what was wrong with credit ratings for securitizations? - evidence from ... was wrong with...

What was wrong with credit ratings for

securitizations? - Evidence from ABS, CDO

and MBS markets

Daniel Rosch a Harald Scheule b,1

aInstitute of Banking & Finance, Leibniz University of Hannover, KonigswortherPlatz 1, 30167 Hannover, Germany, Phone: +49-511-762-4668, Fax:+49-511-762-4670. mailto: [email protected]

bDepartment of Finance, Faculty of Economics and Commerce, University ofMelbourne, Victoria 3010, Australia, Phone: +61-3-8344-9078, Fax:

+61-3-8344-6914, mailto: [email protected]

Preprint submitted to Elsevier 29 April 2009

What was wrong with credit ratings for

securitizations? - Evidence from ABS, CDO

and MBS markets

Abstract

This paper compares and analyzes cross-sectional and time-series characteristics ofcredit rating agency (CRA) ratings, implied impairment rate estimates and realizedimpairment rates of asset portfolio securitizations. This is of highest importance aspast shortcomings may have been instrumental to past, current and possibly futureloss rates of financial institutions in relation to securitizations. The disappointmentof investors manifested in the criticism of models applied by CRAs. Examples areVECTOR from Fitch, CDOROM from Moody’s and CDO Evaluator from Standardand Poor’s.

Our most substantial finding is that rating agencies do not include all factorsexplaining securitization impairment risk. In particular, the state of the economy isnot addressed as CRAs average over the business cycle. The results are threefold.Firstly, CRA ratings are unable to predict impairment risk. Secondly, CRA ratingsover-predict impairment risk in boom years and under-predict impairment risk dur-ing economic downturns. Thirdly, the recent deterioration CRA rating standardscan not be confirmed and may be a result of an economic downturn which was notpredicted by CRA ratings. Additional results are that CRA ratings for securitiza-tions do not fully account for the average credit quality in asset portfolios and donot fully account for structural elements of structured finance transactions. Suchelements include the subordination level, the resecuritization status and principalpayment years.

The inclusion of missing information, in particular the impact of the economy, isparamount to a sound measurement and management of impairment risk. Trans-parency may increase with the accurate econometric forecasting of such risk. Cred-ibility in internal and external models may not be restored otherwise.

Key words: Asset-backed Security, ABS, Collateralized Debt Obligation, CDO,Economic Downturn, Forecasting, Impairment Rate, Loss Rate, Mortgage-backedSecurity, MBS, Structured Finance Rating, Subprime Mortgage Crisis

JEL classification: G20, G28, C51

1 The authors would like to thank the participants of financial seminars at the

2

1 Introduction

This paper analyzes to which degree structured finance ratings have been in-accurate in the measurement of financial risk. The current financial crisis ledto an unprecedented and unexpected increase of impairment and loss rates forsecuritizations. The disappointment of investors manifested in the criticismof models applied by credit rating agencies (CRAs). Examples are VECTORfrom Fitch rating agency (see Fitch Ratings 2006), CDOROM from Moody’srating agency (see Moody’s Investors Service 2006) and CDO Evaluator fromStandard and Poor’s rating agency (see Standard & Poor’s 2005). For a com-parison of these models refer to Fender & Kiff (2004). Similar critique hasbeen put forward after the South East Asian Crisis of 1997 in relation to cor-porate bond issuer and bond credit ratings. For example, Hui et al. (2007)and Leot et al. (2008) find that ratings follow rather than predict the crisis assystematic downgrades have occurred subsequent to the crisis.

Securitizations involve the sale of fixed income assets into bankruptcy-remote,special purpose vehicles, which are funded by investors with different seniori-ties (tranches). Based on the nature of the securitized asset portfolios, impor-tant transactions include Asset-backed Securities (ABSs), Collateralized DebtObligations (CDOs) Home Equity Loan securities (HEL) or Mortgage-backedSecurities (MBS). Despite their name, securitizations are generally over-the-counter contracts and involve a funded or unfunded risk transfer. Therefore,no market prices are available. Counterparties publish accounting values forfunded, and to a limited degree for unfunded transactions. The latter are ac-counted for off-balance sheet and thus only included in the notes of annualreports. Information which is available to measure the risk of securitizationsincludes credit ratings, impairment histories and proxies for the asset port-folio risk such as asset value indices or cash flow indices. The evaluation ofindividual risks, their dependence structure and derivatives (i.e., the fundedor unfunded exposures of investors and guarantors) is complicated by the lowliquidity of the assets, the unavailability of secondary markets and the recentorigination of such transactions.

Two main streams exist in literature on the measurement of financial risks ofsecuritizations. The first stream focuses on the pricing of structured financetransactions where the central issue is to explain observed market prices suchas credit spreads of credit default swap (CDS) indices. The most prominentexamples are the CDX North America index which references a portfolio of125 North American firms and the iTraxx Europe index which references a

Leibniz University Hannover, Melbourne Centre for Financial Studies and The Uni-versity of Melbourne for valuable comments. The financial support of the MelbourneCentre for Financial Studies and the Australian Prudential Regulation Authority isgratefully acknowledged.

3

portfolio of 125 European firms. These indices were originated in 2003 and2004. Credit spreads for the index as well as tranches are available on a dailybasis and recent papers of this direction (compare Longstaff & Rajan 2008,Laurent & Gregory 2005, Hull & White 2004) apply a risk-neutral pricingframework to develop pricing techniques for these spreads. A central point ofthese risk models is the specification of a dependence structure for the portfolioassets, which is often based on a Gaussian copula model or an extension thereof(compare Li 2000).

The second stream is concerned with modeling and estimation of the riskcharacteristics of the underlying asset portfolio without relying on marketprices. The focus is on the derivation of the distribution of future asset values(or losses) based on individual risk parameters. In the case of a loan portfolio,the relevant parameters may be default probabilities, loss rates given default,exposures at default and dependence parameters such as correlations or moregeneral copulas. In this stream, Ohlson (1980), Shumway (2001), Rosch &Scheule (2005), Hamerle et al. (2006) and McNeil & Wendin (2007) modeldefault probabilities for one-year risk horizons and Duffie et al. (2007) modeldefault probabilities for multiple-year risk horizons. Dietsch & Petey (2004),Rosch & Scheule (2005) and McNeil & Wendin (2007) model the correlationsbetween default events. Carey (1998), Cantor & Varma (2005) and Acharyaet al. (2007) develop models for loss rates given default. Altman et al. (2005)and Rosch & Scheule (2005) model correlations between default events and lossrates given default. Gupton et al. (1997) and Credit Suisse Financial Products(1997) aggregate the individual risk parameters to portfolio risk measures.

Within this stream, credit ratings are often used as explanations of financialrisk. Ratings measure the financial risk of corporate bond issuers, corporatebond issues and structured finance securities. In the contemporary climate ofthe US subprime mortgage crisis (2007-2008), the role and importance of rat-ings to all market participants (e.g., issuers, investors and regulators), whilecontroversial, is beyond question. Fundamental issues relating to the generalextent to which credit rating changes convey new information has a rich pedi-gree that is the subject of ongoing academic debate and investigation. For ex-ample, Radelet & Sachs (1998) and Ferri et al. (1999) find that rating changesare pro-cyclical which would suggest that they provide only a limited amountof new information to the market. Ederington & Goh (1993), Dichev & Pi-otroski (2001) and Purda (2007) find that corporate credit rating downgradesdo provide news to the market, although most studies find that rating upgradesdo not. Jorion et al. (2005) show that after Regulation Fair Disclosure, themarket impact of both downgrades and upgrades is significant and of greatermagnitude compared to that observed in the pre-Regulation Fair Disclosureperiod. The relative roles of different CRAs have also been studied. For exam-ple, Miu & Ozdemir (2002) examine the effect of divergent Moody’s and S&Pratings of banks; while Beaver et al. (2006) consider the informational advan-

4

tage of certified (Moody’s) versus non-certified (Egan Jones) rating agencies.Micu et al. (2006) find that ratings announcements have significant effects onfirms’ CDS spreads.

Unfortunately, the literature does not analyze CRA ratings of securitizationsand their accuracy (compare Franke & Krahnen 2008). The present paperaddresses this shortcoming. Based on the rating and impairment data of oneCRA, cross-sectional and time-series characteristics of ratings, implied im-pairment rate estimates and realized impairment rates of asset portfolio se-curitizations are compared and analyzed. This is of highest importance aspast shortcomings may have been instrumental to past, current and possiblyfuture loss rates of financial institutions in relation to securitizations. Themain findings are that credit ratings have omitted idiosyncratic and system-atic information on the underlying asset portfolios as well as securitizationstructures. The hypothesis that the quality of securitization ratings has onlyrecently deteriorated cannot be confirmed. In addition, the previous finding ofliterature, that corporate bond ratings provide limited information, persists,as securitization ratings do not predict impairment risk. It is shown that inparticular economic information is insufficiently reflected in CRA models.

Literature on the current financial crisis is rapidly evolving. Rajan et al. (2008)show that omission of soft information in the ratings can lead to substantialmodel risk. Mayer et al. (2008) find that the decline of house prices was re-sponsible for increasing subprime mortgage delinquency rates. Benmelech &Dlugosz (2008) analyze collateralized loan obligations (CLOs) rated by Stan-dard and Poor’s and find a mismatch between credit ratings and the quality ofthe underlying loan portfolios. Crouhy et al. (2008) point out that CRAs’ rev-enues depend on the number of ratings, which in turn depend on the chancesfor selling the tranches in the market, which is supported by higher ratings.Similarly, Franke & Krahnen (2008) argue that incentive effects have playedan important role in the crisis, particularly associated with the allocation ofequity tranches of securitizations. Hull (2008) and Hellwig (2008) identify de-ficient CRA models as a cause of the current financial crisis.

