and vintage and collateral type fixed effects, indicating that the lack of the ability of reputable underwriters to provide value for their clients is not simply due to poor choices surrounding which types of products to produce, but also rests on the specific types of securities they produced within each of these three important structured finance spaces. Our second prediction is that strategic types may push securities to the market place even 3

5 in the period prior to an impending collapse because they know that their good perceived reputation will be lost once the downturn occurs and the previously created bad securities are revealed. Although debatable, from various public documents regarding the deterioration of mortgages and discussions with industry insiders, we believe that it was clear to industry underwriters at least by January 2007 that mortgage backed securities were extremely problematic. Hence, we examine 2007 issuance volume and find that the volume issued by high reputation players in the first six months of 2007 was nearly identical to and occasionally above that in 2006 despite the less favorable issuance conditions. Finally, our discussion is primarily focused on the actions of strategic players with high reputation, but our model allows for the possibility of commitment types with high reputation. We investigate this across underwriters by examining if there are certain underwriters which both produced high quality CDOs and pulled out of the market when it was clear that the mortgage market was in trouble. We fail to identify any commitment types, as all of the underwriters with perceived high reputation produced poorly performing securities or continued to produce securities all the way to the meltdown in mid Some recent theory papers, like ours, challenge the conventional intuition, though our focus is significantly different. Mathis, McAndrews, and Rochet (2009), Bolton, Freixas, and Shapiro (2012) and Fulghieri, Strobl, and Xia (2010) argue that rating agencies may find it convenient to strategically build and then burn their reputation. Similarly, in the context of issuance, Hartman-Glaser (2012) develops a signaling model where high reputation leads to frequent misreporting of the true value of an asset, because low reputation underwriters have a stronger incentive to invest in building a reputation. These studies do not address the production of securities, the ability of issuers to manipulate the distribution of payoffs, or the incentives to actively produce low quality securities. While we believe our model is most directly applicable to the specific market we study, our empirical results can also be interpreted as support for the class of models that suggest that high reputation can generate 4

6 bad, rather than good, incentives for strategic agents. 2 From the empirical side, there is some interesting related work. Titman and Tsyplakov (2010) find that commercial mortgage originators package worse collateral if they have large negative stock returns in the prior quarter. This is consistent with banks sacrificing their reputation if their organization is in trouble, but is quite different from our prediction and finding that even organizations that are performing well will create poor securities. We find that most high reputation issuers sold poorly performing securities nearly as much in 2005 and 2006 as in There are also some recent papers that show that standards at rating agencies declined in the years prior to the crisis for MBS (Ashcraft, Goldsmith-Pinkham, and Vickery (2010) and Stanton and Wallace (2011)) and CDOs (Griffin and Tang (2012)). He, Qian, and Strahan (2012) focus on the breakdowns in the rating process through market pressure from large issuers. They find that large issuers received larger AAA sizes than small issuers, and that these MBS from larger issuers underperform. Interestingly, in contrast to their channel, we do not find that high reputation underwriters received more favorable initial rating structure Model We posit that there are gains to trade between banks and customers for securities held or obtained by the bank, but that due to risk aversion customers would like the securities to have a specific profile of payoffs in different states of the world. A short-lived customer (player 2) chooses a quantity of a security to purchase, y [0,2]. 2 The most general treatment of this idea can be found in Cripps, Mailath, and Samuelson (2004), who show that, in the generic imperfect monitoring case, all undeserved reputations must uniformly disappear over time. Mailath and Samuelson (2006) provides a thorough treatment of reputation models. 3 This finding is likely due to unobserved deal characteristics as we later discuss. Our model suggests that strategic underwriters will misrepresent the deal features to investors by focus on making similar-looking securities with poor collateral correlation properties. We also find little evidence that structured products from issuers with substantial market power (as measured by league tables) exhibit future underperformance, suggesting that our findings regarding reputation are distinct. 5

