Juvenile Delinquent Mortgages: Bad Credit or Bad Economy?

Federal Reserve Bank of New York

Staff Reports

Andrew Haughwout

Richard Peach

Joseph Tracy

Staff Report no. 341

August 2008

This paper presents preliminary findings and is being distributed to economists

and other interested readers solely to stimulate discussion and elicit comments.

The views expressed in the paper are those of the authors and are not necessarily

reflective of views at the Federal Reserve Bank of New York or the Federal

Reserve System. Any errors or omissions are the responsibility of the authors.

Juvenile Delinquent Mortgages: Bad Credit or Bad Economy?

Andrew Haughwout, Richard Peach, and Joseph Tracy

Federal Reserve Bank of New York Staff Reports, no. 341

August 2008

JEL classification: G21, R21

Abstract

We study early default, defined as serious delinquency or foreclosure in the first year,

among nonprime mortgages from the 2001 to 2007 vintages. After documenting a

dramatic rise in such defaults and discussing their correlates, we examine two primary

explanations: changes in underwriting standards that took place over this period and

changes in the economic environment. We find that while credit standards were important

in determining the probability of an early default, changes in the economy after 2004—

especially a sharp reversal in house price appreciation—were the more critical factor in

the increase in default rates. A notable additional result is that despite our rich set of

covariates, much of the increase remains unexplained, even in retrospect. Thus, the fact

that the credit markets seemed surprised by the rate of early defaults in the 2006 and 2007

nonprime vintages becomes more understandable.

Key words: housing, mortgage default, subprime mortgages

Haughwout: Federal Reserve Bank of New York. Peach: Federal Reserve Bank of New York.

Tracy: Federal Reserve Bank of New York. Address correspondence to Andrew Haughwout

(e-mail: andrew.haughwout@ny.frb.org). The authors wish to thank Jonathan Stewart and

especially Ebiere Okah for excellent assistance with the data. The views expressed in this paper

are those of the authors and do not necessarily reflect the position of the Federal Reserve Bank

of New York or the Federal Reserve System.

Gee, Officer Krupke, we're very upset;

We never had the love that ev'ry child oughta get.

We ain't no delinquents,

We're misunderstood.

Deep down inside us there is good!

“Gee, Officer Krupke” – West Side Story

Rapid increases in US residential mortgage defaults during 2007 and into 2008 captured

the attention of researchers, the public and policy makers, and had a chilling effect on credit

markets worldwide. While these increases were noted originally in the nonprime market,

foreclosure increases have in more recent months begun to spill over into the prime market. This

paper studies a part of this phenomenon, early defaults in the nonprime market.

Historically, four key characteristics (“risk factors” or “underwriting criteria”) have been

thought to determine the probability that a mortgagor will default. Those factors are the loan-to-

value ratio (LTV)

, the debt service-to-income ratio (DTI), the mortgagor’s credit score, and the

extent to which the mortgagor’s income and assets have been verified by third party sources such

as employers, tax returns, and bank account statements. To expand the potential pool of

borrowers, nonprime (subprime and alt-a) mortgages by design relaxed one or more of these

underwriting criteria beyond the margins required for prime mortgage loans. A direct

consequence is that we would expect the default experience of these relatively new mortgage

products to be worse than that of prime loans. Indeed, industry data confirm that the performance

of the very first vintages of nonprime loans was significantly worse than that of prime loans.

Nonetheless, as shown in Figure 1, beginning with the 2005 vintage the performance of

nonprime mortgage loans became notably worse than previous vintages. The performance of the

2006 vintage deteriorated even further. By 12 months following origination, the 2005 vintage

The LTV is calculated by taking the ratio of the mortgage balance to the value of the home. LTVs are

typically expressed as a number ranging from 0 to 100 or higher. If the borrower has “negative equity”

where the mortgage balance exceeds the value of the home, the LTV will exceed 100.

The National Delinquency Survey published by the Mortgage Bankers Association of America (MBA)

is one of the main sources of information on mortgage loan performance, including nonprime loans.

However, it should be noted that mortgages are placed into these categories based on the servicer rather

than the individual loan. Thus, if more than 50 percent of a servicer’s portfolio is nonprime loans, then all

of that firm’s loans are lumped into the subprime category. Alt-a mortgages, according to the MBA, are

divided between the prime and subprime groups. See

http://www.mortgagebankers.org/files/Research/NDSFactSheet.pdf

for details.

had a 90 day or more delinquency rate that was not reached by the 2003 vintage for 20 months,

and the 2006 vintage at 12 months had a rate that was not reached by the 2003 vintage even by

30 months. Moreover, this sharp decline in loan performance was a surprise to investors in these

loans in that to a large extent it seemed unexplained by the observed risk characteristics.

The sharp increase in defaults very early in the life of the loans suggests the moniker

“juvenile delinquents.” In the case of nonprime adjustable rate mortgages (ARMs), defaults often

occurred well before the first interest rate reset while the initial “teaser” rate was still in effect.

We define an “early default” as a mortgage that is 90 or more days delinquent within the first

year after origination. We use this window since performance warranties by originators often

covered the first year. The reasoning was that any serious underwriting problems with mortgages

typically would manifest themselves within the first year. In our data, 10 percent of nonprime

loans originated in 2007 experienced an early default, as compared to 2.7 percent of similar loans

originated in 2003.

The purpose of this paper is to explore potential explanations for the sharp rise in early

defaults of the 2005 through 2007 vintages of nonprime mortgages. We will examine how much

of the deterioration in the early performance of these mortgages can be explained by changing

risk characteristics of nonprime mortgages over time (i.e. “bad credit”). New and existing home

sales peaked in late 2005 in many housing markets, and house prices began to soften and then to

decline as these housing markets cooled. We will also explore the extent to which house price

dynamics over the housing cycle as well as other local economic factors help to explain the early

default behavior of the more recent vintages of nonprime mortgages (i.e. “bad economy”).

Importantly, we will investigate the extent to which the effect of house price dynamics on early

defaults depends on the risk profile of mortgages in a vintage – that is, are there important

interaction effects that help to determine a vintage’s share of juvenile delinquent mortgages in

that vintage.

The next section provides a brief literature review of selective papers that are relevant to

our analysis. We next describe our primary data source and discuss the evolution of the four

basic risk factors for nonprime mortgages from 2001 to 2007. We provide tabulations of these

risk characteristics and early default rates. We then turn to a multivariate analysis of early

defaults. The final section draws insights from our analysis for the current housing policy debate

and concludes.

Review of Past Literature

Residential mortgages are complex financial instruments that confer important options on

the borrower. The extensive body of previous research on residential mortgage default has

adapted option theory to the study of mortgage valuation, since there exist well-developed theory

and empirical methods for valuing financial derivatives and their exercise (Black and Scholes

1973).

An important feature of most residential mortgages is that they are “non-recourse” loans,

either de jure or de facto. This means that in the event of a default, creditors can sell the house to

cover the loan balance, but typically do not legally pursue the borrower for any deficiency.

This

creates a “put” option for the borrower which he/she can exercise if the house value falls

sufficiently relative to the loan balance. In addition to this default option, borrowers may

continue to make the scheduled payments until the mortgage debt is discharged, or prepay the

mortgage either by selling the house and paying off the balance on the mortgage or by

refinancing into a new loan (Kau, Keenan, Mueller and Epperson, 1995). The option to prepay is

often referred to as the “call” option that borrowers hold when they take out a mortgage.

Foote Gerardi and Willen (2008) succinctly summarize the prediction of option theory for

default when they argue that negative equity is a necessary but not a sufficient condition for

default. Borrowers with positive equity ought to rarely if ever default, since (in the event of an

idiosyncratic shock such as illness, loss of job or divorce) they can sell the house or refinance the

mortgage. Borrowers with negative equity, on the other hand, may default in the face of similar

shocks, since the option to refinance and/or sell the house is conditional on being able to raise

cash to cover the difference between the mortgage balance and the proceeds of a sale or a new

mortgage (Foster and Van Order, 1984, Vandell, 1995).

Even borrowers with negative equity, however, default less frequently than simple

models would predict (see Vandell 1995 for a summary of the empirical evidence and Elul 2006

While legal pursuit of borrowers’ assets to cover deficiencies is available in most states, it is often

restricted. In California, for example, deficiency judgments are not allowed for owner-occupied one to

four-family homes (Pence 2003).

for an update). For an owner occupant considering default, transactions costs include moving

costs, the cost of purchasing or renting a new residence, and damage to one’s credit score

resulting in higher future borrowing costs. All told, some authors have argued that these costs

can typically range from 15 to 30% of the value of the house, helping to explain why default

appears to be underexercised relative to the simple option-theoretic prediction (Cunningham and

Hendershott, 1984). Investors face fewer of these transaction costs and therefore may be more

likely to default for a given LTV level.