Generally speaking, the validation of credit ratings is complicated as the useof ratings involves two steps: firstly the ordinal assessments of the financialrisk of issuers or issues by CRAs and secondly the calibration of these ordinalratings to metric credit risk measures such as default rates, loss rates givendefault or unconditional loss rates. This calibration step is generally opaque.

To date, only CRAs make their financial risk measures as well as the respectiverealizations (e.g., impairment histories) available to the general public. Little isknown of the quality of models of other vendors as well as financial institutioninternal models as the respective information is kept private. However, recentearnings announcements of financial institutions suggest that models applied

5

in industry are similar within certain categories of risk portfolios and financialinstruments.

The remainder of this paper is organized as follows. Section 2 provides aframework for the financial risk in securitizations and develops hypotheseswhich are consistent with the current literature in relation to the risk anduncertainty of CRA assessments. Section 3 describes the data used in the studyand analyzes five hypotheses. Section 4 discusses the major ramifications of theempirical results for risk measurement models for securitizations and providessuggestions for a new stability framework for financial markets, institutionsand instruments.

6

2 Model Framework and Hypothesis Development

2.1 Model Framework

Structured finance transactions are investments which are generally linked tothe performance of an asset portfolio which is subject to a retention (alsoknown as attachment level, AL) and a limit (also known as detachment level,DL). The paper refers to the entire transaction as ‘deal’ and the individualinvestment segments as ‘tranche’. In other words, one transaction may consistof one or more tranches of various seniority levels. The asset portfolio of a dealgenerally consists of financial assets (e.g., loans) which are subject to financialrisk (e.g., credit risk). For a description of mechanics of securitizations seeFender & Kiff (2004).

2.1.1 Default process for an individual asset and the portfolio default rate

Following work by Gordy (2003), Heitfield (2005), McNeil & Wendin (2007)and others, we model credit risk by a one factor model for the individual assetreturn based on Merton (1974). Let Skt denote the asset return of borrower kin time period t (k = 1, ..., K; t = 1, ..., T ) and is generated by the followingprocess

Skt =√ρ · Ft +

√1− ρ · Ukt (1)

where Ft and Ukt are independent standard normally distributed systematicand idiosyncratic risk factors and ρ is a parameter denoting the correlationbetween asset returns.

A default event occurs if the asset return Skt falls below a threshold ct whichmay be interpreted as the credit quality. For simplicity of exhibition we assumein the following that all assets in a portfolio are of equal credit quality ct at timet for a given risk segment. Risk segments may be defined by transaction typessuch as asset-backed securities (ABS), collateralized debt obligations (CDO)or mortgage-backed securities (MBS). A borrower default event occurs if theasset return falls below a critical threshold ct

Dkt = 1⇔ Skt < ct (2)

7

where Dkt is an indicator variable with

Dkt =

1 borrower k defaults in t

0 otherwise(3)

The probability of default conditional on the systematic risk factor ft is then

πt(ft) = P (Skt < ct|ft) = Φ

(ct −√ρft√

1− ρ

)(4)

which has expectation (the unconditional probability of default)

πt = P (Skt < ct) =∫ ∞−∞

Φ

(ct −√ρft√

1− ρ

)dΦ(ft) (5)

= Φ(ct) (6)

see Gordy (2000), where Φ(.) denotes the standard normal cumulative distri-bution function.

Kt assets are pooled to an asset portfolio and the portfolio default rate is theaverage over the default indicators

Pt =1

Kt

Kt∑k

Dkt (7)

The default rate of the asset pool then follows a binomial-normal mixtureprobability distribution (see Bluhm et al. 2003)

g(pt) =∫ ∞−∞

Kt

Kt · pt

πt(ft)Kt·pt(1− πt(ft))

Kt(1−pt)dΦ(ft) (8)

For portfolios which consist of infinitely many borrowers whose individualcredit exposures are sufficiently small (see Gordy 2003, for a mathematicaldefinition), the probability density is given by (see Vasicek 1991)

g(pt) =

√1− ρ√ρ· exp

(1

2(Φ−1(pt))

2 − 1

2ρ(ct −

√1− ρ · Φ−1(pt))

2

)(9)

8

The cumulative distribution function is given by

F (pt) = P (Pt < pt) = Φ

(√1− ρΦ−1(pt)− ct√

ρ

)(10)

Note that Pt in Equation (9) and Equation (10) can also be interpreted asthe loss rate (rather than the default rate) of the portfolio when losses givendefault are deterministic and equal to unity.

2.1.2 Impairment process for a tranche within a transaction

In the following, let Pit denote the default rate of asset pool i (i = 1, ..., I), i.e.,of a transaction or deal. A tranche j (j = 1, ..., Ji) of this deal experiences aloss and therefore an impairment if the default rate Pit in the portfolio exceedsthe relative attachment level (or subordination level) ALijt

Dijt = 1⇔ Pit > ALijt (11)

where Dijt is an indicator variable with

Dijt =

1 tranche j of deal i is impaired in t

0 otherwise(12)

The relative attachment level is calculated by the ratio of the attachmentlevel in $ and the deal principal in $ of period t. As a result of this definition,impaired tranches of previous years have reduced both the attachment levelas well as the deal principal. The probability of a tranche impairment is thus

P (Dijt = 1) = P (Pit > ALijt) (13)

Inserting Equation (13) in Equation (10) and replacing ct by Φ−1(πt) resultsin

P (Dijt = 1) = 1− Φ

(√1− ρΦ−1(ALijt)− Φ−1(πt)√

ρ

)(14)

This implies that the tranche impairment probability is a function of the

9

• Average portfolio asset quality;• Asset correlation;• Relative attachment level (subordination level) of a tranche.

A credit rating which serves as an assessment of the impairment risk of a secu-ritized tranche should therefore account for these factors to explain tranche im-pairment probabilities. On the other hand, if a rating excludes or mis-specifiessome information, then additional information besides the rating may explainthe tranche impairment probability. Examples may relate to the asset portfo-lio quality or the securitization structure. This issue will be addressed in thedevelopment of hypotheses.

2.2 Hypotheses development

The following hypotheses aim to answer whether structured finance ratingsare inaccurate and may have been causal for the current financial crisis. Theclaim has been made but no empirical evidence presented (compare Franke &Krahnen 2008). The paper breaks the inaccuracy into idiosyncratic, systematicand a mixture of idiosyncratic and systematic characteristics. The hypothesesare as follows:

• H1a: CRAs mis-evaluate the average asset quality of the asset portfolio;• H1b: CRAs mis-evaluate structured finance transaction characteristics (namely

resecuritization status, subordination level and transaction cash flow struc-ture);• H2: CRAs omit time varying information;• H3a: CRA standards have declined over time;• H3b: CRAs predict impairment risk.

The hypotheses H1a and H1b relate to idiosyncratic, H2 to systematic andH3a and H3b to the interaction between idiosyncratic and systematic riskcharacteristics of securitizations.

H1a addresses characteristics of the asset portfolio. Rajan et al. (2008) findthat securitization risk models omit ‘soft’ information. This implies that theCRA ratings, relying on such models, mis-evaluate the average credit qual-ity of the asset portfolio. Crouhy et al. (2008) suggest that CRAs did notmonitor raw data, were tardy in recognizing the implications of the decliningstate of the subprime market for the ratings of monoline insurers, paid byclients for ratings and that competition is limited by regulation. Importantdrivers of asset portfolio risk may be ratings as well as other asset portfoliocharacteristics.

10

H1b addresses the tranching structure of securitizations and the current dis-cussion on the appropriate specification of the dependence structure of variousassets in a portfolio (compare Hull 2008, Hellwig 2008). Credit rating impliedcorrelations are for example, derived in Gordy (2000). Rosch (2005) and Heit-field (2005) argue that correlations associated with ratings may depend onthe design of the rating system. The probability distribution as well as thepercentiles of losses associated with the pool are particularly sensitive to thecorrelations in the underlying asset pool. Thus, the level of subordination maybe a key driver if correlations are mis-specified and should explain tranche im-pairments even after controlling for credit ratings. This is also evident fromthe model of the previous section.

H2 identifies the degree to which business cycles are included into the riskmodels and relates to the ongoing discussion in relation to corporate issuerand bond issue ratings, and whether CRA ratings include the macro economy(point-in-time rating) or not (through-the-cycle rating). An analysis of bothrating paradigms is given in Loeffler (2004). While through-the-cycle ratingsare often more stable through time, see Nickell et al. (2000), they may reacttoo slowly in economic up- or downturns. Franke & Krahnen (2008) argue thatsensitivities to macroeconomic factors may be higher for securitized tranchesthan for corporate bonds to which the rating grades are applied.

H3a relates to a hypothesis suggested by various authors (e.g., Crouhy et al.2008) that lending standards were declining in recent years. Blume et al. (1998)present a similar hypothesis for corporate bond issuers. Rajan et al. (2008)analyze individual securitized subprime loans and assess the performance ofthe FICO (Fair, Isaac and Company) score and loan-to-value ratio for the pre-diction of mortgage default events. They find that the default model performspoorly in times of higher securitization, i.e., the recent years. Downing et al.(2008) present evidence for declining subordination for commercial MBSs.