7 A long-lived underwriter (player 1) chooses a level of insurance to provide, w (, ), and a markup over the true value of the security, (p v). Positive insurance (w > 0) represents moving payoffs from high cash flow states to low cash flow states, and negative insurance implies the opposite. A positive level of insurance implies that the underwriter has investigated the correlation of assets in a pool and found those that have low correlation. Alternatively, the underwriter could investigate the correlation of assets and choose assets that move together, thus providing negative insurance (w < 0). This will harm a risk averse customer seeking insurance, but will be just as costly to the underwriter as finding the uncorrelated assets (because w enters as a quadratic term in the underwriter s expected payoff). The markup, p v, on the other hand, represents the difference between the unconditional value of the security and the price the underwriter chooses to offer the client. The expected one-shot interaction (stage-game) payoffs for each player are as follows: g 1 (p v,w,y) = y ( a+(p v) w 2) g 2 (p v,w,y) = y(b+w (p v)) 1 2 y2 where (a+b) represents the exogenous amount of gains to trade. The realized payoff to the customer depends on the state, which can be either h or l, with p(h) = π > 1. To keep the following analysis brief, we focus on one particular case 2 designed to simplify algebra. We choose a = 5 4 and b = Realized payoffs are given by: ( ) 1 g2 h (p v,w,y) = y 2 +η (p v) w 1 2 y2 ( 1 g2 l (p v,w,y) = y π 1 π w (p v) π ) 1 π η 1 2 y2 These payoffs capture, in a simplified form, the basic tradeoffs facing an investor in a risky 4 All numerical parameters are selected to simplify algebra and are not important for the results. 6

8 asset. An investor cares more about consumption provided in bad states, either because the security pays off less or because other investments and economic activity pay off less. Thus, the value of the asset is increasing in the underwriter s action w, which moves payoffs from the good state to the bad state. The customer also prefers a lower price relative to the fundamental value of the asset. This is represented by p v, which divides the surplus between the customer and the underwriter. Here, high p v represents an offer price far from the fundamental value (which is only known by the underwriter). An underwriter can accomplish this by overstating the value of the collateral, or, equivalently, seeking out low quality, cheap collateral. Finally, the customer is risk averse and wishes to limit his exposure to the risky asset, which is represented by the quadratic investment cost, y 2. We assume that p v and w are unobservable to the customer, while the realized state is observable. Thus, without loss of generality we can assume the customer observes a signal (v p) w in the good state and (v p)+ 1+π w in the bad state. The assets the 1 π underwriters create are complex in the sense that, while the customer can obviously observe the realized payoff, the customer does not have the capacity to perform the counterfactual analysis to determine what the security would have paid off in the state that did not occur. This assumption is particularly appropriate for structured finance products since the payoff in the rare disaster state depends primarily on the default rate of the underlying assets and the correlations between such assets. Beyond basic summary features, it will be impossible for most investors to have the analytical tools to determine the quality of each underlying loan or the correct correlation between these loans. Note that, while the customer strictly prefers higher values for w, the underwriter s payoff is decreasing in the absolute value of w. Thus, values of w < 0 are both strictly dominated in the stage game and Pareto inferior to w = 0. We also renormalize the action space of the underwriter by defining θ 1 (p v), where θ can be interpreted as the level of truthfulness in the underwriters price offer. We 7

9 assume that θ [0,1], such that an underwriter can choose a markup of no less than zero and no more than 1. This simplifies the expression of the expected stage game payoffs to: ( ) 9 g 1 (θ,w,y) = y 4 θ w2 ( ) 1 g 2 (θ,w,y) = y 2 +w+θ 1 2 y2 which will prove convenient to simplify algebra but involves no substantial loss of generality. In order to be ableto study theeffect of reputation, we assume that the customer believes that there is some type of underwriter who is committed to providing an efficient level of insurance and being fully honest about the value of the security. This type will thus always play θ = 1 and w = 1. The customer believes the underwriter is this commitment type 2 with probability µ 0. This belief could arise, for example, from a history of apparently honest behavior in other sectors of the bank. High reputation implies a greater prior weight on the underwriter being committed to the interests of the customer. The game is repeated twice, with the customer in the second period observing the outcome in the first period. Thus, in the first period, the underwriter must consider how his actions influence the customers belief about his type. In the appendix, we show that the main results of the model go through in an infinite horizon case as well. Thus, the implications of the model are not dependent on their being a final period in which the underwriter burns his reputation Simple Securities In this section, we consider the effect of reputation when securities are simple in the following sense. Given the realized payoff and the realized state, an investor is able to use his information about the security to recover the counterfactual payoff. That is, the customer learns both his actual payoff and the payoff that would have been realized had the other state occurred. Equivalently, the customer is able to observe w and θ separately. An 8

