MBS ratings and the mortgage credit boom

Transcription

1 MBS ratings and the mortgage credit boom Adam Ashcraft (New York Fed) Paul Goldsmith Pinkham (Harvard University, HBS) James Vickery (New York Fed) Bocconi / CAREFIN Banking Conference September 21, 2009 Views expressed in thispresentation areourown own, and do notreflect the opinions of the Federal Reserve Bank of New York or the Federal Reserve System.

2 Net ratings downgrades: average number of notches Alt-A Subprime 0 Net rating downgrade 5 (notches) q1 2003q1 2005q1 2007q1 2001q1 2003q1 2005q1 2007q1 Calendar quarter of deal securitization Source: Authors calculations based on ABSNet and Bloomberg 2

3 This paper Study credit ratings on subprime and Alt A MBS deals issued in the period leading up to the financial crisis ( ). Research question: To what extent were ratings flawed ex ante? Do observed outcomes just reflect an unlucky realization of fundamentals ex post? Two types of analysis: 1. Determinants of credit ratings. How do ratings evolve through time, conditional on risk? 2. Relationship between ratings and realized deal performance. How well did ratings summarize available information? 3

4 Ratings errors: Bad luck or bad modelling? In response to the increase in the riskiness of loans made during the last few years and the changing economic environment, Moody s steadily increased dits loss expectations ti and subsequent levels of credit protection on pools of subprime loans. Our loss expectations and enhancement levels rose by about 30% over the 2003 to 2006 time period Along with most other market participants, however, we did not anticipate the magnitude and speed of the deterioration in mortgage quality (particularly for certain originators) or the rapid transition to restrictive e lending. Michael Kanef, Moodys executive Senate testimony, 9/26/07 / 4

5 Main findings 1. Time variation in ratings, with deterioration in standards at end of mortgage credit boom ( ). Deals become progressively riskier between , but ratings stay flat during this period. 2. Ratings not sufficiently sensitive to credit risk. We construct a simple summary statistic for the credit risk of each deal. This variable strongly predicts worse performance (defaults, losses, rating downgrades), conditional on rating. Implication: High risk deals were over rated. Ratings not a sufficient statistic for level of credit risk in the deal. Results stronger for Alt A A deals and deals with high fraction of low doc loans, where opacity is arguably greater. 5

6 Related Literature Theoretical work on credit rating agencies (CRAs) Mathis, McAndrews and Rochet (JME, forthcoming): Dynamic setting, reputation cycles in credit ratings. Bolton, Freixas and Shapiro (2008); Sangiorgi, Sokobin and Spatt (2009); Mariano (2008); Skreta and Veldkamp (2008) etc. Empirical evidence on credit ratings: Nd Nadauld and Sherlund (2008); Kisgen and Strahan (2009); Becker and Milbourn (2008); Benmelech and Dlugosz (2009); Mason and Rosner (2007); Griffin and Tang (2009). etc. Related work on the subprime crisis: Stanton and Wallace (2009); Coval, Jurek and Stafford (2008); Ashcraft and Schuermann (2008); Gorton (2008). etc. 6

7 What is a credit rating? Ordinal measure of credit risk on a debt security S&P and Fitch: Rating measures the probability of default. Moody's: Closer to a measure of expected loss. Our primary measureof ratings: AAA subordination Fraction of claims on the deal that receive a rating below AAA. Also known as the AAA attachment point. Note: this measure is continuous, even though rating on each individual id lbond dis discrete. 7

8 Data: Nonagency deals 3,144 Alt A and subprime deals 59,995 securities Individual Mortgages Mortgage Pools REMIC Trust RMBS Bonds M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12 M13 M14 M15 M16 M17 M18 M19 M20 M21 M22 M23 M24 M25 M26 M27 M28 M29 M30 M31 M32 M33 M34 M35 M36 M37 M38 M39 M40 M41 M42 M43 M44 M45 M46 M47 M48 M49 M50 M51 M52 M53 M54 M55 M56 M57 M58 M59 M60 M61 M62 M63 M64 M65 M66 M67 M68 M69 M70 M71 M72 M73 M74 M75 M76 M77 M78... M 2000 M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12 M13 M14 M15 M16 M17 M18 M19 M20 M21 M22 M23 M24 M25 M26 M27 M28 M29 M30 M31 M32 M33 M34 M35 M36 M37 M /28 Hybrid ARM Mortgage Pool Fixed Rate Mortgage Special Purpose Vehicle (RMBS Trust) AAA RMBS AA RMBS A RMBS BBB RMBS BBB RMBS M 1000 Residual Source: Gorton (2008) LoanPerformance ( 11m loans) ABSNet / Bloomberg 8