H3b addresses the information degree of credit ratings which has been ex-tensively discussed in literature (see Section 1) and proposed after the SouthEast Asian financial crisis (compare Hui et al. 2007, Leot et al. 2008). Hellwig(2008) argues that the omission of systematic factors related to real-estateprices such as interest rates and the availability of housing finance may haveproduced a bubble of overoptimism leading to valuations and ratings whichare far from reality. Such expectations may be adjusted after the burst of abubble.

11

3 Empirical analysis

3.1 Structured finance data

The paper analyzes a panel data set on structured finance transactions whichare rated by Moody’s Rating Agency. The data covers characteristics of trans-actions, characteristics of tranches, ratings of tranches over time as well as oc-currences of impairment events. The complete data set is described by Moody’sInvestors Service (2008). The time horizon is 1987-2007. Unfortunately, datafor 2008 is not available at the time of writing of this paper.

The central event in the present study is the impairment event. An impairmentevent is defined as follows (compare Moody’s Investors Service 2008):

“Moody’s first introduced the concept of material impairment in 2003 inorder to differentiate the definition of default between corporate and struc-tured finance sectors. Material impairments fall into one of two categories,principal impairments and interest impairments. Principal impairments in-clude securities that have suffered principal write-downs or principal lossesat maturity and securities that have been downgraded to Ca/C, even ifthey have not yet experienced an interest shortfall or principal write-down.Interest impairments, or interest-impaired securities, include securities thatare not principal impaired and have experienced only interest shortfalls.”

The paper focuses on the probability of impairment as the tranche loss givenimpairment is not recorded in the data set, which is equal to the expected lossof the tranche assuming a loss rate given impairment of unity.

Filter rules were applied and the following observations deleted:

• Transaction observations which cannot be placed into the categories ABS,CDO, HEL or MBS. These are mainly asset backed commercial paper, struc-tured covered bonds, catastrophe bonds, and derivative product companies(30,772, 20.5%);• Transaction observations which are not based on the currency USD or trans-

action observations which are not originated in the USA (17,717, 11.8%);• Tranche observations which relate to years prior to 1997 due to a limited

number of impairment events (21,442, 14.3%);• Tranche observations which have experienced an impairment event in prior

years (93, 0.1%).

The numbers in brackets indicate the number of tranche observations whichwere deleted by each filter rule. The percentages in brackets indicate the frac-tion in relation to the initial data sample of 149,859 tranche observations.

12

The resulting data comprises 79,835 annual tranche observations which relateto 15,728 annual transaction observations. Please note that these filter ruleswere chosen in order to create homogeneous risk segments with a limited re-duction of the number of observed impairment events. As a result, the numberof impaired tranche observations is 2,293 while the original data set included2,340 impairment events before the application of filtering rules. This impliesthat structured finance transactions are more common and the number ofimpairment events are more frequent in the US than in non-US regions.

The following categorical variables were generated:

• Impairment (1: impairment, 0: no impairment) indicates that a tranche isimpaired in the observation year;• Rating (1: Aaa to A, 2: Baa, 3: Ba, 4: B, 5: Caa-C) reflects the expected

loss of a tranche and is measured at the beginning of an observation year(compare Moody’s Investors Service 2006);• Principal Payment Year (PPY; 1: principal is repayable, 0: principal is not

repayable) indicates whether principal is repayable in the observation year;• Deal category (ABS: asset backed securities, CDO: collateralized debt obli-

gations, HEL: home equity loans, MBS: mortgage-backed securities);• Resecuritization (1: resecuritization, 0: no resecuritization) indicates whether

a transaction is a resecuritization of previous transactions. These transac-tions are often called ‘squared’ (e.g., CDO-squared).

Table 1 and Table 2 describe the number of observations over time for alltranches as well as per rating category, asset portfolio type, resecuritizationstatus, principal payment year status and subordination level. The overallnumber of rated securitizations has increased at an increasing rate over time(see column 1).

[insert Table 1 here]


Table 1 shows the relative frequency of rating categories (see column 2 tocolumn 6). The average rating quality deteriorates over time which may reflecti) a deterioration of the average asset portfolio quality, ii) a higher average risklevel induced by the securitization structure (e.g., subordination, thickness orfeatures which are not addressed in this paper, such as embedded options)or iii) a change of the rating process (which is analyzed by testing hypothesisH3a). In addition, HELs, i.e., subprime mortgage loan portfolio securitizationshave increased during the observation period (see column 9). This is of highrelevance as the impairment rate for HELs (next to CDOs) has also increasedover-proportionately.

Generally speaking, asset portfolio securitizations are relatively heterogeneous

13

despite the contribution by the International Swap and Derivatives Dealer As-sociation which provides structure templates. Retail asset portfolios generallycomprise a large number (e.g., 100,000) of exposure amounts with small ex-posures (e.g., $100,000) and are mainly exposed to systematic and systemicrisk. Corporate/wholesale asset portfolios comprise a small number (e.g., 100)of exposures with large exposure amounts (e.g., $10 million) and are exposedto idiosyncratic, systematic and systemic risk. ABSs generally comprise retailasset portfolios (e.g., auto, credit card and student loans) as well as corpo-rate/wholesale asset portfolios (e.g., equipment loans and leases). CDOs gener-ally comprise corporate/wholesale asset portfolios (e.g., unsecured or securedcorporate loan exposures). HELs generally relate to retail subprime mortgageportfolios while MBS generally relate to prime mortgage portfolios and relateto commercial and retail borrowers.

Table 2 shows that resecuritizations have increased over time (see column 3)while the relative frequency of principal payment years has decreased in 2006and 2007 as the overall number of tranches has increased (see column 5). Thisis important as both resecuritizations and tranches in their principal paymentyear have experienced a large increase in impairment rates in 2007 (see below).The latter may be interpreted as an important mitigant of impairment riskduring the current financial crisis. In other words, the realized impairmentlosses may have been more pronounced in 2007 if the recent growth in thesecuritization industry would have not occurred as this would imply a largerfraction of securitizations in their principal payment year (i.e., maturity year).In addition, it can be seen that the relative frequency of mezzanine tranches(subordination=3%-12%, see column 7) has increased.

Table 3 and Table 4 are the core of the following analysis and show the im-pairment rates over time for all tranches as well as per rating categories, assetportfolio types, resecuritization status, principal payment year status and sub-ordination level. US securitizations have experienced two economic downturnsin past years: the first one in 2002 subsequent to the US terrorist attacks (aperiod which was characterized by big bankruptcies such as WorldCom) andthe most recent financial crisis. Column 7 shows that the impairment ratehas increased by a factor of 10 within one year between 2006 (minimum oftime series) and 2007 (maximum of time series). Approximately 65% of theimpairment events relate to 2007, the year with the highest impairment rateand highest number of transactions and tranches.



Table 3 shows that impairment rates increase with the degree of financial riskimplied in a rating (i.e., from Aaa-A to Caa, see column 2 to 6). The minimum

14

impairment rates are close to zero while the maximum impairment rates are1% (Aaa-A), 15% (Baa), 43% (Ba), 32% (B) and 75% (Caa). This volatilityshould not exist if ratings explain impairment risk perfectly. It is the aim of thispaper to analyze whether these inaccuracies are random or may be explained.Please note that inconsistencies (e.g., a higher impairment rate for a lower riskrating in a selected year) may reflect model errors or may reflect the stochasticnature of impairment events. The latter is particularly relevant if the numberof observations is low for a given category. These inconsistencies are in linewith reports by the data-providing CRA (compare Moody’s Investors Service2008). The impairment rates are fundamentally different between the variousasset portfolio categories (see column 7 to 10). The economic downturn in2002 relate mainly to CDOs while the recent financial crisis has increased theimpairment ratios for CDOs, HELs and MBSs. HELs may be interpreted assubprime mortgage loans and the impairment risk has increased to a largerdegree than the one of MBSs. It can also be seen that HELs and MBSs didnot experience an economic downturn in 2002.

Table 4 shows that impairment rates of primary securitizations and resecuri-tizations. It is interesting to see that the impairment rate of resecuritizationsincreased to a much larger degree in 2007 (see column 3). The present paperanalyzes funded transactions which are similar to coupon bearing bonds, withthe main difference being that they are collateralized by asset portfolio ratherthan a firm’s assets. Thus, the payment structure involves i) an investment, ii)periodic interest/coupon payments until and including in the maturity yearand iii) repayment of principal in the maturity year. Accordingly, periodic ob-servations may be separated into principal payment years and non principalpayment years. It can be seen that the impairment rate of principal paymentyears increased to a much larger degree in 2007 (see column 5). Impairmentrates have increased in 2007 for equity tranches (subordination=0%-3%, seecolumn 6) and mezzanine tranches (subordination=3%-12%, see column 7).Please note that many transactions are single tranches and that the impair-ment rate for equity tranches is not necessarily higher than for mezzaninetranche, which is not necessarily higher than for senior tranches as these cat-egories relate to different transactions.

3.2 H1a: CRAs mis-evaluate the average asset quality of the asset portfolio

A major concern in the current financial crisis is that current credit portfoliorisk models, including the models used by CRAs, do not capture credit port-folio risk accurately. If credit ratings correctly assess the impairment risk of atranche, then the tranche impairment probability should solely be explainedby the ratings. Other variables may impact upon the likelihood of an impair-ment event taking place, which would indicate that ratings omit information.

