Adverse Selection Dynamics in Privately-Produced Safe Debt Markets

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Adverse Selection Dynamics in Privately-Produced Safe Debt Markets Nathan Foley-Fisher, Gary Gorton, and St´ephane Verani 2020-088 Please cite this paper as: Foley-Fisher, Nathan, Gary Gorton, and St´ephane Verani (2020). “Adverse Selection Dynamics in Privately-Produced Safe Debt Markets,” Finance and Economics Discussion Series 2020-088r1. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2020.088r1. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Adverse Selection Dynamics in Privately-Produced Safe Debt Markets∗ Nathan Foley-Fisher1, Gary Gorton2, and Stéphane Verani1 1Federal Reserve Board 2Yale School of Management and NBER First version: June 2020. This version: October 2020 Abstract Privately-produced safe debt is designed so that there is no adverseselectionintrade. Thisisbecausenoagentfindsitprofitable to produce private information about the debt’s backing and all agents know this (i.e., it is information-insensitive). But in some macro states, it becomes profitable for some agents to produce private information, and then the debt faces adverse selection when traded (i.e., it becomes information-sensitive). We empirically study these adverse selection dynamics in a very important asset class, collateralized loan obligations, a large symbiotic appendage of the regulatedbankingsystem, whichfinancesloanstobelowinvestmentgrade firms. JEL Codes: E44; G14; G23 Keywords: Safe debt, adverse selection, information sensitivity, collateralized loan obligations ∗For providing valuable comments and assistance with data, we would like to thank, without implication, Sriya Anbil, Celso Brunetti, Richard Cantor, Francesca Carapella, Larry Cordell, Burcu Duygan-Bump, Sebastian Infante, Tom King, Brittany Lewis, David Little, Ralf Meisenzahl, Borghan Narajabad, Michael Palumbo, Lubo Petrasek, Coco Ramirez, Algis Remeza, Peter Sallerson, Alex von Hafften, and Guillaume Vuillemey. WearegratefultoSamDreithandNathanHeinrichforexceptionalresearch assistance. Theviewsinthispaperaresolelytheauthors’andshouldnotbeinterpreted as reflecting the views of the Board of Governors of the Federal Reserve System or of any other person associated with the Federal Reserve System.

1 Introduction Thereisaprimitivedemandforsafedebt, debtthatcanstorevaluethrough time safely and can be traded without fear of adverse selection. The government cannot completely meet this demand (see, e.g., Krishnamurthy and Vissing-Jorgensen (2012), Caballero, Farhi and Gourinchas (2017), Gorton, Llewellen and Metrick (2012), and Gorton (2017)). The private sector cannot produce riskless debt, but it can produce a close substitute, debt such that no agent finds it profitable to produce private information about its fundamentals, and all agents know this. This debt is senior, complicated, and opaque, making the cost of understanding it high. It is information-insensitive. It provides protection against adverse selection; it is rated AAA/Aaa (S&P/Moody’s) and it trades at par (minus a small bid-ask spread). But bad public news about the fundamental value of the debt’s backing can prompt sophisticated investors to acquire private information creating adverse selection; see Dang, Gorton and Holmström (2019). We study the dynamics of this switch from information-insensitive to information-sensitive in the collateralized loan market and document the resulting adverse selection. The risk of a switch from information-insensitive to sensitive, and the resultingadverseselection,isauniquerisk. Becausethesecurityisdesigned to be information-insensitive, no investor has an incentive to produce 2

private information about the security’s fundamentals. Ex ante investors do not investigate the fundamentals of the bond and accept the debt at par. Investors are uninformed by design. In particular, the risk of a switch to adverse selection is neither investigated nor priced ex ante. Ex post in the case of bad times, they cannot price the security because they have not invested in the technology to do so. See Hanson and Sunderam (2013), and Monnet and Quintin (2017). So, privately-produced safe debt is a doubleedged sword: It serves its purpose of avoiding adverse selection until it doesn’t. A collateralized loan obligation (CLO) is a legal entity that buys loans from banks and finances them by issuing debt in the capital market.The CLO manager can subsequently buy and sell assets. The CLO liabilities, called “tranches”, have ratings that range from AAA/Aaa to B. CLOs play a very significant role in financing below investment-grade firms, whose loans are called “leveraged loans”. According to Fed Chair Jerome Powell Powell (2019): “Collateralized loan obligations are now the largest [nonbank] lenders, with about 62 percent of outstanding leveraged loans.” The leveraged loan market is about $1.1 trillion, and is used by about 70percentofU.S.companies,includingcompanieslikeBurgerKing,United Airlines, Avis Rent a Car, and Equinox Fitness. Why do banks sell their loans? The answer is that it is profitable to do so. And it reduces the cost of credit to the borrowing firms. Nadauld and 3

Weisbach (2012) found that bank loans that are eligible to be securitized, i.e., sold to a CLO, cost borrowers 17 basis points less than otherwise (100basispointsequalsonepercent). Thebankcanmakeupthisdifference because the AAA/Aaa-rated debt sold by the CLO has a convenience yield, that is investors value it for safety in addition to the interest rate it pays, so its pecuniary coupon rate is lower than it otherwise would be. “Safety” means that it is highly likely to pay off at par at maturity and, if it is sold early, it will sell for (almost) par. It is information-insensitive. In our sample about 65 percent of a typical CLO is rated AAA. This means that to recover the 17 basis points, the convenience yield must be at least 26 basis points. By comparison, Krishnamurthy and Vissing- Jorgensen (2012) find that the yield on U.S. Treasuries over 1926-2008 was, on average, 73 basis points lower than it otherwise would have been, due to the “moneyness” and safety of U.S. Treasury securities. Using higher frequency data, van Binsbergen, Diamond and Grotteria (2019) estimate the convenience yield on Treasuries to be about 40 basis points. Consistent with the AAA tranches having a convenience yield, these tranches are held by insurance companies, pension funds, U.S. and foreign banks, and U.S. investment banks. The mezzanine tranches in the middle of the CLO capital structure are held by hedge funds and asset managers, among others. The equity part of the capital structure is held by private equity firms, the CLO managers, and CLO sponsors. See Liu and Schmidt- 4

Eisenlohr (2019) and DeMarco, Liu and Schmidt-Eisenlohr (2020). To value a CLO tranche the companies whose leveraged loans are in the underlying portfolio have to be studied and the correlations between all the loans in the CLO portfolio have to be determined. This requires credit analysts and a model to simulate outcomes. Also, as we explain below, CLOshavecomplicatedandopaqueinternalstructures. Alltheseattributes make it very expensive for agents to produce private information about the value of the AAA tranche, allowing buyers of this debt to avoid adverse selection because it is very expensive to produce private information. But, in bad times this is exactly the problem! When the COVID-19 pandemic took hold in March 2020, risks to the poolsofloansunderpinningCLOsroseasthepandemicposedanimmediate and prolonged threat to corporate profitability. Facing the prospect of widespread downgrades and possible defaults, we show below that the AAA tranches became information-sensitive. Prior to the pandemic, AAA tranches traded at (almost) par. But in the pandemic the AAA prices fan out below par, despite unprecedented actions by the Federal Reserve to improve financial intermediation.1 The increase in AAA price dispersion, reflecting the differential pricing 1In general, all of the Federal Reserve’s actions potentially had an indirect effect on CLO markets by improving the conditions for financial intermediation. Indeed, we will show that the Federal Reserve improved market liquidity for CLOs. Nevertheless, secondarymarketpricingdidnotreturntopre-pandemiclevels,suggestingthattrading was based on private information about the underlying pool of collateral. 5

based on information, is shown by the boxplots in Figure 1. Figure 1: Trading conditions for AAA-rated CLO debt securities. TheboxplotsinthefigureshowthedistributionofpricesforAAA-ratedCLOtranches. The central boxes show the interquartile range (IQR) of prices bisected by the median asahorizontalline. ThewhiskersshowthedistributionofpricesoutsidetheIQR,upto ±1.5×IQR. For each CLO tranche, we calculate a daily weighted-average price, where theweightsaretransactionvolumes. Wethenaveragethedailypricestoobtainaweekly average weighted by CLO tranche par amount. The dashed line is the imputed round trip cost for agency trades in AAA-rated CLO tranches. Source: Authors’ calculations from data provided by TRACE, Bloomberg LP., Moody’s Analytics, S&P, and Fitch. The central boxes show the interquartile range of prices bisected by the median as a horizontal line. The whiskers show the distribution of prices beyond the interquartile range, excluding outliers. The figure summarizes our argument. Prior to the pandemic there was no price dispersion. AAA tranches traded at par of $100, consistent with information-insensitivity. When the pandemic hits, the price distribution widens out below par as shownbytheboxplots. Importantly, ifitwereaneconomy-wideshockthat introduced a risk premium for AAA-rated CLOs, then their prices would 6

drop uniformly and there would not be price dispersion across CLOs. The lowerdottedlineissimilartoabid-askspread, ameasureofaverseselection (discussed in detail later). This measure increases considerably when the pandemic hits, due to adverse selection. In the analysis presented below, we will provide formal tests of the information shown in Figure 1. To show that when the pandemic struck, some investors started producing information, we show that there is a structural break in the standard deviation of AAA-rated CLO prices with the onset of the pandemic, confirming what is shown in Figure 1 and suggesting that the prices contain tranche-specific information. We also provide the results of a difference-in-differences quantile regression of secondary market prices of AAA-rated CLO tranches on an index of the stock price volatility of industries(e.g., Oil&Gas, Leisure, Retail)thatwereidentifiedbySallerson (2020) as being especially vulnerable to the pandemic. The regression shows that the uncertainty about the vulnerable industries is driving the dispersion in AAA-rated CLO tranche prices. Where do investors get information on the CLOs’ portfolios? We look at AAA-rated CLO tranche price changes just after the release of monthly trustee reports on the CLOs (which occur at different dates during the month) that contain relevant information for investors. Prior to the pandemic, price changes were consistently near zero around trustee report dates, but not so once the pandemic hits. 7

