Comparing patterns of default among prime and subprime mortgages.
Amromin, Gene ; Paulson, Anna L.
Introduction and summary
We have all heard a lot in recent months about the soaring number
of defaults among subprime mortgage borrowers; and while concern over
this segment of the mortgage market is certainly justified, subprime
mortgages account for only about one-quarter of the total outstanding
home mortgage debt in the United States. The remaining 75 percent is in
prime loans. Unlike subprime loans, prime loans are made to borrowers
with good credit, who fully document their income and make traditional
down payments. Default rates on prime loans are increasing rapidly,
although they remain significantly lower than those on subprime loans.
For example, among prime loans made in 2005, 2.2 percent were 60 days or
more overdue 12 months after the loan was made (our definition of
default). For loans made in 2006, this percentage nearly doubled to 4.2
percent, and for loans made in 2007, it rose by another 20 percent,
reaching 4.8 percent. By comparison, the percentage of subprime loans
that had defaulted after 12 months was 14.6 percent for loans made in
2005, 20.5 percent for loans made in 2006, and 21.9 percent for loans
made in 2007. To put these figures in perspective, among loans
originated in 2002 and 2003, the share of prime mortgages that defaulted
within 12 months ranged from 1.4 percent to 2.2 percent and the share of
defaulting subprime mortgages was less than 7 percent. (1) How do we
account for these historically high default rates? How have recent
trends in home prices affected mortgage markets? Could contemporary
observers have forecasted these high default rates?
Figure 1, panel A summarizes default patterns for prime mortgages;
panel B reports similar trends for subprime mortgages. Both use
loan-level data from Lender Processing Services (LPS) Applied Analytics.
Each line in this figure shows the cumulative default experience for
loans originated in a given year as a function of how many months it has
been since the loan was made. Several patterns are worth noting. First,
the performance of both prime and subprime mortgages has gotten
substantially worse, with loans made in 2006 and 2007 defaulting at much
higher rates. The default experience among prime loans made in 2004 and
2005 is very similar, but for subprime loans, default rates are higher
for loans made in 2005 than in 2004. Default rates among subprime loans
are, of course, much higher than default rates among prime loans.
However, the deterioration in the performance of prime loans happened
more rapidly than it did for subprime loans. For example, the percentage
of prime loans that were 60 days or more overdue grew by 95 percent for
loans made in 2006 compared with loans made in 2005. Among subprime
loans it grew by a relatively modest 53 percent.
Home prices are likely to play an important role in
households' ability and desire to honor mortgage commitments.
Figure 2 describes trends in home prices from 1987 through 2008 for the
ten largest metropolitan statistical areas (MSAs). This figure
illustrates the historically high rates of home price growth from 2002
through 2005, as well as the sharp reversal in home prices beginning in
2006. One of the things we consider in this article is whether prime and
subprime loans responded similarly to these home price dynamics.
Although the delinquency rate among prime mortgages is high and
rising fast, it is only about one-fifth the delinquency rate for
subprime mortgages. Unfortunately, however, this does not mean that
total losses on prime mortgages will be just one-fifth the losses on
subprime mortgages. The prime mortgage market is much larger than the
subprime mortgage market, representing about 75 percent of all
outstanding mortgages (International Monetary Fund, 2008), or a total of
$8.3 trillion. (2) Taking the third quarter of 2008 as the starting
point, we estimate that total losses from prime loan defaults will be in
the neighborhood of $133 billion and that total losses from subprime
loan defaults will be about $364 billion. (3)
[FIGURE 1 OMITTED]
Losses on prime mortgages are also distributed very differently
from losses on subprime mortgages. Most prime mortgages for amounts at
or below $417,000 are guaranteed through the government-sponsored
enterprises (GSEs), such as Fannie Mae and Freddie Mac? Losses on these
mortgages that exceed the ability of the GSEs to satisfy their
obligations are ultimately borne by the taxpayer? In contrast, prime
mortgages for amounts greater than $417,000 ("jumbo" loans)
and subprime mortgages were largely securitized privately, and absent
government intervention, investors in asset-backed securities linked to
those mortgages are the ones that are most exposed to declines in their
value due to increasing defaults.