10 example of such a security would be a corporate bond of a publicly traded company. Other information available about the firm, such as accounting information and cash flows, may reveal to the investor whether the firm is healthy or not even if the good state occurs and the bond pays off. The process of actively producing securities from many underlying assets makes doing such a counterfactual analysis difficult, and effectively impossible as the process becomes dependent on detailed underlying data and their correlations which are not revealed to the customer. Here, we show that absent complexity the reputation effect is consistent with the standard results on reputation; an underwriter with a high reputation will be more likely to produce better securities, even when he is strategic rather than actually a commitment type. The proposition establishing this result and the proof can be found in the appendix. All remaining proofs and additional discussion can also be found in the appendix Complex Securities Under complexity, the story is different. The second-period customer is able to observe the payoffs from the first-period security, and will update his prior on the likelihood that he is facing a commitment type based on the outcome of the previous game. Because the security is complex, the customer cannot observe the action of the underwriter in the previous interaction or the counterfactual payoff in the unrealized state. A security with good insurance properties (w > 0) and low markup (p v) cannot be distinguished from a security with bad insurance properties (w < 0) and a high markup, unless the bad state is realized. By moving payoffs from the bad to the good state and at the same time overstating the value of the security by choosing a high markup, the underwriter can mimic the commitment type and thus maintain or build his reputation, risking discovery only when the state turns out to be bad. Our first result here is obvious; if the underwriter has no reputation (µ 0 = 0), he has no incentive to attempt to maintain or build a reputation. Since the horizon of the game is 9

11 finite, such an underwriter will simply play the equilibrium of the stage game twice, 5 and the customer will play the best response to this, y = 1 in period 2 and y = 1 + 3µ in period 1. We now consider what happens when the underwriter starts out with some reputation, i.e. µ 0 > 0. The first step in this analysis is to derive the optimal set of actions for an underwriter when he seeks to preserve his reputation in the high state, sacrificing it in the low state. We refer to this security as a time bomb because it is designed, in effect, to blow up on the customer in the bad state. This intuition builds on the concept in Coval, Jurek, and Stafford (2009a) and Coval, Jurek, and Stafford (2009b) that securities can easily be mispriced and worthless in the bad state with even minor differences in correlation or parameter uncertainty. Lemma 2.1. The optimal action for the underwriter that preserves his reputation in the high state and sacrifices his reputation in the low state is to play θ = 0 and w = 1 2 in the first period. Note that, like the security that would be produced by an underwriter unconcerned with his reputation, the time bomb security involves the underwriter overstating the value of the security as much as possible. The time bomb security, however, has the added negative effect of moving payoffs away from the state in which they are valuable to the customer, and this comes at a cost to the underwriter. Thus, the time bomb security is Pareto dominated by the stage game security. If an underwriter wants to fully imitate the commitment type, he has no choice but to play θ = 1 and w = 1 2 in order for payoffs to match the payoffs from the commitment type in both states. If the underwriter chooses to neglect his reputation, his payoffs are maximized by choosing θ = 0 and w = 0, the stage game equilibrium action. Thus, any 5 We ignore equilibria in which player 2 plays y = 0 and the underwriter chooses to produce very low quality securities, as this equilibrium will not survive a trembling perfect refinement. Admitting such an equilibrium would in fact permit customers to discipline underwriters in the first period, but this technicality is beyond the scope of our analysis. 10