Leveraged loans do not trade frequently. And they do not trade in a central location. They trade over-the-counter. Consequently, CLOs buy third-party prices to mark their loans to “market”. For example, CLOs are required to mark-to-market riskier loans rated CCC and below that they hold in excess of some predetermined contractual limit. We examine the loan prices supplied by two third-party pricing agencies. A measure of the increase in ignorance of the uninformed is the difference in these prices, for the same loan on the same date, from the two pricing agencies. The difference in the distributions of these two loan prices is significant. In other words, disagreement goes up with the onset of the pandemic. We then turn to measuring the resulting adverse selection; see Glosten and Milgrom (1985) and Dang et al. (2019). We show that a measure of the bid-askspread, theimputedroundtripcost(IRC)oftransactions, explained later, increased dramatically, indicating a deterioration in market liquidity. In Figure 1, the dashed line is the IRC in basis points for the AAA-rated CLOtranchesfromJanuary1,2020. Thedeteriorationinsecondarymarket trading conditions began on Monday, March 9. From about 10bps at the beginning of the year, the IRC peaked at over 100bps on March 19. This sudden loss of market liquidity coincided with falling prices. By March 18, the average mid price for transacting the most-senior CLO tranches had already fallen below 95 cents. The literature closest to our work includes Brancati and Macchiavelli 8

(2019), Gallagher, Schmidt, Timmermann and Wermers (2020), Perignon, Thesmar and Vuillemey (2018), and Benmelech and Bergman (2018). Brancati and Macchiavelli (2019) examine the Panic of 2007-2008 and “... provide direct evidence that while in good times bank debt is largely informationally insensitive, it becomes significantly sensitive to information in bad times” (p. 99). Gallagher et al. (2020) study investor information production, money market funds (MMFs) redemptions, and MMF managers’ rebalancing decisions during the Eurozone crisis. They find that there was significant selective information acquisition about fund holdings and, although there were redemptions at all funds, the reductions were largest for funds with the most sophisticated investors, who did produce information. “Under these circumstances, MMF shares become information-sensitive because MMFs’ risk exposures are suddenly differentiated following the acquisition of information” (p. 1). Perignon et al. (2018) study the wholesale CD market in Europe during 2008-2014. Lenders to banks moved their money from bad banks (those whose future performance decreases) to good banks, apparently on the basis of private information that they produced. The uninformed faced adverse selection in that they did not move their money. Benmelech and Bergman (2018) study U.S. corporate bonds. They show that when there is a drop in a bond’s price, the bond becomes more illiquid, suggesting adverse selection. Other, more distantly, related literature is surveyed in Dang et al. (2019). We 9

differ from this previous research by focusing on privately-produced safe debt. We explicitly show that when AAA-rated debt becomes informationsensitive, it coincides with the arrival of adverse selection, consistent with Dang, Gorton and Holmström (2018). The paper proceeds as follows. In Section 2 we briefly describe the data, provide some background on CLOs, and some background on the loans in CLO portfolios. Section 3 analyzes the switch from informationinsensitive to information-sensitive. When the pandemic hit some agents became informed while others did not. Section 4 studies these two groups: the informed and the uninformed. That this switch coincided with the arrival of adverse selection is studied in Section 5. Section 7 concludes. 2 Background and Data In this section we first discuss the dataset that we constructed to assess the information sensitivity of CLOs during the pandemic. We combined several large datasets, including transactions on individual CLO debt tranches, tranche-level credit ratings, loan-level information about the underlying pools of collateral, and loan transactions by CLO managers. The combination of these data create a window to study how information about CLO tranches can affect the trading decisions of CLO investors. Then we provide some detail about how CLOs work, and examine the 10

underlying loans in CLO portfolios. Finally, we summarize the loans held by CLOs. 2.1 Data We identify CLO tranches using Bloomberg’s Backoffice data on assetbacked securities and match these data by CUSIP identifiers to dealer transaction data reported in the regulatory version of the Trade Reporting and Compliance Engine (TRACE), created by the Financial Regulatory Authority (FINRA).2 Bloomberg provides information on individual CLO tranche characteristics, including offering amount, offering yield, and amount outstanding. After merging these two large data sets, we add Moody’s, Fitch, and Standard & Poor’s credit ratings and credit watch list information for each CLO tranche at a daily frequency. Our data on secondary market over-the-counter trading of CLO debt is from TRACE. Under regulations introduced in 2002 by FINRA, dealers are required to file detailed reports of their transactions, including trade time, quantity, price, andcounterparty. FINRAmadeasset-backedsecurity secondary market trading data, including CLO debt, available from 2011 as part of an effort to increase transparency in this over-the-counter market after the 2007-09 financial crisis. The limitation on data availability means that, unfortunately, wecannotcomparetradingduringthepandemictothe 2CUSIPstandsforCommitteeonUniformSecuritiesIdentificationProceduresandis a unique identifier for most financial instruments, including privately-placed securities. 11

financial crisis 2007-09. We follow standard procedures for cleaning these data.3 We use confidential regulatory data with dealer bank identifiers, which allows us to match trades by buyer, seller, amount, and trade time while removing duplicates. The regulatory version of TRACE allows us to identify dealers and observe the trading behavior of every dealer. We construct daily aggregates for each CLO tranche using the transaction-level data. The daily trading price is defined as the weighted average price on transactions, where the weights are the transaction volumes. This measure is designed to approximate the mid price for each CLO tranche as the vast majority of trades are matched in advance. We also calculate the total volume transacted. The data we use to study the loans held by CLOs are from Moody’s Analytics, which are based on trustee reports issued by each CLO each month and vastly augmented by Moody’s proprietary leveraged loan analytics. These data contain details of the entire portfolio of assets held, including type of asset, credit rating, maturity, par value, market value, creditor information, as well as multiple individual loan prices provided by Markit, Reuters, and the CLO trustees. Moody’s also provides information about the performance of the CLO, such as details of the internal triggers for cash flows, discussed below. Table 1 presents summary statistics from the baseline dataset where 3See, for example, Dick-Nielsen (2009) and Bao, O’Hara and Zhou (2018). 12

the unit of observation is CLO tranche i on day t, conditional on at least one transaction occurring. Our sample data cover the period from January 1, 2020 onwards. We separate the period into two subsamples: The dates prior to March 1, 2020 are in the “pre-pandemic period” and the dates thereafter are in the “pandemic period.” The tests of equality of the means and standard deviations between the two subsamples are reported in the columns labeled “p-value”. During the pandemic there was a drop in the average transaction prices, together with an increase in the dispersion of transaction prices, while there was no change in aggregate trading volumes. Thesestatisticssuggestthatinvestors’tradingbehaviourchanged significantly as investors sought information about the underlying risks across CLO tranches. We investigate the role of information production below. Dang et al. (2018) have no prediction about the trading volume once there is a switch to information-sensitive. While the market might collapse, as in Akerlof (1970), uninformed agents may choose to trade anyway, accepting the adverse selection. And, in reality, fearing that the AAArated tranches may be downgraded may motivate some institutions to sell. 2.2 CLO Background CLOs are securitization vehicles created typically by large asset managers to finance pools of leveraged loans extended to below investment grade 13

Table 1: Summary statistics This table reports summary statistics for the CLO tranche secondary market trading data used in the analysis. Trade_price is it the weighted average transaction price, where the weights are the transaction volumes. Trade_volume is the sum of transaction volumes. Our sample begins on January 1, it 2020 and is divided on March 1, 2020 into the pre-pandemic period and the pandemic period. Source: TRACE, Bloomberg LP., Moody’s Analytics, S&P, and Fitch. Variable Statistic AAA-rated tranches Class E tranches (units) Pre Post p-value Pre Post p-value Trade_price Mean 99.91 96.48 0 95.69 72.08 0 it ($ per 100 face) SD 0.38 3.17 0 4.9 14.87 0 N 1,125 4,000 581 1,222 Trade_volume Mean 9.77 11.36 0.95 5.3 5.33 0.56 it ($’000s) SD 30.11 24.33 1 5.03 4.83 0.88 N 1,125 4,000 581 1,222 firms. These loans are predominantly originated by banks as part of a syndication (Bruche, Malherbe and Meisenzahl, 2020). A CLO manager first buys a pool of leveraged loans on the secondary market. The initial pool of loans may remain unchanged in a “static” CLO. More often, the loans are actively traded by the CLO manager in accordance with the terms of the prospectus, which specifies inter alia the currency, maturity, industry, and rating of the loans that may enter the pool. The cash flows from the pool of loans are used to pay the interest and principal on a set of securities (tranches) issued by the CLO vehicle, which are structured as a hierarchy of claims on the underlying collateral. Each security has a priority claim on the capital structure. CLOs allocate the effects of default risk in the underlying pool of 14

loans in several ways. First, there is subordination. For example the AAA-rated tranche is protected by the mezzanine (middle) tranche and the junior tranche. Second, there are over-collateralization tests. Overcollateralization tests are calculated by dividing the principal balance of the portfolio by the total cumulative balance of the tranche (and, if it is not the senior tranche then all tranches senior to it). The numerator is adjusted when some of the loans face stress: they become more risky, default, or are worth significantly less than their face value. There are specific haircuts for loans in these categories. When an over-collateralization test is violated, there is a reallocation of excess spread. Excess spread is the interest earned on the portfolio of loans in excess of the interest due on the CLO tranches. If there are no defaults and the CLO can make its obligated payments to the note holders, then the excess spread flows to the equity holders (on a monthly basis). But, if the portfolio experiences stress in the form of defaults, ratings downgrades, etc., suchthatover-collateralizationtestsfail, thenexcessspreadisdirected away from equity, and sometimes away from junior tranches, and used to pay down principal on the AAA debt. CLO managers are large sophisticated entities that usually manage multiple CLOs. According to the fourth-quarter 2019 manager rankings from CreditFlux, the average CLO manager had 10 CLOs with a total par value of $5.3 billion. CLO managers are often affiliated with private equity 15