As noted by many authors, exercise of the default option will be related to the value of

the prepayment option, regardless of the borrower’s equity in the property (Schwartz and Torous

1993, Vandell 1993, Elul 2006). This suggests that in evaluating the prevalence of default, we

must account for the value of the option to sell the house or refinance the mortgage. The typical

approach to this problem is to estimate a duration model that simultaneously accounts for the

competing risks of prepayment and default (Deng, Quigley and Van Order 2000). An additional

advantage of this approach is that it allows insight into the value of a particular mortgage, or

mortgage backed security. While our ambitions in the current study are more modest, we must

remain attentive in both the specification of our models and the interpretation of our results to the

fact that even early defaults may be affected by the availability of prepayment (Deng and Gabriel

2006).

In addition, Kau, Keenan and Kim (1993) argue that the apparent underexercise of the

default option is partly explained by the fact that continuing to make payments preserves the

borrower’s ability to default or sell the house in the future. That is, exercise of the default option

depends on the future value of both living in the house and selling the house, either in the

marketplace or through default (effectively selling the house to the lender).

Much of the empirical research conducted on mortgage default over the last two decades

has focused on the large market for prime/conforming or FHA mortgages. Most relevant for our

study is recent work on the FHA market segment, which serves borrowers who are similar to the

nonprime sample we study.

Besides not incurring any moving costs, investors may also not face the same increase in future

borrowing costs in the event of a default. If the housing investment is held in a limited liability

corporation, then a default would not affect the owner’s personal credit rating. In this case, we are more

confident that the borrower would be identified as an “investor” in the LoanPerformance data.

Deng and Gabriel (2006) and An, Bostic, Deng and Gabriel (2007) present competing

risk estimates for a sample of FHA purchase loans originated between 1992 and 1996. In these

studies, lower FICO scores are associated with higher default rates and somewhat lower

prepayment rates. Higher local unemployment rates have little effect on default rates. However,

higher local unemployment is associated with lower prepayments among FHA borrowers.

Finally, higher LTVs – or measures of the probability that the put option is “in the money” –

raise both the default risk, particularly for borrowers with low FICO scores.

Very recently, studies of subprime mortgages have become more common, as their

market share has expanded. Pennington-Cross and Ho (2006) provide a detailed analysis of the

performance of subprime mortgages over the period leading up to the crash in the housing

market. Their analysis provides insights into the behavior of these mortgages in a general

environment of rising house prices. They use LoanPerformance data on subprime mortgages that

were originated between 1998 and 2005. Loans were followed for up to five years or the end of

2005. Fixed rate mortgages are contrasted to the hybrid 2/28 adjustable mortgage. Like the two

FHA studies, Pennington-Cross and Ho find that borrowers with lower credit scores are more

likely to default, and that local unemployment rates seem to have little effect on default. In an

interesting contrast with the FHA results, Pennington-Cross and Ho find that among subprime

borrowers, higher LTVs raise the default risk but lower the prepayment risk.

Gerardi, Shapiro and Willen (2007) take a non-traditional approach to analyzing

mortgage default. Using Massachusetts deed records from January 1987 to August 2007, they

compile a panel data set tracking purchases, refinances, sales and foreclosures on all residential

properties in the state. The strength of this data is the ability to follow the same property though

different owners, as well as across different mortgages for a given owner. This is in contrast to

the typical loan level data which tracks a given mortgage over its life, but does not permit linking

mortgages over time for the same borrower or the same property.

A weakness of the deed based data is that it lacks information on the characteristics of the

borrower and some characteristics of the mortgage. The deed records indicate the mortgage

originator. The authors identify subprime mortgages by matching the originator to a Department

of Housing and Urban Development list of subprime lenders, and estimate a competing risk

model of the outcomes of distinct “ownership experiences.” Starting with a home purchase, they

As noted earlier, they control for average household income for the census tract.

follow the household until the home is sold or goes into foreclosure. A key finding by Gerardi et

al (2007) is the important role of house price appreciation on the likelihood of a foreclosure.

Cumulative price appreciation since the date of the house purchase exerts a sizeable downward

effect on the probability that the ownership experience ends in a foreclosure. The authors impose

symmetry in the effect of price appreciation and price depreciation. It would be useful to know if

nominal losses are relatively more important at generating defaults than nominal gains are at

preventing defaults. The authors do not include a negative equity indicator for the current

mortgage due to worries over the endogeneity of this indicator variable.

Gerardi et al (2007) report that ownership experiences that begin with a subprime

mortgage are much more likely to end in a default than observably similar ownership

experiences that begin with a prime mortgage. Here it is important to keep in mind that the

authors are not able to control for some key characteristics of the borrower such as income and

credit score.

It is not clear, then, how much of the difference in default rates across ownership

experiences could be explained by differences in these borrower specific risk factors. Foote,

Gerardi and Willen (2008) use the same data set to examine the specific role of negative equity

in default behavior. The authors confirm many of the results in Gerardi et al (2007), including

the higher likelihood of default among subprime borrowers, and find additional evidence that

borrowers in a negative equity position – whether the mortgage is prime or not – are more likely

to default. Our focus on the nonprime market segment will allow us to determine the effect of

negative equity on borrowers in particularly high cost mortgages.

Demyanyk and van Hemert (2008) adopt an approach similar to ours in their analysis of

subprime mortgages. Using LoanPerformance data, which we discuss in more detail below, they

examine the likelihood that a mortgage is either 60 or more days delinquent or in foreclosure

within the first twelve months following origination. Their principal aim is to see to what extent

changes in the distribution of risk factors can explain the deterioration in the early performance

of subprime mortgages in 2005 and 2006. We will contrast our findings to theirs in greater detail

later in the paper. The authors conclude that declining underwriting standards, particularly

reflected in increasing LTVs at origination, are the dominant force explaining the rapid rise in

early delinquency and defaults among subprime borrowers.

One important difference between our approach and that of Demyanyk and van Hemert

involves the treatment of house price appreciation. While we follow previous literature by

controlling for an updated LTV using current house price information, Demyanyk and van

Hemert control for the initial LTV and treat house price appreciation as an independent effect on

the default probability. This approach misses what we believe is an important interaction

between house price dynamics and origination LTV, and complicates the evaluation of the

borrower’s put option. We discuss the separate effects of updated LTV and house price

appreciation below. In addition, Demyanyk and van Hemert impose that house price increases

and decreases have a symmetric impact on defaults. We test for asymmetric effects.

A second difference in approaches is that Demyanyk and van Hemert include the

mortgage rate as a control variable in their empirical specifications. The mortgage rate has the

largest standardized marginal effect in their specifications. The coefficient on the mortgage rate

reflects the impact on early performance from variation in mortgage rates that is orthogonal to

the other risk factors in their specification. One possibility is that this residual rate variation

reflects risk factors observed in the underwriting but not fully reflected in the data by the

recorded risk measures. Alternatively, this variation could reflect the degree of competition in

the local lending markets, differences in bargaining skill across different borrowers, or lagged

performance of mortgages in the local lending market. It is not clear, then, what specifically the

mortgage rate is capturing in their analysis. We do not include the loan-specific mortgage rate in

our empirical specifications, since our aim is to evaluate the extent to which standard observable

risk factors can account for the rise in early defaults.

Nonprime Mortgage Data and Tabulations on Early Defaults

Loan Performance Data

Our mortgage data come from FirstAmerican CoreLogic’s Loan Performance Data, a

proprietary data base which, as of June 2008, provides loan-level information at a monthly

frequency on approximately seven million active, securitized subprime and alt-a loans, carrying

balances of over $1.6 trillion.

Subprime mortgages are small loans (compared to alt-a loans)

and are often made to borrowers with some blemish on their credit history, or who are willing to

commit large shares of their incomes to debt service. Alt-a mortgages are typically larger value

See http://www.loanperformance.com/data-power/default.aspx

loans made to more credit-worthy borrowers who, for a variety of reasons, may choose not to

provide the income or asset verification required to obtain a prime mortgage. Both types of

nonprime mortgages are typically higher-cost than prime conforming loans.

The database consists of information on securitized subprime and alt-A mortgage loans.

A large share of outstanding subprime and alt-A mortgages are securitized, with the balance

remaining in lender portfolios, and LoanPerformance captures upwards of 90% of the securitized

loans (Mayer and Pence 2008). Pennington-Cross (2002) argues that securitized subprime

mortgages differ systematically from those retained in portfolio. Since our data are limited to

securitized loans, any inferences should be limited to this set of loans.

The LoanPerformance dataset is a rich source of information on the characteristics of

these securitized loans. The dataset includes information on the date of origination, the zip code

in which the collateral property is located, details of the mortgage contract (including term,

initial interest rate, and rate adjustment schedule), and underwriting information (including

borrower credit score, debt to income ratio, the loan to value ratio for senior and junior liens, and

the extent of income and asset verification provided by the borrower). Also included are monthly

updates of “dynamic” information including the current interest rate, mortgage balance and the

borrower’s payment record.

We analyze a one percent random sample of the first-lien subprime and Alt-A loans

reported in the data as of our most recent monthly update, for April 1, 2008.

The universe from

which we sample includes all loans, whether they are currently active or have been paid off.