15

Table 5 presents two Probit models which link the impairment of tranche j(j = 1, ..., Ji) of pool i (i = 1, ..., I) in time t (t = 1, ..., T ) with observableinformation

P (Dijt = 1) = Φ (β′xijt) (15)

xijt is a vector of observable and thus known variables. β is the respectivevector of sensitivities and includes an intercept. The models may be usedfor forecasting as the CRA ratings, which are time-varying covariates, aremeasured at the beginning of the observation year.

Table 5 shows the parameter estimates for the two models. Model 1 (see col-umn 1) takes the ratings provided by CRAs into account. Aaa-A is the refer-ence category. As measures for in-sample accuracy of the models the Pseudo-R2, re-scaled R2, and the area under the receiver operating characeristic (AU-ROC) are given (see Agresti 1984). The parameter estimates increase fromrating Aaa-A to Caa which show the predictive power of ratings. Model 2(see column 2) includes the ratings as well as the type of the underlying assetportfolios which shows that the impairment likelihood of CDOs is larger thanHELs, which is larger than ABSs, which is larger than MBSs. 2


In summary, CRAs do not take all available asset portfolio information intoaccount. Important ramifications are that i) CRAs may have to include assetportfolio characteristics into their rating models and ii) investors should applyasset portfolio specific impairment rates to ratings when interpreting CRAratings.

Figure 1 shows the real-fit diagram of Model 1 and Model 2. Neither model isable to explain the cyclical fluctuations of the impairment rate over time.

[insert Figure 1 here]

3.3 H1b: CRAs mis-evaluate the structured finance transaction characteris-tics

In order to test the hypothesis whether CRAs mis-specify structured financetransaction characteristics, we include three additional variables:

2 We estimated Logit models as a robustness check (i.e., nonlinear models with alogistic link function). The results were comparable.

16

• Resecuritization: the variable indicates whether the transaction consists ofa resecuritized asset portfolio. Hull (2008) suggests that resecuritizationscontributed to the increased number of impaired structured finance trans-actions. Resecuritizations were originated to create a market for mezzaninetranches which are generally less popular amongst investors. Thus a resecu-ritization often involves the tranching of a portfolio of mezzanine investmenttranches. To date, no empirical evidence was presented as to whether rat-ings for resecuritizations have the same information content as ratings forprimary securitizations;• Subordination: the variable represents the subordination level of the ob-

served tranche and relates to an ongoing discussion of whether CRAs applyreasonable levels of asset correlations. Asset correlations measure the de-pendence between the asset performances of the portfolio underlying thetransaction and are an important input parameter in the risk models ofCRAs as well as many financial institutions;• Principal payment year (PPY): securitizations may be more likely to be

impaired in the year of the principal payment than in the years of interestpayments. The impact of this variable on the impairment risk of tranches hasnever been formally tested. However, it is important as structured financetransactions are fundamentally different to other credit risk exposures, suchas corporate bonds and retail loans:· Retail loan repayments are generally structured as annuities and aligned

with the income of a borrower. Early years relate mainly to interest pay-ments while later years relate mainly to principal repayments. Retail bor-rowers who are unable to meet payment obligations may avoid impairmentby refinancing or renegotiating their debt. Please note that restructuringmay or may not be a default criteria in risk models.· Corporate loans or bonds have the traditional interest-maturity profile

where the largest amount (i.e., interest and principal) is due at matu-rity. However, bonds are issued by large corporations which are financedby a portfolio of equity, hybrid capital and debt of various maturities.Corporate borrowers who are unable to meet payment obligations mayavoid default by refinancing or renegotiating their debt. Please note thatrestructuring may or may not be a default criteria in risk models.· Structured finance transactions are fundamentally different from retail

and corporate loans, in the sense that the life of the special purpose ve-hicles and thus the liabilities are generally termed. This implies that theliquidation value of the assets has to be sufficient to meet all contractualinterest and principal payments in the final period. Impairment occurs perdefinition if this is not the case.

Table 6 confirms that impairment risk increases significantly, if i) a transactionis a resecuritization (Model 3, see column 1), ii) the subordination increases(Model 4, see column 2) and iii) in principal payment years (Model 5, see

17

column 3). 3 The significance of the subordination may imply that the CRArisk models do not properly include the subordination level or alternatively,the distribution of portfolio losses is mis-specified. In the case of a CRA riskmodels, this may imply that the Gaussian copula model may not reflect theempirical data or that a dependence parameter such as the asset correlationor the correlation between default events and loss given default may be mis-specified. Thus, CRAs do not take all available information on structuredfinance transaction characteristics into account. Similar to hypothesis H1a,important ramifications are that i) CRAs may have to include structuredfinance transaction characteristics into their rating models and ii) investorsshould apply transaction structure-specific impairment rates to CRA ratings.


Figure 2 shows the accuracy of Model 3, Model 4, Model 5 and Model 6 overtime and suggests that systematic elements are missing in these models.

[insert Figure 2 here]

3.4 H2: CRAs omit time varying information

The previous analysis suggests that indicates finance ratings omit informationabout systematic risks in securitizations. CRAs are known to rate ‘through-the-cycle’ for corporate bonds (see e.g., Loeffler 2004) and include mainlyidiosyncratic characteristics. Cyclical effects or macroeconomic informationare not included as the assessment of credit quality should only reflect a bor-rower’s ability to pay based on firm fundamentals and should not be affectedby cyclical swings in order to avoid too frequent up- and downgrades. Omit-ting business-cycle information from ratings for securitizations may lead to theobserved mismatch of the time-constant rating and true risk, which is clearlycyclical (compare Figure 1 to Figure 2).

Following McNeil & Wendin (2007) and others, we extend the models by atime-specific, contemporaneous and thus random effect and analyze to whichextent ratings of securitizations include or omit macroeconomic information.The conditional probability of impairment (given a random variable Ft) isgiven by a probit transformation of a linear combination of an intercept αplus a random component Ft which depends only on time and not on thetranche

3 Model 6 (see column 4) confirms the robustness of the results by including allthree variables.

18

P (Dijt = 1|Ft) = Φ (α + b · Ft) (16)

Ft is a proxy for unobserved economic influences driving the impairments. Weassume that Ft is standard normally distributed and the tranche exposure toFt is denoted by b. The parameters are estimated via the Maximum-Likelihoodmethod.

In Model 7 we estimate the sensitivity to systematic risk only. The first versionis based on the whole data set and serves as a base case (see column 1). Theestimate for the intercept is -2.1255, the estimate for the exposure to the latentfactor is 0.2734 and highly significant. This shows that tranche risks (overall)are driven by the macroeconomy. Columns 2 to 6 estimate Model 7 for subdata sets of the same rating grade. Higher rated tranches are more sensitiveto the economy than lower rated ones. Vice versa, this means that if economicinformation is omitted in the ratings, a change in macroeconomic conditionswill lead to a higher discrepancy of rating and true risk for the better ratedtranches. All macroeconomic exposures are significant at the 10 per cent level.

Franke & Krahnen (2008) hypothesize that better rated tranches have higherexposure to the macro economy than bonds. Using the same model, Rosch& Scheule (2009) estimate bond default rate sensitivities to systematic riskwith 0.491 (investment grade rating), 0.245 (Ba), 0.355 (B) and 0.364 (Caa).These numbers are similar to the asset correlations (which are equal to b2)specified by the Basel Committee on Banking Supervision for deriving theminimum capital requirement of corporate credit risk exposures under theInternal Ratings-based Approach (compare Basel Committee on Banking Su-pervision 2006). A comparison of securitizations and bonds shows that thesystematic sensitivities are i) similar for investment grade ratings, ii) largerfor medium rated securitization exposures and iii) lower for low rated securi-tization exposures.


Table 8 presents the estimation results of Model 7 for different deal types. Allmacroeconomic exposures are highly significant and there is a clear differencebetween asset pool types. CDOs (see column 1) and ABS s(see column 3)have a similar medium-high sensitivity, HELs (see column 4) have the highestexposure and MBSs (see column 2) the lowest exposure. This implies the ob-servation that sub-prime mortgage loans (HEL) are more sensitive to economicdownturns than other securitization categories.


In Table 9, we investigate whether there are differences between deals which

19

are resecuritized and those which are not. Such structures are also known as‘tranches of tranches’ or ‘squared’ products. The intuition is that resecuriti-zation eliminates idiosyncratic risk and therefore tranches of tranches shouldbe exposed to systematic risk more strongly, see Hull (2008). The results con-firm that the exposure of resecuritized tranches is 0.5771 (see column 3) andthus more than twice as much as for unsecuritized tranches where it is 0.2703(see column 1). As most resecuritized tranches are CDOs, we estimate sep-arate models for this asset pool type. The relatively higher macroeconomicexposure also holds for CDOs but is less pronounced (see columns 2 and 4) .


After analyzing the macroeconomic exposure of tranches in general, we even-tually test if, and how much, business cycle information is included in theCRA ratings. We extend the random effects probit model and include ratingdummies (Model 8) as before

P (Dijt = 1|Ft) = Φ (β′xijt + b · Ft) (17)

The exposure b of the latent factor should no longer be significant after con-trolling for ratings if all information on the economy affecting the impairmentprobabilities is captured by the rating. Table 10 shows the estimates for thebase case model (Model 7 without ratings, see column 1) and the model whichcontrols for the ratings (Model 8, see column 2). It is clear that the exposuredoes not vanish and is still significant, showing that there is cyclical move-ment beyond ratings which supports our hypothesis that CRA ratings omittedtime-varying systematic information about the economy. In other words, CRAratings do not explain the cyclically of financial risk in securitizations.


