Are CLO Collateral and Tranche Ratings Disconnected?

Abstract

Between March and August 2020, S&P and Moody’s downgraded approximately 25|$\%$| of collateral feeding into CLOs and only 2|$\%$| of tranche values, with rating actions concentrating in junior tranches. Both S&P and Moody’s modeling indicate that the impacts should have been considerably larger, especially for higher-rated tranches. Neither changes in correlation nor the accumulation of pre-COVID-19 protective cushions can explain the downgrade asymmetry on upper tranches. Instead, CLO managers repositioned their collateral pools to dampen the negative credit shock and rating agencies incorporated qualitative adjustments in their CLO ratings. Important potential policy and market implications from these findings are discussed.

Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online.

COVID-19 imposed significant financial stress across a wide swath of companies with market forecasts originally calling for bankruptcy levels to potentially exceed that of 2009.¹ Such claims were more pronounced for firms backed by speculative-grade debt. One particularly active source of such financing is the corporate leveraged loan market, whose dramatic growth in the years leading up to the COVID-19 crisis has drawn the attention and concern of policy makers (Warren 2018). Importantly, leveraged loans also serve as the underlying collateral backing collateralized loan obligation (CLO) claims. However, various CLO credit enhancement features transform a collateral pool (often carrying a “B” credit rating) into a tranched set of claims that vary dramatically in risk characteristics, credit ratings (from “AAA” to “CCC” or unrated), and likely end-investors. The COVID-19-induced economic shock provides a window into credit rating agency (CRA) responsiveness and processes in both the leveraged loan and CLO tranche credit rating markets, which while interconnected, differ in their structure, opacity, and demand for financial stability.

An intuitive tenet of rating agency methodologies is that the credit risk of the underlying collateral pool is a key driver in the overall credit risk of a CLO’s tranches. If assets in a CLO pool suddenly experience an increase in credit risk, this will decrease the value of the underlying assets and thus the value of tranches built upon these assets. However, an array of credit-enhancing features that include collateral diversification, overcollateralization, excess spreads earned on collateral relative to tranches, and cash flow diversion mechanisms may serve to concentrate a change in risk within a subset of a CLO’s capital structure.

To this end, we begin by modeling the relationship between an increase in collateral risk and tranche risk through the lens of a simulation approach based on S&P’s methodology. An important feature of rating agency modeling is that collateral risk is based solely on the underlying collateral’s credit ratings. Interestingly, the effects of collateral deterioration (downgrades) are broad based, with a relatively uniform impact on expected rating changes across tranches. While slightly less sensitive, modeling also suggests sizeable rating actions taken against AAA-rated tranches. We confirm these patterns using rating sensitivity analysis collected from Moody’s disclosures. Intuitively, the relatively uniform nature by which credit risk is spread across the tranche structure stems from the effect of collateral downgrades on the pool’s loss distribution. Increasing collateral risk shifts the entire distribution outward, increasing not only the mean but also the right tail of the loss distribution. As the senior tranche ratings reflect extremely low impairment probabilities, which are only realized in the states of the world corresponding to the right tail of the loss distribution, the small nominal increase in risk due to the loss distribution’s outward shift translates into an expected rating downgrade for a nontrivial share of senior tranches.

Having established general expectations, we examine post-COVID-19 CRA actions taken on underlying collateral and tranches for $591 billion of CLO debt issued between 2014 and 2019 from 1,185 CLOs.² Rating actions in March and April suggest a substantial increase in collateral risk, with S&P and Moody’s downgrading 30|$\%$| and 26|$\%$| of collateral, respectively. While ratings stabilized by mid-June, the credit risk for the mean CLO in our sample increased by 11|$\%$| between January 2020 and August 2020. For this increase in collateral risk, our modeling (corroborated by reported values from Moody’s disclosures) predicts rating downgrades for 27.3|$\%$| of tranches. In contrast, we find rating agency responses were far more subdued, with only 2.0|$\%$| (3.5|$\%$|⁠) of S&P and 1.75|$\%$| (5.5|$\%$|⁠) of Moody’s par-weighted tranches being downgraded (or placed on negative watch) as of August 2020. Interestingly, counter to inferences drawn from our simulations and Moody’s guidance, rating actions were concentrated squarely on subordinate tranches. Given the observed level of collateral deterioration, 9.2|$\%$| of AAA-rated tranches are expected to be downgraded according to [JN3]Moody’s guidance, whereas neither AAA nor AA-rated tranches were downgraded or placed on negative watch.

For a subset of 394 S&P-rated deals, monthly trustee reports disclose the results generated by S&P’s model for the AAA tranche. Consistent with the previous results, test results in June 2020 indicate that AAA-rated tranche in 12.5|$\%$| of the reporting CLOs are not able to withstand the level of collateral defaults possible in a scenario in which a AAA-rated tranche should be able to survive. Notwithstanding some improvement by August 2020, 8|$\%$| of AAA tranches continue to fail this evaluation test. This finding indicates that the lack of senior tranche downgrades is at least partially driven by rating agencies placing weight on nonmodel considerations in determining ratings.

Next, we consider additional forces that potentially mitigated rating actions taken on upper-tranches or reduced overall tranche downgrades relative to expectations given the severity of the economic shock. We consider three other potential contributing factors: (a) changes in underlying collateral correlations, (b) accumulation of protective cushions built up from prior performance, and (c) preemptive trading by CLO managers.

In modeling credit risk, CRAs assume an asset correlation structure that determines the shape of the pool’s loss distribution. Intuitively, a larger assumed asset correlation results in a thicker right tail. If COVID-19 served as a decoupling event that reduced asset correlations, this would effectively increase the diversification of the collateral pool, which disproportionally benefits senior tranches. Instead, 12-month rolling correlation estimates of underlying loan prices indicate that asset correlations within a collateral pool substantially increased at the onset of COVID-19. Furthermore, despite some variation across deals, the asset correlation increased for every CLO in our sample, inconsistent with rating inaction in senior tranches reflecting a reduction in asset correlation for select deals.

Another possible explanation for the scarcity of tranche downgrades is the accumulation of a protective cushion since issuance. This could take the form of either growing collateral pools, which bolster overcollateralization rates, or precrisis increases in the credit quality of collateral pools.³ Inconsistent with the accumulation of a protective cushion prior to the COVID-19 crisis, we find that senior tranche overcollateralization rates remain unchanged, while underlying pools generally increased in risk between issuance and January 2020.

The final explanation we consider is that CLO managers respond in such a way as to actively reduce risk. Consistent with this idea, we find trading behavior reduces model-implied credit risk in two ways. First, similar to the risk management strategy of a lender, managers actively tilt collateral pools toward shorter-maturing collateral from March to August 2020. Intuitively, a credit rating’s default probability is monotonically increasing in a loan’s time to maturity, a relationship reflected in CRA methodologies. Second, managers actively reduce credit risk through the disposition of lower-rated collateral and purchase of safer collateral during the COVID-19 crisis. These results highlight the role of active management in reducing model-implied collateral risk relative to a static CLO structure.

Taken together, our findings suggest that delayed tranche downgrades reflect nonmodel considerations and that active CLO management also mitigated COVID-19’s effect on tranche ratings. In a final analysis, we gauge the relative magnitude of these factors by modeling collateral risk under a set of counterfactual scenarios. While the senior tranches in 8|$\%$| of CLOs do not pass S&P’s modeling criteria for a AAA rating in August 2020, we find that this share increases to 13.5|$\%$| when removing the risk-mitigating benefit of collateral maturity shortening, 10.8|$\%$| when managers do not actively trade out of risky collateral, and 18.8|$\%$| when both active management channels are shut down. Overall, this suggests that model-implied tranche risk would be considerably greater had managers not repositioned their portfolios in response to COVID-19’s impact on the economy.

