Default Rates On Prime And Subprime Mortgages

RESEARCH REVIEW(MSA) from the Federal Housing FinanceAgency (FHFA)—an independent federalagency that is the successor to theOffice of Federal Housing EnterpriseOversight (OFHEO) and othergovernment entities.7 We use the FHFAall transactions House Price Index (HPI)that is based on repeat sales information.Trends in loan and borrower characteristicsMany commentators (see, forexample, Demyanyk and Van Hemert,2009) have noted that subprimelending standards became more laxduring the period we study, meaningthat the typical borrower may havereceived less scrutiny over time and itbecame easier for borrowers to getloans overall, as well as to get largerloans. Table 1 summarizes loancharacteristics for each year from2004 through 2007 for both prime andsubprime mortgages.Consistent with prior work, we alsodocument declining borrower qualityover time in the subprime sector. Forexample, whereas the average FICOscore for subprime borrowers in 2004was 617, it had declined to 597 by2007. 8 By contrast, when we look atprime loans, the decline in lendingstandards is less obvious. The averageFICO score among prime borrowerswas 710 in 2004 and 706 in 2007, adecline of less than 1 percent.Our data also allow us to look at theprevalence of different mortgagetransactions, such as purchases orrefinancings. We are particularlyinterested in refinancings that extracthome equity (a cash-out refinance). Bytaking out equity in a refinancing, ahousehold may end up being morevulnerable to future home pricedeclines, especially if its new mortgagehas a high loan-to-value ratio. To theextent that the practice of cash-outrefinancing was common over theperiod in our study, the increases inhome prices may be associated withconstant or even increasing leveragerather than with safer loans and abigger cushion against future pricedeclines. In this way, greater prevalenceof cash-out refinancing transactionsmay be indicative of the increasing riskin the universe of existing loans.As indicated in table 1, mortgageservicers assign many refinancingtransactions to the ambiguous categoryof “refinancing with unknown cash-out.”Nevertheless, among prime loans madein 2004, 12 percent were known toinvolve cash-outs. By 2005, thispercentage had risen to about 21percent, and it remained at this levelthrough 2007. For subprime loans madein 2004, 35 percent involved cash-outs;for those made in 2005, 43 percent; forthose made in 2006, 47 percent; and forthose made in 2007, a staggering 57percent. Put differently, cash-out loansaccounted for at least 82 percent(0.575/0.7) of all subprime refinancingtransactions in 2007! Another loancharacteristic that might be an importantdeterminant of subsequent defaults iswhether the interest rate is fixed for thelife of the contract or allowed to adjustperiodically (as in adjustable-ratemortgages, or ARMs). When an ARMresets after the initial defined period(which may be as short as one year or aslong as seven), the interest rate and,consequently, the monthly mortgagepayment, may go up substantially. Higherpayments may put enough stress onsome families that they fall behind ontheir mortgages. While these loans seemattractive at first because of lowintroductory interest rates (and low initialpayments), they expose borrowers toadditional risk if interest rates go up or ifcredit becomes less available in general.Among subprime mortgages, ARMsaccounted for 73 percent in 2004, 69percent in 2005, and 62 percent in2006. By 2007, the ARM share hadfallen to 39 percent, since the availabilityof these types of loans declined in thesecond half of the year. Importantly,nearly all subprime ARMs haveintroductory periods of three years orless, which makes borrowers with theseloans very dependent on the ability torefinance.9 In contrast, loans to primeborrowers are predominantly made asfixed-rate contracts (about 75 percent ofall prime loans), and the majority of primeARMs have introductory periods of five toseven years.One oft-mentioned culprit for thesubprime crisis is the growth of lenderswho followed the “originate-to-distributemodel” (see, for example, Keys et al.,2010, and Calomiris, 2008). Theselenders sold virtually all of the mortgagesthey made, typically to privatesecuritizers. Because such lenders donot face a financial loss if thesemortgages eventually default, they haverelatively little incentive to screen andmonitor borrowers. In addition to sellingloans to private securitizers, the lenderscan hold loans on their own portfolios orsell them to one of the GSEs. However,only loans that meet certain criteria(borrower with a FICO score of at least620, loan value of less than 417,000,and a loan-to-value ratio [LTV] of 80percent or less) can generally be sold tothe GSEs. Most subprime loans cannotbe sold to GSEs and must be eitherprivately securitized or held in portfolio.The extent of loan securitization is oneof the striking facts in table 1. Recall thatthe LPS data comprised loans servicedby the large mortgage servicers. As aresult, LPS overstates the actual extentof securitization somewhat, as it is morecommon for smaller banks to hold loansin portfolio and also to service theminternally. That being said, the LPS dataindicate that within the first month oforigination, about half of primemortgages made in 2004 remained intheir originators’ portfolios. This figuredeclined to about 40 percent in each ofthe subsequent years in our data. Bycomparison, many fewer subprime loanswere retained by their originators evenfor the first month: just over 40 percentof loans made in 2004 and less than 30percent made in the following years.However, by the end of the first yearsince origination, the share of originatedProfitwise News and ViewsSeptember 20103