Figure 2: Leveraged loans outstanding and held by CLOs. The time series in the figure below show the amount of outstanding U.S. leveraged loans. The bars in the figure show the aggregate amount of U.S. leveraged loans held by U.S. CLOs (grey) and European CLOs (black). Source: S&P and Moody’s Analytics. funds, hedge funds, asset managers, banks, or insurance companies. These firms have large teams of credit analysts and loan traders, which contribute to the CLO managing and trading its portfolios. CLOs are an important way for banks to offload large amounts of risky loans that would otherwise reside on their balance sheets. Figure 2 compares the amounts of leveraged loans and CLOs outstanding. As of 2019, accordingtotheLoanSyndicationsandTradingAssociation(LSTA), there were more than $1.1 trillion of U.S. leveraged loans outstanding. U.S. CLOs, which are the largest nonbank investors in U.S. leveraged loans, amounted to over $500 billion.4 4These CLOs may hold other types of investments, e.g., junk bonds, but are predominantly backed by U.S. leveraged loans. 16

2.3 CLO Loan Portfolios We use Moody’s data largely drawn from CLO trustee reports to examine the typical CLO loan portfolio characteristics. We then use the same data to describe how prices of leveraged loans evolved during the pandemic. We show that variation in the distributions of loan prices depend on whether the firms belonged to industries that were more vulnerable to the pandemic shock. Table 2 presents summary statistics across CLOs from the information provided in the trustee reports. The unit of observation is a CLO. The first three columns show statistics for 1,627 CLOs during the pre-pandemic period and are calculated using the last trustee report for each CLO published before February 15, 2020. The second three columns cover 1,599 CLOs during the pandemic period, calculated using the first trustee report published after April 1, 2020. Data constructed by Moody’s indicate that the average attachment point for the AAA-rated tranches in these CLOs before the pandemic is about 37 percent. In other words, on average 63 percent of a CLO is rated AAA. The data confirm that loans dominate the asset portfolios of CLOs, with the typical portfolio containing about 370 loans and a total market value of roughly $420 million. Bonds account for less than 3 percent of the principal value of the typical portfolio. The average loan in the portfolio is worth $1-1.5 million and has a residual 17

maturity of 4.5-5 years. Table 2: Summary statistics of CLO loan portfolios. For the prepandemicsampleof1,627CLOs,thestatisticsarecalculatedusingthelasttrusteereport published before February 15, 2020 for each CLO. The pandemic sample of 1,599 CLOs use the first trustee report published after April 1, 2020. Source: Authors’ calculations from data provided by Moody’s Analytics. Pre-pandemic,1,627CLOs Pandemic,1,599CLOs Characteristic Mean Median SD Mean Median SD AAAtrancheattachmentpoint(%) 36.9 35.7 5.5 36.7 35.6 5.9 LoansperCLO 369.6 344 233 390.3 349 285 BondsperCLO 7.7 0 25 9.8 0 35 CLOmarketvalue(loans,$mn) 421.4 408.9 175.7 367.7 353.7 157.7 CLOprincipalvalue(nonloans,$mn) 10.7 1.4 24.2 12.1 1.1 28.2 Meanloanvalue($mn) 1.5 1.3 0.9 1.3 1.2 0.8 Medianloanvalue($mn) 1.2 1 0.9 1 0.9 0.7 Meanloanmaturity(yrs) 4.5 4.8 0.9 4.5 4.8 0.9 Medianloanmaturity(yrs) 4.7 5 0.9 4.7 4.9 0.9 Principalvalueinbonds(%) 2.7 0 8.6 3.4 0 11.4 Vulnerableloans(marketvalue,%) 18.4 18.1 9.6 16.4 16.2 9 The last line of Table 2 shows that the typical CLO loan portfolio has about 18 percent of its market value in industries that Moody’s identified as vulnerable to the pandemic shock. Sallerson (2020) identified seven industries vulnerable to the pandemic shock: Automotive, Consumer goods: Durable, Energy: Oil & Gas, Hotel, Gaming & Leisure, Retail, Transportation: Cargo, and Transportation: Consumer. Appendix 8 describes how we map these industries to other data. We use these sectors throughout our analysis to tease out a differential impact of the pandemic. Table 3 presents summary statistics from the leveraged loan transactions reported in the trustee reports. We separate the transactions since January 1, 2020 into two time periods: The pre-pandemic period 18

up to March 1 and thereafter the pandemic period. The table reports both the number and average transaction value of sales and purchases per CLO. The table is divided into three parts. The first three columns provide summary statistics for all leveraged loan transactions. The middle three columns provide statistics for transactions on loans to the sectors identified by Moody’s as vulnerable to the pandemic shock. The final three columns provide statistics for transactions on the non-vulnerable loans. The statistics generally indicate that loan portfolio turnover (sales and purchases) increased, including for loans to firms in vulnerable sectors. There was no change in the average value of the loan transactions per CLO. As a robustness check on the summary statistics reported in the table, we excluded loan transactions that took place during a CLO’s ramp-up period and report the results in Table 8 in appendix 9. The trustee reports also reveal greater deterioration in the market value of leveraged loans in the seven vulnerable sectors. CLO managers provide monthly updates through the trustee reports on the market value of the leveraged loans in their portfolios.5 We use these data to construct distributions of loan market values over time. For each month, we plot the distributions separately for loans to firms in the vulnerable and not vulnerable industries. Figure 3 shows that the values of loans to firms 5Wecomparedtheloanmarketvaluesinthetrusteereportstothebidpricesprovided by Reuters and Markit. These pricing services obtain their data using a combination of polling traders and modelling. We found that the vast majority of CLO managers report market values identical to Markit bid prices. 19

Table 3: Loan transactions summary statistics. The table shows summary statistics for leveraged loan transactions per CLO in the pre-pandemic period (January 1, 2020—March 1, 2020) and the pandemic period (March 2, 2020—June 30, 2020). Source: Authors’ calculations from data provided by Moody’s Analytics. Variable Statistic Alltransactions Vulnerablesector Non-vulnerablesector (units) Pre Post p-value Pre Post p-value Pre Post p-value Numberofsales Mean 47.16 82.79 0 6.83 10.99 0 45.09 78.77 0 (perCLO) SD 59.33 136.29 0 16.41 23.03 0 53.49 127.75 0 N 1,581 1,625 493 603 1,579 1,624 Meansalesvalue Mean 0.8 0.72 1 0.57 0.53 0.82 0.81 0.73 1 ($mnperCLO) SD 0.76 0.78 0.1 0.64 0.62 0.25 0.76 0.79 0.08 N 1,581 1,625 493 603 1,579 1,624 Numberofpurchases Mean 62.66 102.43 0 6 12.73 0 58.79 94.8 0 (perCLO) SD 63.55 146.42 0 13.52 29.2 0 55.96 130.94 0 N 1,549 1,592 1,001 961 1,549 1,591 Meanpurchasevalue Mean 1.16 0.84 1 1.1 0.77 1 1.17 0.85 1 ($mnperCLO) SD 0.91 0.86 0.01 1.01 1 0.37 0.92 0.86 0 N 1,549 1,592 1,001 961 1,549 1,591 in the vulnerable industries (shaded boxplots) fell further and were more variable than the values of loans to firms in other industries. 3 The Switch to Information-Sensitive Debt In this section we study the switch from information-insensitive debt to information-sensitive debt. First, we show that the onset of the pandemic coincided with a regime switch in the standard deviation of AAA tranche prices. Then we adopt a quantile regression approach to show that uncertainty about the vulnerable industries is responsible for differentiation in transaction prices across AAA-rated CLO tranches. The lowest transaction prices for AAA CLOs became correlated with an index of the volatility of the vulnerable industries’ stock prices, while the highest 20