We now briefly discuss qualitative forces possibly represented in nonmodel factors that rating agencies might consider. The announcement of Fed intervention improved corporate debt prices, liquidity, and credit spreads (O’Hara and Zhou 2021; Haddad, Moreira, and Muir 2021; Kargar et al. 2021), generating a reversal of capital flows back into bond funds (Falato, Goldstein, and Hortaçsu 2021). While not codified in rating methodology, delayed tranche rating actions might reflect expectations of similar Fed intervention in the form of highly rated debt purchases. CRAs might also place increased emphasis on rating “through the cycle” for structured finance products relative to underlying collateral. However, this explanation is inconsistent with the observed heterogeneity in rating actions for senior relative to junior CLO tranches. Alternatively, if downgrades raise doubts over the stability of senior ratings going forward, a possible reduction in the demand for new issuance to be rated (Coval, Jurek, and Stafford 2009a) serves as a strong economic incentive not to downgrade. Finally, senior tranche downgrades, which reflect small nominal increases in default risk, could raise concerns about financial fragility if these downgrades trigger institutional sales, such as insurance or pension funds, which often must hold certain percentages of AAA or AA debt implicitly, thereby creating the potential for fire sales (Ellul, Jotikasthira, and Lundblad 2011; Nanda, Wu, and Zhou 2019; Elkamhi and Nozawa 2022).

Our paper is related to a larger body of work analyzing structured finance credit ratings leading up to the last financial crisis, and a smaller set of postcrisis rating analyses. In the 2000 to 2007 period, evidence was found in both RMBS and CDO markets for rating inflation (Ashcraft, Goldsmith-Pinkham, and Vickery 2010; Griffin and Tang 2012) and conflicts of interest (He, Qian, and Strahan 2012; Griffin, Nickerson, and Tang 2013; Kedia, Rajgopal, and Zhou 2014; Efing and Hau 2015).⁴ Substantial scrutiny and effort were spent on CRA reforms as part of the 2010 Dodd-Frank Act. Nevertheless, Baghai and Becker (2018) find that S&P issued higher ratings to regain fusion CMBS market share in 2012, while Flynn and Ghent (2018) find rating catering by new entrants in the CMBS market from 2009 to 2014. Our work contributes to this literature by suggesting that rating agencies may exhibit different standards regarding the timeliness of rating actions, complementing prior work examining different default probability standards across asset classes (Cornaggia, Cornaggia, and Hund 2017). We believe our paper provides the first academic glimpse of the state of structured finance credit ratings actions in the recent COVID-19 crisis. Finally, our paper is related to the recent work of Cordell, Roberts, and Schwert (Forthcoming), which finds that CLO equity tranches earned positive abnormal returns through the 2008 financial crisis. While at first blush these findings might appear at odds with each other, we view our paper as offering a different perspective able to speak to the nature of credit ratings rather than the realized performance of equity claims. The forward-looking nature of credit ratings, the small probability of default reflected in senior tranches, and the systematic nature of such right-tail events make it difficult to evaluate CLO credit ratings over two recessionary events.⁵

Ultimately, the precise motives for the delay in tranche downgrades in the COVID-19 crisis remains unclear because such motivations are not clearly discussed by rating agencies. One benefit of this paper is to document the use of nonmodel considerations. Increased rating agency disclosure of qualitative factors would likely aid market participants in assessing the validity of nonmodeled considerations and their capacity to increase or distort economic efficiency (Goldstein and Huang 2020). The more favorable and delayed credit rating actions toward tranches may incentivize the off-loading of systematically risky assets into structured products. These findings may have important implications for other areas of structured finance that also exhibit opacity in credit ratings. While market reforms like Dodd-Frank seek to decrease reliance on ratings, current Fed actions, which segment debt markets (including static CLOs) by credit ratings, further formalize their importance. These issues are practically important to policy makers and investors as substantial research has shown ratings’ impact on the allocation of capital and economic decisions.

1. Data, Sample and Market Conditions

1.1 Data and sample selection

Our main data are from Bloomberg and collected from CLO trustee reports. Monthly trustee reports provide information on collateral holdings and trades, deal and tranche characteristics, and performance metrics. From Bloomberg, we gather tranche rating actions, collateral pricing information, and supplementary data on collateral pool holdings.

To construct a recent sample of CLOs, we begin by collecting the collateral holdings data for all CLOs with closing dates from 2014 through 2019. Holdings for these deals are collected (typically at the monthly level) from January 2020 to August 2020. These data include details about security and issuer names, loan quantities, and reported ratings from Moody’s, S&P, and Fitch, among other information. We restrict the sample to holding snapshots in which one of Moody’s or S&P rates at least 85|$\%$| of the par value of collateral, yielding a sample of 1,185 CLOs. For these CLOs, we also collect all trades from January to August 2020.

Next, we collect holdings data on collateral pools from Bloomberg, which provides additional fields not present in our primary holdings data. Specifically, Bloomberg collateral data typically includes reported Moody’s and S&P industries as disclosed in the trustee reports and unique identifiers (i.e., FIGIs). We construct a matching algorithm based on commonality in holdings, quantities, maturity dates, and to some extent security names to establish a link between trustee report data and Bloomberg, with a final par-weighted matching rate of 97.7|$\%$|⁠. Details of the procedure can be found in the Internet Appendix.

When examining a CRA’s tranche rating actions we require the CRA to rate at least 75|$\%$| of a CLO’s tranches by par weight. Requiring Bloomberg coverage, as well as this restriction, reduces our sample to 1,110 CLOs, consistent with the occasional instance in which neither S&P nor Moody’s rated the majority of the tranches in a deal. Of this set, 782 deals are rated by Moody’s, whereas 406 deals are S&P rated (with 78 deals rated by both CRAs). Additionally, we estimate a cash flow model based on S&P’s rating approach in some tests, which requires additional data on a CLO’s deal features. This additional data requirement, along with the requirement that the CLO be rated by S&P, results in a sample of 345 CLOs. Finally, we also collect data on 947 CLOs from Moody’s rating announcements obtained from Factiva.

1.2 Summary statistics

With our sample in hand, we begin by briefly characterizing the structure of a CLO and the trading behavior of collateral managers. Table 1 reports summary statistics at the deal, holding snapshot, and trade level. The average deal in our sample has a total tranche par, which includes the equity tranche, of slightly less than $500M, yielding an aggregate par size of $591B for our sample of 1,185 CLOs.⁶ This sample represents approximately 70|$\%$| of global CLO debt outstanding as of year-end 2019.⁷ The average deal in our sample has 7.4 tranches, with a AAA tranche size of 60.6|$\%$| and an equity slice of 10.3|$\%$|⁠. This is lower than the 70|$\%$| AAA share documented by Benmelech and Dlugosz (2009) for prefinancial crisis CLOs. Backing these tranches is a collateral pool with an aggregate par of $483M, made up of 285 loans from 255 obligors. The collateral is relatively diverse in its industry origins, with an inverse HHI of 22.9 (based on S&P’s industry classifications). We see that Moody’s evaluation of the collateral is slightly more favorable than S&P, with average ordinal ratings of 14.43 for Moody’s and 14.73 for S&P.⁸ Finally, we see that the average collateral manager is quite active in trading, as 429,000 trades equates to approximately 47 trades per deal-month.

We continue by graphically depicting the composition of ratings for collateral and tranches of a typical CLO. Figure 1 shows the average collateral rating composition (panel A) and average tranche rating structure (panel B) by vintage. While the majority of collateral carries a B rating, the majority of tranche par is rated AAA. Moreover, the tranche structure is relatively stable across vintages, with a slightly larger AAA share and smaller equity share in more recent years.

1.3 Market conditions

We begin by documenting COVID-19-induced changes in rating-implied collateral risk, a key building block in determining tranche risk. Figure 2 examines the relative change in risk by initial rating. Specifically, we first map the credit rating of each underlying loan to the 10-year asset default rate for the corresponding rating agency, scaled by 10,000. Using rating agency terminology, this corresponds to the ‘rating factor’ used by Moody’s, from which a collateral pool’s WARF (the par-weighted average rating factor) is derived. We then partition collateral based on January credit ratings and recompute the par-weighted rating factor for each initial rating group through time, weighted by January 2020 holdings. Panel A reports the time-series evolution of WARF by initial rating, scaled by the initial value. The panel shows that S&P downgraded collateral from March through May. By June, WARF increased between 10|$\%$| and 20|$\%$| across rating classes, with the largest relative decline among BB-rated loans and the smallest increase in BBB-rated loans. Moody’s also started to downgrade collateral in early March, although the overall deterioration in collateral quality was not as severe. Instead, B and BB-rated loans continued to experience downgrades in July and August. In Figure IA.1, we compare the rating agency actions for all pieces of collateral that had the same rating in January and find that while S&P is generally more aggressive in their downgrading, this is not always the case.