RESEARCH REVIEWTable 1: Loan characteristics at originationPrime Loans20042005Subprime Loans200620072004200520062007% default in first 12 % default in first 18 % default in first 21 %HPI growth (12 months 1.52%-3.94%HPI growth (21 months 80%5.15%4.70%4.45%4.80%5.28%4.83%4.55% 50,065 49,486 48,417 48,221 45,980 44,965 43,790 43,817 173,702 200,383 211,052 205,881 167,742 172,316 179,003 172,667Unemployment rate (12months following orig.)Median income in zip code(in 100,000)Origination amountFICO4.81%710715708706617611607597Loan to value 56%Debt to income ratio Interest rate at origination5.6%6.0%6.7%6.5%7.1%7.5%8.5%8.4%Margin rate (rate increase atreset for 3.16%12.93%73.31%69.49%61.78%38.92%reset 3 et 3 Prepayment 2%Debt to income not available(fraction of loans)Fraction of loans that are:Adjustable rate mortgages(ARMs)Purchase 46%Refinancing 2%28.89%24.47%23.67%13.40%Loan sold to an sold to .51%64.07%42.60%Loan held in %24.88%Cash out refinancing loansRefinancing withoutcashoutRefinancing with unknowncashoutInvestment property loansConforming loanAs recorded at originationAs recorded at 12 months since originationLoan sold to an sold to .89%88.59%55.06%Loan held on umber of loans in 8,562SOURCE: FHFA for HPI growth, BLS for unemployment rate and median income, LPS for all other variables.4Profitwise News and ViewsSeptember 2010

RESEARCH REVIEWloans kept on portfolio drops to lowsingle digits for both prime and subprimemortgages. Not surprisingly, nearly allsubprime mortgages are securitized byprivate investors, whereas GSEsdominate the securitization of primemortgages. However, by the second halfof 2007, the private securitization markethad all but disappeared. The GSEs tookup much of the slack, accounting forabout 40 percent of all subprimesecuritizations in that year.10Estimates of defaultIn this section, we estimate empiricalmodels of the likelihood that a loan willdefault in its first 12 months. This allowsus to quantify which factors make defaultmore or less likely and to examine howthe sensitivity to default varies over timeand across prime and subprime loans.Econometric modelMortgages can have multiple sourcesof risk—for example, low credit quality,high loan-to-value ratios, and adjustablerates with short introductory periods andhigh spreads to the reference rate. Totake into account these and other factorsthat might influence default rates, weestimate a number of multivariateregression models that allow us toexamine the effect of varying one riskfactor while holding others fixed.The analysis sample includes loansthat do not default and are observed for12 months after origination and loansthat default (become 60 or more dayspast due) within 12 months of origination.We drop loans that get refinanced ortransferred within their first 12 months.While this may bias our results, keepingearly refinanced and transferred loans inthe sample would understate the shareof actual defaults, since by definitionthese loans are current for the durationof their (short) presence in the sample.Our goal is to evaluate the relativestrength of associations between loandefault and observable borrower, loan,and macroeconomic characteristics indifferent market segments and differentyears. To that end, we estimate thefollowing regressions:Prob (default within 12 months)i,j,k Φ(β1Loani,j,k, β2Borroweri,j,k, β3Econj,k, β4Dk),(1)where the dependent variable is anindicator of whether a loan to borrower i,originated in an MSA j in state kdefaulted within the first 12 months. Wemodel this probability as a function ofloan and borrower characteristics, MSAlevel economic variables (unemployment,home price appreciation, and income),and a set of state dummy variables (Dk)that capture additional aspects of theeconomic and regulatory environment.We estimate the model as a standardmaximum likelihood probit with statefixed effects.To retain maximum flexibility inevaluating the importance of covariatesfor prime and subprime defaults, we carryout separate estimations of equation (1)for prime and subprime loans. To achievesimilar flexibility over time, we furthersubdivide each of the prime andsubprime samples by year of origination(2004 through 2007). Finally, we attemptto account for unobserved heterogeneityat the state level by incorporating statefixed effects in our econometricspecification.The economic variables include boththe realized growth in the FHFA HPI andthe average realized unemployment rate.Both of these variables are measured atthe MSA level, and both are computedover the first 12 months since loanorigination. Consequently, they match theperiod over which we are tracking loanperformance. In contrast to all of theother regressors, this information wouldclearly not be available to the analyst atthe time of loan origination. We can thinkof the model described in equation 1 asthe sort of analysis one would be able todo at the end of 2005, after all loansoriginated in 2004 had gone throughtheir first 12 months, and one is able toobserve what happened to home pricesand unemployment rates over the sameperiod. The same exercise can beperformed for loans originated in 2005at the end of 2006, for loans originatedin 2006 at the end of 2007, and so on.ResultsThe results of the estimation aresummarized in table 2. The first fourcolumns of data depict estimates forprime loans originated in each of the foursample years, and the next four columnscontain the estimates for subprime loans.The juxtaposition of the data for the twomarket segments allows us to easilycompare the importance of certainfactors. The table presents estimates ofthe marginal effects of the explanatoryvariables, rather than the coefficientsthemselves. The marginal effects tell ushow a one unit change in eachexplanatory variable changes theprobability that a loan defaults in its first12 months, holding fixed the impact ofthe other explanatory variables.The defaults of both prime andsubprime loans are strongly associatedwith the FICO score, the LTV, and theinterest rate in every estimation year foreach loan type. For instance, an increaseof 100 points in the FICO score of primeborrowers in 2004 and 2005 isassociated with about a 1.2 percentagepoint decrease in default likelihood (theestimated marginal effect of –0.00012multiplied by 100). To gauge the strengthof this effect, note that in those years thebaseline rate of default was about 2.2percent. The point estimates of marginaleffects for 2006 and 2007 increaseabout twofold for prime loans, but sodoes the baseline sample default rate.For subprime loans, the estimatedmarginal effects are a full order ofmagnitude higher, implying that animprovement in FICO scores generates agreater decline in subprime defaults, atleast in absolute terms.Similarly, higher LTV values have astrong positive association with defaultsfor both loan types originated in 2005,2006, and 2007. For subprime loans, aProfitwise News and ViewsSeptember 20105