Figure 3: Loan collateral price distribution The boxplots in this figure show the distributions of loan prices reported in CLO trustee reports of U.S. leveraged loans. TheshadedboxplotsincludeloanstofirmsintheindustriesidentifiedbyMoody’s as vulnerable to the pandemic shock. The unshaded boxplots contain loans to firms in all other industries. Source: Moody’s Analytics. transaction prices remained relatively uncorrelated with the same index. 3.1 AAA CLO Price Dispersion During the Pandemic Figure 1 above displayed the fanning out of post-pandemic prices, suggestingthatinvestorsweredifferentiallypricingtheAAAtranchesusing information from trustee reports most likely. “Fanning out” corresponds to an increase in measures of prices dispersion. Panel (a) of figure 4 shows the time series of two measures of the dispersion in the prices of AAA-rated CLO tranches. Also shown is the Bank of America-Merrill Lynch High Yield Index Option-Adjusted Spread. This index is appropriate for comparison because CLOs hold 21

loans to below investment-grade firms. Compared to this index are the two measures of dispersion across CLO tranches: the standard deviation and the interquartile range.6 The two measures of dispersion are highly correlated and mirror the movements in the High Yield Index. Figure 4: Dispersion of prices, the high yield spread, and vulnerable sector stock price volatility. Thisfigurecomparesthetimeseries of dispersion in the prices of AAA-rated CLO tranches to the Bank of America-Merrill Lynch High Yield Index Option-Adjusted Spread (left-hand panel) and the volatility of the stock prices of firms in the sectors identified by Moody’s as vulnerable to the pandemic shock (right-hand panel). We calculate two measures of dispersion across CLO prices: the standard deviation and the interquartile range. Source: Authors’ calculationsfromdataprovidedbyTRACE,FRED,BloombergLP,Moody’sAnalytics, S&P, and Fitch. (a) High yield spread (b) Vulnerable stock price volatility As a complement to panel (a), we replaced the high-yield spread with the stock price volatility in the industries that Moody’s identified as being most vulnerable to the pandemic shock. The solid line in panel (b) of figure 4 is the weighted average daily difference between the high and low log prices on the seven vulnerable industries identified by Moody’s, where the weights are the transaction volumes. Bank loans and bonds are 6To calculate the dispersion measures we first calculate a daily weighted-average price, where the weights are transaction volumes. 22

alternative sources of funding for firms. Based on the time-series of data since January 1, 2020, the correlation between the high-yield spread and the volatility index is 0.61. We can further formalize this increase in dispersion by estimating structural breaks in the standard deviation of AAA-rated CLO prices. We do this by applying the method of Bai and Perron (2003) for identifying multiple structural breaks in a single time series. For each tranche, we calculate a daily weighted-average price, where the weights are transaction volumes. We then calculate the standard deviation across tranches. The optimal number of breaks to explain the time series is determined by the Bayesian Information Criterion. Figure 5 shows the results. The shaded regions show 95 percent confidence intervals for the location of the structural breaks. AAA-rated tranche prices show little dispersion prior to the pandemic, consistent with the AAA-rated tranches being information-insensitive. From mid-March until early May the dispersion is high and variable, before settling down somewhat, but still above the pre-pandemic levels. The first structural break is estimated to have occurred between March4andMarch11, 2020, whentheWorldHealthOrganizationdeclared the novel coronavirus outbreak a pandemic.7 The timing of this structural break coincides with disruptions to a wide range of financial markets, 7https://www.who.int/news-room/detail/27-04-2020-who-timeline---Covid-19 23

Figure 5: Structural breaks in the standard deviation of AAArated CLO prices. Thisfigureshowstheestimatedstructuralbreaksfromapplying themethodofBaiandPerron(2003)foridentifyingmultiplestructuralbreaksinasingle time series. Source: Authors’ calculations from data provided by TRACE, Bloomberg LP, Moody’s Analytics, S&P, and Fitch. including Treasuries (Logan, 2020) and corporate debt (Kargar, Lester, Lindsay, Liu, Weill and Zuniga, 2020; Haddad, Moreira and Muir, 2020). The second structural break is estimated to have occurred between May 6 and May 18, with a point estimate of May 7. The window of the break is wider than the first, and is likely a response to the combined impact of several announcements by the Federal Reserve. In particular, on May 4, the Board of Governors announced that the Primary Market Corporate Credit Facility (PMCCF) and Secondary Market Corporate Credit Facility (SMCCF) would begin purchasing corporate bonds in “early May”. The SMCCF is the only emergency facility with the authority to purchase corporate debt below investment grade. Following through, 24

the New York Fed announced on May 11 that the SMCCF would begin purchasingexchangetradedfundsonMay12. AlsoonMay12, theBoardof Governors announced that the Term Asset-Backed Securities Loan Facility (TALF)wouldacceptanarrowly-definedsetofhighly-ratedCLOtranches.8 Alloftheseannouncementshadthepotentialtodirectlyimproveconditions in the CLO market. In addition, several further announcements in the time window indirectly affected conditions by modifying the liquidity coverage ratio (May 5) and the supplementary leverage ratio (May 15). These modifications were intended to ease banks’ ability to provide credit and financial intermediation services. As such, they had a potentially indirect effect on conditions in the CLO market. 3.2 Quantile Regression We delved deeper into the relationship between the dispersion of AAA CLO tranche prices and uncertainty about vulnerable industries using quantile regression with CUSIP fixed effects. The dependent variable Trading price is the weighted-average price of CLO tranche i on day t, it where the weights are the transaction volumes. The variable Covid takes t the value 0 before March 1, 2020 and 1 thereafter. Volatility is the volumet weighted average daily difference between the high and low log prices on seven vulnerable industries (Sallerson, 2020). We estimate the conditional 8See https://www.federalreserve.gov/newsevents/pressreleases/files/ monetary20200512a1.pdf 25

quantile functions Q (τ|Covid ,Volatility ) of the response of Trading price t t it the t-th observation on the i-th CLO tranche’s Trading price given by it Q (τ|Covid ,Volatility ) = αi +β (τ)Covid Trading price t t 1 t it +β (τ)Volatility 2 t +β (τ)Covid ×Volatility , (1) 3 t t with quantile τ ∈ {0.25,0.5,0.75} and where αi is the CUSIP fixed effect. TheCUSIPfixedeffectsabsorballtime-invariantcross-sectionaldifferences in the CLOs that were traded during the period. The quantile fixed effect regressions are implemented using the penalized fixed-effects estimation method proposed by Koenker (2004). Table4showstheresults. Thecolumnsinthetablerefertothedifferent CLO tranches by seniority. The industry calls the most senior debt tranche the A Class and the most junior debt tranche the E Class. Class A tranches are designed to attract a AAA rating by a credit rating agency at issuance. Therefore, the majority of AAA-rated tranches are Class A debt securities. In our sample of CLO trades, about 92 percent of the CLO tranches rated AAA by at least one of the main credit rating agencies (S&P, Moody’s, and Fitch) are Class A tranches. The remaining AAA CLO tranches in our sample are below Class A. In the table, the first column is all AAA tranches. Columns 2 through 6 follow the CLO capital structure from the 26

most senior debt tranches (Class A) to the most subordinate debt tranches (Class E). The regression reveals how sensitive the CLO prices within a tranche group are to the vulnerable industries volatility index. The table includes a row reporting a χ2 test of the null hypothesis that the 25th and 75th percentile coefficients are the same. The statistical significance of the difference between the two coefficients increases monotonically with the seniority of the tranches and loses significance lower in the capital structure (E class). Percentiles of the distribution of transaction prices are responding heterogeneously to uncertainty about the vulnerable industries. The variation is strongest for the tranches that were information insensitive in the pre-pandemic period. Before the pandemic, the distribution of transaction prices of AAA tranches was uniformly uncorrelated with the vulnerable industries volatility index. During the pandemic, the lowest transaction prices for AAA-rated CLOs became correlated with an index of the volatility of the vulnerable industries’ stock prices, while the highest transaction prices remained relatively uncorrelated with the same index. Looking at the AAA tranches, the difference in the coefficients between the 25th and the 75th percentile is economically meaningful. The counterfactual price of a AAA CLO tranche that moved from the 75th percentile to the 25th percentile would have been 180 bps lower, given 27