Panel B of Figure 2 further decomposes the change in collateral ratings. The panel reports the distribution of rating changes from January to August across initial ratings. Both S&P and Moody’s rating actions are quite widespread across initial ratings with most rating categories of collateral experiencing between 20|$\%$| and 35|$\%$| downgrades.

Figure 2 provides a first glimpse at the potential increase in collateral risk brought on by the COVID-19 crisis. CLO collateral pools also experienced a negative shock during the 2009–2009 financial crisis. While we do not have granular collateral ratings going back to the financial crisis, to put the current collateral deterioration into context, we turn to S&P/LSTA’s BB-rated loan price index. Figure IA.2 plots the index price from 2005 to 2021, showing that loan prices fell to $60.2 in December 21, 2008, whereas the index only decreased to a low of $78 in March 22, 2020. Overall, the financial crisis was a considerably more stringent stress test for CLOs than the 2021 COVID-19 crisis, although it is important to note that, as with any crisis, there was considerable uncertainty, and the initial economic forecasts in March 2020 were commonly more dire than what materialized.

2. Modeling the Relationship between Collateral and Tranche Ratings

2.1 Collateral risk, tranche risk, and S&P’s methodology

To create a CLO, underlying corporate loans are pooled together and then the collateral pool’s cash flows are distributed among different tranches with differing levels of seniority. The credit risk of the underlying assets and the asset correlation structure are the two most important deal features as the diversification effect reduces idiosyncratic risk, leaving senior tranches mainly exposed to systematic risk (Coval, Jurek, and Stafford 2009b). An attractive feature of this setting is that collateral credit risk is inferred from credit ratings, and thus reflect CRA beliefs over such factors as the speed of an economic recovery, policy intervention, and general preferences to rate “over the cycle.” As the credit risk increases for many of the individual assets, this increases the average risk of the collateral pool, finally increasing the credit risk exposure of the CLO’s tranches. However, is the increasing credit risk distributed approximately evenly across tranches, or disproportionately concentrated in upper or lower tranches? This question is made more unclear given the different forms of credit enhancement present in a CLO. To answer this question, we first provide intuition by evaluating a single CLO using a simulation approach designed to replicate the primary features of S&P’s rating philosophy. In doing so, we are able to demonstrate the effect of collateral deterioration across the tranche structure.

S&P’s approach revolves around the comparison of two key metrics capturing: (a) credit risk of the collateral pool, and (b) a tranche’s resilience to defaults occurring in the collateral pool. The metric representing collateral credit risk, referred to as a Scenario Default Rate (SDR), is equivalent to the value-at-risk of the collateral pool’s default distribution. This collateral default distribution is estimated from Gaussian Copula Monte Carlo simulations, which rely on underlying credit ratings and S&P’s asset correlation assumptions. Importantly, S&P’s approach yields rating-contingent SDR values (e.g., AAA, AA).⁹

The metric capturing a tranche’s resilience, referred to as the “break-even default rate” (BDR), represents the share of the collateral pool that must default before the tranche holder is unlikely to be made whole. A tranche’s BDR is derived from a series of simulations using a CLO’s cash flow model, which incorporates the extent to which various credit enhancement features offer protection for a specific tranche. At the pool level, this cash flow model takes into consideration a pool’s excess spread, overcollateralization, and weighted-average recovery rate. Excess spread represents the difference in the weighted-average coupon rate of the underlying collateral relative to tranche notes. Intuitively, the greater the excess spread, the greater the ability of a collateral pool to incur some defaults and still have enough interest proceeds to satisfy tranche coupon payments. Overcollateralization represents the degree to which the total par value of the collateral pool exceeds the par value of the tranches. Thus, a greater degree of overcollateralization allows a collateral pool to suffer more defaults and still be able to make interim interest payments and pay down tranche principle. A final pool-level feature incorporated into the cash flow model is a pool’s weighted-average recovery rate. Intuitively, a greater recovery rate leads to more proceeds recouped from defaulted collateral and reinvested in other collateral. Finally, the cash flow model takes into consideration any interest coverage (IC) ratio and overcollateralization (OC) ratio tests present in a CLO. These tests, which may apply to one or more tranche classes, divert interest and principle payments to more senior tranches if the ratio of collateral pool-to-tranche interest payments (IC) or par values (OC) fall below predetermined “trigger” values.¹⁰

Ultimately, credit ratings are determined by comparing a CLO’s rating-contingent SDRs and tranche-specific BDRs. A rating is typically earned when the collateral pool’s expected default rate in an extreme circumstance (SDR) is less than what the tranche can withstand (BDR). To illustrate this process, we first evaluate a single CLO (Mariner 2015-1) using our replication of S&P’s process. This involves replicating S&P’s Monte Carlo approach to estimate the full set of SDRs (as outlined by Nickerson and Griffin 2017) and engineering a cash flow model based on S&P’s methodological disclosures designed to capture the different credit enhancements present in a given CLO.¹¹

Panel A of Figure 3 shows the status of the deal modeled using data as of February 2020. The solid blue line represents the distribution of underlying collateral defaults generated from Monte Carlo simulations. From this distribution, rating-contingent scenario default rates (SDRs) are computed, shown as dashed red lines in the figure. For example, in an extremely distressed state of the world (which S&P refers to as a “AAA” scenario, such as the Great Depression), approximately 60|$\%$| of the collateral pool is predicted to default. Likewise, in the less extreme and marginally more likely “AA” scenario, 53|$\%$| of the collateral pool is expected to default. From the cash flow model, we estimate the BDR for the most junior tranche in each class, reported as dotted blue lines. For example, the Class A tranche has a BDR of 63|$\%$|⁠. This suggests that the Class A can withstand a 63|$\%$| default rate in the underlying collateral pool before suffering delayed interest payments or losses. As only 60|$\%$| of the collateral pool is expected to default in a “AAA” scenario, this implies that the Class A tranche is able to achieve a AAA rating. In contrast, the 56|$\%$| BDR for Class B implies that the tranche would not receive timely and complete interest and principle payments given the expected 60|$\%$| default rate in a “AAA” scenario; therefore the tranche does not merit a AAA rating. Reported in brackets are the initial credit ratings assigned to each tranche, consistent with the comparison of SDR and BDR across tranches.

2.2 Changes in credit risk across tranches

What do changes in collateral credit risk imply about expected rating actions? Panel B of Figure 3 examines the potential effect across tranches of a 15|$\%$| increase in the default probability of the collateral pool, reporting the new default distribution (solid blue line) relative to the previous distribution (dashed-dotted grey line). Following a negative shock to the collateral pool, the distribution of possible defaults shifts outward and the right tail thickens. The result is an increase in the full set of rating-contingent SDRs. In contrast, BDRs are unchanged since each tranche is still able to withstand the same level of collateral defaults before facing impairment.¹² Following the thickening of the loss distribution, the AAA SDR exceeds the Class A’s BDR suggesting the tranche no longer warrants a AAA rating. An examination of the remaining tranches (with the exception of Class E) produces similar inferences with SDR exceeding BDR. While illustrative of a deteriorating collateral pool’s potential impact across tranche seniority levels, we cannot draw general inferences from the modeled effects of one CLO.

We now examine the relationship between collateral deterioration and expected rating actions across our full sample of S&P-rated deals. For each CLO, we start with the deal’s characteristics as of February 2020, increase the default probability of the loan pool by 15|$\%$|⁠, and reestimate credit ratings. Panel A of Figure 4 plots the expected change in rating partitioned by the initial credit ratings. Interestingly, following a 15|$\%$| increase in default risk, a significant share of tranches with initial ratings of AAA to BB- are expected to experience rating downgrades (of one or two notches). Of note, a small share of tranches are predicted to experience a rating upgrade. This reflects our modeling choice to calibrate the cash flow model for each CLO such that the average (prestressed) model-predicted cushion across tranches matches the reported AAA cushion in the February 2020 trustee report, possibly overstating the (unobservable) cushion on subordinate tranches in some CLOs. Overall, it appears that collateral deterioration has large effects both on junior and senior tranches.