RESEARCH REVIEWTable 2: Probability of defaulting within 12 months of loan origination (probit regressions with state fixed effects)Prime 005(6)default in12default in12default in12default in12default in12default 700.1310.2180.743***-0.749**-0.476Estimation sample meanHPI growthUnemployment rateMedian income in zipcodeOrigination amountFICO scoreLoan to value ratio (LTV)Debt to income ratio (0 ifmissing)Missing DTI dummyInterest rate at originationARM w/ reset 3 yrsdummyARM w/reset 3 yrs dummyCash out refinancing dummyPurchase loan dummyInvestment property dummyConforming loan dummyGSE-securitized loan dummyPrivate label securitized loanObservationsR-squared2006(7)default in12 default 0.09260.0745SOURCE: FHFA for HPI growth, BLS for unemployment rate and median income, LPS for all other variables.NOTES: ***, **, and * denote statistical significance at the 1, 5, and 10 percent levels, respectively.62007(8)-0.00149Margin rate (0 if FRM)Prepayment penalty dummySubprime LoansProfitwise News and ViewsSeptember 2010

RESEARCH REVIEWTable 3.A: Average marginal effects from changes in key explanatory variablesPrime2004 - 2007VARIABLESMeanStd. deviationCHANGE ( )Baseline predicteddefault rateHPI growthFICO scoreLoan to value ratio(LTV)Debt to income ratio(if not missing)4.871076.137.710.36216.814.910 ppt50 points10 ppt10 ppt2004(1)2005(2)2006(3)2007(4)default in12default in12default in12default %Interest rate atoriginationMargin rate (0 18**-0.0004-3%4%4%-1%SOURCE: FHFA for HPI growth, BLS for unemployment rate and median income, LPS for all other variables.NOTES: ***, **, and * denote statistical significance at the 1, 5, and 10 percent levels, respectively.rise in LTV generates a strongerabsolute increase in loan defaults. Itmust be noted that the effect of theleverage on the likelihood of default maybe understated by the LTV measure thatwe have. A better measure of howleveraged a borrower is on a givenproperty would be the combined loanto-value ratio (CLTV) that also takes intoaccount second-lien loans on theproperty. This variable is not available inthe LPS data, however. If the practice ofobtaining such “piggyback loans” ismore prevalent in the subprime market,then the estimated coefficient for LTVfor subprime loans may be lower than itstrue value.At first glance, the interest rate atorigination is similar to LTV and FICOscore in having a strong statistical andeconomic effect on both prime andsubprime loan defaults in eachorigination year. What stands out is thesheer magnitude of the estimatedeffects. However, one must be cautiousin interpreting hypothetical marginaleffects of the interest rate. While LTVand FICO score cover fairly wide rangesfor both prime and subprime loans,interest rate values are relatively tightlydistributed.11 This means that adifference of even 1 percent in loaninterest rate makes it look quitedifferent from loans with otherwiseidentical characteristics (e.g., FICOscore, LTV, DTI). In such cases, a likelyexplanation is that the lender hasadditional information about the creditquality of the borrower and is charging ahigher interest rate to take into accountadditional risk factors – hence, thestrong positive association with eventualdefault rates.There are also a number of notabledifferences between the prime andsubprime samples. Perhaps the mostinteresting finding is the differentsensitivity of defaults to changes in homeprices. For subprime loans, defaults aremuch lower when home price growth ishigher for three out of the four sampleyears. This relationship is particularlystriking for 2006 loan originations, manyof which experienced home pricedeclines over their first 12 months. Forprime loans, 2006 is notable as the onlyyear of origination in which changes inhome prices ar

Nevertheless, among prime loans made in 2004, 12 percent were known to involve cash-outs. By 2005, this percentage had risen to about 21 percent, and it remained at this level through 2007. For subprime loans made ), , )