12 equilibrium involves playing one of these three actions described. The following proposition establishes the threshold above which the underwriter will choose to produce time bombs, and the volume of such securities that are produced. Proposition 2.2. There exists an equilibrium where time bombs are created with probability 1 if and only if µ (1 π). The quantity of securities produced is given by y 1 = 2µ 0, such that the volume of time bombs is increasing in reputation. The strategic underwriter will rely on the time bomb strategy exclusively when his reputation is sufficiently high. The possibility that the underwriter will fully imitate the commitment type in the first period only occurs when the underwriter s reputation is high and the disaster state is more likely than not to occur. This is clearly not the parameter region of interest. Proposition 2.3. There exists a pure strategy equilibrium where the underwriter fully imitates the commitment type if and only if µ and π µ 0. Thus, for π > 11 27, there is no equilibrium in which the underwriter fully imitates the commitment type. This leaves as possible pure strategies only that underwriters play the stage game equilibrium in the first period. This turns out not to be an equilibrium as shown in the Appendix. We are left with a substantial region in which there is no pure strategy equilibrium. To keep things simple we confine attention to the case where π > 1, such that the normal state 2 is in fact modal. We conclude the analysis of the model by deriving the mixed strategy between time bombs and full imitation of the commitment type, which will be the equilibrium for underwriters with intermediate reputation levels, and showing that, for a significant region, the quality of the security issued in the first period is decreasing in the reputation of the underwriter. We derive the mixed strategy equilibrium in the appendix and present the result on monotonicity below: 11

13 Corollary 2.4. The probability of a time bomb is increasing in reputation as long as µ 32 27(1 π) ( ) π 11 3 (1 π) 1, 2 up to the point above which time bombs are produced with probability Empirical Implications Our analysis of the empirical implications of reputation depends on comparing the behavior of strategic players with high reputation to strategic players with low reputation. Firms that have developed a good reputation for dealing with clients honestly in other areas of the bank will be perceived as more likely to try to induce their employees who work in structured finance to preserve that reputation by avoiding opportunistic behavior. Uncertainty remains, however, since outside investors cannot observe or interpret the details of compensation contracts within financial firms, and providing good incentives is neither easy nor necessarily in the interest of the firm as a whole. In particular, the relatively recent movement of financial firms from partnerships, under which many developed their reputations, to publicly traded firms has major implications for the incentives of employees within the firm (see Morrison and Wilhelm (2007). These implications were likely not obvious to investors and still are not completely clear. Both our theoretical and our empirical analysis assume that investors view highly reputable underwriters as more likely to induce their structured finance departments to behave reputably. Our model predicts that strategic underwriters with high reputation produce time bombs that perform worse than the securities of low reputation underwriters. Our model does not explicitly speak to the proportion of strategic and commitment types with high reputation. However, under the assumption that players with high reputation are mostly strategic, we 12

14 obtain our first and most central empirical hypothesis: Hypothesis 3.1. Securities produced by low reputation underwriters will outperform securities produced by high reputation underwriters in the crisis period. Strategic players with high reputation will continue to issue time bombs even when it becomes evident that the bad state is imminent. This is because they know they have produced poor securities and will soon lose their reputation anyway. This leads to our second empirical hypothesis: Hypothesis 3.2. If high reputation underwriters are in fact strategic, they will not voluntarily decrease the volume of issues in the period immediately prior to the crisis. Finally, our model is mainly focused on strategic types. Among underwriters with high reputation, there may be commitment types that issue appreciably better securities. While our model does not speak to the percentage of strategic players, if high reputation players are committed only to their profits, they will burn their reputation. Within the high reputation players, we explore whether there is evidence of any commitment types that make appreciably better securities, or if all high reputation players appear to be strategic. 4. Data, Reputation Measure, and Background 4.1. Data We collect issuance and rating history data for all structured finance securities issued from January 2000 to December 2010 that are available on the Bloomberg system. According to Bloomberg these data represent the entirety of actual issuances. The data are broadly classified as CLO, non-agency MBS, ABS, and CDO, where CDOs are generally defined as collateralized debt securities backed by MBS, ABS, and other CDO securities. Therefore, 13