The ramifications are that the exposure to the business cycle increases i) witha CRA rating which indicates lower financial risk, ii) for subprime mortgageloans and iii) for resecuritizations. Thus, CRA ratings should reflect the dif-ferent embedded degrees of systematic risk or alternatively, investors shouldassign time-varying impairment rates, which control for asset portfolio typeand resecuritization status to CRA ratings.

3.5 H3a: CAR standards have declined over time

The next hypothesis addresses the critique that rating standards of CRAshave declined over time. The deterioration of rating quality by CRAs hasbeen identified as a possible reason for the current financial crisis. This should

20

be reflected in a declining quality of rating standards or deteriorating qualityof risk forecasts from credit ratings, particularly in the years prior to thesubprime crisis. In other words, the implied impairment probability and thusobserved impairment rate of a given rating grade may have increased overtime, because the grade contains more and more entities with decreasing creditquality.

We test this hypothesis using a random effects model (Model 9) of the form

P (Dijt = 1|Ft) = Φ (α + β · t+ b · Ft) (18)

where t = (year−1996) counts the number of years from the beginning of theobservation period and is thus a year effect. If the sensitivity of the variableis positive, then the Impairment probability of a rating grade increases overtime, that is, in year t + 1 the same rating grade (e.g., rating Baa) exhibitshigher impairment risk than in year t. 4 Table 11 shows the results for Model9. For none of the rating grades we find a significant time trend of impairmentprobabilities. However, the parameters are positive in five out of six models(the exception are Baa-rated securitizations, see column 3. Therefore, we can-not reject the hypothesis that the rating standards have not changed overtime.


3.6 H3b: CRAs predict impairment risk

The hypotheses addresses the information content of credit ratings, in particu-lar the ability to predict impairment risk. The hypothesis is embedded in manyapplications as ratings are generally applied as proxies for future impairmentrisk. Literature has analyzed the quality of rating standards for corporatebonds (see Blume et al. 1998). However, to the best of our knowledge, noevidence for CRA ratings on securitizations has been presented.

We test how the forecasting power of credit ratings behaved by an approachrelated to Rajan et al. (2008). Our approach proceeds in three steps. First weestimate a probit regression for each year

P (Dijt = 1) = Φ (β′xijt) (19)

4 We included nonlinear transformations of the time-variable as a robustness check.The results were comparable.

21

where xjit are dummy variables for the ratings which are observed at thebeginning of the observation period. Next, we calculate the linear predictorfor the subsequent year

ηijt+1 = β′xijt+1 (20)

and the impairment probability prediction for the subsequent year

pijt+1 = Φ(β′xijt+1) (21)

using the estimated coefficients β from (19). Finally, we assess the forecastingpower by running a probit regression (Model 10)

P (Dijt+1 = 1) = Φ (γ0 + γ1ηijt+1) (22)

If the rating provides perfect forecasts, then γ0 = 0 and γ1 = 1 which caneasily be tested. As a robustness check we also run a linear regression (Model11)

Dijt+1 = δ0 + δ1 · pijt+1 + εijt+1 (23)

so that E(Dijt+1) = P (Dijt+1) = δ0 + δ1 · pijt+1 where δ0 = 0 and δ1 = 1.

We repeat all steps for each year from 1999 to 2007 where in the probit regres-sion (19) all data up to year t are used. Table 12 shows parameter estimatesfrom each regression Model 10, i.e., Equation (22). Table 13 contains the es-timation results from each regression Model 11, i.e., Equation (23).



It can be seen that in most years, both coefficients of either regression arestatistically significantly and thus different from their ideal values (see columns1 and 2). Moreover, the respectiveR2 neither increase nor decrease throughout.This implies that the rating quality has neither declined nor improved. 5 Whilefor most years, the evidence of underprediction or overprediction is mixed,years 2002 and 2007 exhibit a significant underestimation of risk by the ratings.If ratings predicted impairment risk accurately, they should have anticipated

5 A comparison of the R2 shoud be interpreted with care since each year has adifferent number of sample observations.

22

the downturns in 2002 and 2007 and should have downgraded the transactionsaccordingly. However, the observation that the estimates of γ0 and δ0 aregreater than zero and estimates for γ1 and δ1 are greater than one indicatesthat impairment risk has been under-predicted by the ratings in these years.In summary, the analysis shows that the rating quality has neither declinednor improved through time in general. However, the results show that therehas been a mix of years of overprediction and years of underprediction ofimpairment risk. This indicates that CRA ratings have a very limited abilityto predict impairment risk.

The ramification is that CRAs do not predict periods of higher impairmentrisk and that investors relying on predictions of future levels of impairmentrisk may have to build private models.

23

4 Discussion and Outlook

To date, literature has not analyzed the accuracy of risk models for securi-tizations (compare Franke & Krahnen 2008). The article’s main objective isto analyze the impact of idiosyncratic and systematic risk characteristics onimpairment risk of securitizations.

Our most substantial finding is that rating agencies do not include all fac-tors explaining securitization impairment risk. In particular, the state of theeconomy is not addressed as CRAs average over the business cycle. The re-sults are threefold. Firstly, CRA ratings are unable to predict impairmentrisk. Secondly, CRA ratings over-predict impairment risk in boom years andunder-predict impairment risk during economic downturns. Thirdly, the recentdeterioration CRA rating standards can not be confirmed and may be a resultof an economic downturn which was not predicted by CRA ratings. Additionalresults are that CRA ratings for securitizations do not fully account for theaverage credit quality in asset portfolios and do not fully account for struc-tural elements of structured finance transactions. Such elements include thesubordination level, the resecuritization status and principal payment years.

In response to the presented hypotheses, CRA ratings for securitizations

• Do not fully account for the average credit quality in asset portfolios;• Do not fully account for the structure of structured finance transactions;• Reflect the average impairment risk over the business cycle: risks are as-

sessed through-the-cycle rather than point-in-time;• Are based on rating standards which have not systematically declined over

time;• Do not predict impairment risk.

However, the findings should not be interpreted as a critique of the valuablework CRAs provide. Please note that the major CRAs cover a large numberof rated debt issuers and issues per year (e.g., Moodys: 100 sovereign nations;12,000 corporate issuers; 29,000 public finance issuers; 96,000 structured fi-nance obligations) with a limited number financial analysts (e.g., Moodys:more than 1,000 analysts). These ratings may provide useful information onthe average idiosyncratic impairment risk over the business cycle.

The findings should rather be seen as an encouragement for future improve-ment of industry models. Other risk models may have similar features andthus similar shortcomings. Structured finance markets were originated in re-cent years and only years of economic downturns may lead to loss experienceswhich are required to identify such shortcomings. It should also be noted thatCRAs are amongst the most transparent providers of risk models. No datais available for the risk measures and the respective loss experiences of other

24

market participants such as financial institutions and commercial model ven-dors.

Going forward, the paper may provide a stimulus for an increase of trans-parency in relation to risk modeling for securitizations The following comple-mentary suggestions are supported by the paper and may contribute additionalelements to a new stability framework for financial markets, institutions andinstruments:

• Refinement of risk models: contrary to existing securitization model, modelsshould be point-in-time and be able to predict the credit risk for futureperiods with a reasonable degree of accuracy. This may require the following:· Regulation should address pro-cyclicality. It has to be determined whether

financial institution capital should be pro-cyclical, neutral or counter-cyclical. While regulators in the past have tried to avoid pro-cyclicality,the current crisis may support dynamic capital adequacy regulations whichanticipate economic downturns.· Existing regulations should be homogeneous across countries and indus-

tries. Past securitizations were often inspired by differences in regulationsbetween countries, industries and other categories. The homogenizationmay i) improve the transparency, ii) regulate industries closely linked toregulated industries and avoid transactions which aim to exploit regula-tory differences (regulatory arbitrage).· In addition, stricter rules for information disclosure beyond Basel II may

have to be implemented and potentially high risk strategies limited. Oneexample may be the limitation of securitizations to older and thus betterknown underwriting years.

• Pooling and publication of information: one major challenge to pricing andresearch is the availability of data for various stages of the business cycles.As a result to this limitation, past research has often focused on CRA datasets which are publicly available. It may be important to make additionaldata, in particular credit loss histories in relation to other risk segmentsavailable to researchers, financial institutions and their regulators. The datashould be pooled but not aggregated over lenders. A global and compulsorydata warehouse for credits may be a rewarding approach. The data may befiled anonymously and be validated by the national regulators. In addition,models of various participants may have to be transparent and available inan open source environment.• Evaluation of model risk and stress-testing of risk models: For pricing and

risk measurement of complex financial products, models have to be used.Each model relies on a number of assumptions and therefore provides only asimplification of reality. Recent research has shown that risk measures andimplied prices of structured products are highly sensitive to the underlyingmodel and the assumptions with respect to the parameters used, than forexample, straight bonds. Model risk persists as the financial institutions and

25

model provider industries are highly concentrated. Thus, banking supervi-sors may sponsor independent research centers which provide an open sourceenvironment for data and models and the opportunity to assess model riskand stress-test existing risk models.