Although calibrated to match February 2020 reported cushions, one concern is that our modeling approach misses an aspect of S&P’s methodology that has systematic ramifications.¹³ Fortunately, for a large portion of rated CLOs, Moody’s discloses guidance on expected credit rating implications associated with collateral deterioration scenarios (usually an increase in WARF of 15|$\%$| or 30|$\%$|⁠). Put simply, Moody’s performs an analogous exercise at issuance to the approach used in the previous panel.¹⁴ Panel B of Figure 4 graphically summarizes the expected effect of collateral deterioration across tranche seniority levels for a hand-collected sample of 947 announcements released by Moody’s. This set of announcements demonstrates a large overlap (616 deals) with our primary sample of 1,110 CLOs. Interestingly, following a 15|$\%$| increase in WARF, the majority of tranches with initial ratings between Aa1 down to Baa3 are predicted to experience a one or two notch downgrade according to Moody’s guidance, with 29.9|$\%$| of Aaa tranches experiencing a one notch downgrade. While comparable in magnitude, this final projection is slightly more aggressive than estimates from our modeling of S&P’s approach, where 20.3|$\%$| of AAA-rated tranches are expected to experience a downgrade. Overall, while modeling based on S&P’s approach seems to yield slightly more uniform downgrades across seniority levels, both approaches indicate considerable downgrades across junior and senior tranches.

Finally, we seek to more formerly quantify the effect of collateral deterioration on the collateral pool’s tail credit risk (e.g., the SDR), and the resultant implications for tranches with differing seniority levels. Doing so has the benefit of precisely measuring the rating-specific change in SDR resulting from an increase in underlying collateral risk.¹⁵ For each CLO-month of holdings data from January to August 2020, we first simulate a series of rating downgrades, and then compute the resultant set of SDRs. Table 2 presents the results when regressing rating-specific SDRs on the change in a collateral pool’s default probability (because of our simulated downgrades). The coefficient of 0.295 for the AAA SDR indicates that a 10|$\%$| increase in an underlying collateral pool’s default probability would result in a 2.95 percentage point increase in the AAA SDR |$(0.10 \times 0.295 =.0295)$|⁠. In comparison, the most sensitive rating class (BBB) is only 10|$\%$| more sensitive than the AAA SDR. Interestingly, sensitivity is not monotonic, with the B-rated SDR exhibiting the second lowest coefficient.¹⁶

One thing of note, the previous analysis based on S&P’s approach assumes that BDR values are constant. These tranche-specific break-even rates, which are based on cash flow modeling, are a function of a CLO’s overcollateralization, excess spread, and loan recovery rates and hence could change over time. In fact, Cordell, Roberts, and Schwert (Forthcoming) attribute the abnormal returns of CLO equity to a manager’s ability to both invest in higher yielding collateral and refinance senior tranches at lower rates, postfinancial crisis. If this mechanism is present in our setting, it would manifest as an increase in BDR values.

While non-AAA tranche BDR levels are not disclosed, trustee reports do report AAA BDR values for our sample of S&P-rated deals. Figure 5 plots the monthly change in BDR relative to February 2020 values and indicates considerable cash flow deterioration after the onset of COVID-19. BDR values decreased a full 2.0|$\%$| by July 2020 relative to February 2020 values. Multiple factors possibly contribute to this difference. First, if managers disposed deteriorating collateral and used the proceeds to purchase safer replacement collateral, this would likely reduce the collateral pool’s aggregate par value and thus overcollateralization ratios.¹⁷ Second, the benefit of a higher-yielding environment is generally limited to reinvested proceeds from maturing collateral, as managers would otherwise face a lower price when selling collateral prior to maturity. Third, while the refinancing of rated tranches at lower rates benefits equity holders, this is not reflected in tranche credit ratings, which only pertain to the current tranche structure rather than the refinanced (new) CLO. Fourth, the postfinancial crisis period represents a single draw of possible realization of credit spreads relative to collateral defaults. In contrast, the forward-looking nature of credit ratings are intended to capture all possible future states of the world.

2.3 Changes in CLO-level credit risk

We now examine how tranche sensitivity to collateral risk aggregates to the deal level across a range of collateral deterioration levels. Specifically, we reestimate the S&P-based approach for collateral risk increases ranging from 1|$\%$| to 25|$\%$|⁠, generate the predicted set of tranche downgrades, and aggregate the result to the deal level. Figure 6 shows the distribution of expected par-weighted tranche shares with a downgrade for a given increase in underlying collateral default risk. The simulated increase in underlying collateral risk is shown on the x-axis and the deal-level tranche downgrade share is on the y-axis. The figure reports the median value (dashed black line), and percentiles from 10|$\%$| to 90|$\%$| as shaded regions where each shade represents a 5|$\%$| band of the distribution.

For the median CLO that experiences a 15|$\%$| increase in default risk, approximately 26|$\%$| of tranches warrant a downgrade. The distribution is positively skewed, with a mean downgrade rate of 39.1|$\%$| across CLOs. This predicted value aligns closely with Moody’s sensitivity disclosures, where the mean par-weighted tranche downgrade rate is 42|$\%$| following a 15|$\%$| increase in default risk. For the 10th percentile (upper-most band), the predicted tranche-share downgraded in our simulations increases dramatically as one approaches a 7|$\%$| increase in collateral risk. As the (AAA-rated) Class A tranche typically makes up a large share of all rated tranches, this sudden jump in tranche-share represents the point at which the AAA tranche is downgraded. Variation in this large jump across percentile bands is consistent with significant cross-sectional resilience of AAA tranches to collateral downgrades, with the AAA tranche in the median CLOs being able to withstand more than a 25|$\%$| increase in collateral risk. Overall, while CLOs vary in their sensitivity to increases in collateral risk, an increase in collateral risk of 10|$\%$| to 25|$\%$| is generally associated with an expected downgrade in a nontrivial share of tranches.

3. How Do Collateral and Tranche Ratings Line Up?

We now turn from simulation-based predicted tranche downgrading behavior to understanding the relation between realized COVID-19 rating actions taken on collateral and CLO tranches. After examining collateral and tranche downgrades, we contrast these realized actions against expectations from Moody’s guidance and S&P’s reported AAA model results.

3.1 Collateral and tranche downgrades

To better understand the connection between realized collateral and tranche rating actions in 2020, panel A of Figure 7 plots the par-weighted percent of tranche rating actions (y-axis) against the increase in a CLO’s underlying collateral risk (x-axis). To measure the change in collateral risk, we compute the percentage change in a pool’s WARF as of August 2020 relative to January 2020. The figure reports the par-weighted share of tranches downgraded (hollow circles) as well as placed on credit watch (x markers). While downgrades do not demonstrate a noticeable relation with collateral deterioration, tranche rating actions that include negative credit watches taken by S&P are generally increasing in the change in underlying collateral risk. While a 5|$\%$| to 10|$\%$| increase in collateral risk is associated with tranche rating actions of approximately 2.5|$\%$|⁠, a 20|$\%$| increase in collateral risk is associated with tranche rating actions of 8|$\%$|⁠. Moody’s actions appear fairly similar, albeit slightly stronger, with 11|$\%$| of tranches being downgraded or placed on negative watch among CLOs experiencing the largest increases in collateral risk.¹⁸ However, the overall rate of tranche downgrade activity contrasts heavily with the modeled projections in Figure 6, where 26|$\%$| of par-weighted tranches are expected to experience a negative rating action in the median CLO following a 15|$\%$| increase in collateral risk. Overall, panel A of Figure 7 demonstrates an extremely modest relationship between collateral deterioration and tranche rating activity, one not anticipated when considering the analysis in Section 2.

3.2 Effects across tranches

The previous results are silent regarding the extent to which rating agency actions exhibit heterogeneity or uniformity across tranche classes. Panel B of Figure 7 reports tranche rating actions taken by S&P and Moody’s conditional on initial rating. Interestingly, the majority of the downgrade actions occur among more subordinate tranches, which represent less than 10|$\%$| of CLO capital. Between 25|$\%$| and 45|$\%$| of tranches rated BB or B have been placed on credit watch, while between 10|$\%$| and 20|$\%$| have been downgraded, with more aggressive actions for CCC-rated tranches. In contrast, tranches initially rated at or above A experience almost no rating changes, with very few watch list actions. In fact, no AAA tranches have been placed on credit watch or downgraded, while 0.4|$\%$| of AA-rated par has been placed on negative watch by Moody’s.