9 Summary statistics Subprime Alt A A All Number of deals Deal size, mean ($m) Deal size, median ($m) Total number of securities Securities per deal, median AAA securities per deal, median Credit enhancement Mean fraction of AAA securities Percent ofdealswithbondinsurance insurance Average value of insurance (%FV) Excess spread at origination (%), avg Number of CRAs that rated deal (%) One Two Three Four

10 Non agency MBS issuance Subprime Alt-A # of Deals q1 2002q1 2003q1 2004q1 2005q1 2006q1 2007q1 2008q1 2009q1 2001q1 2002q1 2003q1 2004q1 2005q1 2006q1 2007q1 2008q1 2009q1 Year-Quarter Year-Quarter # of Deals Orig. Amt ($ Bn) # of Deals Orig. Amt ($ Bn) IMF Volume ($ Bn) IMF Volume ($ Bn) 10

11 Trends in Subordination (below AAA, A, BBB) Subordination over Time Subprime Alt-A q1 2002q1 2003q1 2004q1 2005q1 Year-Quarter 2006q1 2007q1 2008q1 2001q1 2002q1 2003q1 2004q1 2005q1 Year-Quarter 2006q1 2007q1 2008q1 AAA Subordination BBB Subordination AA/A Subordination AAA Subordination BBB Subordination AA/A Subordination 11

12 Hypotheses Hypothesis 1: Ratings stability. Level of ratings remains constant through time, after controlling for level of credit risk and structural features of the deal. Hypothesis 2: Informational efficiency. Other credit risk variables do not systematically ypredict future deal performance, after conditioning on ratings. Benchmark: Ratings are an efficient forward looking summary statistic for credit risk; reflect all available information. Analogy: Rational expectations forecasts (e.g. Sargent, 1987). 12

13 Step 1: Default model Baseline: simple logit model (10% LoanPerformance sample). Similar structure to Demyanyk & Van Hemert (2009). Dependent variable: 90+ delinquent, REO, prepaid with loss after 12 months Key predictors of default: Trailing house price appreciation (OFHEO, past 12 months), unemployment. Underwriting : FICO score; debt to income (DTI); combined loan to valuation (CLTV) ratio; documentation of borrower income; investor dummy, SATO. Loan type variables (ARM, FRM, interest only, balloon, refinancing, prepayment penalty dummy, loan size). Estimate model dlrecursively every six months between Calculate expected default rate for each deal, based only on ex ante historical i data (i.e. data available at time deal was rated). 13

14 Predicted and realized mortgage default rates Fraction loans 90+ delinquent, prepaid with loss, REO 12 months after deal issuance 30 Subprime Alt-A q1 2002q1 2003q1 2004q1 2005q1 Year-Quarter 2006q1 2007q1 2008q1 2001q1 2002q1 2003q1 2004q1 2005q1 Year-Quarter 2006q1 2007q1 2008q1 Actual SD Forecast Model Actual SD Forecast Model 14

15 Determinants of AAA attachment point Dependent variable: ln(1+subordination below AAA class in percentage points). Subprime Alt A Credit risk ln(1+projected default rate) 0.751*** 0.680** 0.727*** 0.651*** (0.231) (0.254) (0.231) (0.186) ln(1+projected default rate) ** (0.0676) (0.0705) (0.0870) (0.0707) Joint significance: F Test (p value) *** *** *** *** Other deal characteristics Bond insurance (1=yes) 0.473*** 0.478*** (0.100) (0.100) (0.0395) (0.0376) Fraction of deal with bond insurance ** ** * ( ) ( ) ( ) ( ) Wihtd Weighted average coupon rate *** *** (0.0408) (0.0405) (0.0145) (0.0148) Weighted mortgage interest rate * ** * (0.0231) (0.0233) (0.0368) (0.0341) Geographic concentration of loans 1.897*** 1.677*** 0.406*** 0.399*** (0.212) (0.263) (0.134) (0.117) Year x quarter dummies Yes Yes Yes Yes F test: ratings decline over ? (p value) a *** *** *** Include aggregated loan level variables No Yes No Yes Joint significance: F Test (p value) *** N R