Table 4: Information sensitivity of CLO debt tranches – quantile fixed effect regression. This table shows that highly-rated CLO debt tranches became information-sensitive during the pandemic. The dependent variable is the weightedaveragepriceofCLOtrancheiondayt,wheretheweightsarethetransaction volumes. Covid takes the value 0 before March 1, 2020 and 1 thereafter. Volatility t t is the weighted average daily difference between the high and low log prices on the seven vulnerable industries identified by Moody’s, where the weights are the transaction volumes. Column 1 includes only the CLO tranches rated AAA by at least one of the three main credit rating agencies. The quantile fixed effect regressions are implemented using the penalized fixed-effects estimation method proposed by Koenker (2004). Percentiles are indicated in the square parentheses. Clustered bootstrappedstandarderrors(1,000replications)areimplementedusingthegeneralized bootstrapofChatterjeeandBose(2005)withunitexponentialweightssampledforeach individualCUSIP,andreportedinparentheses. Source: Authors’calculationsfromdata provided by TRACE, BarChart, Bloomberg LP., Moody’s Analytics, S&P, and Fitch. ∗∗∗ p<0.01, ∗∗ p<0.05, ∗ p<0.1 Dep. var.: Debt class Trading price AAA A B C D E (1) (2) (3) (4) (5) (6) [0.25]Covid -0.59∗∗∗ -0.59∗∗∗ -1.23∗∗ -3.72∗∗∗ -12.54∗∗∗ -22.09∗∗∗ t (0.21) (0.22) (0.58) (0.89) (1.33) (1.89) [0.25]Volatility -2.41 -4.97∗∗ -24.88∗∗∗ -53.56∗∗∗ 0 -69.14∗∗ t (1.94) (2.25) (7.64) (13.32) (12.01) (31.54) [0.25]Covid×Volatility -79.04∗∗∗ -83.59∗∗∗ -150.69∗∗∗ -181.15∗∗∗ -232.75∗∗∗ -197.09∗∗∗ t t (8.03) (8.48) (23.16) (31.24) (37.71) (61.1) [0.5]Covid -1.23∗∗∗ -1.23∗∗∗ -2.1∗∗∗ -3.83∗∗∗ -10.97∗∗∗ -18.75∗∗∗ t (0.18) (0.17) (0.34) (0.6) (1.14) (1.88) [0.5]Volatility -1.35 -1.58 -15.87∗∗∗ -23.11∗∗ 1.53 -49.01∗ t (1.03) (1.45) (6.07) (11.44) (7.14) (25.69) [0.5]Covid×Volatility -36.24∗∗∗ -40.42∗∗∗ -78.8∗∗∗ -106.52∗∗∗ -149.41∗∗∗ -100.54∗ t t (6.29) (6.26) (13.54) (22.88) (32.74) (58.66) [0.75]Covid -1.55∗∗∗ -1.54∗∗∗ -2.15∗∗∗ -3.7∗∗∗ -8.91∗∗∗ -13.58∗∗∗ t (0.16) (0.15) (0.33) (0.5) (0.85) (1.82) [0.75]Volatility 1.3 0.22 -11.13∗ -7.74 0.41 -26.57 t (2.19) (1.58) (5.96) (9.71) (7.82) (25.3) [0.75]Covid×Volatility -12.46∗∗ -16.05∗∗∗ -44.23∗∗∗ -54.05∗∗∗ -81.18∗∗∗ -51.83 t t (5.47) (5.04) (12.02) (18.41) (20.82) (48.65) CUSIP FE Y Y Y Y Y Y Observations 16,417 18,208 5,899 6,734 8,386 6,017 H : [0.25]Covid×Volatility = [0.75]Covid×Volatility 0 t t t t χ2 test statistic 76.78∗∗∗ 77.67∗∗∗ 18.89∗∗∗ 13.59∗∗∗ 16.54∗∗∗ 5.04∗∗ 1 28

a one standard deviation increase in the vulnerable volatility index during the pandemic period. That change is almost two dollars per 100 face value, a huge difference. For some perspective on the 180 bps decrease, note that the standard deviation of AAA CLO tranche prices in the pre-pandemic period was 8 bps. Figure 6 shows the AAA tranche prices by the top and bottom quartiles, with one standard deviation bounds. This graphically shows the differentiation between CLO tranches. ThatinvestorsaredistinguishingbetweendifferentCLOtranchesmeans that they are distinguishing between different loans. We can also look at the loans in CLO portfolios. Panel (a) of Figure 7 shows loan prices by quartile. It is clear that, indeed, investors were differentiating been good andbadloans. AndthiswasattherootofdifferentialAAAtranchepricing. But note that it is not the case that the price dispersion is simply due to the pandemic shock to the vulnerable industries. Panel (b) of Figure 7 shows vulnerable loan prices by quartile. 4 The Informed and the Uninformed According to a Bloomberg article, on May 20, 2020, J.P. Morgan sent an extraordinary email to its clients warning them of “information 29

Figure 6: AAA CLO tranche prices by quartile. The figure shows the median AAA-rated tranche price ± one standard deviation for the tranches that are traded in the top and bottom quartiles of the price distributions. Source: Authors’ calculations from data provided by TRACE, Bloomberg LP, Moody’s Analytics, S&P, and Fitch. asymmetries” inthefast-movingCLOmarket.9 Thecitedcauseforconcern was originators, sponsors, and even managers that were interested in buying their own CLO tranches based on nonpublic information about “the potential effects of Covid-19 on the portfolio” that “may constitute insider information”. In this section we examine the risk to investors from such information asymmetries. Investors produce information about investment-grade debt securities when it is profitable to do so. After a corporate bond has been issued there are few subsequent analyst reports. Resources are only devoted to information production when the bond starts to deteriorate. Johnson, 9“JPMorganWarnsonInsiderTradingRiskinFast-MovingCLOMarket” byAlastair Marsh and Sally Bakewell, published May 20, 2020, 9:58AM EDT. 30

Figure 7: Loan prices by quartile. The figure shows the median loan price ± one standard deviation for the leveraged loans in the top (high) and bottom (low) quartiles of their price distributions. The left panel shows the prices for all leveraged loans and the right panel shows the prices for loans to firms in industries identified by Moody’s as vulnerable to the pandemic shock. Source: Authors’ calculations from data provided by Moody’s Analytics. (a) All loans (b) Loans to vulnerable firms Markov and Ramath (2009) write that: “the amount of resources devoted to debt research depends on the debt’s price sensitivity to information about the value of the asset. Intuitively, the sensitivity of the price of debt determines how much one can profit from information about the company’s assets in the debt market” (p. 92). In the case of AAA-rated CLO tranches there are never analysts’ reports. ButtherearemonthlyCLOtrusteereports. Inthefirstsubsection, we ask whether the information in those reports was used once the pandemic started. Then we look at a measure of disagreement between third-party loan price suppliers. These agents see the trustee reports but have no direct knowledge of the loan market. They proxy for uninformed agents. To be clear, informed agents are those who have produced private 31

information, meaning credit analysis of the loans in a CLO portfolio. These agents need to invest significantly in the technology required to produce the information. As Guggenheim Partners put it: “[analyzing a CLO] requirestheexpertisetoperformrigorousbottom-upresearchonindividual bank loans . . . managers must have significant corporate credit research capabilities”.10 This costly technology is the reason that investment firms manage multiple CLOs and may well have other credit products. Bain Capital, for example, has 310 employees in Bain Capital Credit.11 4.1 The Informed The first step in producing information about CLO tranche is finding out what loans are in the portfolio. This must be checked repeatedly because theCLOmanagercantradetheloansintheportfolio. Portfolioinformation comes from studying trustee reports, which are published monthly for each CLO. Each CLO has a trustee, a fiduciary, who carries out a number of tasks including reporting on the CLO’s portfolio composition, as well as its compliance with the many requirements of the CLO’s indenture, e.g., checking over-collateralization tests, the loan ratings, and the account balances. The trustee also maintains the CLO assets in custody and is responsible for paying funds to investors on coupon dates. 10See https://www.guggenheimpartners.com/perspectives/ portfolio-strategy/collateralized-loan-obligations-clos. 11See https://www.baincapitalcredit.com/approach. 32

The trustee reports are informative but, as mentioned above, they are onlythestartingpoint. Thesecondstepiscreditanalysisofthefirmswhich borrowed money and whose loans are in the portfolio. Credit analysts are necessary to take the information in the trustee reports and make it meaningful. We do not observe this second step except insofar as the AAA-tranche price changes. Further, and perhaps most importantly, the correlations between loans in the portfolio have to be studied. Figure 8 shows the percentage change in the price of AAA-rated tranches on the day the trustee report is released and the day after. Prior to the pandemic, the percentage price change was essentially zero (the tall bar), while after the pandemic the price changes are spread out, suggesting that there was valuable information in the trustee reports. Note that sometimes the percentage price change is positive. This can happen if the CLO manager took positive trading actions or a loan improved. 4.2 The Uninformed Bonds and loans are traded over-the-counter so there is no price discovery in the sense that a price aggregates many agents’ information sets. Like bond mutual fund managers, CLO managers engage a third-party pricing service to track the value of their loans, that is obtain a “price” for each loan. Loans do not trade frequently so these prices are essentially informed guesses. Pricing services use models and rely heavily on communication 33

Figure 8: Price changes around trustee report dates. The histogram shows the distribution of percentage price changes of AAA-rated CLO tranches traded on the same day or the day after each trustee report was released. Source: Authors’ calculations from data provided by TRACE, Moody’s Analytics, Bloomberg LP, Moody’s Analytics, S&P, and Fitch. from traders and the trustee reports. In other words, it is hard to come up with such prices. This difficulty has been shown by Cici, Gibson and Merrick (2011) who study bonds. In the case of bonds, the “prices”, called “marks”, are supplied by dealers using different methodologies. These are not transaction prices. Cici et al. (2011) study the dispersion of month-end mark-to-market prices for identical bonds held by many bond mutual funds. The marks from different dealer banks differ substantially (in normal times), even for AAArated bonds and the dispersion increases the lower the rating. In the case of loans, Markit and Reuters provide loan marks, “prices”. The basis for these prices is proprietary, but likely involve anecdotal 34

evidence from loan traders and models. These companies do not have large teams of credit analysts. Because these companies also see the trustee reports, they can proxy for uninformed traders who cannot analyze or understand the reports. To measure the degree to which the pandemic created problems and/or disagreements between these essentially uninformed agents we can measure the percentage difference between Markit and Reuters prices for the same vulnerable leveraged loan on the same date and compare that distribution to the distribution of the mark difference in pre-pandemic times. Figure 9 shows that disagreement increased significantly once the pandemic started. But there is no statistical difference between the disagreement about vulnerable and non-vulnerable industries. 5 Information-Sensitivity, Adverse Selection TheprecedinganalysisshowedthatAAACLOtranchesbecameinformationsensitive. In this section we argue that the switch from insensitive to sensitive coincided with the arrival of adverse selection or the fear of adverse selection. Prior to the pandemic, investors had not been producing information. There was no need to and it would not pay to produce information. In fact, they may not know how to produce information. It is not sufficient to just look at the portfolio. An investor must understand 35