From this set of loans, we select those that originated after January 1, 2001 and for which we

observe at least twelve months of performance; thus, our “youngest” loans originated during

April 2007. For each origination, we use the payment history for the first twelve months in order

to determine whether it defaulted during that time period. The result is a dataset consisting of

about 117,000 loans. We classify these loans as subprime or Alt-A based on the designation of

the security in which they were packaged. Approximately 2,000 of our loans are missing this

designation in the data, bring our total for analysis to around 115,000 loans.We combine this

loan-level information with economic data measured at the metropolitan area level. These data

Since observations in the LoanPerformance dataset are loans coded to the zip code, we choose our

panel based on first-lien loans only. This avoids the possibility of double counting subordinate lien loans

on the same property.

include measures of house price appreciation (the widely-used OFHEO repeat-sale index

) and

labor market conditions from the Bureau of Labor Statistics. The principal characteristics of our

data are described in Table A-1.

Tabulations on Early Default

The value of the default put option discussed earlier will likely depend on the initial LTV

on a mortgage and the pattern of house price appreciation since the origination of the loan. Table

1 provides information on the initial LTV for subprime and alt-a mortgages originated from 2001

to 2007. For each type of mortgage, we provide descriptive statistics on the distribution of the

cumulative LTV at origination of the first-lien mortgage. The cumulative LTV reflects the total

mortgage balance of the first-lien mortgage and any subordinate lien loans if they exist at

origination. We also report the fraction of transactions that involved one or more 2

-lien loans.

What is striking for both classes of mortgages is the significant rise in the incidence of 2

-liens

from 2001 to 2006. For subprime mortgages over this period the incidence of 2

-liens rises from

3.2 percent to 29.4 percent, and for alt-a mortgages the incidence rises from 2.2 percent to 43.9

percent. A consequence is that 30.1 percent of subprime mortgages originated in 2006 had an

initial cumulative LTV of at least 100 – that is, the borrower at origination did not have any

equity in the house. Excluding the 2

-lien loans, for subprime mortgages in 2006 the median

LTV was 80, while the the 90

percentile LTV was 95.

While from 2004 to 2007 the incidence

of 2

-liens was higher for alt-a mortgages, the distribution of initial LTV was not as

significantly affected. It is also clear that the number of nonprime mortgages in our data rises

sharply over most of the time period, reflecting both increases in originations and more complete

coverage by LoanPerformance (Mayer and Pence 2008).

Over the period from 2003 to 2007, the incidence of early defaults more than quadrupled

for both subprime and alt-a mortgages. Comparing subprime to alt-a mortgages in Table 2, for

any given initial range of LTV the early default rate for subprime mortgages tended to be higher

than for alt-a mortgages. As the average early default rates were rising, the data show for both

classes of mortgages increases in the early default rates within each LTV range. For alt-a

See http://www.ofheo.gov/hpi.aspx for details.

For alt-a mortgages in 2006 based just on the first lien mortgage the median through 90

percentile

LTV was 80.

mortgages originated from 2003 to 2007, the data also show a rise in the average early default

rate as one moves from lower to higher initial LTV intervals. This pattern is less consistent for

subprime mortgages over this same time period.

The housing boom in the first half of this decade was concentrated in a small number of

states. Figure 2 shows a scatter plot of the average annual house price appreciation rate from the

third quarter of 2001 to third quarter of 2006 on the vertical axis and appreciation rate from the

first quarter of 2007 to the first quarter of 2008 on the horizontal axis. Nine states experienced

double digit house price appreciation sustained over a five year period that were followed by

house price declines in the year following. In four of these states (AZ, CA, FL, and NV) the

reversal has been especially sharp.

Three states (IN, MI and OH) presents a different picture

where instead of experiencing a housing boom and bust, these states have been suffering from

relative economic weakness and soft housing markets during this entire decade.

The rapid house price increases in the boom/bust states prior to the downturn would act

to keep the put option for default out-of-the-money. Even where the lender finances most or all

of the borrower’s down payment with a 2

lien loan, twelve months of double-digit house price

appreciation will generate more than sufficient equity to cover the transactions costs of selling

the house. Similarly, in cases where a borrower in a boom/bust state suffers a job loss, divorce or

significant health problem during the boom period, we would not expect to see this result in a

default. The borrower would have a financial incentive to sell the house and prepay the mortgage

rather than default. Finally, as discussed earlier, owners may be less likely to exercise the default

put option than investors other things equal.

We take a preliminary look in Table 3 at the likely interplay between house price

dynamics, local economic conditions, investor status and initial LTV. For owners in

economically depressed states, the incidence of early defaults tends to be higher than the overall

average for both classes of mortgages. In contrast, for subprime owners in those states that

experienced a house price boom, the incidence of early defaults is low relative to owners in the

“other” states except for the highest LTV interval. This pattern does not emerge for alt-a owners

where the data indicate that the default rate in the boom/bust states was generally higher than for

Both the OFHEO and the other widely-used house price index, S&P/Case-Shiller (CS), are derived

using the repeat sales methodology. However, there are important differenced in the construction of these

indices which result in quite different assessments of the behavior of home prices, particularly at the

national level. See Leventis (2008) for details.

alt-a owners in the “other” states. Consistent with investors having a lower threshold for

exercising the default put option, the overall early default rate is higher for investors than for

owners for initial LTVs greater than 90. Comparing investors across the boom/bust states and the

economically weak states, nonprime investors have higher early default rates in the economically

weak states, with the exception of alt-a investors who start out with no equity in the deal.

The second risk factor is the DTI ratio. This ratio is meant to capture a borrower’s

capacity to pay even in the face of transitory shocks to his/her personal finances.

Table 4 gives

the evolution of the DTI ratios for nonprime mortgages over our sample period. The first thing to

note is that subprime borrowers relative to alt-a borrowers have a distribution of DTI that is less

concentrated in the DTI range below 30 and more concentrated in the DTI range above 40. That

said, the distributions of DTIs were reasonably constant over the time period with the exception

of 2006/2007 where there was a noticeable increase in high DTI subprime and alt-a mortgages.

The early default performance by different initial DTI intervals and year is presented in Table 5.

Subprime mortgages display a stronger relationship than alt-a mortgages between DTI and early

defaults across the years in our sample. Similar to our finding for initial LTV, from 2003

onwards the incidence of early defaults rises over time within each DTI range for both types of

mortgages, but in this case the change is relatively common across the DTI intervals.

The “willingness” to pay is captured by the borrower’s FICO score. Table 6 provides

information on the distribution of FICO scores over time for the nonprime mortgages in our

sample. Similar to the DTI risk measure, subprime borrowers are more concentrated than alt-a

borrowers in the below 600 interval of FICO scores and much less concentrated in the above 660

inverval. The likelihood that a subprime borrower had a FICO score below 600 was decreasing

from 2001 to 2007. While few alt-a borrowers had FICO scores below 619, the incidence was

declining as well over this time period. In contrast to the initial LTV and DTI risk factors, the

distributions of nonprime FICO scores were not deteriorating over the period leading up to the

sharp rise in early defaults.

Table 7 shows early default rates across time for the different FICO ranges. In each year

early defaults typically are a declining function of FICO scores. Generally, borrowers with a

FICO score of less than 600 are at least three times more likely to experience an early default as

A possibility is that DTI is less predictive for early defaults than it is for overall defaults, as the

likelihood of a significant financial shock is less over the first year of a mortgage as compared to the

expected life of the mortgage.

borrowers with a FICO score of over 660. The exception is for subprime borrowers in 2006 and

2007 where the early default rates tended to converge across the range of FICO scores. The data

do not display this same convergence in early defaults for alt-a mortgages. For subprime

borrowers, the relative increase in early defaults since 2003 within a range of FICO scores was

larger for the two higher FICO score ranges than for the two lower ranges. The relative increase

in early defaults within a FICO interval for alt-a borrowers is more varied across the different

ranges of credit scores.

The final standard risk factor is the level of documentation used in the underwriting of

the mortgage. The data classify underwriting into one of three categories: full documentation,

low documentation (“limited-doc”) and no documentation (“no-doc”). Table 8 provides the

distribution of the documentation level over time for subprime and alt-a mortgages. The use of

lower documentation was much more prevalent for underwriting alt-a mortgages than for

subprime mortgages. Despite the focus in the press made on no-doc mortgages, in each year the

incidence of no-doc mortgages was in single digits, and was declining over the sample period.

What is more notable is the shift in composition from fully documented to limited documented

underwriting. From 2001 to 2006, the share of fully documented subprime mortgages fell from

77.8 percent to 61.7 percent, while the share of fully documented alt-a mortgages fell from 36.8

percent to 18.9 percent.

Table 9 gives the yearly average early default rates for subprime and alt-a mortgages

broken down by the level of documentation. In each year for subprime mortgages, early defaults

are more prevalent for limited as compared to fully-documented mortgages. For alt-a mortgages,

the incidence of early defaults in each year generally increases as one moves from fully-

documented to limited doc mortgages, and from limited doc to no-doc mortgages. From 2005 to

2007 as the overall incidence of early defaults among alt-a mortgages was rising, the incremental

effect on early defaults associated with moving from fully documented mortgages to limited

documented mortgages was higher than the incremental effect associated with moving from

limited documented mortgages to mortgages with no documentation.