16 Actual and predicted subordination Subprime Decomposition Subprime q1 2001q3 2002q1 2002q3 2003q1 2003q3 2004q1 2004q3 2005q1 2005q3 2006q1 Year-Quarter, Deal Origination 2006q3 2007q1 2007q q4 2005q2 2005q4 2006q2 2006q4 2007q2 Year-Quarter 95% Lower Bound 95% Upper Bound Unconditional Subordination Linear Model Unconditional Subordination Model, SD Forecast Fixed Model, All Char Fixed Model, HPA Fixed Alt-A 2001q1 2001q3 2002q1 2002q3 2003q1 2003q3 2004q1 2004q3 2005q1 2005q3 2006q1 2006q3 2007q1 2007q3 Year-Quarter, Deal Origination Alt-A 2004q4 2005q2 2005q4 2006q2 2006q4 2007q2 Year-Quarter 95% Lower Bound 95% Upper Bound Unconditional Subordination Linear Model Unconditional Subordination Model, SD Forecast Fixed Model, All Char Fixed Model, HPA Fixed 16

17 Hypothesis 2: Informational efficiency Relationship between ratings and ex post performance: Performance = a. credit rating + b. credit enhancement + c. model projected default rate + e Null hypothesis: Projected credit risk and other variables do not systematically predict performance after conditioning on rating. (i.e. test c = 0, and also a 0). Performance measured by: (i) ()df defaults; (ii) () rating downgrades; d (iii) realized ex post losses. 17

18 Ratings and ex post default: Subprime deals A. Subprime Dependent variable: Fraction of deal in default 12 months after deal is issued model, rating and baseline rating only model only model & rating loan covariates ln(1+% subordination below AAA) 0.285*** 0.112*** 0.112*** (0.0396) (0.0340) (0.0305) ln(1+% subordination below BBB ) 0.110*** *** *** (0.0211) (0.0157) (0.0144) ln(1+projected default rate) 1.076*** 0.941*** 1.004*** (0.0566) (0.0622) (0.0567) Other deal characteristics Yes Yes Yes Yes Yes Year x quarter dummies Yes Yes Yes Yes Yes Loan covariates aggregated to deal No No No No Yes F test: Aggregated loan covariates [p val] *** N R

19 Ratings and ex post default: Alt A deals B. Alt A Dependent variable: Fraction of deal in default 12 months after deal is issued model, rating and deal covariates baseline rating only model only model & rating ln(1+% subordination below AAA) 0.356*** 0.198*** (0.0531) (0.0378) (0.0344) ln(1+% subordination below BBB ) 0.201*** 0.144*** *** (0.0638) (0.0423) (0.0241) ln(1+projected default rate) 1.556*** 1.470*** 1.523*** (0.0604) (0.102) (0.0422) Other deal characteristics Yes Yes Yes Yes Yes Year x quarter dummies Yes Yes Yes Yes Yes Aggregated mortgage characteristicsno No No No Yes F test: Deal lvl covariates [p value] *** N R

20 Other determinants of default: Subprime (Alt A similar) Include projected default rate No Yes Yes Credit boom interactions No No Yes Rating strategy One Rating 0.509*** 0.213*** 0.305*** (0.126) (0.0343) (0.0392) Two Ratings *** *** * 0505* (0.0171) (0.0149) (0.0268) Four Ratings *** 0.106*** (0.0274) (0.0290) (0.0275) Aggregate g loan level covariates (FICO, IO not presented) LTV *** *** ( ) ( ) ( ) HPA ** ( ) ( ) ( ) Low doc *** 00725*** *** 00681*** *** 00327*** ( ) ( ) ( ) Investor ** * *** ( ) ( ) ( ) Loan level covariate interactions Projected delinquency rate * boom 0.317*** (0.0939) Low doc * boom *** ( ) Investor * boom *** *** ( ) Less Than Three Ratings* boom ** (0.0295) 20