Figure 9: Loan price disagreement by industry vulnerability. The figure shows monthly distributions of the percentage difference between Markit and Reuters prices for the same leveraged loan on the same date. Source: Authors’ calculations from data provided by Moody’s Analytics. how the loans to different industries are correlated and understand the internal workings of the CLO. Without this the investor risks facing a better informed trader. In this section we look at evidence of adverse selection. In the first subsection below we look at the imputed roundtrip trading cost (IRC), a measure of the bid-ask spread. Glosten and Milgrom (1985) show that the bid-ask spread widens when there is an increase in adverse selection.12 Then in the second subsection we look at the evidence for adverse selectionusingtheIRC.Thebid-askspreadisacommonmeasureofadverse selection in fixed income markets. For example, Benmelech and Bergman 12There are also interest costs and regulatory costs associated with holding an inventory. But agency trades are intraday. And, in any case, these did not change over our sample period. 36

(2018) use it in their study of corporate bonds and Wittenberg-Moerman (2008) uses it in her study of the secondary loan market. 5.1 Trading Costs WemeasuretradingcostsinthesecondarytradingmarketforCLOtranches using the method proposed by Feldhütter (2012). The imputed roundtrip trading cost (IRC) is the difference in the price paid by a dealer to purchase a bond from a client and the price charged by a dealer to sell the same amount of the same security to a client.13 Measuring the time between trades is important for markets with relatively thick trading and where the dealer is willing to carry risk in so-called “principal” trades. When trading CLO tranches, dealers almost always match clients in advance, in “agency” trades, and make very few principal trades. For this reason, the liquidity measure based on principal trades proposed by Choi and Huh (2017) is uninformativeinoursetting. Theratioofagencytradestoprincipaltrades, bothbynumberandbyvolume, isclosetoone.14 Weidentifyagencytrades as those with two or three trades in a given CLO tranche with the same trade size that take place within a calendar day.15 13We also estimated the realized bid-ask spread, calculated as the difference between the volume-weighted average prices paid to dealers by their clients and the volumeweighted average prices paid by dealers to clients. The two measures of liquidity are highly correlated. 14This measure underestimates the number and volume of agency trades because it treatsasprincipaltradesthosetransactionsthatthedealersplitovertwoormoreclients. 15Dick-Nielsen, Feldhütter and Lando (2012) calculate a similar measure of agency trades using corporate bond transactions within one day 37

To calculate the IRC, we first identify within-day roundtrip trades, composed of a sale from a client to a dealer and a purchase by a client from a dealer, potentially with inter-dealer trades. Following Dick-Nielsen et al. (2012), if we observe two or three trades in a given CLO tranche with the same trade size on the same day, and there are no other trades with the same size on that day, we define the transactions as part of a roundtrip trade. For each such trade, we calculate the IRC as the percentagedifferencebetweenthemaximumandminimumpricescontained in this roundtrip trade. We then calculate a daily weighted-average IRC over all roundtrip trades, where the weights are the volume of each trade. 5.2 Adverse Selection Figure 10 shows the estimated structural breaks in the IRC time series from applying the method of Bai and Perron (2003) for identifying multiple structural breaks in a single time series. The shaded areas are 95 percent confidence intervals. The first structural break is estimated to have occurred between March 3 and March 12, 2020, overlapping with the first break in the standard deviation of trading prices (Figure 13) and coinciding with the March 11 declaration by the World Health Organization of the pandemic. The second structural break in the IRC occurred between March 24 and April 23, 2020, with a point estimate of April 2. This break likely 38

Figure 10: Structural breaks in the imputed roundtrip cost (IRC). Thisfigureshowstheestimatedstructuralbreaksfromapplyingthemethodof Bai and Perron (2003) for identifying multiple structural breaks in a single time series. Source: Authors’ calculations from data provided by TRACE, Bloomberg LP, Moody’s Analytics, S&P, and Fitch. reflects the significant actions taken by the Federal Reserve to improve financial intermediation. In particular, on March 20, the PDCF became operational and accepted AAA-rated CLO tranches. On March 23, the Board of Governors announced the establishment of the TALF, PMCCF, and SMCCF as well as expanding the role of the Money Market Mutual Fund Liquidity Facility (MMLF) and Commercial Paper Funding Facility (CPFF). On April 1, the Board announced that it was relaxing the supplementary leverage ratio requirements “to allow banking organizations to expand their balance sheets as appropriate to continue to serve as financial intermediaries”. The third, and final, structural break in the IRC occurred between 39

May 13 and June 17, with a point estimate of May 14. The timing of this break overlaps (partly) with the second structural break in the standard deviationoftradingprices. Twoproximateactionswereintendedtofurther improve banks’ financial intermediation services. The Board of Governors modified the liquidity coverage ratio (May 5) and the supplementary leverage ratio (May 15). Figure11showsthatimputedroundtripcostsweredifferentiallyaffected by the pandemic. We separate the CLOs in our sample into two groups: Those above and those below the median holdings of loans to firms in vulnerable industries identified by Sallerson (2020). To avoid any potential confounding effect of secondary loan market trading, we calculate each CLO’s exposure to vulnerable industries from the last trustee report prior to the declaration of the pandemic. We find generally higher and more variable imputed roundtrip costs of trading CLOs that were more exposed to vulnerable industries during the pandemic. Nonparametric tests confirm that the two distributions are statistically different. We implemented both Anderson-Darling and Kruskal-Wallis rank tests using the trade data after the WHO declared the pandemic. In both cases, we reject with a p-value less than 5 percent the null hypothesis that the samples are drawn from a common distribution. Full details and the results of the tests are provided in appendix 9. We provide further evidence consistent with an increase in adverse 40

Figure 11: The distributions of imputed roundtrip costs by industry vulnerability. The figure shows weekly distributions of imputed roundtrip costs (IRC). The CLOs are separated into two groups that are determined by the portfolio of loans held as collateral. Source: Authors’ calculations from data provided by TRACE, Moody’s Analytics, Bloomberg LP, S&P, and Fitch. selection during the pandemic using quantile regressions of IRC on transaction prices (see Benmelech and Bergman (2018)). The dependent variable IRC is the imputed roundtrip cost of trading CLO tranche i on it day t, conditional on a trade taking place. Trading price is the weightedit average price of the CLO tranche i on day t, where the weights are the transaction volumes. Covid takes the value 0 before March 1, 2020 and 1 t 41

thereafter. The specification is given by: Q (τ|Covid ,Trading price ) = αi +β (τ)Covid IRC t it 1 t it +β (τ)Trading price 2 it +β (τ)Covid ×Trading price , (2) 3 t it with quantile τ ∈ {0.25,0.5,0.75} and where αi is the CUSIP fixed effect. Table 5 shows the results. Column1oftable5showsthatthenegativecorrelationbetweentheIRC and the transaction price for AAA-rated CLOs is consistent with adverse selection and the quantile regression reveals that the negative correlation is stronger for more illiquid (higher IRC quantile) CLOs. Moreover, the χ2 test of the null hypothesis of equality between the 25th and 75th percentile coefficientsontheinteractionterms(Covid ×Trading price )indicatesthat t it the negative correlation became stronger for those illiquid CLOs during the pandemic. By contrast, column 2 of the same table shows that the information sensitive E-class tranches did not experience the same shift. Of course, these results cannot be interpreted causally because the pandemic likely affected the liquidity of CLOs through several channels, not only through the decline in prices. Nevertheless, the findings are consistent with AAA-rated CLOs becoming information-sensitive during the pandemic. 42

Table 5: Adverse selection in AAA-rated CLOs during the pandemic – quantile fixed effect regression. This table shows there was an increase in the negative correlation between the imputed roundtrip cost (IRC) and the price of AAA-rated CLOs during the pandemic. The dependent variable is the weightedaveragepriceofCLOtrancheiondayt,wheretheweightsarethetransaction volumes. Covid takes the value 0 before March 1, 2020 and 1 thereafter. Column 1 t includesonlytheCLOtranchesratedAAAbyatleastoneofthethreemaincreditrating agencies. Thequantilefixedeffectregressionsareimplementedusingthepenalizedfixedeffects estimation method proposed by Koenker (2004). Percentiles are indicated in the square parentheses. Clustered bootstrapped standard errors (1,000 replications) are implemented using the generalized bootstrap of Chatterjee and Bose (2005) with unit exponential weights sampled for each individual CUSIP, and reported in parentheses. Source: Authors’calculationsfromdataprovidedbyTRACE,BloombergLP.,Moody’s Analytics, S&P, and Fitch. ∗∗∗ p<0.01, ∗∗ p<0.05, ∗ p<0.1 (1) AAA (2) E [0.25]Covid -120.56 -1.33 t (147.47) (44.9) [0.25]Traded price -3.26∗∗∗ -0.68 it (1.43) (0.45) [0.25]Covid ×Traded price 1.18 -0.03 t it (1.48) (0.46) [0.5]Covid 38.67 62.82 t (161.16) (49.86) [0.5]Traded price -3.92∗∗∗ -0.72 it (1.62) (0.48) [0.5]Covid ×Traded price -0.42 -0.75 t it (1.61) (0.52) [0.75]Covid 319.23∗ 29.99 t (183.09) (70.74) [0.75]Traded price -4.94∗∗∗ -2.03∗∗∗ it (1.78) (0.7) [0.75]Covid ×Traded price -3.22∗ -0.46 t it (1.83) (0.72) CUSIP FE Y Y Observations 4,551 3,013 χ2 test statistic† 13.22∗∗∗ 0.44 1 †H : [0.25]Covid ×Traded price = [0.75]Covid ×Traded price 0 t it t it 43