Determinants of Early Default

Econometric specification

For each mortgage in our data, the outcome of interest is an indicator for whether the

mortgage experienced an early default. We adopt a linear index model in order to examine the

determinants of early default. For each mortgage i originated in year and month

t in

metropolitan area j, we assume that there is a latent index

ijt

that captures the net benefit to the

borrower from an early default. We specify this latent index as a linear index of the observable

risk factors (X) and local economic conditions (Z).

ijt it jt jt ijt

IX Z

δα ε

+++

Given the paucity of data that we have to control for local economic conditions, we

assume that the performance of mortgages originated in a given metropolitan area and

year/quarter may be affected by a common random error component

that captures any

unobserved economic shocks that impacted this local housing market. We account for the

possible presence of the error component

in calculating the standard errors of our estimates.

Let

ijt

denote our observed indicator variable for whether a mortgage experiences an

early default.

We assume that this occurs whenever the unobserved latent index takes on a

positive value.

1 if 0

0 otherwise

ijt

≥

The probability of an early default is given by the following.

Pr( 1| , ) Pr( 0| , )

ijt it jt ijt it jt

XZ I XZ==>

It is important to distinguish this early default indicator from a default hazard in a competing risk

specification. If a mortgage prepays in the first year, this would censor the default hazard at the

prepayment date. In contrast, in our setup an early prepayment is treated in the same manner as a

mortgage that is still ongoing without an early default after 12 months. Both outcomes would generate a

value of zero for the early default indicator.

We report linear probability estimates which facilitate calculating the decomposition of the

change in aggregate defaults into credit and economy effects.

The first risk factor is the LTV. We start with the initial combined LTV at origination. If

there is a 2

-lien loan, we include it in the calculation of this initial LTV. To update the LTV we

take into account any paydown of the principal from the 1

lien mortgage over the next year. We

then use the metro area OFEHO repeat-sale house price index to update the house value and

LTV for each of the next four quarters following the origination. We take a simple average of

these updated LTVs. To allow for potential nonlinear effects of LTV on early default, we enter

the updated LTV as a series of indicators for different intervals. The left-out interval covers all

LTVs below 80. We include indicators for the following LTV intervals: 80 to 84, 85 to 89, 90 to

94, 95 to 99, and 100 or higher (100+). To test whether investors react differently to the put

option for default, we interact the investor indicator with the two highest LTV indicators, 95 to

99 and 100+.

The next risk factor is the DTI. LoanPerformance includes the “back-end ratio” which is

calculated as the ratio of the sum of the annual mortgage principle and interest, property taxes

and insurance, and other debt (such as car loans and student loans) to the income of the

borrower. This variable is missing for 37% of the data. Appendix table A2 provides the results

from estimating a probit on an indicator for whether DTI is missing on the other explanatory

variables we use in this study. No-doc loans are much more likely to have a missing DTI, which

suggests that lenders were less likely to code the variable if it is a “stated” item. There is a

pattern to the missing DTI suggesting that as other risk factors of a loan deteriorate (ie higher

LTV and/or lower FICO), the DTI is more likely to be reported in the data. We treat the missing

DTI in two different manners to check for robustness. First, we include an indicator for whether

DTI is missing and the DTI for those mortgages where it is reported. Second, we regress the

reported DTI values on the other explanatory variables and replace the missing DTI values with

the predicted DTI value from this regression. We also allow for nonlinearity in the effect of DTI

on early defaults by entering DTI as a series of indicators for the following intervals: 40 – 44, 45

– 49, 50+. The left-out interval is for DTI less than 40.

We tested for whether creating subintervals of DTI below 40 would improve the fit. The data supported

collapsing these into a single interval.

Next we turn to the credit or FICO score for the borrower. Freddie Mac uses three broad

intervals for the FICO score for classifying mortgages: under 620, 620 to 680, and over 680.

Their analysis suggests that only movements between these intervals are materially important in

predicting performance of mortgages. We use a finer set of intervals that allow us to test whether

variation in FICO scores within those broad intervals are informative for predicting early

defaults for nonprime mortgages. We create a series of indicator variables for the following

FICO intervals: less 560, 560 to 589, 590 to 619, 620 to 649, 650 to 679, and 680 to 719. The

left-out interval is for FICO scores of 720 or higher.

The last of the four primary risk factors discussed in the introduction is the level of

documentation used in underwriting the mortgage. We code an indicator for a limited-

documentation mortgage and an indicator for a mortgage with no documentation. It is important

to keep in mind when interpreting the coefficient estimates on these indicators for documentation

level that the estimation procedure treats the stated risk factors for these lower documentation

mortgages as being equivalent to the verified risk factors for fully documented mortgages.

We include a few additional variables related to the mortgage in our credit (X) vector.

While the interest rates for all mortgages are fixed during the first twelve months that we follow

their performance, we include an indicator for a fixed-rate mortgage. We include two indicators

for mortgages that result from a refinance as opposed to a purchase – one for when no cash is

taken out, and a second for when cash is taken out. As noted earlier, we include an indicator for

an investor property as well as an indicator for a 2

home.

All of our specifications include six

indicators for different property types for the underlying collateral: condo, two-four unit,

townhouse, planned-urban-development, manufactured housing, and other.

Turn now to the variables we use to capture changes in the local economic conditions (Z)

that may affect early defaults. For each mortgage, we take the average of the metro area

unemployment rate over the twelve months following the origination. This variable should pick

up income stresses that arise through job loss. We noted earlier that we control for the updated

LTV based on changes in metropolitan house prices over the year following origination. To see

It is likely that the investor variable is underreported in the data. It is reasonable to assume that

misreporting is one-sided. That is, some investors hide this fact and are coded as owner-occupied; but

owner-occupied borrowers are unlikely to misreport themselves as an investor. Appendix table A2 reports

the results from running a probit of the investor indicator on our other control variables.

Single family houses are the left-out property type.

if there are additional effects of changes in house prices beyond how they affect the LTV, we

include the change in metropolitan house prices over the year following the origination of the

mortgage, and we allow house price increases and decreases to have asymmetric effects on early

defaults. In addition, we interact this house price appreciation with the investor indicator to test if

investors react to house price movements in a different manner from owner-occupied borrowers.

We also include year effects in all of our specifications.

Our specification is a reduced-form representation of the competing risks to which loans

are exposed in their first year of life. To control for the risk of prepayment, we experimented

with variables measuring changes in the interest rate environment borrowers faced when

considering refinancing their loans. Specifically, we calculate for each month subsequent to

mortgage origination the ratio of prime 30-year (and nonprime) mortgage rates to those

prevailing at origination. We take the minimum of this ratio as a measure of the opportunity to

refinance into a lower cost mortgage. In addition, we control for the presence of a prepayment

penalty on the mortgage, which affects the probability of default by reducing the attractiveness

of prepayment. Only the prepayment penalty was significantly related to early default behavior.

In interpreting our results, it is important to bear in mind that the effects of house price declines

on early default subsume both direct effects – determining the value of the put option - and

indirect effects - reducing the availability of the prepayment option and thus extending borrower

exposure to the risk of default (Caplin, Freeman and Tracy 1997). We return to these issues

below.

Empirical Results

The linear probability model results are given in Table 10. The first specification is

estimated with our sample of subprime mortgages and the second specification is estimated with

our sample of alt-a mortgages. Descriptive statistics for each of these samples are provided in

appendix table A1. The standard errors for both specifications use clustering on mortgages

originated in the same metropolitan area, year and quarter.

The current LTV exerts a strong influence on early defaults for both classes of

mortgages. The LTV marginal effects are relative to mortgages with LTVs below 80. The data

indicates that as the LTV increases, the likelihood of an early default rises by a similar amount

for subprime and alt-a mortgages, with the incremental effect magnified for those mortgages that

have negative equity. Compared to subprime borrowers with a current LTV below 80, borrowers

with negative equity have an 6.8 percentage point higher early default rate (relative to an average

early default rate of 6.6 percent). For alt-a borrowers, those with negative equity have a 6.9

percentage point higher early default rate compared to borrowers with a current LTV below 80

(relative to an average early default rate of 2.1 percent). While the absolute size of the negative

equity effect is similar across subprime and alt-a loans, the relative impact is considerably larger

for alt-a loans.

As discussed in the literature review, there may be differences between how “ruthless”

investors are versus owners in exercising the default option on the mortgage. To test this

hypothesis, we interact the indicator variables for the two highest LTV intervals with the investor

indicator variable. For borrowers with negative equity, the data indicate that investors appear to

be much more likely than owners to default. The point estimate for the incremental effect on the

default rate is over 24.6 percentage points for subprime investors and 20.3 percentage points for

alt-a investors.