[FIGURE 2 OMITTED]
In this article, we make use of loan-level data on individual prime
and subprime loans made between January 1, 2004, and December 3 l, 2007,
to do three things: l) analyze trends in loan and borrower
characteristics and in the default experience for prime and subprime
loans; 2) estimate empirical relationships between home price
appreciation, loan and borrower characteristics, and the likelihood of
default; and 3) examine whether using alternative assumptions about the
behavior of home prices could have generated more accurate predictions
of defaults. Throughout the analysis, we divide the loans into eight
groups based on two characteristics: prime versus subprime and the
"vintage," the year in which the loan was made.
First, we describe trends in loan and borrower characteristics, as
well as the default experience for prime and subprime loans for each
year from 2004 through 2007. Next, we estimate empirical models of the
likelihood that a loan will default in its first 12 months. This allows
us to quantify which factors make default more or less likely and to
examine how the sensitivity to default varies over time and across prime
and subprime loans. Finally, we use these results to examine whether
market participants could have forecasted default rates more accurately,
that is, could have made predictions that were closer to actual default
rates, by using alternative assumptions about the behavior of home
prices. This article draws on much of the very informative literature on
the performance of subprime loans, including, Bajari, Chu, and Park
(2008); Chomsisengphet and Pennington-Cross (2006); Demyanyk and Van
Hemert (2009); Dell' Ariccia, Igan, and Laeven (2008); DiMartino
and Duca (2007); Foote et al. (2008); Gerardi, Shapiro, and Willen
(2008); Gerardi et al. (2008); and Mian and Sufi (2009). By including
prime loans in the analysis, our intention is to complement the existing
literature on subprime loans.
By looking at prime and subprime loans side by side, we also hope
to refine the possible explanations for the ongoing mortgage crisis.
Both prime and subprime loans have seen rising defaults in recent years,
as well as very similar patterns of defaults, with loans made in more
recent years defaulting at higher rates. Because of these similarities,
it seems reasonable to expect that a successful explanation of the
subprime crisis--the focus of most research to date--should also account
for the patterns of defaults we observe in prime mortgages.
We find that pessimistic forecasts of home price appreciation could
have helped to generate predictions of subprime defaults that were
closer to the actual default experience for loans originated in 2006 and
2007.6 However, for prime loans this would not have been enough.
Contemporary observers would have also had to anticipate that default
among prime loans would become much more sensitive to changes in home
prices. Among prime loans originated in 2006 and 2007, defaults were
much more correlated with changes in home prices than was the case for
prime loans originated in 2004 and 2005. While this pattern is
straightforward to document now, it would have been difficult to
anticipate at the time.
Loan and borrower characteristics
In this section, we discuss trends in loan and borrower
characteristics, as well as the default experience for prime and
subprime loans for each year from 2004 through 2007.
Data
The loan-level data we use come from LPS Applied Analytics, which
gathers data from a number of loan servicing companies. (7) The most
recent data include information on 30 million loans, with smaller, but
still very large, numbers of loans going back in time.
The data cover prime, subprime, and AlbA loans, (8) and include
loans that are privately securitized, loans that are sold to the GSEs,
and loans that banks hold on their balance sheets. Based on a comparison
of the LPS and Home Mortgage Disclosure Act (HMDA) data, we estimate
that the LPS data cover about 60 percent of the prime market each year
from 2004 through 2007. (9) Coverage of the subprime market is somewhat
smaller, but increases over time, going from just under 30 percent in
2004 to just under 50 percent in 2007.