To put current rating actions into context and to understand whether AAA CLO tranches from the financial crisis also had few downgrades, we examine tranche downgrades for 707 precrisis CLOs from a sample studied in Griffin, Nickerson, and Tan (2013). Figure IA.6 shows that rating actions in the last crisis were considerably more severe, particularly among senior tranches. Moody’s and S&P downgraded 57|$\%$| and 71|$\%$| of AAA-rated tranches as of June 2010, respectively, with more frequent rating downgrades among tranches initially rated AA and A.

The concentration of rating actions taken in more subordinate tranches stands in stark contrast to the relatively broad-based rating downgrades predicted from our modeling approach and Moody’s disclosures. To this end, we now attempt to more directly compare expected and realized rating actions. Recall that panel B of Figure 4 plots the expected change in rating for a 15|$\%$| increase in collateral default probability based on Moody’s sensitivity analysis. Figure 8 overlays these projections on the realized tranche rating actions for the set of CLOs in which the underlying collateral has experienced more than a 15|$\%$| increase in WARF (260 CLOs) as of August. The figure illustrates the stark contrast between initial guidance provided by Moody’s and future rating actions, which are predominantly concentrated among subordinate CLO claims initially rated Baa1 or worse. To provide economic content, given the WARF increases observed in the data, Moody’s guidance would predict downgrades for 9.2|$\%$| of Aaa-rated tranches.¹⁹ We assess the statistical significance of the difference in expected versus realized rating actions for each rating category using a variation of the Fisher-Pitman permutation tests for independent samples (Kaiser 2007). Intuitively, the test compares the difference in observed downgrades relative to Moody’s projected downgrades to a distribution of differences occurring under the null that the two samples are drawn from the same distribution. The corresponding t-statistics, which are highly significant for the senior rating categories, are reported in brackets.²⁰ Overall, while Moody’s guidance suggests expected actions that are broad-based across most tranches, rating actions instead concentrate among junior tranches.

3.3 S&P’s model output

While the previous comparison of anticipated and realized rating actions is based on Moody’s disclosed sensitivity analysis, all S&P-related inferences up to this point are based on our replication of S&P’s disclosed methodology. However, unlike the SDR, insufficient detail in S&P’s methodological publications and lack of data on minor deal-specific waterfall considerations prevents an exact replication of the cash flow modeling needed to compute a tranche’s BDR. Fortunately, for a subset of 364 S&P-rated deals, monthly trustee reports disclose both the AAA SDR and BDR (but not values for other rating categories).

Figure 9 contrasts the difference between SDR and BDR (SDR – BDR) in February, June, and August 2020, reporting the kernel density as solid red, blue, and green lines, respectively.²¹ In February, 2.05|$\%$| of deals have a reported SDR that exceeds the BDR. By June 2020, this share substantially increased to 12.50|$\%$| of deals. Put differently, in June the AAA tranche in one-eighth of the reporting CLOs were not able to withstand the level of collateral pool defaults that S&P’s methodology estimates could occur in a scenario which a AAA-rated tranche should be able to survive. In August, 8|$\%$| of CLOs continued to have an SDR that exceeds the corresponding BDR, approximately 5 months after the start of the COVID-19 crisis. While only AAA SDR and BDR values are reported, recall that the simulations underlying Figure 4 suggest a relatively uniform downgrading behavior across the capital structure. This would suggest a nontrivial share of tranches initially rated AA and below face a similar situation where BDR does not exceed SDR. Similar to Moody’s, no rating actions have been taken by S&P on AAA-rated tranches as of August. Overall, both the lack of downgrading by S&P in instances in which AAA BDRs fall below their respective SDR values and divergence between Moody’s projected and realized rating actions are consistent with the reliance on nonmodeled considerations.

4. Why Are CLO Tranches Not Experiencing More Downgrades?

Counter to a relatively uniform set of expected rating actions across seniority levels, the observed concentration of rating actions on subordinate tranches appears consistent with the incorporation of nonmodeled considerations. However, downgrading behavior possibly reflects forces that mitigated actions on upper tranches. Additionally, economic forces may provide intuition for any nonmodeled considerations. We now consider three forces possibly related to the observed downgrading behavior: (a) changes in asset correlations or differences in the cross-sectional distributions of correlations across CLOs, (b) accumulation of protective cushions built up from prior performance, and (c) preemptive trading by CLO managers. We conclude the section with a discussion of the remaining potential explanations for the findings and possible implications.

4.1 The potential role of asset correlations

Default correlations among assets in CLO pools play a pivotal role in the transformation of risk and the concentration of tail risk. Credit rating agencies use internal default correlation assumptions to model tail risk, which is pivotal in determining upper tranche sizes but also extremely sensitive to correlation assumptions (Nickerson and Griffin 2017). Given the importance of asset correlations, it is possible the lack of upper tranche downgrades reflects either changes in asset correlations during the crisis or cross-sectional differences in correlation across CLOs.

To this end, we begin by considering the possibility that rating agencies revised asset correlation assumptions. If rating agencies viewed COVID-19 as a decoupling event that decreased pairwise asset correlations and revised assumptions downward in response, this predicts both (1) a reduction in the share of tranches downgraded relative to prior expectations and (2) a thinning of the right tail of the collateral pool’s default distribution, disproportionately benefiting more senior tranches from increased diversification of the underlying collateral pool.

To test this hypothesis, we turn to the observed correlation in loan valuations. Although we are unable to directly observe rating agency correlation assumptions, loan prices provide a market-based measure of the extent to which credit risk comoves across loans. Since observed transactions occur relatively infrequently, we turn to Bloomberg’s valuation algorithm (BVAL), which provides loan valuations for our sample at a weekly interval from January 2014 through the end of our sample period. Inconsistent with a post-COVID-19 reduction in asset correlation, March 2020 marks a significant increase in the average pairwise correlation, which reaches its greatest value of 0.57 in August 2020 (as shown in Figure IA.9).²²

However, the average correlation could be masking significant heterogeneity across individual CLO collateral pools. We now examine the cross-section of changes in loan price comovement across CLOs. For each CLO-month in our sample, we compute the average pairwise 52-week correlation for the loan pool. Figure 10 illustrates the distribution of this asset correlation across CLOs relative to January 2020 values. Interestingly, all CLOs in our sample exhibit an increase in loan price correlation by April 2020.²³ As of August 2020, the median CLO experiences a 50-pp increase in loan price correlation relative to January 2020, with a minimum increase of 11 pp across all CLOs in the sample. Overall, the results in the figure are inconsistent with a lowering of rating agency correlation assumptions able to explain the attenuated response of rating agencies with respect to tranche downgrade actions. Nickerson and Griffin (2017) note that in response to the financial crisis, Moody’s and S&P made upward revisions to default correlation assumptions. It will be of interest to see whether similar responses are taken later in this context as well.

One advantage when considering a change in the weighted-average default probability of the collateral pool, as we do in Section 3, is that it represents a summary measure able to capture the intensity with which the underlying collateral was affected by a negative shock, whether because of overexposure to a severely hard-hit industry or because of stress more broadly felt across industries. However, differential changes in asset correlations might enter into tranche rating determinations in a different fashion. Specifically, Standard and Poor’s (2019b) finds that a greater degree of rating diffusion (“Default Rate Dispersion”) is associated with a smaller SDR. If less diversified CLOs had above-average exposure to sectors with more severe actions, then the resultant “barbell” distribution would demonstrate a greater degree of rating dispersion.

To gauge the economic importance of this alternative, we again adapt S&P’s simulation approach used to model collateral default risk (SDR). Intuitively, our approach involves contrasting the change in SDR due to an equivalent increase in weighted-average default probability due to (a) many small rating changes possibly experienced by a more diversified pool and (b) fewer, but more severe (two- or three-notch) downgrades possibly experienced by a less diversified pool. For each CLO’s holdings as of February 2020, we first perform a series of simulations where we randomly seed one-notch rating downgrades. We then compute the change in the pool’s weighted-average default probability and the full set of rating-contingent SDRs for each simulation.²⁴ We then repeat the process when seeding two- and three-notch downgrades. The resultant panel of SDRs and changes in average pool risk allows us to contrast the relationship between collateral risk changes and SDR changes due to many small downgrades versus a few large downgrades.