The second risk factor is the DTI which measures the ability of the borrower to make the

mortgage payments. For each sample of mortgages, the coefficient on the linear DTI effect was

similar across the two methods for handling the missing value for DTI. We report the

specification in Table 10 that includes an indicator variable for a missing DTI. The data indicated

that there is little effect of changes in DTI below 40 on early defaults. Increases in DTI above 40

have a small effect on increasing the likelihood of early default for subprime loans. Subprime

borrowers who appear to be financially stretched as indicated by a DTI above 50, have a rate of

early defaults that is 1.3 percentage points higher than borrowers who start out with a DTI below

40.

For alt-a loans, variations in DTI have a small and inconsistent pattern of effects on early

default rates.

If the investor variable is mismeasured along the lines discussed earlier, then these estimated

differential effects for negative equity on early defaults are likely to be conservative.

We also check to see if there is any interaction between high DTI and high LTV. We interact the

indicator for DTI exceeding 50 with the indicators for LTV of 95-99 and LTV of 100+. Neither

interaction was significant.

The third risk factor is the FICO credit score which is meant to capture a borrower’s

“willingness to pay.” We tested whether the three broad intervals used by Freddie are appropriate

for nonprime mortgages in predicting early defaults. The data strongly rejected the hypothesis

that variation of the FICO score within the Freddie intervals was not predictive for early defaults.

FICO scores exert a strong most influence on the early default behavior for owners. The reported

marginal effects for each interval are relative to borrowers who have a FICO score of 720 or

higher. The data indicates that as the FICO score declines below 680 for subprime loans the

likelihood of early default rises. Subprime borrowers with a FICO score of 560 to 589 have a

early default rate that is 6.9 percentage points higher than borrowers with a FICO above 720.

This differential early default rate increases to 10.3 percentage points as we move to FICO scores

below 560.

The data indicate a similar pattern of marginal effects of FICO scores on early

default rates for alt-a owner-occupied loans. However, for alt-a investor loans there is no

significant impact of variation of FICO scores on likelihood of a loan experiencing an early

default.

The final basic risk factor is the degree of documentation carried out in underwriting the

mortgage. Controlling for the “stated” risk factors in these mortgages, the data indicate that low-

doc underwriting is associated with a higher early default rate of around 3 percentage points for

subprime loans and 1.3 percentage points for alt-a loans. Both of these effects exceed the simple

weighted average of the early default differences listed in table 9. This is consistent with a

degree of systematic bias in the statement of the risk profile of these mortgages relative to the

risk profiles for fully documented mortgages.

The next set of variables reported in table 10 refers to characteristics of the mortgage and

whether the mortgage was for a purchase or a refinance. Controlling for the observed risk

factors, borrowers who select a fixed rate mortgage have from a 0.6 to a 1.2 percentage point

We explore whether the poor performance for low FICO borrowers is affected by the prevailing house

price appreciation in the local housing market. We interact the house price appreciation with the three

indicators for FICO scores below 620. The interaction is only statistically significant for the lowest FICO

score interval, but the magnitude was small relative to the direct effect of this FICO score interval.

We also tested for differences between owner-occupied and investor subprime loans on the impact of

FICO scores and did not find them to be statistically significant.

An obvious risk factor that may be biased downward is the DTI. Note, though, in appendix table A2

that DTI is often missing for low and no documentation mortgages.

lower incidence of early default.

We control for whether the mortgage has a prepayment

penalty. The prepayment penalty applies if the mortgage is paid off during the first year, but is

immaterial in the case of an early default. The data indicate that prepayment penalties are

associated with a 0.7 percentage point higher early default rate for subprime loans, but have no

significant impact on the early default rate for alt-a loans.

Over 60 percent of the subprime

mortgages were initiated as a refinance rather than for a purchase. The data indicate that, holding

constant the observed risk factors, early defaults were less likely for subprime mortgages that

involved a refinance regardless of whether the borrower extracted cash or not.

For alt-a

mortgages, only the cash out refinances were associated with a significantly lower early default

rate, though the magnitude of the effect is smaller than for subprime cash out refinances.

Many of the discussions for providing assistance to preventing foreclosures exclude

investors and owners of 2

homes from the proposals. Justifications offered include that the

costs of foreclosure differ for a household that is living in the home versus an investor or a

vacation home. In addition, concerns over moral hazard from any government intervention tend

to be magnified for investors and 2

homes. We include indicators for investors and for 2

homes to see if their average early default rates differ from owner-occupied primary residences.

The data indicate that controlling for observed risk factors investors are more likely to default in

the first year. On average, subprime investors have an early default rate that is 2.7 percentage

points higher than for owner-occupied primary residences, while alt-a investors have a 1.3

percentage point higher early default rate. The data does not find any significant difference in

early default rates for 2

homes.

The final variables reported in Table 10 attempt to capture differences in the local

economies and housing markets. We control for the change in house prices over the year since

origination using the OFEHO repeat-sale house price for the MSA. We also control for the

average unemployment rate in the MSA over the course of the year. It is important to keep in

Future research should explore the extent to which this may reflect a selection effect on the borrowers

and/or on the lenders. We are currently working with Professor Chris Mayer of Columbia University to

merge in borrower characteristics by matching the LoanPerformance data with the HMDA data.

One interpretation of this finding is that in the absence of the prepayment penalties some of the early

subprime defaults might have refinanced instead during the first year.

The data does not support the idea of “strategic early defaults” where borrowers sensing the turn in the

housing markets extract all of the equity possible through cash-out refinances and then quickly default.

mind that we adjusted the initial LTV for the house price appreciation.

Even after factoring in

the house price appreciation into the borrower’s LTV, the data suggests a strong independent and

asymmetric effect of price appreciation on early defaults. If house prices over the year rise by 10

percentage points in a local market, early defaults are reduced by 1.4 percentage points for

subprime mortgages held by owners, and are reduced by 2.7 percentage points for subprime

mortgages held by investors .

In contrast, if house prices decline by 10 percentage points in a

local market, early defaults rise by 4.8 percentage points for subprime owners, and rise by 10.3

percentage points for subprime investors.

For alt-a owners, only house price declines exert an

independent impact beyond the current LTV on early defaults. Finally, while we control for the

initial DTI, we do not observe shocks to the borrower’s income over time. To proxy for this, we

include the average local unemployment rate. The data indicate that a 1 percentage point increase

in the local unemployment rate is associated with only a quarter of a percentage point rise in

early defaults for subprime mortgages, and no significant difference in early defaults for alt-a

mortgages. This is consistent with the muted response of early defaults to the initial DTI.

Bad credit or bad economy?

Table 11 examines the question of the relative importance of credit effects versus

economy effects in explaining the sharp rise in early defaults. The first column of the table

reports differences from 2003 in the average early default rate by year starting with the rise in

early defaults in 2005 (note that 2003 has the lowest average rate of early defaults for our sample

period, 3.4% of subprime loans and 0.7 % of alt-a loans). This is the overall change in early

defaults that the empirical model is trying to explain. The second column reports the difference

in average early defaults predicted by our linear probability model, while the third column

In contrast, Demyanky and Van Hemert (2008) control for the initial LTV and subsequent house price

appreciation. Given data limitations, we have to assume that each borrower experienced the average

house price appreciation for the msa based on the price index. If we had access to the estimated variances

for these MSA price indices, we could generate a distribution of updated LTVs and calculate the

probability that the borrower’s LTV is in each of our intervals.

Demyanky and Van Hemert (2008) report a similar marginal effect of house price appreciation. This is

surprising since in their empirical specification the marginal effect of house price appreciation includes

the indirect effect through the loan-to-value. Generally, we find the larger marginal effects for credit

factors than they report.

The incremental effect of house price decline for subprime investors relative to subprime owners is

large but not precisely estimated in the data.

reports the fraction of the total change in average early default rates that is predicted by our

model.

The final two columns disaggregate the explained rise in early defaults into components

that are attributable to differences across years in underwriting standards and to economic

conditions. The LTV indicators reflect aspects of both credit – the initial LTV – and economy –

the effect of house price appreciation on the current LTV. To separate out these two influences,

we create a set of counter factual LTV averages where for each year we take the initial LTV for a

mortgage and adjust it using the house price appreciation from 2003 for the same metropolitan

area and quarter of origination. Differences in these counter factual averages for the LTV

indicators from their 2003 average reflect only the initial distribution of LTV in each year – a

credit effect. The difference between the actual LTV averages for a given year and the counter

factual averages for that year reflect only the differences in house price appreciation rates for that

year and 2003 – an economy effect.

The major difference between 2003 and 2005-2007 was a dramatic change in house price

appreciation. After rising nearly 14% in 2003, the OFHEO index accelerated to 16% in 2004

before slowing and eventually reversing. For 2005-2007, OFHEO grew 10%, 1% and –4%

respectively.

The decomposition indicates that changes in economic variables, particularly this

reversal in house price appreciation, from 2003-2007 account for the bulk of our explanation for

observed increases in early defaults. In 2006, we estimate that changes in the economy added 2.4

percentage points to the average early default rate for subprime loans, while in 2007 that figure

rises to 4.1 percentage points.