6 CLOs during the Pandemic During the pandemic, CLOs have continued to be issued. But their structure has been altered significantly to avoid adverse selection. A key issue is how much discretion the portfolio manager has to trade the portfolio, meaning that loans go in and out (see Table 3). More trading would require constant monitoring of new loans. As reported by Millar (2020) “New issues were largely either static or had one-year reinvestment periods in April...” In other words, the managerial discretion to alter the portfolio has been significantly reduced. This shift is shown in Figure 12. The dark shaded part of a bar is the number of CLOs that had a five-year reinvestment period, i.e., the manager could trade the portfolio during the first five years. Another key issue is how incentive compatible the manager’s stake is in the CLO. Deals have been restructured to decrease managerial discretion and also to increase the incentives of managers not to engage in insider trading of the sort J.P. Morgan warned about. Also, new deals are being driven by affiliated equity (equity from elsewhere in the firm that is not consolidated with the CLO managers). Small CLO managers do not have access to affiliated equity and so cannot issue new CLOs. Only those with reputation risk are able to issue, using affiliated equity. 44

Figure 12: US CLO count by reinvestment period. The bar chart shows the monthly count of newly-issued CLOs by the period over which managers havereinvestmentdiscretion. Source: Authors’calculationsfromdataprovidedbyS&P Global Market Intelligence. 7 Concluding remarks We traced the adverse selection dynamics in the CLO market, focusing on AAA tranches, before and after the pandemic. When securities are information-insensitive, it is not profitable to produce private information about their fundamentals ex ante and everyone knows this. AAA CLO tranches trade at par. Adverse selection is avoided. When AAA-rated CLO tranches become information-sensitive, as they did in the pandemic, uninformed traders are not prepared to produce information (see Hanson and Sunderam (2013)). They had not invested in the technology to do so (e.g., buying data, hiring analysts). The price of the security plummeted in part because uninformed investors do not know what the price should be. They face adverse selection. But informed investors do know what 45

the price should be. The price goes down but the dispersion of prices goes up because of the informed traders. In this paper we documented these dynamics of adverse selection in the CLO market when the pandemic hit. Note that if it were an economy-wide shock that introduced a risk premium for AAA-rated CLOs, then their prices would drop uniformly and there would not be dispersion across CLOs. The information about each CLO’s exposure to vulnerable industries in the trustee reports is not sufficient to explain the dispersion in prices. Importantly, it is one thing to know the exposure and another thing to know how loans/industries will perform. We showed in Figure 7 that there is dispersion in loan prices within the vulnerable/not industries and we showed in Figure 9 that there is price disagreement within the vulnerable/not industries. For these reasons, we believe that investors are producing private information about the loans held by CLOs. In normal times, investors don’t produce this private information. We used the variation in vulnerable/not industries to tease out further evidence for adverse selection. The quantile regression (Table 4) showed that only the lowest transaction prices for AAA-rated CLOs became correlated with uncertainty about the vulnerable industries during the pandemic. At the same time, the IRC increased more for those CLOs that were heavily exposed to vulnerable industries. The combination of these results is difficult to rationalize outside an environment where investors are 46

producing private information. When adverse selection sets in, the market becomes less liquid in the precise sense that the securities are information-sensitive; there is adverse selection. Such price drops have often been referred to as “fire sales” or “selling pressure”. These notions seem to imply that there are no buyers with available cash to invest, though no evidence has been shown that this is the case. It would also seem to imply that all the prices should drop more or less uniformly. We showed, contrary to these ideas, that the prices drop and become more dispersed due to adverse selection. References Akerlof, George, “TheMarketfor‘Lemons’: QualityUncertaintyandthe MarketMechanism,” Quarterly Journal of Economics,1970,84,488–500. Bai, Jushan and Pierre Perron,“ComputationandAnalysisofMultiple Structural Change Models,” Journal of Applied Econometrics, 2003, 18 (1), 1–22. Bao, Jack, Maureen O’Hara, and Xing Alex Zhou, “The Volcker Rule and Market-Making in Times of Stress,” Journal of Financial Economics, 2018. Benmelech, Efraim and Nittai Bergman, “Debt, Information, and Illiquidity,” Working Paper 25054, National Bureau of Economic Research September 2018. Brancati, Emanuele and Marco Macchiavelli, “The Information- Sensitivity of Debt in Good Times and Bad Times,” Journal of Financial Economics, 2019, 133 (1), 99–112. 47

Bruche, Max, Frederic Malherbe, and Ralf R Meisenzahl,“Pipeline Risk in Leveraged Loan Syndication,” The Review of Financial Studies, 03 2020. Caballero, Ricardo, Emmanuel Farhi, and Pierre-Olivier Gourinchas, “The Safe Assets Shortage Conundrum,” Journal of Economic Perspectives, 2017, 31 (3), 29–46. Chatterjee, Snigdhansu and Arup Bose, “Generalized Bootstrap for Estimating Equations,” The Annals of Statistics, 2005, 33 (1), 414–436. Choi, Jaewon and Yesol Huh, “Customer Liquidity Provision: Implications for Corporate Bond Transaction Costs,” Finance and Economics Discussion Series 2017-116, Federal Reserve Board 2017. Cici, Gjergji, Scott Gibson, and John Merrick, “Missing the Marks? Dispersion in Corporate Bond Valuations across Mutual Funds,” Journal of Financial Economics, 2011, 101, 206–2262. Dang, Tri Vi, Gary Gorton, and Bengt Holmström, “Ignorance, Debt and Financial Crises,” Working Paper, Columbia University 2018. , , and ,“TheInformationViewofFinancialCrises,” WorkingPaper 26074, National Bureau of Economic Research July 2019. DeMarco, Laurie, Emily Liu, and Tim Schmidt-Eisenlohr, “Who Owns U.S. CLO Securities? An Update by Tranche,” Technical Report, Federal Reserve Board June 2020. Dick-Nielsen, Jens, “Liquidity Biases in TRACE,” Journal of Fixed Income, 2009, 19 (2), 43–55. , Peter Feldhütter, and David Lando, “Corporate bond liquidity before and after the onset of the subprime crisis,” Journal of Financial Economics, 2012, 103 (3), 471–492. Feldhütter, Peter, “The Same Bond at Different Prices: Identifying SearchFrictionsandSellingPressures,” The Review of Financial Studies, 2012, 25 (4), 1155–1206. 48

Gallagher, Emily, Lawrence Schmidt, Allan Timmermann, and Russ Wermers, “Investor Information Acquisition and Money Market Fund Risk Rebalancing during the 2011-2012 Eurozone Crisis,” Review of Financial Studies, 2020, 33 (4), 1445–1482. Glosten, Lawrence and Paul Milgrom, “Bid, Ask and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders ,” Journal of Financial Economics, 1985, 14, 71–100. Gorton, Gary, “The history and economics of safe assets,” Annual Review of Economics, 2017, 9 (1). , Stefan Llewellen, and Andrew Metrick, “The Safe Asset Share,” American Economic Review, 2012, 102 (3), 101 – 106. Haddad, Valentin, Alan Moreira, and Tyler Muir, “When Selling BecomesViral: DisruptionsinDebtMarketsintheCOVID-19Crisisand theFed’sResponse,” WorkingPaper27168,NationalBureauofEconomic Research May 2020. Hanson, Samuel and Adi Sunderam, “Are There Too Many Safe Securities? Securitization and the Incentives for Information Production,” Journal of Financial Economics, 2013, 108, 565–584. Johnson, Rick, Stanimir Markov, and Sundaresh Ramath, “Sell- Side Analysts,” Journal of Accounting and Finance, 2009, 47, 91–107. Kargar, Mahyar, Benjamin Lester, David Lindsay, Shuo Liu, Pierre-Olivier Weill, and Diego Zuniga, “Corporate Bond Liquidity During the COVID-19 Crisis,” mimeo, 2020. Koenker, Roger, “Quantile Regression for Longitudinal Data,” Journal of Multivariate Analysis, 2004, 91 (1), 74–89. Krishnamurthy, Arvind and Annette Vissing-Jorgensen, “The Aggregate Demand for Treasury Debt,” Journal of Political Economy, 2012, 120 (2), 233–267. Liu, Emily and Tim Schmidt-Eisenlohr, “Who Owns U.S. CLO Securities?,” Technical Report, Federal Reserve Board July 2019. 49