21 Vintage analysis: Subprime deals (Alt A estimates similar) Vintage All Years A. Subprime deals Baseline (just deal controls; same as Column 1 of Table 6) R Baseline & rating Subordination below AAA 0.343** *** 0.369** 0.512** 0.474*** 0.285*** (0.0606) (0.0503) (0.0884) (0.0216) (0.0913) (0.137) (0.0634) (0.0393) Subordination below BBB ** *** (0.314) (0.184) (0.0530) (0.0264) (0.0269) (0.0494) (0.0673) (0.0209) R Baseline & model prediction Projected delinquency rate 1.072*** 0.864*** 0.718*** 0.927*** 1.428*** 0.825** 1.224*** 1.076*** (0.0249) 0249) (0.113) (0.106) (0.129) (0.0202) 0202) (0.228) (0.0252) 0252) (0.0561) 0561) R Baseline & rating & model prediction Subordination below AAA ** * 0.196* 0.112*** (0.0817) (0.0526) (0.0875) (0.0390) (0.0561) (0.136) (0.0767) (0.0337) Subordination below BBB ** ** *** (0.0288) (0.101) (0.0471) (0.0143) (0.0286) (0.0347) (0.0647) (0.0156) Projected delinquency rate 1.070*** 0.889*** 0.684*** 0.753*** 1.327*** 0.697* 0.942*** 0.941*** (0.106) (0.146) (0.109) (0.127) (0.0340) (0.275) (0.0275) (0.0617) R N

22 Determinants of rating downgrades Subprime Alt A Deal subordination below AAA 0.923*** 0.902*** * (0.246) (0.258) (0.227) (0.282) Deal subordination below BBB 0.630*** 0.442** 1.489*** 1.640*** (0.206) (0.178) (0.420) (0.440) Projected default rate 0.817* *** 3.066*** (0.472) (0.597) (0.909) (0.948) Other deal characteristics Yes Yes Yes Yes Rating strategy No Yes No Yes Aggregated loancovariates No Yes No Yes Year x quarter dummies Yes Yes Yes Yes N R

23 Determinants of realized losses to date A. Dependent variable: Realized losses to date rating only model only Subprime model & rating model, rating and deal covariates rating only model only Alt A model & rating model, rating and deal covariates ln(1+subn. below AAA) * *** 0.290*** (0.0264) 0264) (0.0328) 0328) (0.0263) 0263) (0.0676) 0676) (0.0672) 0672) (0.0502) 0502) ln(1+subn. below BBB) 0.164*** 0.163*** 0.124*** 0.129** ** ** (0.0260) (0.0258) (0.0191) (0.0503) (0.0374) (0.0189) Projected ddf default rate 0.157*** *** 0.972*** 0.841*** 0.684*** (0.0546) (0.0728) (0.0668) (0.139) (0.145) (0.0726) Other deal characteristics Yes Yes Yes Yes Yes Yes Yes Yes Year x quarter dummies Yes Yes Yes Yes Yes Yes Yes Yes Deal level mortgage characteristics No No No Yes No No No Yes F test: Deal level covariates [pvalue] *** *** N R for internal use only 23

24 Remarks Low doc deals particularly underperform their rating. Consistent with theoretical work that opacity is related to degree of rating bias (Skreta and Veldkamp, JME, 2009). Also consistent with Rajan, Seru and Vig (2009). In several dimensions, our results are stronger for Alt A deals than subprime deals. Overall downgrades also greater for Alt A. Robustness checks: Alternative measures of performance: (i) default at 24 months; (ii) default to date. More complex default models (interactions between different underwriting i variables ibl etc.). 24

25 Summary Time variation in rating standards; erosion between (Deals become riskier, but ratings stay flat.) Ratings are informative, but surprisingly less predictive of expost performance than naïve summary statistic. High risk deals as measured by summary statistic, perform significantly worse ex post (defaults, downgrades, losses). Particularly true for Alt A A deals, low doc deals. True over whole sample, not just during the crisis. Caveat: Although h some suggestive evidence, our results are not conclusive as to whether observed limitations in ratings reflect agency problems, or innocent shortcomings in methodology. 25