“Bad Credit,” on the other hand, contributes less to our explained rise in average early

defaults. Had the economy continued to produce unemployment and house price appreciation

rates in 2005 through 2007 like those in 2003, our model predicts that changes in the credit

profiles of new nonprime mortgages in each year would result in an increases in average early

default rates for subprime loans of less than a percentage point in each year. For example, the

model predicts that the average early default rate in 2006 for subprime loans would have

increased by 94 basis points due to credit related factors. Of this, 44 basis points reflects changes

in the initial LTV distribution; 8 basis points reflects worsening DTI by borrowers; changes in

These figures refer to national average growth rates, but we use the MSA-level growth rates to conduct

the experiment.

the distribution of FICO scores would have reduced early defaults by 17 basis points; shifts away

from fully documented underwriting would have contributed 16 basis points; and the balance of

43 basis points reflects changes in the other credit factors.

Since our benchmark year of 2003 produced remarkably strong house price appreciation

that was subsequently sharply reversed, it may not seem surprising that we find a large role for

the economy and a relatively minor role for changes in credit standards in our explanation of the

rise in early defaults. After all, very high house price appreciation is sufficient to offset even a

substantial upward shift in the initial LTV distribution. For example, the 2003 growth rate was

sufficient to bring a property with an initial LTV between 100 to 109 to a current LTV below 95

where the incremental effect of LTV changes on early default is minimal. This raises the

question of whether our result that “bad credit” played a relatively minor role in explaining early

defaults would continue to hold if we used a more “normal” house price appreciation experience

as our benchmark.

We tested this proposition by replicating the experiment described above, but subjecting

each property value in both the base year (2003) and the comparison years to the 1985-2000

annual average OFHEO growth rate of 4.2 percent. This change only marginally affects the role

of “bad credit” as an explanation for the rise in early defaults. In 2006, for example, the overall

rise in defaults predicted by changes in credit standards rises from 94 basis points (shown in

Table 11) to 100 basis points. This points out that even relatively modest increase in house prices

would have been sufficient to have offset much of the upward drift in the distribution of initial

LTV in 2006 and 2007.

Conclusion

We use loan-level data on securitized nonprime mortgages to examine what we refer to as

“juvenile delinquency”: default or serious delinquency in the first year following a mortgage’s

origination. Early default became much more common for loans originated in 2005-2007. Two

complementary explanations have been offered for this phenomenon. The industry-standard

Similar to Demyanyk and van Hemert (2008), we find that changes in the distribution of LTV accounts

for a larger share of the credit related rise in early defaults than FICO, DTI and the level of

documentation. However, in our results the contribution of LTV is smaller than that of changes in house

prices.

explanation of default behavior focuses attention on a relaxation of lending standards after 2003.

We see evidence of this in our data, as some underwriting criteria, particularly loan-to-value

ratios at origination, deteriorated. At the same time, however, the housing market experienced a

sharp and pervasive downturn, a factor which has received attention in recent research. Our

results suggest that while both of these factors – bad credit and bad economy - played a role in

increasing early defaults starting in 2005, changes to the economy appear to have played the

larger role.

Perhaps as important a finding is that, in spite of the set of covariates we control for, our

model predicts at most 43 percent of the annual increase in subprime early defaults during the

2005-2007 period. Observable changes in standard underwriting standards and key economic

measures appear to be unable to explain the majority of the run-up in early defaults.

The fact,

noted in our introduction, that many participants in the industry appeared to have been surprised

by the degree of the increase in early defaults is in some sense verified here: observable

characteristics of the loans, borrowers and economy seem to leave much unexplained, even with

the benefit of hindsight. The difference between what we predict, conditional on observables,

and what we actually observe is the difference between a bad few years for lenders/investors and

a full-blown credit crunch.

The data does indicate a significant difference in behavior between owners and investors,

especially in terms of how they respond to downward movements in house prices and negative

equity situations. This has implications for underwriting. First, there may be payoffs to increased

efforts at determining the true occupancy status of the borrower as part of the underwriting

process. Second, originators may want to require additional equity up front from investors to

reduce the likelihood that future house price declines could push the investor into negative

equity.

An aim of our future research will be to improve the ability of the model to track the

changes in average early default rates. A first step is to add more extensive demographic controls

for the borrower. Second, given the estimated nonlinear response of the current LTV on early

defaults, being able to estimate a distribution of current LTVs for each loan may be quite

The model does a better job of explaining the rise in early defaults for alt-a mortgages in 2005 and 2006

with the model capturing two-thirds of the increase. However, in 2007 only 42 percent of the change is

explained by the model.

important during periods of house price declines.

A third area of investigation is the possibility

that a significant number of the investors misrepresented their status as an “owner”.

The data

indicate that investors appear to have a much stronger reaction in their early default decisions to

negative equity and to declines in house prices. Identifying likely cases where the investor status

is misrepresented could lead to significant improvements in the ability of the estimated model to

track the rise in early defaults. Finally, it is possible that part of the rise in early defaults reflects

changes in the composition of unobserved risk factors of borrowers. Further progress on this

possibility would be facilitated by panel data that follows a borrower across multiple mortgages.

How much of the rise in juvenile delinquent mortgages reflected bad credit or a bad

economic environment? Based on our evidence to date, a definitive answer is still elusive since

too much of the rise in early defaults is remains unexplained. More work is needed to narrow the

gap between actual and explained performance for nonprime mortgages. Of what is explained,

Officer Krupke might have to concede that many of these juvenile delinquent mortgages “never

had the house price appreciation that ev’ry mortgage oughta get.”

We are in discussions with OFEHO regarding the use of the metro-area house price variances.

Fitch reports that in a small sample of subprime loans that defaulted 66% of the borrowers

misrepresented their occupancy status.

References

An, Xudong, Raphael W. Bostic, Deng Yongheng, and Stuart A. Gabriel. "GSE Loan Purchases,

the FHA, and Housing Outcomes in Targeted, Low-Income Neighborhoods." In

Brookings-Wharton Papers on Urban Affairs, edited by Gary Burtles and Janet

Rothenberg Pack, 205-240. Washington DC, Brookings Institution Press, 2007.

Black, Fischer, and Myron Scholes. "The Pricing of Options and Corporate Liabilities." Journal

of Political Economy

81 (May 1973): 637-654.

Caplin, Andrew, Charles Freeman, and Joseph Tracy. "Collateral Damage: Refinancing

Constraints and Regional Recessions." Journal of Money, Credit, and Banking 29

(November 1997): 496-516.

Cunningham, Donald F., and Patric H. Hendershott. "Pricing FHA Mortgage Default Insurance."

Housing Finance Review 3 (October 1984): 373-392.

Demyanyk, Yuliya, and Otto Van Hemert. "Understanding the Subprime Mortgage Crisis."

Working Paper. New York University, Stern School of Business, February, 2008.

Deng, Yongheng, and Stuart Gabriel. "Risk-Based Pricing and the Enhancement of Mortgage

Credit Availability Among Underserved and Higher Credit-Risk Populations." Journal of

Money, Credit and Banking

38 (September 2006): 1431-1460.

------, John M. Quigley, and Robert van Order. "Mortgage Terminations, Heterogeneity and the

Exercise of Mortgage Options." Econometrica 68 (March 2000): 275-307.

Elul, Ronel. "Residential Mortgage Default." Federal Reserve Bank of Philadelphia Business

Review

, Third Quarter 2006, 21-30.

Fitch Ratings. The Impact of Poor Underwriting Practices and Fraud in Subprime RMBS

Performance

. US Residential Mortgage Special Report, 2007.

Foote, Christopher L., Kristopher Gerardi, and Paul S. Willen. "Negative Equity and

Foreclosure: Theory and Evidence." Public Policy Discussion Papers 08-3. Federal

Reserve Bank of Boston, 2008.

Foster, Chester, and Robert van Order. "An Option-Based Model of Mortgage Default." Housing

Finance Review

3 (October 1984): 273-291.

Gerardi, Kristopher, Adam Hale Shapiro, and Paul S. Willen. "Subprime Outcomes: Risky

Mortgages, Homeownership Experiences, and Foreclosures." Working Paper No. 07-15.

Federal Reserve Bank of Boston, December, 2007.

Kau, James B., Donald C. Keenan, and Taewon Kim. "Transaction Cost, Supoptimal

Termination, and Default Probabilities for Mortgages."

Journal of the American Real

Estate and Urban Economics Association

21 (Fall 1993): 247-263.

------, Donald C. Keenan, Walter J. Muller, and James F. Epperson. "The Valuation at

Origination of Fixed Rate Mortgages with Default and Prepayment."

Journal of Real

Estate Finance and Economics

11 (July 1995): 5-36.

Leventis, Andrew. "Revisiting the Differences Between the OFHEO and S&P/Case-Shiller

House Price Indexes: New Explanations." Research Paper. Office of Federal Housing

Enterprise Oversight, January, 2008.