15 CLO, MBS and ABS arise from a first round of securitization of individual debt claims, while CDOs arise from a second round of securitization, also known as repackaging. In this spirit, we order the data in this manner (CLO, MBS, ABS and CDO) because we believe that this reflects an increasing order of collateral complexity. In our model complexity has a well-defined meaning inherently related to the investor s ability to collect information about the actual distribution of cash-flows and hence is reflected in our ordering. The collateral pool of a CLO is typically composed of 50 to 100 corporate loans that usually come with enough detailed information to make possible a partial assessment of the quality of the loans. Because of the high level of transparency, we include collateralized corporate bond obligations, CBO, in the CLO group. Because the original debtor is a corporation, many sources of information can be used to evaluate not only the quality of the individual claims but also the correlation structure of the collateral pool. Differently from CLOs, MBS and ABS are usually composed of thousands of individual claims whose original debtor is either a physical individual or a small legal entity (in the case or CMBS). MBS consists of both residential (RMBS) and commercial mortgages (CMBS) that are largely outside of the standards for securitization used by the government sponsored enterprises. We do not examine agency RMBS because they are implicitly secured by the government. Non-agency RMBS consists of prime, first-lien fixed, Alt-A, and adjustable rate loans. ABS consists of autos, credit card loans, equipment, home equity, manufactured housing, student loans, and other. ABS Home equity is distinct from MBS in that it contains various forms of non-standard residential housing, including subprime, home equity, and second-lien loans. While largely standardized, it would be difficult to check the quality of each mortgage supporting an MBS and, therefore, investors must rely on a few reported statistics. For ABS securities, in addition, the collateral is typically less standardized (i.e., auto loans, credit cards receivables, and subprime) making it even more difficult to perform an effective 14

16 valuation, not to mention an accurate assessment of the pool correlation structure. Finally, we reserve our CDO category for collateral obligations constructed using an asset that is from another structured finance product (ABS, MBS, or another CDO), and therefore relatively harder to evaluate. In Table 1 we detail various summary statistics for CLO, MBS, ABS, and CDOs. First, we note that we gather information at the deal and tranche level for 22,450 deals and $16.9 trillion in underlying collateral. However, in our empirical analysis we lose deals because of their lack of ratings or the absence of the reputation measure (which we detail in the next section). The Table details these losses for each category. Most of the losses come from the MBS category, both from unrated securities as well as underwriters with no reputation score. Overall, in most of our empirical analysis we examine 13,008 deals representing over 10.3 trillion USD in underlying collateral. The collateral is most concentrated in the non-agency residential MBS at $3.7 trillion. ABS consists of $4.4 trillion with almost $2 trillion from Home equity. The CLO and CDO categories are considerably smaller and similar in size. The structured finance CDOs consists of $679 billion in rated securities with underwriter reputation, with the majority of the collateral from CDOs of ABS. Appendix Table A.1 details these issuance figures by subcategories through time. The average deal size for each category is typically over 300 million USD with both residential and commercial MBS having an average size of over a billion dollars. Across collateral types, at least 60% of the securities that comprise a deal are rated at AAA. RMBS, auto loans, equipment, and student loans have as much as 85% of their securities rated at AAA. We believe that our large dataset is unparalleled in breadth and should enable us to accurately test our main hypotheses Reputation Measure We evaluate the relative performance of underwriters based on their reputation score in the IPO market. Time-series of reputation scores for major players in the IPO market are available for the period from 1984 throughout 2009 from Professor Jay Ritter s website. The 15

17 score is a measure of underwriters prestige that is computed based on the original proposal by Carter and Manaster (1990): it gives a higher score to underwriters that appear more prominently on the tombstone of an IPO prospectus. Although the measure ranges between one and nine, with nine the highest ranking, the vast majority of banks are in the 6-9 range. For our empirical analysis, we separate underwriters into three groups: underwriters with a score bigger or equal to 9 are labeled Top Reputation ; underwriters with a score bigger or equal to 8 and less than 9 are labeled as Prime Reputation ; underwriters with a score lower than 8 are labeled as Limited Reputation. For our first empirical measure for our regression analysis, we construct a variable Reputation that assigns a value of three to underwriters with top reputation, two to those with prime reputation, and one to those with limited reputation. Underwriters in our sample for which there is no reputation score (456 out of 626) are excluded from the main analysis resulting in a loss of 9,442 deals from the sample (approximately six trillion USD worth of securities). For deals that have more than one underwriter, the reputation score is calculated as the average score across underwriters. Appendix Table A.2 provides a list of all banks with reputations used in our empirical analysis as well as large banks without reputation scores. Conceptually, we have in mind the idea that the reputation of a bank is shared across divisions of the bank and depends on how much customers trust the bank to look after their interests. A customer who finds himself holding a worthless MBS security is unlikely to trust the issuer in the IPO market, and vice versa. The tombstone measure is a measure of underwriter prestige relative to other banks and is more strongly related to future IPO performance than market share, as shown by Carter, Dark, and K. (1998). This measure serves our purposes well because it is exogenous to decisions made in the market for structured products and can thus assess the prior beliefs about underwriters. At a deeper level, the likelihood that a particular desk or individual within a bank will choose to treat clients 16