Figure IA.10 reports the results for three rating-contingent SDRs: AAA, A, and BB. Panel A reports the effects on AAA SDR, yielding two interesting findings. Consistent with Standard and Poor’s (2019b), tail risk is less responsive to a change in default risk when due to fewer but more severe downgrades. Second, this effect appears to be economically modest. Interestingly, the remaining panels demonstrate that rating dispersion plays a smaller role for more junior tranches. Overall, Figure IA.10 suggests an economically small effect played by the magnitude of rating downgrades.

Finally, we examine the extent to which either a change in pool diversification or resultant rating dispersion can explain the small number of tranche downgrades observed. Table 3 reports the results of OLS regressions, which regress CLO-level tranche rating actions on the change in the collateral pool’s weighted-average default probability, and three different deal-level measures of collateral diversification/rating dispersion.²⁵

Panel A of Table 3 considers the par-weighted percent of tranches downgraded or placed on negative watch as of August 2020. The first specification serves as a base case, examining the relation between Moody’s collateral deterioration and tranche downgrades. Consistent with the attenuated relationship between collateral deterioration and tranche downgrades depicted in Figure 7, the coefficient indicates that a 10|$\%$| deterioration in underlying collateral risk is associated with rating actions being taken on 2.8|$\%$| of a CLO’s tranches. While the inclusion of the diversification and dispersion measures in Specifications 2, 3, and 4 has a modest effect on the relation between collateral deterioration and tranche rating actions, in no specification does the coefficient approach the predicted effects from Section 2. The largest coefficient across the three specifications suggests that a 15|$\%$| increase in collateral credit risk is associated with rating actions taken on 5.9|$\%$| of a CLO’s tranches. For comparison, Moody’s disclosure data indicates that the average deal is expected to have a rating action taken on 42|$\%$| of its tranches following a 15|$\%$| increase in collateral risk.²⁶ The remaining columns examine S&P’s rating actions and finds that collateral deterioration and tranche rating actions continue to fall well short of expected actions discussed in Section 2. Panel B repeats the previous exercise when considering tranche rating actions (but not credit watch events), with similar inferences being drawn from the results.²⁷ Overall, there is little evidence that changes in asset correlations during the crisis, or differences in the cross-sectional distributions of correlations across CLOs explain our findings.

4.2 Potential pre-COVID-19 accumulation of a protective buffer

Structured products are typically rated “at-the-edge” when issued, meaning that they maximize senior tranche sizes allowable by rating agency models (Griffin and Tang 2012) in order to maximize revenue. However, an increase in collateral credit risk might erode tranche value, but not to the point of affecting ratings if current deals have accumulated a protective cushion since issuance. This could be in the form of growing collateral pools, which bolster overcollateralization rates or precrisis increases in the credit quality of collateral pools. If true, this would suggest that pre-COVID-19 ratings were overly conservative and were only brought back in line during the crisis. While this alternative could not explain the lack of AAA downgrades by S&P in instances where BDR exceeds SDR (as reported SDR and BDR values reflect such cushions), it might partially explain deviations by Moody’s or the degree of heterogeneity more generally observed in downgrades across seniority levels.

First, we consider the potential change in credit risk of collateral pools. Figure 11 reports the change in WARF since issuance for yearly vintages of CLOs from the start of 2018 to present. The positive values as of January 2018 indicate that the average collateral in older vintages has increased in risk since issuance. Moreover, between July 2018 and January 2020 the average credit risk for each yearly vintage has trended upward, inconsistent with the accumulation of a buffer due to increasing collateral quality.²⁸

Second, the active management component of CLOs offers a second avenue through which deals can accumulate a cushion, as managers sell collateral that has improved in quality to finance a larger share of collateral with risk characteristics in line with the collateral pool. The result of this process would be an increase in overcollateralization ratios over time. Figure 11 also examines this possibility, reporting the difference in OC ratios for different rating classes relative to the ratio at issuance. Contrary to the accumulation of a cushion, January 2020 overcollateralization ratios are generally slightly lower than their levels at issuance. In fact, only the B-rated class exhibits an average OC ratio, which improves from issuance to January 2020. Taken together, these results do not support the hypothesis that a general lack of downgrades, or fewer downgrades among more senior tranches, were due to the accumulation of a protective cushion since issuance.

We now turn to market prices as an alternate measure of collateral credit risk. Panel A of Figure 12 reports the distribution of the value-weighted price of CLO collateral pools at weekly intervals. The average price of CLO loans started around $97.50 before experiencing a precipitous drop in March, settling near $83. Such an immediate and large drop might serve to pressure rating agencies to downgrade underlying collateral, whereas the opacity and illiquidity in the secondary market for CLO tranches potentially afforded rating agencies a greater ability to delay rating actions. Following this decline, average prices recovered to slightly more than $90 by the beginning of July. Ultimately, after approximately 6 months at the end of our sample, the average price of a CLO’s collateral pool stood roughly $5 lower than its pre-COVID-19 value. To the extent that this partial recovery informs rating agencies of a change in economic conditions, while not reflected in subsequent collateral upgrades, this might serve to forestall tranche rating actions.

To provide additional economic content on the relation between collateral value changes and tranche risk, we now present a simple measure motivated by the idea of an OC ratio. Recall, the OC ratio represents the par value of a collateral pool relative to the liability of a tranche and all tranches more senior. For example, a CLO with a collateral par value of $150 million, a $100 million AAA tranche, and a $20 million AA tranche will have a AAA OC ratio of 1.5 and a AA ratio of 1.25 (150/(100+20)). We adapt this idea in two ways. First, we consider the collateral pool’s market value rather than the par value. Second, if tranche liabilities exceed the aggregate collateral value, we instead consider the ratio of the aggregate collateral value in excess of all tranches more senior (as these tranches would be paid off first) relative to the tranche’s value. The result is a ratio we refer to as the “tranche liability coverage ratio.” A ratio less than one indicates that if the CLO pool were to be liquated at current market prices, the owner of the tranche in question would not be made whole. However, as long as pools are not forced to liquidate, a tranche with a ratio below one can still have considerable value as it can continue to collect interest and principle payments in the future.

Panel B of Figure 12 contrasts the distribution of tranche liability coverage ratios in June and August (in green) against February ratios (in blue). The first thing to note is that tranche liability coverages declined considerably by June 2020. While some AA, A, and BBB tranches would be impaired if faced with liquidation under June prices, all AAA tranches would be made whole. At the same time, liability coverage ratios experienced a partial reversal by August 2020. While these ratios remain below February 2020 values, indicative of a reduction in the cushion protecting senior tranches of some CLOs, the stabilization and increase in coverage ratios from the height of the COVID-19 crisis might serve as motivation for a delay in rating actions taken against CLO tranches.

4.3 Active trading and a reduction in model-implied credit risk

The final class of explanations that we consider relate to manager trading actions.²⁹ One way in which a CLO manager may strategically respond to the onset of the COVID-19 crisis is to adjust trading behavior to reduce the maturity of the collateral pool. Such actions may represent a form of risk management, similar to the risk management strategy of a lender. In the context of CRA methodologies, this possibility is reflected in the use of the current collateral pool’s weighted-average maturity (WAL) when determining the pool’s default probability.³⁰ As such, the ability to trade into shorter maturity loans provides a potential lever available to a collateral manager to reduce the model-implied risk of a collateral pool.

Panel A of Table 4 reports the results of OLS regressions examining this possibility, contrasting the remaining years to maturity of loans purchased against those sold through time. The key variables of interest are a set of trading month dummy variables interacted with an indicator for the transaction being a purchase. The inclusion of trade-month fixed effects subsumes the average remaining maturity of loans sold in each trading month. In the first column, the coefficient for 1(Purchased) indicates that loans purchased are, on average, 0.899 years longer-lived than those sold in the base case month of January 2020. This is to be expected, as capital is rolled over during the reinvestment period. Interestingly, the remaining point estimates indicate that the difference in the life span of purchased versus sold loans narrows considerably in March and continues through the end of the sample in August. From an economic standpoint, the difference is quite large, with point estimates ranging from a 0.325-year decrease (March) to a 0.72-year decrease (July). These inferences remain relatively unchanged when including deal-month fixed effects in the second column, as well as the inclusion of rating-month fixed effects in the final two specifications to control for potential differences in the distribution of maturities across loans of different credit quality.