Mayer, Christopher, and Karen Pence. "Subprime Mortgages: What, Where, and to Whom?"

FEDS Working Paper 2008-29. Board of Governors of the Federal Reserve System, June,

2008.

Pence, Karen M. "Foreclosing on Opportunity: State Laws and Mortgage Credit." FEDS

Working Paper 2003-16. Board of Governors of the Federal Reserve System, May, 2003.

Pennington-Cross, Anthony. "Subprime Lending in the Primary and Secondary Markets."

Journal of Housing Research 13, no. 1 (2002): 31-50.

------, and Giang Ho. "The Termination of Subprime Hybrid and Fixed Rate Mortgages."

Working Paper 2006-042A. Federal Reserve Bank of St Louis, Research Division, July,

2006.

Schwartz, Eduardo S., and Walter N. Torous. "Mortgage Payment and Default Decisions: A

Poisson Regression Approach." Journal of the American Real Estate and Urban

Economics Association

21 (Winter 1993): 431-449.

Vandell, Kerry D. "How Ruthless Is Mortgage Default? A Review and Synthesis of the

Evidence." Journal of Housing Research 6, no. 2 (1995): 245-264.

Table 1. Combined initial LTV

Subprime

Year 10

% w. 2

2001 63 75 80 89 90 3.16 3,984

2002 65 75 80 90 97 3.56 6,636

2003 64 75 84 90 100 7.40 11,210

2004 65 78 85 95 100 16.07 17,093

2005 66 80 87 100 100 24.55 19,816

2006 66 80 90 100 100 29.36 14,756

2007 65 78 87 95 100 17.93 1,556

Alt-A

Year 10

% w. 2

2001 64 75 80 90 100 2.17 1,107

2002 60 73 80 90 100 3.28 2,013

2003 51 67 80 90 100 12.36 3,746

2004 60 74 80 95 100 29.95 7,613

2005 60 74 80 95 100 36.39 12,164

2006 62 75 85 95 100 43.88 11,556

2007 60 75 80 95 100 34.82 2,079

Notes: Loan Performance data, authors calculations

Table 2. Early Defaults by Initial LTV and Year

Subprime

Year Overall < 80 80 – 89 90 – 94 95 – 99 100+

2001 6.70 6.97 6.98 5.50 5.37 8.99

2002 4.73 4.76 5.51 3.76 1.85 5.30

2003 3.40 3.23 3.45 3.42 2.95 4.03

2004 4.70 3.76 4.77 5.11 5.44 5.30

2005 6.60 5.35 6.15 7.15 7.83 7.60

2006 11.26 6.95 11.03 12.91 11.23 14.19

2007 14.52 9.20 14.18 16.94 14.58 19.94

Alt-A

Year Overall < 80 80 – 89 90 – 94 95 – 99 100+

2001 2.98 0.99 3.80 8.05 2.80 1.54

2002 1.39 0.74 1.75 2.68 1.18 1.72

2003 0.67 0.33 0.38 1.05 1.57 1.72

2004 0.85 0.28 0.69 0.80 1.64 2.13

2005 1.34 0.58 1.09 1.41 2.10 2.99

2006 3.21 1.21 2.32 3.64 4.51 6.47

2007 6.93 2.47 4.79 8.56 11.11 15.61

Notes: LoanPerformance data, authors calculations

Table 3. Early Defaults by Initial LTV, Investor status and Geography

Subprime

< 80 80 – 89 90 – 94 95 – 99 100+

Owner 5.07 6.26 6.58 6.71 8.99

Boom/bust (AZ, CA, FL, NV) 4.22 5.47 5.40 6.19 9.39

Weak economy (IN, MI, OH) 7.82 9.21 10.24 9.68 9.74

Other 5.47 6.33 6.57 6.43 8.63

Investor 4.64 6.18 9.17 9.30 11.79

Boom/bust (AZ, CA, FL, NV) 2.90 3.51 6.18 8.51 3.45

Weak economy (IN, MI, OH) 11.56 12.65 18.33 13.43 22.81

Other 4.31 5.92 7.02 8.44 9.09

Alt-A

< 80 80 – 89 90 – 94 95 – 99 100+

Owner 0.70 1.65 2.36 2.95 4.34

Boom/bust (AZ, CA, FL, NV) 0.66 1.56 2.46 3.83 5.28

Weak economy (IN, MI, OH) 0.00 3.68 2.68 3.97 4.15

Other 0.81 1.58 2.22 2.26 3.63

Investor 0.97 1.40 3.15 3.45 8.62

Boom/bust (AZ, CA, FL, NV) 0.67 1.30 2.58 5.42 17.58

Weak economy (IN, MI, OH) 2.74 3.62 4.76 7.69 5.71

Other 1.17 1.22 3.34 1.99 6.51

Notes: LoanPerformance data, authors calculations

Table 4. Distribution of Debt-to-Income Ratios – by Year

Subprime

Year < 30 30 – 34 35 – 39 40 +

2001 49.05 9.06 9.76 32.13

2002 46.88 8.88 10.76 33.48

2003 42.05 9.49 11.01 37.45

2004 37.85 9.12 12.34 40.69

2005 40.54 7.82 11.00 40.64

2006 29.54 8.28 11.70 50.47

2007 37.53 7.78 10.54 44.15

Alt-A

Year < 30 30 – 34 35 – 39 40 +

2001 78.05 5.33 6.87 9.76

2002 76.45 6.36 8.49 8.69

2003 75.04 6.54 7.82 10.60

2004 65.87 8.05 10.84 15.24

2005 65.42 7.88 11.16 15.54

2006 57.61 8.19 14.17 20.04

2007 58.01 6.97 13.56 21.45

Notes: LoanPerformance data, authors calculations

Table 5. Early Defaults by Debt-to-Income and Year

Subprime

Year < 30 30 – 34 35 – 39 40 +

2001 6.09 7.76 6.94 7.27

2002 4.56 3.90 5.04 5.09

2003 3.20 3.10 3.57 3.64

2004 4.73 3.59 4.74 4.90

2005 6.48 5.87 5.69 7.09

2006 10.03 10.15 10.60 12.31

2007 13.87 14.88 15.24 14.85

Alt-A

Year < 30 30 – 34 35 – 39 40 +

2001 3.59 0 1.32 0.93

2002 1.43 1.56 1.17 1.14

2003 0.71 0.41 0.34 0.76

2004 0.70 1.14 1.09 1.21

2005 1.12 1.15 1.84 2.01

2006 3.05 3.59 3.24 3.50

2007 5.72 4.83 7.09 10.76

Notes: LoanPerformance data, authors calculations

Table 6. Distribution of FICO Scores – by Year

Subprime

Year < 600 600 – 619 620 – 659 660+

2001 51.53 13.50 19.78 15.19

2002 45.16 13.32 21.47 20.04

2003 39.68 12.58 23.68 24.06

2004 38.06 13.53 24.50 23.91

2005 35.59 14.34 25.76 24.30

2006 35.75 16.22 27.00 21.03

2007 39.91 17.16 24.04 18.89

Alt-A

Year < 600 600 – 619 620 – 659 660+

2001 1.81 2.89 17.43 77.87

2002 2.24 2.78 14.41 80.58

2003 0.83 1.63 13.93 83.61

2004 0.71 1.29 14.93 83.07

2005 0.41 1.18 14.03 84.37

2006 0.14 0.84 16.79 82.23

2007 0.00 0.43 15.92 83.65

Notes: LoanPerformance data, authors calculations

Table 7. Early Defaults by FICO Scores and Year

Subprime

Year < 600 600 – 619 620 – 659 660+

2001 9.16 5.02 4.82 2.31

2002 7.21 2.60 3.65 1.73

2003 5.51 3.55 1.85 1.37

2004 7.24 3.81 3.84 2.03

2005 9.92 5.91 5.35 3.45

2006 13.31 10.82 11.14 8.25

2007 15.14 15.36 15.24 11.56

Alt-A

Year < 600 600 – 619 620 – 659 660+

2001 15.00 3.13 6.74 1.86

2002 2.22 3.57 2.76 1.05

2003 12.90 3.28 0.96 0.45

2004 1.85 4.08 1.58 0.66

2005 6.00 5.56 2.34 1.09

2006 12.50 10.31 4.95 2.77

2007 0.00 22.22 13.29 5.64

Notes: LoanPerformance data, authors calculations

Table 8. Distribution of Documentation Level – by Year

Subprime

Year Full Low None

2001 77.84 21.76 0.40

2002 71.13 28.30 0.57

2003 67.02 32.52 0.46

2004 65.37 34.34 0.29

2005 62.28 37.47 0.24

2006 61.71 38.00 0.29

2007 64.20 35.48 0.32

Alt-A

Year Full Low None

2001 36.77 55.56 7.68

2002 40.64 51.96 7.40

2003 35.50 57.26 7.23

2004 37.75 55.72 6.53

2005 31.11 64.44 4.46

2006 18.92 76.56 4.53

2007 16.84 77.49 5.68

Notes: LoanPerformance data, authors calculations

Table 9. Early Defaults by Documentation Level and Year

Subprime

Year Full Low None

2001 6.58 7.15 6.25

2002 4.49 5.43 0.00

2003 3.31 3.54 5.77

2004 4.62 4.87 2.04

2005 6.05 7.54 0.00

2006 9.26 14.55 4.65

2007 11.91 19.20 20.00

Alt-A

Year Full Low None

2001 0.98 3.41 9.41

2002 0.98 1.34 4.03

2003 0.45 0.75 1.11

2004 0.77 0.94 0.60

2005 0.77 1.57 2.03

2006 2.06 3.47 3.63

2007 3.14 7.57 9.32

Notes: LoanPerformance data, authors calculations

Table 10. Probability of an Early Default

Subprime Alt-A

Variable (1) (2)