26

References

Acharya, V. V., Bharath, S. T. & Srinivasan, A. (2007), ‘Does industry-widedistress affect defaulted firms? - evidence from creditor recoveries’, Journalof Financial Economics 85, 787–821.

Agresti, A. (1984), Analysis of Ordinal Categorical Data, Wiley, New York.Altman, E., Brady, B., Resti, A. & Sironi, A. (2005), ‘The link between default

and recovery rates: Theory, empirical evidence and implications’, Journalof Business 78, 2203–2227.

Basel Committee on Banking Supervision (2006), ‘International convergenceof capital measurement and capital standards: A revised framework, com-prehensive version’.

Beaver, W., Shakespeare, C. & Soliman, M. (2006), ‘Differential properties inthe ratings of certified vs. non-certified bond rating agencies’, Journal ofAccounting and Economics 42, 303–334.

Benmelech, E. & Dlugosz, J. (2008), ‘The alchemy of CDO credit ratings’,Working Paper, Harvard University .

Bluhm, C., Overbeck, L. & Wagner, A. (2003), An Introduction to Credit RiskModeling, Chapman/CRC, London.

Blume, M., Lim, F. & MacKinlay, A. (1998), ‘The declining credit quality ofu.s. corporate debt: Myth or reality?’, Journal of Finance 53, 1389–1413.

Cantor, R. & Varma, P. (2005), ‘Determinants of recovery rates on defaultedbonds and loans for North American corporate issuers: 1983-2003’, Journalof Fixed Income 14, 29–44.

Carey, M. (1998), ‘Credit risk in private debt portfolios’, Journal of Finance53, 1363–1387.

Credit Suisse Financial Products (1997), ‘CreditRisk+ - a credit risk manage-ment framework’.

Crouhy, M., Jarrow, R. & Turnbull, S. (2008), ‘The Subprime Credit Crisis of07’, Working Paper, University of Houston, Natixis and Cornell University.

Dichev, I. & Piotroski, J. (2001), ‘The long-run stock returns following bondratings changes’, Journal of Finance 56, 173–203.

Dietsch, M. & Petey, J. (2004), ‘Should SME exposures be treated as retailor corporate exposures? a comparative analysis of default probabilities andasset correlations in french and german SMEs’, Journal of Banking andFinance 28, 773–788.

Downing, C., Stanton, R. & Wallace, N. (2008), ‘Volatility, mortgage de-fault and CMBS subordination’, Working Paper, University of Californiaat Berkeley .

Duffie, D., Saita, L. & Wang, K. (2007), ‘Multi-period corporate default pre-diction with stochastic covariates’, Journal of Financial Economics 83, 635–665.

Ederington, L. H. & Goh, J. C. (1993), ‘Is a bond rating downgrade bad news,good news, or no news for stockholders?’, Journal of Finance 48, 2001–2008.

27

Fender, I. & Kiff, J. (2004), ‘Cdo rating methodology: Some thoughts on modelrisk and its implications’, BIS Working papers 163.

Ferri, G., Liu, G. & Stiglitz, J. (1999), ‘The procyclical role of rating agencies:Evidence from the East Asian Crisis’, Economics Notes 28, 335–355.

Fitch Ratings (2006), ‘Exposure draft: Introducing the Fitch VECTOR De-fault Model Version 3.0’.

Franke, G. & Krahnen, J. (2008), ‘The future of securitization’, Working Pa-per, Frankfurt Center for Financial Studies .

Gordy, M. (2000), ‘A comparative anatomy of credit risk models’, Journal ofBanking and Finance 24, 119–149.

Gordy, M. (2003), ‘A risk-factor model foundation for ratings-based bank cap-ital rules’, Journal of Financial Intermediation 12, 199–232.

Gupton, G., Finger, C. & Bhatia, M. (1997), ‘CreditMetrics technical docu-ment’.

Hamerle, A., Liebig, T. & Scheule, H. (2006), ‘Forecasting credit event fre-quency - empirical evidence for West German firms’, Journal of Risk 9, 75–98.

Heitfield, E. (2005), ‘Dynamics of rating systems’, Basel Committee on Bank-ing Supervision, Studies on the Validation of Internal Rating Systems,Working Paper No. 14.

Hellwig, M. (2008), ‘Systemic risk in the financial sector: An analysis of thesubprime-mortgage financial crisis’, Max Planck Institute for Research onCollective Goods, Working Paper.

Hui, C., Wong, T., Lo, C. & Huang, M. (2007), ‘Ratings versus market-basedmeasures of default risk of east asian banks’, Working Paper, Hong KongMonetary Authority .

Hull, J. (2008), ‘The credit crunch of 2007: What went wrong? why? whatlessons can be learned?’, Working Paper, University of Toronto .

Hull, J. & White, A. (2004), ‘Valuation of a CDO and nth to default CDSwithout monte carlo simulation’, Journal of Derivatives 12, 8–23.

Jorion, P., Liu, Z. & Shi, C. (2005), ‘Informational effects of regulation fd:Evidence from rating agencies’, Journal of Financial Economics 76, 309–330.

Laurent, J. & Gregory, L. (2005), ‘Basket default swaps, CDOs and factorcopulas’, Journal of Risk 18, 103–122.

Leot, P., Arber, D. & Schou-Zibell, L. (2008), ‘Securitisation in east asia, asiandevelopment bank working paper series on regional economic integration’,Asian Development Bank Working Paper Series on Regional Economic In-tegration .

Li, D. X. (2000), ‘On default correlation: A copula function approach’, Journalof Fixed Income 9, 43–54.

Loeffler, G. (2004), ‘An anatomy of rating through the cycle’, Journal of Bank-ing and Finance 28, 695–720.

Longstaff, F. & Rajan, A. (2008), ‘An empirical analysis of the pricing ofcollateralized debt obligations’, Journal of Finance 63, 529–563.

28

Mayer, C., Pence, K. & Sherlund, S. (2008), ‘The rise in mortgage defaults:Facts and myths’, forthcoming in Journal of Economic Perspectives .

McNeil, A. & Wendin, J. (2007), ‘Bayesian inference for generalized linearmixed models of portfolio credit risk’, Journal of Empirical Finance 14, 131–149.

Merton, R. C. (1974), ‘On the pricing of corporate debt: The risk structure ofinterest rates’, Journal of Finance 29, 449–470.

Micu, M., Remolona, E. & Wooldridge, P. (2006), ‘The price impact of ratingannouncements: Which announcements matter?’, BIS Working Paper No.207 .

Miu, P. & Ozdemir, B. (2002), ‘Rating banks, risk and uncertainty in anopaque industry’, American Economic Review 92, 874–888.

Moody’s Investors Service (2006), ‘CDOROM v2.3 user guide’.Moody’s Investors Service (2008), ‘Default & loss rates of structured finance

securities: 1993-2007’.Nickell, P., Perraudin, W. & Varotto, S. (2000), ‘Stability of rating transitions’,

Journal of Banking and Finance 24, 203–227.Ohlson, J. (1980), ‘Financial ratios and the probabilistic prediction of

bankruptcy’, Journal of Accounting Research 18, 109–131.Purda, L. (2007), ‘Stock market reactions to anticipated versus surprise rating

changes’, Journal of Financial Research 30, 301–320.Radelet, S. & Sachs, J. (1998), ‘The east asian financial crisis: Diagnosis,

remedies, prospects’, Brookings Papers 28, 1–90.Rajan, U., Seru, A. & Vig, V. (2008), ‘The failure of models that predict fail-

ure: distance, incentives and defaults’, Working paper, University of Michi-gan, University of Chicago and London Business School .

Rosch, D. (2005), ‘An empirical comparison of default risk forecasts fromalternative credit rating philosophies’, International Journal of Forecasting25(1), 37–51.

Rosch, D. & Scheule, H. (2005), ‘A multi-factor approach for systematic de-fault and recovery risk’, Journal of Fixed Income 15(2), 63–75.

Rosch, D. & Scheule, H. (2009), ‘Forecasting downturn credit portfolio risk’,forthcoming in Financial Markets, Institutions and Instruments .

Shumway, T. (2001), ‘Forecasting bankruptcy more accurately: a simplehazard-rate model’, Journal of Business 74, 101–124.

Standard & Poor’s (2005), ‘CDO Evaluator Version 3.0: Technical document.’.Vasicek, O. (1991), Limiting loan loss probability distribution, Working paper,

KMV Corporation.

29

Figures

Fig. 1. Real impairment rate and estimated impairment rate for Model 1 and Model2, 1997-2007

Notes: The real impairment rate is the mean impairment indicator (1: impair-ment, 0: no impairment) per year. The estimated impairment rate is the meanestimated tranche impairment probability given Model 1 and Model 2.

30

Fig. 2. Real impairment rate and estimated impairment rate for Model 3, Model 4,Model 5 and Model 6, 1997-2007

Notes: The real impairment rate is the mean impairment indicator (1: impair-ment, 0: no impairment) per year. The estimated impairment rate is the meanestimated tranche impairment probability given Model 3, Model 4, Model 5and Model 6.

31

Tables

32

Tab

le1

Tot

alnu

mbe

rof

obse

rvat

ions

,pe

rcen

tage

per

rati

ngca

tego

ryan

dpe

ras

set

port

folio

type

,19

97-2

007

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Yea

rA

llA

aa-A

Baa

Ba

BC

aaA

BS

CD

OH

EL

MB

S

1997

889

91.9

0%3.

82%

3.15

%1.

12%

0.00

%47

.24%

0.45

%12

.71%

39.6

0%

1998

1,03

491

.39%

4.45

%3.