If managers are tilting collateral pools toward shorter-lived assets as a means of reducing credit risk, it is plausible that this action would be exacerbated in deals experiencing a greater deterioration of collateral. Panel B of Table 4 explores potential heterogeneity in the previous result. To focus on the cross-section of the effect, all specifications include the interaction of trade direction (purchase vs. sell) and trade-month, thus subsuming the time-series relation documented in panel A of Table 4. The variable of interest in the first four columns is the interaction of 1(Purchase) with a measure of recent collateral performance, Change WARF, which equals the month-over-month percentage change in the WARF of a collateral pool. Here, a negative coefficient implies that CLOs experiencing a bigger decline in collateral quality (increase in WARF) tilt their purchases more toward short lived collateral. When measuring WARF using Moody’s ratings, we see a negative and statistically significant coefficient for the interaction term. In the second specification, which accounts for potential heterogeneity in manager propensity to buy longer- versus shorter-lived assets with a deal-by-trade direction (purchase vs. sell) fixed effect, the point estimate of |$-$|1.163 indicates that deals experiencing a 10|$\%$| relative increase in WARF reduce the average life of loans purchased by 1.4 months (⁠|$-1.163 \times.10 \times 12$| months). We find similar effects when measuring collateral deterioration using S&P’s credit ratings. In the fourth specification, the point estimate of |$-$|1.491 (t-statistic |$= -$|5.41) indicates that a 10|$\%$| increase in collateral risk is associated with a 1.78 month decrease in the relative years to maturity of purchased loans compared to sold loans. Note, the previous specifications focus on deteriorating collateral, which may not capture the overall change in CLO tranche risk (e.g., if cash flow protections are also changing). Fortunately, for a subset of S&P-rated CLOs we are able to observe both components, SDR and BDR. In the final two columns, we shift focus to the difference between AAA SDR and BDR (measured in percentage points). Here, although not comparable in magnitude to an increase in WARF, an increase in the difference continues to correspond to an increase in the AAA tranche risk. In the final specification, the coefficient of |$-$|0.018 (t-stat |$= -$|1.90) on the interaction term indicates that a 1|$\%$| decrease in the AAA cushion (a 1|$\%$| increase in SDR – BDR) is associated with a 0.22-month relative decrease in the life of purchased collateral (⁠|$-0.018 \times 12$|⁠).

Beyond actively reducing the maturity of the loan pools, managers may also actively reduce the pool’s risk profile through the replacement of lower-rated collateral with higher-rated loans. Figure IA.12 examines this possibility, with panel A contrasting the par-share of collateral purchases with sells across rating categories. Pre-March buying and selling propensities, represented by solid circles, are relatively balanced for a given credit rating. In contrast, post-March trades are consistent with a net purchasing of higher-rated collateral and net selling of lower-rated collateral (B and below). This suggests that managers are actively purchasing higher-rated assets through their trading activity. However, while this action mitigates credit risk, the sale of risky collateral to finance the purchase of safer assets likely reduces the aggregate pool par value. While these actions reduce a deal’s collateral risk profile (e.g., SDR) they may do so by undermining the cash flow protection of the deal through a reduction in overcollateralization. This effect is consistent with the decrease in BDR found in Figure 5.

Another related possibility is that CLO managers made attempts to boost cash flows available to tranche and equity holders by “reaching for yield” within rating buckets, in a similar manner as documented by Becker and Ivashina (2015) for insurance companies. Instead, we find that managers purchase lower (rating-adjusted) yielding loans in the disruptive part of the COVID-19 crisis (from March through June 2020) as shown in Table IA.2.³¹

Taken together, the results in this section shed light on potential factors resulting in an attenuated response of tranche ratings to COVID-19-induced collateral deterioration relative to what one might expect. While we find no evidence for a reduction in collateral correlation or accumulation of a protective buffer, the results presented in Table 4 indicate that managers tilted collateral portfolios toward shorter-lived assets during the COVID-19 crisis. Though the effect cannot fully explain a lack of tranche rating actions, through the lens of rating agency methodologies this action reduced the model-implied default probability of the underlying collateral. We revisit this idea in Section 5, where we estimate the relative economic magnitude of different factors that influence tranche risk. However, before doing so, we discuss other potential economic forces at play and the impacts thereof.

4.4 Discussion of potential qualitative considerations

Credit rating agencies rapidly downgraded collateral in March and April of 2020 but took minimal actions on CLO tranches, which appears partially inconsistent with Moody’s and S&P models. While not formally incorporated as quantitative factors in rating agency methodology, potential explanations for this divergence include Federal Reserve policy on structured credit, differential emphasis of rating through the cycle, the economic ramifications of downgrading, and the financial incentives of rating agencies. Goldstein and Huang (2020) provide an interesting framework with which to evaluate each explanation, modeling the real effects of credit rating agency (CRA) actions in a rational environment.³² In the model real effects arise from two key sources, informational effects and strategic effects, which have different implications for economic welfare. Informational effects are those that arise when the CRA makes rating choices that are not a function of the CRA’s expected profitability, which lead to efficient investment. Strategic effects occur when the credit rating decision instead acknowledges the real effects caused by a rating and thus also reflects the resultant expected profit, yielding more favorable ratings.³³ Strategic effects are often harmful as firms can use the resultant cheaper cost of capital to fund value-destroying investment.

One potential explanation for a delay in tranche rating actions is that it might reflect CRA expectations over Federal Reserve actions (e.g., TALF), which previously had a measured impact on the corporate bond market.³⁴ While TALF expansions only went so far as to include AAA-rated static CLOs, CRAs might have plausibly expected a future expansion to include actively managed CLOs, such as those in our sample. To examine this possibility, we leverage insights from Flanagan and Purnanandam (2020), which finds that the Federal Reserve was more likely to purchase corporate bonds used as repo collateral and held in bond funds. Given this, it is plausible that the Fed would follow suit for CLOs if the TALF was ever expanded to include actively managed deals. To examine this possibility, we utilize repo collateral data reported by money market mutual funds in Form N-MFP-2 and fund holdings from the CRSP holdings data set. For each CLO, we compute the share of rated tranches used as repo collateral (Repo share) or held in bond funds (Fund share), averaged over the 12-month period prior to February 2020. We test for a differential effect of collateral deterioration on tranche rating actions as a function of the level of repo usage or fund holdings. We do not find any statistically significant evidence of a differential response for S&P or Moody’s, either when including rating watches or when including only rating actions (as shown in Table IA.3).

Alternatively, CRAs may have also considered the potential market disruption and financial fragility effects as a result of senior tranches downgrades. Many holders of AAA CLO securities, such as pension funds, insurance companies, and money market funds, are required or strongly prefer to hold such highly rated instruments and might need to sell if a rating is downgraded. In addition to the transaction costs from asset sales, downgrading could create a short-term market disruption and fire sales among downgraded tranches (Ellul, Jotikasthira, and Lundblad 2011; Nanda, Wu, and Zhou 2019; Elkamhi and Nozawa 2022). Consequently, this could generate concerns regarding financial fragility (Greenwood and Thesmar 2011). Thus, rating agencies might be particularly reluctant to downgrade AAA instruments given the economic costs this might impose to participants, particularly if economic conditions could improve. This preference for increased rating stability is somewhat related to the well-known CRA practice of rating “through the cycle.” While CRAs may possibly place increased emphasis on rating through the cycle for CLO tranches relative to underlying collateral, this explanation is only able to explain a differential response across, rather than within, the two asset classes. As such, this explanation is not consistent with the heterogenous downgrading behavior shown between senior and junior tranches. Moreover, while we cannot rule out differential preferences across asset classes, such practices seem somewhat at odds with recent CRA disclosures.³⁵

However, beyond the pure economic costs to market participants, the financial incentives of the CRA are closely intertwined with any feedback effect due to ratings. To the extent that senior tranche downgrades have the capacity to create a market disruption, thereby decreasing subsequent demand for CLOs (senior tranches in particular), future CRA fees will suffer as a consequence. In the context of Goldstein and Huang (2020), this explanation represents a strategic effect whereby a CRA takes into consideration the feedback effect of its rating choice, resulting in inflated ratings. Skreta and Veldkamp (2009) and Sangiorgi and Spatt (2017) show that financial incentives may increase in relative importance in opaque sectors, such as structured finance products, which potentially allows for more rating inflation and catering. Consistent with this idea, Cornaggia et al. (2017) show more rating inflation in more complex asset classes. Rather than taking rating actions against senior tranches, which represent a small nominal increase in default risk with large economic ramifications (e.g., AAA to AA), CRAs might be slow to downgrade upper tranches. A CRA’s consideration of rating downgrades on future profits is consistent with previously observed episodes of economic crisis. Following a 2015 settlement, DOJ (2015) notes: “As S&P admits under this settlement, company executives complained that the company declined to downgrade underperforming assets because it was worried that doing so would hurt the company’s business. While this strategy may have helped S&P avoid disappointing its clients, it did major harm to the larger economy.”