LTV: 80 – 84 0.95

(0.30) 0.26 (0.24)

85 – 89 1.40

(0.28) 1.24

(0.26)

90 – 94 2.02

(0.36) 1.38

(0.29)

95 – 99 3.07

(0.42) 2.91

(0.38)

100+ 6.84

(1.26) 6.92

(0.89)

Investor • 95 – 99 1.78 (2.45) 3.77

(1.61)

• 100+ 24.59

(12.35) 20.29

(5.96)

DTI missing 0.41 (0.22) 0.37

(0.16)

DTI: 40 – 44 0.71

(0.28) 0.59

(0.31)

45 – 49 0.77

(0.28) 0.47 (0.44)

50+ 1.26

(0.39) –0.90

(0.42)

FICO: <560 10.35

(0.42) 11.81

(4.57)

560 – 589 6.92

(0.40) 6.23

(2.29)

590 – 619 4.47

(0.36) 4.77

(1.14)

620 – 649 3.32

(0.35) 2.98

(0.33)

650 – 679 1.72

(0.32) 1.75

(0.23)

680 – 719 0.09 (0.34) 0.68

(0.14)

Investor • <560 –11.84

(4.59)

• 560 – 589 –6.97

(2.37)

• 590 – 619 –4.68

(1.18)

Limited documentation 2.96

(0.22) 1.27

(0.15)

No documentation 0.73 (1.08) 2.14

(0.38)

Fixed rate mortgage –1.17

(0.20) –0.57

(0.17)

Prepayment penalty 0.70

(0.21) 0.28 (0.18)

Refinance – no cash –2.65

(0.35) –0.11 (0.20)

Refinance – cash out –3.11

(0.25) –0.61

(0.16)

Investor 2.73

(0.65) 1.29

(0.35)

– home 1.37 (0.96) 0.51 (0.44)

House price appreciation − positive (10%) –1.42

(0.14) –0.02 (0.11)

House price appreciation − negative (10%) 4.80

(1.03) 1.66

(0.69)

Investor • house price appr − positive (10%) –1.32

(0.39) –0.44

(0.18)

Investor • house price appr − negative (10%) 5.57 (4.57) –0.74 (1.25)

Local unemployment rate 0.25

(0.08) –0.04 (0.06)

Root mean square error 0.243 0.139

Mean early default rate 6.61 2.06

Observations 75,051 40,278

Notes: Linear probability estimates with standard errors given in parentheses. Standard errors use

clustering at the msa•year•quarter level. LoanPerformance data, 1 percent random sample. Year

effects and six property type fixed effects are included.

** significant at the 1 percent level

* significant at the 5 percent level

Table 11. Decomposition of Rise in Early Default Rates

Subprime

Early Default Percent Due to:

Year Difference Explained Explained Credit Economy

2005 3.20 0.90 28.1 0.58 0.32

2006 7.86 3.39 43.1 0.94 2.45

2007 11.13 4.80 43.2 0.72 4.08

Alt-A

Early Default Percent Due to:

Year Difference Explained Explained Credit Economy

2005 0.67 0.44 65.8 0.20 0.24

2006 2.54 1.70 66.8 0.55 1.14

2007 6.26 2.61 41.7 0.49 2.12

Notes: All differences expressed relative to 2003. LoanPerformance data, authors calculations

using the estimates reported in table 10 and sample means by year for the variables.

Table A1: Descriptive Statistics

Subprime Alt-A

Mean

Standard

Deviation

Mean

Standard

Deviation

Early default 6.61 24.84 2.06 14.20

Loan size ($000) 189.9 123.4 291.8 225.3

Initial interest rate 7.86 1.43 5.60 2.18

LTV 74.37 19.69 73.21 19.73

80 – 84 0.105 0.306 0.079 0.269

85 – 89 0.122 0.327 0.095 0.293

90 – 94 0.098 0.297 0.099 0.299

95 – 99 0.114 0.318 0.103 0.304

100+ 0.018 0.133 0.028 0.165

DTI 28.92 19.68 16.41 19.10

40 – 44 0.150 0.357 0.098 0.298

45 – 49 0.188 0.390 0.051 0.221

50+ 0.083 0.276 0.018 0.132

FICO 617.49 60.54 708.37 48.10

< 560 0.184 0.388 0.001 0.034

560 – 589 0.147 0.354 0.003 0.052

590 – 619 0.188 0.391 0.010 0.101

620 – 649 0.193 0.395 0.104 0.305

650 – 679 0.138 0.345 0.177 0.381

680 – 719 0.091 0.288 0.298 0.457

Limited documentation 0.344 0.475 0.654 0.476

No documentation 0.003 0.058 0.054 0.226

Fixed rate mortgage 0.232 0.422 0.436 0.496

Prepayment penalty 0.722 0.448 0.368 0.482

Refinance – no cash 0.083 0.276 0.162 0.369

Refinance – cash out 0.547 0.498 0.327 0.469

Investor 0.073 0.259 0.213 0.409

– home 0.010 0.100 0.034 0.182

House price appreciation 9.559 9.138 8.402 10.108

Local unemployment rate 5.126 1.309 4.851 1.260

Origination year: 2001 0.053 0.224 0.027 0.163

2002 0.088 0.284 0.050 0.218

2003 0.149 0.356 0.093 0.290

2004 0.228 0.419 0.189 0.391

2005 0.264 0.441 0.302 0.459

2006 0.197 0.397 0.287 0.452

2007 0.021 0.142 0.052 0.221

Note: LoanPerformance data, 1 percent random sample. Sample sizes: 75,051 subprime loans and

40,278 alt-a loans.

Table A2. Probability of an Investor and Missing DTI

Variable

Investor

(1)

Missing DTI

(2)

LTV: 80 – 84 –0.35 (0.30) –2.50

(0.51)

85 – 89 –1.41

(0.23) –2.67

(0.48)

90 – 94 –6.53

(0.16) –2.29

(0.52)

95 – 99 –7.03

(0.13) –4.31

(0.54)

100+ –8.48

(0.12) –8.29

(1.02)

Investor • 95 – 99 –4.15 (2.13)

Investor • 100+ 0.28 (6.08)

DTI missing –1.40

(0.19)

DTI: 40 – 44 –3.34

(0.22)

45 – 49 –3.01

(0.23)

50+ –2.25

(0.33)

FICO: <560 –9.97

(0.13) –14.88

(0.52)

560 – 589 –8.74

(0.13) –10.44

(0.57)

590 – 619 –8.23

(0.15) –9.93

(0.52)

620 – 649 –6.94

(0.16) –10.52

(0.48)

650 – 679 –4.99

(0.18) –7.84

(0.48)

680 – 719 –2.21

(0.20) –4.29

(0.48)

Limited documentation –0.55

(0.17) 1.89

(0.32)

No documentation –4.98

(0.29) 56.12

(0.63)

Fixed rate mortgage 2.17

(0.20) 16.51

(0.34)

Prepayment penalty –1.24

(0.18) –12.34

(0.31)

Refinance – no cash –3.38

(0.20) –2.91

(0.50)

Refinance – cash out –5.03

(0.19) –3.58

(0.36)

Investor 0.62 (0.66)

– home 1.89 (1.09)

House price appreciation –0.16

(0.01) –0.24

(0.02)

Investor • house price appr 0.10

(0.05)

Local unemployment rate 0.05 (0.07) –0.69

(0.13)

Mean of dependent variable 12.21 37.57

Notes: Probit marginal effects with standard errors given in parentheses. Left-

out year is 2003. LoanPerformance data, 1 percent random sample. Year effects

and six property effects are included. Each marginal effect reflects a percentage

point change in the dependent variable.

significant at the 1 percent level

significant at the 5 percent level

Figure 1. Nonprime 90+ Days Delinquencies – by vintage

0 6 12 18 24 30

Months since origination

Percent

2003

2004

2005

2006

Notes: FirstAmerican CoreLogic LoanPerformance

Figure 2. House Price Appreciation Over Time – by State

Source: Office of Federal Housing Enterprise Oversight

KYKS

-12 -10 -8 -6 -4 -2 0 2 4 6 8

% Change -Annual Rate (2007Q1-2008Q1)

% Change - Annual Rate (2001Q3-2006Q3)