Logan, Lorie, “The Federal Reserve’s Recent Actions to Support the Flow of Credit to Households and Businesses,” 2020. Speech at the Foreign Exchange Committee, FRB New York, New York City. Millar, Luke, “CLO market remains resilient in Q3, though concerns persist,” S&P Global, 2020. Monnet, Cyril and Erwan Quintin, “Rational Opacity,” Review of Financial Studies, 2017, 30, 4317–4348. Nadauld, Taylor and Michael Weisbach, “Did Securitization Affect the Cost of Corporate Debt?,” Journal of Financial Economics, 2012, 105 (2), 332–352. Perignon, Christophe, David Thesmar, and Guillaume Vuillemey, “Wholesale Funding Dry-Ups,” Journal of Finance, 2018, 73, 575–617. Powell, Jerome, “Business Debt and Our Dynamical Financial System,” 2019. Speech at “Mapping the Financial Frontier: What Does the Next Decade Hold?”, 24th Annual Financial Markets Conference, sponsored by the Federal Reserve Bank of Atlanta, Amelia Island, Florida. Sallerson, Peter, “CLO Deal & Manager Exposure Analysis by Industries Impacted Due to Coronovirus (COVID-19),” Moody’s Analytics Research, March 2020. van Binsbergen, Jules, William Diamond, and Marco Grotteria, “Risk-free Interest Rates,” Working Paper 26138, National Bureau of Economic Research August 2019. Wittenberg-Moerman, Regina, “The role of information asymmetry and financial reporting quality in debt trading: Evidence from the secondry loan market,” Journal of Accounting and Economics, 2008, 44, 240–260. 50

8 Data appendix – For Online Publication This appendix provides the details on data construction to replicate the analysis in the paper. 8.1 Bloomberg We begin our data construction using the asset-backed security Backoffice data from Bloomberg. We identify the universe of collateralized loan obligations (CLOs) using the identifier “mtg_deal_typ” column in the Backoffice data. Bloomberg provides cross-sectional information for every tranche issed by each CLO, including CUSIP identifiers. 8.2 Regulatory TRACE We use the CUSIP variable from the Bloomberg data to identify individual transactions on CLO tranches in the Trade Reporting and Compliance Engine (TRACE), created by the Financial Regulatory Authority (FINRA). 8.3 Ratings WealsousetheCUSIPvariablefromtheBloombergdatatoidentifyratings actions by Moody’s, S&P, and Fitch. Our rating information comes from direct daily feed from Moody’s, S&P, and Fitch that are added to the 51

initial historical baseline refresh that cover every security in our sample. The rating actions include when an agency places a tranche on watch to forewarn investors of a potential change in rating. 8.4 Moody’s loan-level data We combine loan-level information about the CLOs in our analysis from Moody’s. The majority of this information is obtained from monthly trustee reports prepared by each CLO and processed by Moody’s. In addition, Moody’s provides further information from pricing specialists (e.g. Loanx and Reuters) and in-house assessments. We link the loan-level data to the other datasets using a cross-walk provided by Moody’s to the Bloomberg identifier “mtg_deal_name”. 8.5 Barchart data Download the seven sectors listed in the table below from the website (https://www.barchart.com/stocks/sectors/rankings) and stack the files in a single .csv file. Our measure of volatility is the difference between the intra-day high and low log prices. We combine the seven sectors into a single weighted-average measure of volatility in the vulnerable sectors, where the weights are the volumes. 52

Table 6: Mapping Moody’s vulnerable sectors to Barchart sectors. The table below provides the seven sectors we used to calculate our measure of volatility. Moody’s Sector Barchart Sector Automotive Automobiles Consumer Goods: Durable Consumer Goods: Durable Household Products Energy: Oil & Gas Oil & Gas Producers Hotel, Gaming & Leisure Hotels Retail Retail Transportation: Cargo Industrial Transportation Transportation: Consumer Transportation Services 53

9 Additional results – For Online Publication 9.1 Testing the difference between IRC distributions Figure 11 showed that the imputed roundtrip cost (IRC) of trading a CLO depends on whether that CLO is more or less exposed to the industries identified by Moody’s as vulnerable to the pandemic shock. We formally test this hypothesis using the Anderson-Darling and Kruskal-Wallis rank tests for whether k samples are drawn from a common distribution. We divide all the IRC observations in the month following the declaration of the pandemic into two samples: Above and below the median market value of loans in the CLO collateral pool. Table 7 reports that there were roughly equal number of observations in the two samples. The left-hand panel and right-hand panels of Table 7 report the results from the Anderson-Darling and Kruskal-Wallis tests, respectively. The tables report both the asymptotic and simulated p-values, as well as two versions of the Anderson-Darling test that differ in how they treat “ties” i.e. identical values in a sample. In all cases, we can reject the null hypothesis that the two samples are drawn from a common distribution at less than 5 percent. 54

Table 7: Testing for significant differences between distributions of imputed roundtrip costs by vulnerability. Panels A and B show the results from Anderson-Darling and Kruskal-Wallis rank tests for differences in two distributions of imputed roundtrip costs (IRC). The two distributions are formed by separating CLOs into those above and those below the median share of their market value that is exposed to the industries identified by Moody’s as vulnerable to the pandemic shock. We calculate the exposure from the last trustee report prior to the pandemic. Source: Authors’ calculations from data provided by TRACE, Bloomberg LP, and Moody’s. (a) Anderson-Darling Test (b) Kruskal-Wallis Test Number of samples: 2 Number of samples: 2 Sample sizes: 314, 395 Sample sizes: 314, 395 Number of simulations: 10,000 Number of simulations: 10,000 asympt. sim. test asympt. sim. AD T.AD p-value p-valsuteatistic p-value p-value version 1: 3.967 3.908 0.009 0.00865.101 0.0239 0.0243 version 2: 3.980 3.931 0.009 0.0087 55

9.2 Loan transaction summary statistics without rampup period This table repeats the analysis in Table 3 excluding the transactions that occurred during the ramp-up period between the CLO closing date and the completion of the initial portfolio purchases. Because the data do not include a date for the end of the ramp-up period, we excluded all transactions that occurred in the two months after a CLO’s closing date. The ramp-up period typically lasts one or two months, so this is a conservative approach.16 9.3 Structural breaks in Class E CLO tranches Figure 13 presents the structural breaks in both the AAA rated CLO tranches (Panel A) and the Class E CLO tranches (Panel B). These results complement the presentation of Panel A only in Figure 5. The Class E tranches in Panel B are the most junior tranches that are debt. CLO equity is junior to the Class E tranche and does not trade. The results from the quantile fixed effects regression in Table 4 showed that Class E tranches were information-sensitive prior to the pandemic. 16https://www.pinebridge.com/insights/investing/2019/09/ clo-beyond-the-complexity 56

Table 8: Loan transactions summary statistics. The table shows summary statistics for leverage loan transactions per CLO in the prepandemic period (January 1, 2020—March 1, 2020) and the pandemic period (March 2, 2020—June 30, 2020). These summary statistics exclude transactions in a CLO’s ramp-up period by removing transactions in the two months immediately after a CLO’s closing date. The statistics are provided for all leveraged loans and separately for the sectors that Moody’s identifiedasvulnerabletothepandemicshock. Source: Moody’sAnalytics. Variable Statistic Alltransactions Vulnerablesector Non-vulnerablesector (units) Pre Post p-value Pre Post p-value Pre Post p-value Numberofsales Mean 47.16 82.79 0 6.83 10.99 0 45.09 78.77 0 (perCLO) SD 59.33 136.29 0 16.41 23.03 0 53.49 127.75 0 N 1,581 1,625 493 603 1,579 1,624 Meansalesvalue Mean 0.80 0.72 1 0.57 0.53 0.82 0.81 0.73 1 ($mnperCLO) SD 0.76 0.78 0.1 0.64 0.62 0.25 0.76 0.79 0.08 N 1,581 1,625 493 603 1,579 1,624 Numberofpurchases Mean 62.66 102.43 0 6 12.73 0 58.79 94.8 0 (perCLO) SD 63.55 146.42 0 13.52 29.2 0 55.96 130.94 0 N 1,549 1,592 1,001 961 1,549 1,591 Meanpurchasevalue Mean 1.16 0.84 1 1.1 0.77 1 1.17 0.85 1 ($mnperCLO) SD 0.91 0.86 0.01 1.01 1 0.37 0.92 0.86 0 N 1,549 1,592 1,001 961 1,549 1,591 57

Figure 13: Structural breaks in the standard deviation of CLO prices. This figure shows the estimated structural breaks from applying the method of Bai and Perron (2003) for identifying multiple structural breaks in a single time series. For each tranche, we calculate a daily weighted-average price, where the weights are transaction volumes. We then calculate the standard deviation across tranches. The optimal number of breaks to explain the time series is determined by the Bayesian Information Criterion. The solid line is the standard deviation of daily prices. The blue dot-dashed line is the fitted values of the regression including the structural breaks. The vertical dashed lines are the locations of the structural breaks. Source: Authors’ calculations from data provided by TRACE, FRED, and Bloomberg, LP. (a) AAA rated (b) Class E 58

9.4 Attachment points Figure 14: The distributions of AAA-rated CLO tranche attachment point by CLO vintages for the current population of CLO outstanding and the population of CLO traded in 2020. Roughly half of the triple A CLO population is traded in 2020 (779 cusip outof1684tripleAaretradedin2020). Source: Authors’calculationsfrom data provided by Bloomberg LP, Fitch, Standard & Poor’s and Moody’s. 59