18 fairly should depend on the tangible compensation structure and the intangible culture of the institution, both of which would be established at higher levels of the bank s leadership. Our approach has some difficulties as well. First, there is some idea that certain banks are considered stronger in either fixed income or equities, and thus the equity reputation may not perfectly correspond to the reputation in the fixed income space. However, there is no such relative prestige measure in fixed income that we could find, and for a relatively novel product like a CDO it seems that the overall reputation of the bank is most relevant. Another empirical issue is that, because most of our observations are from banks that belong to the top and prime reputation groups, our empirical regression analyses will by default put the majority of the weight on the distinction between these two groups. However, we are not sure that in practice there is much distinction between the top and prime banks. Highly regarded banks in 2006 and 2007, such as Lehman Brothers, Bear Sterns, Barclays Bank, and UBS have reputation scores of 8 (as shown in Appendix Table A.2), whereas banks such as Credit Suisse, 6 HSBC, Santander, and Nomura are listed as top banks even though they are less well-known or well-regarded in many circles. To address this issue, we also construct an indicator variable High Reputation which is equal to one for institutions that belong to either the top or the prime reputation group Background Information Figure 1 shows the marked rise in the market for structured products and the subsequent collapse following the onset of the financial crisis. The collapse in the market was most pronounced for the more complex ABS securities and structured finance CDOs. This drop in volume is consistent with our model if the crisis revealed that a higher than expected proportion of underwriters were, in fact, strategic types. There are, of course, other expla- 6 Credit Suisse paid the SEC a $100 million dollar fine for rebating IPO profits. Although several top investment banks paid the SEC fines for illicit IPO activity around this period, this was by far the largest fine paid as their activity was more egregious. 17

19 nations for why the market would come to a halt. 7 Still, it is worth noting that the marked downward revision in beliefs about the value of these securities does not necessarily imply that fewer should be created and sold, as the decrease in value did not necessarily imply a decrease in the gains to trade available from these securities. Further it is interesting that most of the ABS and CDO market is denominated in US dollars, whereas the CLO market and to a lesser extent the non-agency MBS market have a more sizable non-us component, which also continued to lead to some issuances in 2009 and Figure 2 shows that top and prime reputation underwriters dominate the market for the four main types of structured products. Their dominance is most prominent for CLOs and ABS, where limited reputation underwriters only issue 5% and 4% of the volume, respectively. The elite reputation underwriters have the largest relative share of the CDO space with 59%, but this is also the market in which the lowest reputation underwriters are most represented with 14%. We next examine how certain deal characteristics (measured at issuance) are related to their underwriter reputation. Table 2 reports estimation results of regression models in which the dependent variables are: the size of the deal (Amount), the deal s initial rating (Initial Rating) and the proportion of tranches that is rated AAA (AAA Fraction). We estimate regressions across CLO, MBS, ABS, and CDOs with vintage and collateral type fixed effects. The vintages are defined as semi-annual issuance groups and the collateral types are defined in Table 1. We include control variables: an indicator variable that is equal toonewhenthedealisdenominatedinusd(usddeal), thesizeofthedeal, statedmaturity (Maturity), the initial rating, the initial fraction AAA, and whether the deal includes some derivatives (Synthetic). This test allows us to consider whether deals could have been perceived to be different, 7 For example, Glode, Green, and Lowery (2011) and Easley and O Hara (2010) argue that, with endogenous expertise and ambiguity aversion, respectively, markets may break down in response to increases in uncertainty. 18