The total number of loans originated in the LPS data in each year
of the period we study ranges from a high of 6.2 million in 2005 to a
low of 4.3 million in 2007. (10) The mortgage servicers reporting to LPS
Applied Analytics give each loan a grade of A, B, or C, based on the
servicer's assessment of whether the loan is prime or subprime. We
label A loans as prime loans and B and C loans as subprime loans. (11)
To make the analysis tractable, we work with a 1 percent random sample
of prime loans made between January 1, 2004, and December 31, 2007, for
a total of 68,000 prime loans, and a 10 percent random sample of
subprime loans made during the same time period, for a total of 62,000
subprime loans.
The LPS data include a wide array of variables that capture
borrower and loan characteristics, as well as the outcome of the loan.
The variables that we use in the analysis are defined in box 1. In terms
of borrower characteristics, important variables include the
debt-to-income ratio (DTI) of the borrower (available for a subset of
loans) and the borrower's creditworthiness, as measured by his Fair
Isaac Corporation (FICO) score. (12) Some of the loan characteristics
that we analyze include the loan amount at origination; whether the loan
is a fixed-rate mortgage (FRM) or adjustable-rate mortgage (ARM); the
ratio of the loan amount to the value of the home at origination (LTV);
whether the loan was intended for home purchase or refinancing and, in
case of the latter, whether it involved equity extraction (a
"cash-out refinance"); and whether the loan was sold to one of
the GSEs, privately securitized, or held on the originating bank's
portfolio.
The outcome variable that we focus on is whether the loan becomes
60 days or more past due in the 12 months following origination. We
focus on the first 12 months, rather than a longer period, so that loans
made in 2007 can be analyzed the same way as earlier loans, as our data
are complete through the end of 2008. (13)
We augment the loan-level data with information on local economic
trends and trends in local home prices. The economic variable we focus
on is the local unemployment rate that comes from U.S. Bureau of Labor
Statistics monthly MSA-level data. Monthly data on home prices are
available by MSA from the Federal Housing Finance Agency (FHFA)--an
independent federal agency that is the successor to the Office of
Federal Housing Enterprise Oversight (OFHEO) and other government
entities. (14) We use the FHFA's all transactions House Price Index
(HPI) that is based on repeat sales information.
Trends in loan and borrower characteristics
Many commentators (see, for example, Demyanyk and Van Hemert, 2009)
have noted that subprime lending standards became more lax during the
period we study, meaning that the typical borrower may have received
less scrutiny over time that it became easier for borrowers to get loans
overall, as well as to get larger loans. These trends have been
particularly well documented for subprime loans, but there has been less
analysis of prime loans. Table 1 summarizes mortgage characteristics for
each year from 2004 through 2007 for prime and subprime mortgages.
Consistent with prior work, we also document declining borrower
quality over time in the subprime sector. For example, the average FICO
score for subprime borrowers in 2004 was 617, but it had declined to 597
by 2007. (15) By contrast, when we look at prime loans, the decline in
lending standards is less obvious. The average FICO score among prime
borrowers was 710 in 2004 and 706 in 2007, a decline of less than 1
percent.
Another potential indicator of the riskiness of a mortgage is the
reason for taking out the loan: to buy a house or to refinance an
existing mortgage. People who are buying a home include first-time home
buyers who tend to be somewhat riskier, perhaps because they have
stretched to accumulate the necessary funds to purchase a home or
perhaps because they tend to be younger and have lower incomes. While we
do not have data on whether loans for home purchase go to first-time
home buyers or to individuals who have owned a home before, we do know
that the fraction of home purchase loans among prime mortgages is
roughly 50 percent and stays at about that rate throughout the 2004-07
period. Among subprime mortgages, about 40 percent of loans are for home
purchase in 2004-06; this share drops to just under 30 percent of
subprime loans made in 2007.