09%

1.06

%0.

00%

50.6

8%0.

68%

12.5

7%36

.07%

1999

1,55

685

.28%

5.72

%6.

94%

1.80

%0.

26%

48.2

0%3.

66%

20.3

1%27

.83%

2000

2,22

082

.30%

9.77

%4.

55%

2.39

%0.

99%

50.1

8%9.

23%

15.4

1%25

.18%

2001

3,18

379

.80%

10.4

6%6.

60%

2.67

%0.

47%

48.6

3%12

.38%

15.0

5%23

.94%

2002

4,57

679

.28%

12.0

0%6.

29%

1.88

%0.

55%

38.3

5%12

.39%

25.4

2%23

.84%

2003

6,41

379

.84%

12.5

5%4.

74%

2.06

%0.

81%

33.0

0%8.

79%

33.9

0%24

.31%

2004

8,51

979

.43%

13.5

1%4.

08%

1.91

%1.

06%

31.6

0%9.

32%

40.0

6%19

.02%

2005

11,3

5278

.64%

15.4

2%3.

70%

1.47

%0.

77%

27.1

3%8.

60%

45.7

7%18

.50%

2006

16,0

1076

.98%

16.6

8%4.

65%

1.12

%0.

57%

22.5

3%7.

58%

50.5

0%19

.39%

2007

24,0

8376

.01%

17.2

9%5.

50%

0.83

%0.

37%

15.4

8%7.

94%

54.6

8%21

.91%

Tot

al79

,835

81.9

0%11

.06%

4.85

%1.

66%

0.53

%37

.55%

7.37

%29

.67%

25.4

2%

Not

es:

This

table

show

sth

atth

enum

ber

ofra

ted

tran

ches

has

incr

ease

dat

anin

crea

sing

rate

.T

he

rati

ng

qual

ity

ofra

ted

tran

ches

has

gener

ally

dec

reas

edov

erti

me.

The

rela

tive

freq

uen

cyof

HE

L,i.e.

,su

bpri

me

mor

tgag

elo

anp

ortf

olio

shas

incr

ease

d.

33

Tab

le2

Tot

alnu

mbe

rof

obse

rvat

ions

,pe

rcen

tage

per

rese

curi

tiza

tion

stat

us,

per

prin

cipa

lpa

ymen

tye

arst

atus

and

per

subo

rdin

atio

nle

vel,

1997

-200

7(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)

Yea

rA

llR

esec

.=0

1P

PY

=0

1Sub.=

0%-3

%3%

-12%

12%

-100

%

1997

889

96.4

0%3.

60%

95.8

4%4.

16%

29.9

2%12

.15%

57.9

3%

1998

1,03

495

.55%

4.45

%93

.33%

6.67

%32

.59%

11.4

1%56

.00%

1999

1,55

696

.27%

3.73

%94

.41%

5.59

%34

.90%

12.0

8%53

.02%

2000

2,22

097

.48%

2.52

%95

.41%

4.59

%35

.18%

12.7

9%52

.03%

2001

3,18

397

.61%

2.39

%93

.21%

6.79

%32

.36%

15.8

3%51

.81%

2002

4,57

696

.42%

3.58

%91

.48%

8.52

%29

.90%

17.0

0%53

.10%

2003

6,41

395

.63%

4.37

%88

.18%

11.8

2%27

.34%

18.1

5%54

.51%

2004

8,51

994

.91%

5.09

%89

.46%

10.5

4%27

.95%

19.5

0%52

.55%

2005

11,3

5294

.85%

5.15

%88

.05%

11.9

5%27

.98%

21.4

2%50

.60%

2006

16,0

1095

.12%

4.88

%89

.78%

10.2

2%26

.82%

23.8

7%49

.31%

2007

24,0

8393

.78%

6.22

%90

.01%

9.99

%26

.24%

26.1

8%47

.58%

Tot

al79

,835

95.8

2%4.

18%

91.7

4%8.

26%

30.1

1%17

.31%

52.5

9%

Not

es:

This

table

show

sth

atth

enum

ber

ofra

ted

tran

ches

has

incr

ease

dat

anin

crea

sing

rate

.T

he

rese

curi

tiza

tion

leve

lhas

gener

ally

incr

ease

dov

erti

me.

The

rela

tive

freq

uen

cyof

pri

nci

pal

pay

men

tye

ars

has

dec

reas

edas

the

over

all

num

ber

oftr

anch

eshas

incr

ease

d.

The

rela

tive

freq

uen

cyof

mez

zanin

etr

anch

es(s

ub

ordin

atio

n=

3%-1

2%)

has

incr

ease

d.

34

Tab

le3

Impa

irm

ent

rate

sfo

ral

lob

serv

atio

ns,

per

rati

ngca

tego

ryan

dan

dpe

ras

set

port

folio

type

,19

97-2

007

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Yea

rA

llA

aa-A

Baa

Ba

BC

aaA

BS

CD

OH

EL

MB

S

1997

1.01

%0.

00%

2.94

%17

.86%

30.0

0%0.

00%

0.00

%0.

00%

6.19

%0.

57%

1998

0.87

%0.

21%

4.35

%12

.50%

9.09

%0.

00%

0.19

%0.

00%

6.15

%0.

00%

1999

1.86

%0.

30%

7.87

%5.

56%

32.1

4%75

.00%

0.53

%0.

00%

6.65

%0.

92%

2000

1.35

%0.

11%

0.46

%7.

92%

9.43

%63

.64%

1.17

%3.

41%

2.34

%0.

36%

2001

1.79

%0.

24%

4.80

%7.

14%

18.8

2%26

.67%

1.55

%7.

36%

0.63

%0.

13%

2002

3.45

%0.

30%

4.74

%27

.78%

30.2

3%60

.00%

4.16

%12

.70%

0.52

%0.

64%

2003

2.25

%0.

18%

4.10

%14

.47%

29.5

5%36

.54%

3.97

%6.

38%

0.60

%0.

71%

2004

2.10

%0.

46%

2.35

%10

.63%

19.6

3%57

.78%

4.64

%4.

03%

0.18

%0.

99%

2005

0.76

%0.

01%

0.46

%4.

05%

15.5

7%39

.08%

0.68

%2.

36%

0.31

%1.

24%

2006

0.60

%0.

00%

0.30

%1.

88%

13.9

7%53

.26%

0.89

%1.

15%

0.28

%0.

87%

2007

6.21

%1.

04%

15.0

1%42

.60%

31.6

6%59

.55%

0.48

%14

.70%

8.25

%2.

08%

Tot

al2.

02%

0.26

%4.

31%

13.8

5%21

.83%

42.8

6%1.

66%

4.74

%2.

92%

0.77

%

Not

es:

This

table

show

sth

atim

pai

rmen

tra

tes

are

hig

hin

2002

and

2007

and

low

in19

97,

1998

and

2006

.Im

pai

rmen

tra

tes

per

rati

ng

cate

gory

fluct

uat

ean

dim

pai

rmen

tra

tes

eras

set

por

tfol

ioty

pe

incr

ease

in20

02fo

rC

DO

san

din

2007

for

AB

Ss,

CD

Os

and

HE

Ls.

35

Tab

le4

Impa

irm

ent

rate

sfo

ral

lobs

erva

tion

s,pe

rre

secu

riti

zati

onst

atus

,per

prin

cipa

lpay

men

tye

arst

atus

and

per

subo

rdin

atio

nle

vel,

1997

-200

7

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Yea

rA

llR

esec

.=0

1P

PY

=0

1Sub.=

0%-3

%3%

-12%

12%

-100

%

1997

1.01

%1.

05%

0.00

%1.

06%

0.00

%3.

01%

0.93

%0.

00%

1998

0.87

%0.

91%

0.00

%0.

93%

0.00

%1.

48%

0.85

%0.

52%

1999

1.86

%1.

94%

0.00

%1.

91%

1.15

%1.

84%

5.32

%1.

09%

2000

1.35

%1.

34%

1.79

%1.

42%

0.00

%2.

43%

2.11

%0.

43%

2001

1.79

%1.

80%

1.32

%1.

85%

0.93

%3.

01%

3.77

%0.

42%

2002

3.45

%3.

58%

0.00

%3.

75%

0.26

%8.

41%

3.47

%0.

66%

2003

2.25

%2.

20%

3.21

%2.

49%

0.40

%3.

37%

3.69

%1.

20%

2004

2.10

%2.

00%

3.92

%2.

31%

0.33

%2.

73%

3.49

%1.

25%

2005

0.76

%0.

61%

3.42

%0.

75%

0.81

%1.

64%

0.95

%0.

19%

2006

0.60

%0.

58%

0.90

%0.

63%

0.37

%1.

33%

0.65

%0.

18%

2007

6.21

%5.

46%

17.5

0%4.

17%

24.5

9%18

.09%

4.74

%0.

47%

Tot

al2.

02%

1.95

%2.

91%

1.93

%2.

62%

4.30

%2.

72%

0.58

%

Not

es:

This

table

show

sth

atim

pai

rmen

tra

tes

are

hig

hin

2002

and

2007

and

low

in19

97,

1998

and

2006

.T

he

impai

rmen

tra

tehas

incr

ease

din

2007

for

rese

curi

tiza

tion

s,pri

nci

pal

pay

men

tye

ars

and

equit

ytr

anch

es(s

ub

ordin

atio

n=

0%-3

%)

asw

ell

asm

ezza

nin

etr

anch

es(s

ub

ordin

atio

n=

3%-1

2%).