20 ex ante, relative to the underwriter reputation. In other words, it is important to establish whether high reputation underwriters were packaging deals that had different profiles to begin with. As one might expect, deal size is increasing in reputation. On average there is no support for higher reputation underwriters receiving more favorable ratings or a higher fraction of AAA tranches at closing. This is only in apparent contrast with current findings in the finance literature. He, Qian, and E. (2011) and He, Qian, and Strahan (2012) find that large MBS issuers obtain more favorable ratings from the credit rating agencies than small issuers. We obtain different results for a few reasons: first, our reputation variable is not perfectly correlated with league table rankings. 8 In fact the correlation is almost zero with the ranking position, and slightly below 50% with the underwriter s market share in the fixed income space. Second, our model would suggest that rather than pushing rating agencies to increase the proportion of AAA rated tranches of an average collateral pool over what might have been considered the standard, underwriters would have been better off by obtaining that level with a worse collateral pool. In summary, our theory would suggest that a strategic underwriter might prefer to pick poor collateral or highly correlated collateral that it can source cheaply and package into an MBS or CDO that looks like other good securities during good states of the economy, but behaves very differently in bad states. This poorer collateral quality would be unobserved in our empirical analysis. In Panels B-E, we further examine detailed results for CLOs, MBS, ABS, and structured finance CDOs. The results confirm that there is no consistent pattern of high reputation underwriters receiving more AAA and better ratings in the MBS market; the pattern is in the opposite direction in the ABS market and generally not present in the CLO and CDO markets. 8 For example, He, Qian, and Strahan (2012) show that in most years, the top two MBS issuers in the 2000s are generally Countrywide and General Motors. 19

21 5. Do Reputable Underwriters Issue Better Securities? We now examine the relation between underwriter reputation and deal performance and test our first hypothesis that high reputation underwriters issue worse performing securities. We construct two measures of performance at the deal level. Both measure are based on the rating history of the securities that are issued as a result of the creation of collateralized obligations. A few remarks are worth making at this stage. For each security (i.e., tranche), we observe the history of rating updates from the three main rating agencies (Moody s, Standard & Poor s, and Fitch). When faced with more than one rating option, we choose the lowest (worst grade). Not all rating histories are complete. Because rating agencies oftendo notreport thelast update when thesecurity isindefault, we consider securities with a rating below CC (Ca for Moody s) to be effectively in default. Moreover, often rating agencies interrupt the rating service by effectively withdrawing the rating. This might happen for two reasons: the security might be in default and the deal manager stops paying the rating fee, or the deal is fully re-paid. Obviously those two outcomes have very different implications for our analysis. We adopt the following rule to disentangle the two cases. If the final balance of the security (which is also available through Bloomberg) reports both a positive outstanding value and losses, and all ratings are withdrawn, then we consider the security to be in default. If the balance is zero and there are no reported losses, we consider the security to be repaid in full with a rating equal to the last known one. About 5,000 deals drop out of the sample because they are not rated by any rating agency at any point in time. Our first measure (Deal Rating) is constructed by calculating a value-weighted average rating score of all the tranches that belong to each particular deal. Our second measure (Proportion of Tranches in Default) is constructed by calculating, at any point in time, the ratio of the nominal value of tranches that are in default to the total nominal value of the deal. In most of our analyses, we then measure the performance of the deal by looking 20

22 at time-series changes of two variables. We consider, as our main dependent variable, the change in the rating and the change in the proportion of tranches in default from the beginning of the crisis (June 2007) through the end of the sample (December 2010). For both measures, a larger number indicates worse performance Changes in Ratings Since our theory predicts that the underperformance will not appear in good states of the economy, in the first three columns of Table 3 (Panel A) we examine the relation between reputation and performance from January 1, 2007 throughout May 31, As expected, both with and without controls, underwriter reputation is insignificantly related to performance, indicating that performance in good states is similar for deals with high or low underwriter reputation. In the columns (4-9) of Panel A, we consider changes in ratings from June 1, 2007 (before large scale credit rating downgrades began) through December 31, The regression model reported in columns (4) and (6) are without any controls and ask if, on average, reputable banks are making better securities while preserving complete discretion to structure any type of security they please. From an investor standpoint, the regression asks how investors fared if they bought structured products from a reputable investment bank without knowing anything else about the security. This regression is most relevant for unsophisticated investors who bought structured products because they wanted to hold structured products and not because they had any particular mandate to purchase a CLO, ABS, MBS, or CDOs. The classical view of the impact of reputation on the production of securities suggests that investors will benefit from dealing with high reputation underwriters. Interestingly, the regression shows that securities from more reputable underwriters did not outperform. They were actually downgraded relatively more. To get a more precise estimate of the economic magnitude of the effect, in column (7) we compare high reputation issuers (with top and prime reputation) to limited reputation underwriters. We observe 21