BOX 1
Definitions of variables
Variable Description
Default (%) first 12 Share of loans that are 60 days or more
months delinquent, in foreclosure, or real estate
owned within 12 months of origination
Default (%) first 18 Share of loans that are 60 days or more
months delinquent, in foreclosure, or real estate
owned within 18 months of origination
Default (%) first 21 Share of loans that are 60 days or more
months delinquent, in foreclosure, or real estate
owned within 21 months of origination
Fair Isaac Corporation Credit score at time of origination (range
(FICO) score between 300 and 850, with a score above 800
considered very good and a score below 620
considered poor)
Loan-to-value ratio Face value of the loan divided by the
(LTV) appraised value of the house at time of loan
origination
Interest rate at Initial interest rate of loan at origination
origination
Origination amount Dollar amount of the loan at origination
Conforming loan Dummy variable equal to 1 for loans that
satisfy the following conditions: FICO score
of at least 620, LTV of at most 80 percent,
and loan amount at or below the time-varying
limit set by the Federal Housing Finance
Agency; 0 otherwise
Debt-to-income ratio Ratio of total monthly debt payments to
(DTI) gross monthly income, computed at
origination
DTI missing Dummy variable equal to 1 if DTI is not
available from the mortgage servicer; 0
otherwise
Cash-out refinance Dummy variable equal to 1 if the loan
refinances an existing mortgage while
increasing the loan amount; 0 otherwise
Purchase loan Dummy variable equal to 1 if the loan is
used for a property purchase; 0 otherwise
Investment property Dummy variable equal to 1 if the loan is for
loan a non-owner-occupied property; 0 otherwise
Loan sold to govern- Dummy variable equal to 1 if the loan is
ment-sponsored enter- sold to a GSE; 0 otherwise
prise (GSE)
Loan sold to private Dummy variable equal to 1 if the loan is
securitizer sold to a non-GSE investor; 0 otherwise
Loan held on portfolio Dummy variable equal to 1 if the loan is
held on originator's portfolio; 0 otherwise
Prepayment penalty Dummy variable equal to 1 if the loan is
originated with a prepayment penalty; 0
otherwise
Adjustable-rate Dummy variable equal to 1 if the loan's
mortgage (ARM) interest rate is adjusted periodically, and
the rate at origination is kept fixed for an
introductory period; 0 if it is a fixed-rate
mortgage (FIRM), a loan whose rate is fixed
at origination for its entire term
Margin rate Spread relative to some time-varying
reference rate (usually London interbank
offered rate, or Libor), applicable after
the first interest rate reset for an ARM
House Price Index Change in
(HPI) growth metropolitan-statistical-area-level
(MSA-level) housing price index in the 12
months after origination, reported by the
Federal Housing Finance Agency
Unemployment rate Average change in the MSA-level unemployment
rate in the 12 months after origination,
reported by the U.S. Bureau of Labor
Statistics
Median annual income Median annual income in the zip code where
in zip code property is located, as reported in the 2000
U.S. Decennial Census
Like home purchase loans, refinancing transactions probably include
both individuals who are less likely to default after they refinance and
those who are more likely to default. For example, a household that
refinances the existing balance on its original mortgage to take
advantage of falling interest rates will have lower monthly payments
that should be easier to maintain, even if it experiences a period of
economic hardship. In contrast, a household that refinances its mortgage
to extract equity (a cash-out refinance) when the value of its home
increases may end up being more vulnerable to future home price
declines, especially if its new mortgage has a higher loan-to-value
ratio. To the extent that the practice of cash-out refinancing was
common over the period we study, increases in home prices may be
associated with constant or even increasing leverage rather than with
safer loans and a bigger cushion against future price declines. In this
way, greater prevalence of cash-out refinancing transactions may be
indicative of increasing risk in the universe of existing loans. The
percentage of loans that involved refinancing together with cashing out
some of the built-up equity is much lower for prime loans than for
subprime loans, but it increases for both over the 2004-07 period.
As indicated in table 1 (p. 23), mortgage servicers assign many
refinancing transactions to the ambiguous category of "refinancing
with unknown cash-out." 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 (the share of unclassified refinancing transactions
remained fairly constant over time). For subprime loans made in 2004, 35
percent were refinancing transactions involving known cash-outs; for
those made in 2005, 43 percent; for those made in 2006, 47 percent; and
for those made in 2007, a staggering 57 percent. Put differently,
cash-out loans accounted for at least 82 percent (0.575/0.7) of all
subprime mortgage refinancing transactions in 2007.