36

Table 5Parameter estimates for Model 1 and Model 2

(1) (2)

Variable Model 1 Model 2

Intercept -2.6429*** -3.2526***

(0.0211) (0.0421)

Baa 1.1195*** 1.0897***

(0.0277) (0.0288)

Ba 1.8127*** 1.8273***

(0.0310) (0.0323)

B 1.8711*** 2.0300***

(0.0469) (0.0498)

Caa 2.6692*** 2.7945***

(0.0612) (0.0634)

ABS 0.5329***

(0.0447)

CDO 0.8764***

(0.0442)

HEL 0.7219***

(0.0394)

Pseudo R2 0.0733 0.0795

R2 re-scaled 0.3196 0.3466

AUROC 0.8787 0.9078

Notes : Table shows parameter estimates from the probit modelP (Dijt = 1) = Φ (β′xijt); standard errors in parentheses; ***: signifi-cant at 1%, **: significant at 5%, *: significant at 10%; AUROC is the areaunder the receiver operating characteristic .

37

Table 6Parameter estimates for Model 3, Model 4, Model 5 and Model 6

(1) (2) (3) (4)

Variable Model 3 Model 4 Model 5 Model 6

Intercept -3.3572*** -3.0131*** -3.4332*** -3.3111***

(0.0436) (0.0447) (0.0447) (0.0491)

Baa 1.1107*** 0.8893*** 1.1447*** 0.9654***

(0.0292) (0.0316) (0.0300) (0.0333)

Ba 1.8757*** 1.6219*** 1.8375*** 1.6818***

(0.0329) (0.0349) (0.0333) (0.0366)

B 2.1414*** 1.8460*** 2.0768*** 2.0159***

(0.0507) (0.0512) (0.0511) (0.0535)

Caa 2.9197*** 2.6040*** 2.8838*** 2.8361***

(0.0643) (0.0645) (0.0647) (0.0666)

ABS 0.6015*** 0.5563*** 0.5921*** 0.6711***

(0.0457) (0.0450) (0.0456) (0.0468)

CDO 0.5324*** 0.8714*** 0.9702*** 0.6041***

(0.0499) (0.0445) (0.0454) (0.0519)

HEL 0.8017*** 0.7160*** 0.7348*** 0.8144***

(0.0406) (0.0396) (0.0404) (0.0420)

Resecuritization 0.7887*** 0.8273***

(0.0508) (0.0525)

Subordination -1.0547*** -1.0426***

(0.0916) (0.0909)

PPY 0.7506*** 0.8067***

(0.0297) (0.0307)

Pseudo R2 0.0822 0.0816 0.0861 0.0913

R2 re-scaled 0.3585 0.3558 0.3755 0.3978

AUROC 0.9117 0.9162 0.9118 0.9210

Notes : Table shows parameter estimates from the probit modelP (Dijt = 1) = Φ (β′xijt); standard errors in parentheses; ***: signifi-cant at 1%, **: significant at 5%, *: significant at 10%; AUROC is the areaunder the receiver operating characteristic .

38

Table 7Parameter estimates for Model 7 (random effect) given ratings

(1) (2) (3) (4) (5) (6)

Rating All Grades Aaa-A Baa Ba B Caa

# 79,835 62,528 11,811 3,907 1,113 476

Intercept -2.1255*** -3.0155*** -1.9228*** -1.2268*** -0.7946*** 0.02115

(0.0848) (0.1560) (0.1641) (0.1622) (0.0867) (0.0949)

b 0.2734*** 0.4653*** 0.5091*** 0.5145*** 0.2255** 0.1999*

(0.0601) (0.1313) (0.1183) (0.114) (0.0742) (0.0956)

Notes : Table shows parameter estimates from the random effects probitmodel P (Dijt = 1|Ft) = Φ (α + b · Ft); standard errors in parentheses; ***:significant at 1%, **: significant at 5%, *: significant at 10%.

Table 8Parameter estimates for Model 7 (random effect) given asset pool type

(1) (2) (3) (4)

Deal Type CDO MBS ABS HEL

# 6,692 17,230 21,334 35,579

intercept -1.7045*** -2.4341*** -2.2912*** -2.1937***

(0.1460) (0.0816) (0.1325) (0.1725)

b 0.4121*** 0.205*** 0.4111*** 0.5563***

(0.1123) (0.0688) (0.1066) (0.1242)


39

Table 9Parameter estimates for Model 7 (random effect) given asset pool type and resecu-ritization status

(1) (2) (3) (4)

resecuritization 0 1

Deal Type All CDO All CDO

# 75,835 3,622 4,010 3,070

intercept -2.1366*** -1.8302*** -2.1439*** -1.8881***

(0.0839) (0.1562) (0.2226) (0.2523)

b 0.2703*** 0.4346*** 0.5771** 0.5801**

(0.0594) (0.1203) (0.1823) (0.2073)


40

Table 10Parameter estimates for Model 7 (random effect) and Model 8 (random effect con-trolling for rating)

(1) (2)

Model 7 Model 8

# 79,835 79,835

intercept -2.1255*** -2.8533***

(0.0848) (0.1350)

Baa 1.1013***

(0.0401)

Ba 1.8208***

(0.0583)

B 2.0696***

(0.0727)

Caa 2.9167***

(0.0983)

b 0.2734*** 0.3541***

(0.0601) (0.0672)

Notes : Table shows parameter estimates from the random effects probitmodel P (Dijt = 1|Ft) = Φ (β′xijt + b · Ft); standard errors in parentheses;***: significant at 1%, **: significant at 5%, *: significant at 10%.

41

Table 11Parameter estimates for Model 9 (random effect controlling for year) given ratingcategory

(1) (2) (3) (4) (5) (6)

All grades Aaa-A Baa Ba B Caa

# 79,835 62,528 11,811 3,907 1,113 476

c -2.2819*** -3.0197*** -1.8121*** -1.277*** -0.8287*** -0.0260***

(0.1816) (0.3519) (0.3800) (0.3637) (0.2367) (0.3572)

beta 0.02543 0.000674 -0.01725 0.008043 0.004828 0.005858

(0.0263) (0.0505) (0.0536) (0.0522) (0.0311) (0.0428)

b 0.2634*** 0.4653*** 0.5061*** 0.5135*** 0.2252** 0.1972*

(0.0575) (0.1314) (0.1176) (0.1140) (0.0742) (0.0975)

Notes : Table shows parameter estimates from the random effects probitmodel P (Dijt = 1|Ft) = Φ (α + β · t+ b · Ft); standard errors in parentheses;***: significant at 1%, **: significant at 5%, *: significant at 10%.

42

Table 12Parameter estimates for Model 10 (prediction model, Probit Regression)

(1) (2) (3) (4) (5)

Prediction year γ0 γ1 Pseudo R2 R2 Rescaled AUROC

1999 -0.3623* 0.7647** 0.0416 0.266 0.872

(0.2106) (0.1001)

2000 -0.4889*** 0.9471 0.0633 0.4743 0.941

(0.1555) (0.0998)

2001 -0.2884* 0.8176** 0.0438 0.2664 0.879

(0.1505) (0.0726)

2002 0.7389*** 1.1986*** 0.1035 0.399 0.917

(0.126) (0.0658)

2003 -0.0389 0.9906 0.072 0.3723 0.918

(0.0933) (0.053)

2004 -0.227*** 0.8775*** 0.0609 0.3303 0.867

(0.0804) (0.0427)

2005 -0.2617** 1.2778*** 0.0415 0.4875 0.97

(0.1047) (0.0833)

2006 0.0432 1.5492*** 0.0392 0.5544 0.948

(0.114) (0.0998)

2007 1.2847*** 1.1665*** 0.124 0.3331 0.864

(0.0531) (0.0224)

Notes : Table shows results of out-of-sample prediction probit regressionsP (Dijt+1 = 1) = Φ (γ0 + γ1ηijt+1); standard errors in parentheses; ***:significant at 1%, **: significant at 5%, *: significant at 10%; Hypotheses aretested that γ0 = 0 and γ1 = 1.

43

Table 13Parameter estimates for Model 11 (prediction models, Linear Regression)

(1) (2) (3)

Prediction year δ0 δ1 Adj. R2

1999 0.0035 0.7808*** 0.0733

(0.0034) (0.0699)

2000 -0.0049** 0.72835*** 0.2913

(0.0022) (0.0241)

2001 0.0068*** 0.62567*** 0.0688

(0.0024) (0.0407)

2002 0.007*** 1.60006*** 0.1785

(0.0026) (0.0507)

2003 0.0020 0.87975*** 0.1483

(0.0018) (0.0263)

2004 -0.0020 0.0015 0.1979

(0.0015) (0.0218)

2005 -0.0055*** 0.6465*** 0.1849

(0.0008) (0.0127)

2006 -0.0068*** 0.7879*** 0.2397

(0.0006) (0.0111)

2007 0.0352*** 2.16958*** 0.1253

(0.0015) (0.0369)

Notes : Table shows results of out-of-sample prediction linear regressionsDijt+1 = δ0 + δ1 · pijt+1 + εijt+1; standard errors in parentheses; ***: significantat 1%, **: significant at 5%, *: significant at 10%; Hypotheses are tested thatδ0 = 0 and δ1 = 1.

44

what was wrong with credit ratings for securitizations? - evidence from ... was wrong with...

Documents