23 even more distinction between the high reputation banks and those with limited reputation. On average, high reputation underwriters issue worse securities: investors who bought structured products from highly reputation underwriters experienced an average downgrade of 1.58 more notches than investors who dealt with low reputation underwriters. In the next two specifications (columns 5-6 and 8-9, respectively), we examine the regressions with controls for certain features of the deal (as introduced in Table 2). In the final specification (columns 6 and 9) we include vintage and collateral type fixed effects. Vintages are defined as semi-annual issuance groups and collateral types are defined in Table 1. Through these controls, one implicitly assumes that underwriters did not have complete discretion over the general characteristics of the securities they issued or did not understand the nature of the types of the underlying collateral in these markets. In both the specification without and with the fixed effects, reputable underwriters issue securities that perform worse. In the dummy variable specification, column (9), high reputation underwriters indeed issue securities which underperform those without reputation by 1.3 rating notches. In Panels B-E, we examine these same specifications for CLO, MBS, ABS, and CDOs. In none of these markets do high reputation issuers produce better securities for their clients. We focus on the June 2007 to December 2010 period. For CLOs, there is no significant relationship with underwriter reputation. For MBS, the general reputation specification is insignificant but for the dummy variable specification the high reputation underwriters have considerably worse performance through the crisis. Interestingly, a negative relationship between high reputation and negative future performance is present unconditionally(without controls) within the MBS, ABS, and CDO market, as well as with controls and fixed effects. This suggests that the lack of the ability of reputable underwriters to provide value for their clients is not simply due to poor choices about which general types of securities to produce, but rests on the specific types of securities they produced within each of these three important structured finance spaces. 22

24 5.2. Percentage of Defaulted Tranches Rating changes rely on rating agencies, and ratings were shown to be quite imperfect, at least prior to the crisis. Default is typically a hard event and hence a rating agency might be forced to classify a tranche as in default, making this rating less subjective. Hence, as an alternative measure of performance we examine the percentage of defaulted tranches. For background information, Appendix Figure A.3 shows the evolution of defaults through time for CLO, ABS, MBS, and CDOs, with defaults increasing in that ordering. The largest difference is between ABS tranches, with over 40% in default as of December 2010, versus CDOs, with a 70% default rate. 9 Figure 3 plots relative default performance for top, prime, and limited reputation underwriters. We report results for CLO, MBS, ABS and CDOs separately. Within CLOs, there is considerably less default and little difference between reputation groups. In contrast, top and prime reputation underwriters have similar and higher levels of default than limited reputation underwriters in the MBS, ABS, and CDO markets. We study the relation between the proportion of defaulted tranches and reputation more rigorously in Table 4(Panel A). As expected, we find that default is not related to reputation in the pre-crisis period. Without controls, we find that default from June 2007 to December 2010 is increasing in reputation. Investors who purchase securities from more reputable underwriters end up with a higher proportion of defaulted deals. When controlling for security characteristics and fixed effects, the relation is only slightly weaker. What we learn from column (9), is that high reputation underwriters issue securities that subsequently end up with 7.4% more capital in default even after controlling for the subtype and vintages of the collateral. Across Panels B-E, we find no evidence that reputation is related to better security performance through the crisis. The high reputation underwriters issue no 9 For a sample of 727 structured finance ABS CDOs, Cordell, Huang, and Williams (2011) estimate that 65% of the original issuances will end up being lost with 70% of these losses having occurred as of March Appendix Figure A.4 shows default numbers for the collateral groups within each market. 23