Becker and Benmelech (2021) document that unlike the bond market, syndicated loan issuance decreased in the second-half of 2020. Given that CLO issuance in 2020 was roughly similar to previous years (S&P 2020a), the decrease in syndicate loans may have been considerably greater if CLO issuance had declined. Chakraborty, Goldstein, and MacKinlay (2020) document that Fed QE MBS purchases coming out of the financial crisis stimulated residential mortgage lending at the expense of commercial lending. Future research might examine whether the strong CLO market moved lending away from other markets.

5. Economic Assessment of Manager Trading Activities

In sum, we find support for two theories seeking to reconcile the disconnect between underlying collateral downgrades and the lack of tranche downgrades. First, we find evidence consistent with weight being placed on nonmodel considerations. Second, our findings are consistent with managers undertaking risk mitigating actions in response to collateral deterioration, manifesting as a flight to shorter-maturing and higher-rated loans.

We briefly benchmark the relative importance of these findings under a common framework to gauge the relative economic magnitude. Intuitively, we contrast the observed outcome (which reflects nonmodel considerations and the two active trading channels) with one in which we shut down the two managerial action channels, one at a time. We do so by adapting S&P’s approach to accommodate our set of counterfactual scenarios. Specifically, to gauge the impact of managers trading into safer assets, we reestimate the SDR for each deal-month when freezing the collateral pool as of January 2020. To measure the effect of managers tilting asset portfolios toward shorter-maturing debt, we recompute each SDR when fixing the weighted-average life of collateral at its January 2020 value.³⁶

Panel A of Figure 13 depicts the monthly kernel density of the difference in AAA SDRs under each counterfactual and when using the true collateral portfolio through time. We see little average difference as of March in the true SDR and that of either the adjusted life or static CLO counterfactual. This may be a partial artifact of the March distribution including CLOs reporting at any point in March (as opposed to end-of-month values). SDR differences under both counterfactuals begin to turn positive in April. By June, both managerial actions have an economically meaningful impact on reducing a collateral pool’s SDR, with a median increase of 0.5|$\%$| had pool composition been frozen in January and a 0.7|$\%$| increase had pool maturity not decreased. By August 2020, each effect exhibits a slight increase in economic magnitude, with approximately a 1|$\%$| greater SDR when fixing the CLO’s WAL and a 0.6|$\%$| increase had pools been fixed at their January composition. For reference, the average AAA cushion (BDR – SDR) in August is 5.12|$\%$|⁠. Thus, the effect of decreasing collateral maturities for the median deal is equivalent to 19.5|$\%$| of August cushions.

Panel B of Figure 13 illustrates this point in a different manner, reporting the kernel density of SDR – BDR differences under each counterfactual. Specifically, we take the true SDR – BDR difference disclosed in trustee reports and add the difference in SDRs reported in panel A.³⁷ Following this change, we see a considerable increase in the number of CLOs that would have an SDR greater than the share of defaults a AAA tranche is able to withstand before suffering losses. While 12.5|$\%$| of tranches have a reported SDR that exceeds its BDR in June, this share increases to 17.3|$\%$| when removing the modeling effects associated with the shortening of a collateral pool’s maturity. In contrast, had managers not actively traded out of riskier collateral (but still reduced the average maturity of the collateral pool), 16.2|$\%$| of CLOs would have an SDR that exceed their break-even rate. Interestingly, the share of CLOs with an SDR exceeding the corresponding BDR decreases to 8|$\%$| as of August 2020. Yet, 13.5|$\%$| of deals would surpass their BDR thresholds had managers not traded into shorter maturing collateral, while 10.8|$\%$| would exceed their BDR threshold if managers had not actively trade out of risky collateral. This increases to 18.8|$\%$| in a counterfactual where both active management channels are shut down.³⁸ This suggests that both actions undertaken by collateral managers help contribute to reducing model-implied collateral pool risk.

6. Conclusion

This paper examines the functioning of the CLO market in the COVID-19 crisis by comparing the downgrading activity of CLO collateral to tranches. Although both rating agencies took considerable downgrading actions on CLO loans, their approach to tranche downgrades were considerably more lenient. The small proportion of tranche rating actions do not align with either rating agency’s methodology, based on prior disclosures or current model outputs. A second important finding of our paper is that more than twice as many CLO tranches would be considered failing by S&P standards had managers not engaged in trading that reduced model-implied credit risk.

While our research explores the relation between collateral and tranche ratings, other important issues should be considered regarding the accuracy of CLO ratings. These include the proper level of CLO correlations (Nickerson and Griffin 2017), parameter uncertainty (Coval, Jurek, and Stafford 2009a), the strategic interaction between ratings and active management, collateral rating informativeness, recovery rate assumptions (Dubitsky 2020), and other modeling choices. These may have further consequences for future CLO rating performance, especially since the current COVID-19 crisis was likely not the full stress test possibly realized without Fed intervention.

Our results have substantial implications for current market conditions and regulators. First, the incorporation of nonmodel considerations suggests potential benefits of more rating agency disclosure, aiding market participants in assessing the validity of qualitative factors being considered. Second, credit ratings that are affected by active CLO management or that deviate from quantitative risk metrics could have considerably different forward-looking implications for investor risk management. Third, since banks, pension funds, and other institutions often face implicit or explicit rating-based constraints, downgrades that reflect small changes in credit risk could contribute to financial fragility concerns. Fourth, while Dodd-Frank and similar policies have largely emphasized reducing reliance on credit ratings, COVID-19-related Fed policy has moved in the opposite direction by operationalizing the purchasing of assets according to credit ratings (e.g., AAA static CLO tranches), whereas our analysis shows inconsistencies in CRA methods. Finally, a differential response that favors CLO credit ratings relative to underlying collateral has potential forward-looking implications for corporate investment decisions. This divergence provides incentives to pool levered loans with high systematic exposure into structured finance products, possibly resulting in the misallocation of capital akin to the recent evidence of Chen, Cohen, and Gurun (2021) in mutual fund bond ratings. An interesting avenue for future research is to further examine the real effects of structured finance credit ratings.

Acknowledgement

We thank Itay Goldstein (the editor), two anonymous referees, Scott Bauguess, Effi Benmelech, Jess Cornaggia, Andra Ghent, Itay Goldstein (the editor), Frank Partnoy, Michael Schwert, Phil Strahan, Dragon Tang, two anonymous referees, anonymous industry participants, and seminar/conference participants at MIT-Sloan, the Washington DC Federal Reserve, and the University of Texas-Austin for comments. We thank Prateek Mahajan and Eric Vance for excellent research assistance, as well as Integra Research Group for research support. Griffin is an owner of Integra Research Group, Integra FEC, and Integra REC, which engage in research and financial consulting on a variety of issues related to investigating fraud, including credit ratings and various types of bonds. Griffin and Nickerson consult on issues on credit ratings and bonds with Integra FEC. Additional results are available in an Internet Appendix at http://www.JordanNickerson.com. Supplementary data can be found on The Review of Financial Studies web site.

Footnotes

References