Another loan characteristic that might be an important determinant
of subsequent defaults is whether the interest rate is fixed for the
life of the contract or allowed to adjust periodically (as in
adjustable-rate mortgages). When an ARM resets after the initial defined
period (which may be as short as one year or as long as seven), the
interest rate and, consequently, the monthly mortgage payment, may go up
substantially. Higher payments may put enough stress on some households
so that they fall behind on their mortgages. While these loans seem
attractive because of low introductory interest rates (and low initial
payments), they expose borrowers to additional risk if interest rates go
up or if credit becomes less available in general. Some ARMs have
relatively long introductory periods of five to seven years before the
contract interest rate increases. Other ARMs have short introductory
periods of one to three years. (16) With longer introductory periods,
borrowers have more time to build up equity in their homes before they
need to refinance to avoid the interest rate reset.
The percentage of subprime ARMs was 73 percent in 2004, 69 percent
in 2005, and 62 percent in 2006. By 2007, it had fallen to 39 percent,
since the availability of these types of loans declined in the second
half of the year. Importantly, nearly all subprime ARMs have
introductory periods of three years or less, which makes borrowers with
these loans very dependent on the ability to refinance. In contrast,
loans to prime borrowers are predominantly made as fixed-rate contracts
(about 75 percent of all prime loans), and the majority of prime ARMs
have introductory periods of five to seven years. The decline in the
share of ARMs in 2007, evident in both the prime and subprime markets,
mirrors the virtual disappearance of the securitization market for ARMs
with introductory periods of three years or less in the second half of
2007.
One off-mentioned culprit for the subprime crisis is the growth of
lenders that followed the "originate-to-distribute model"
(see, for example, Keys et al., 2010, and Calomiris, 2008). These
lenders sold virtually all of the mortgages they made, typically to
private securitizers. Because these lenders do not face a financial loss
if these mortgages eventually default, they have relatively little
incentive to screen and monitor borrowers. In addition to selling loans
to private securitizers, the lenders can hold loans on their own
portfolios or sell them to one of the GSEs. Only loans that meet certain
criteria (borrower with a FICO score of at least 620, loan value of
$417,000 or less, and an LTV of 80 percent or less) can generally be
sold to the GSEs. (17) Most subprime loans cannot be sold to GSEs and
must be either privately securitized or held on portfolio.
One of the striking facts in table 1 (p. 23) is the extent of loan
securitization. The LPS data overstate the actual extent of
securitization somewhat because the data are made up of loans serviced
by the large mortgage servicers (see note 7). It is more common for
smaller banks to hold loans on portfolio and also to service them
internally. Portfolio loans are therefore underrepresented in the LPS
data. (18) That being said, the LPS data indicate that within the first
month of origination, about half of prime mortgages made in 2004
remained in their originators' portfolios. This figure declined to
about 40 percent among the prime loans made in each of the subsequent
years in the data. The level of "rapid" securitization has
been consistently higher for subprime loans, whose originators retained
just over 40 percent of loans made in 2004 and less than 30 percent of
them made in the following years. The observed differences in the speed
of turning the loan over to outside investors do not translate to
differences in the extent of eventual securitization. Indeed, by the end
of the first year since origination, the share of loans kept on
portfolio drops to low single digits for both prime and subprime
mortgages. Not surprisingly, nearly all subprime mortgages are
securitized by private investors, and GSEs dominate the securitization
of prime mortgages. However, by the second half of 2007, the private
securitization market had all but disappeared. The fraction of subprime
loans originated in 2007 that were privately securitized was just 55
percent, with most of these loans being made in the first half of the
year. The GSEs took up much of the slack, accounting for about 40
percent of all subprime securitizations. (19)
Estimates of default
In this section, we estimate empirical models of the likelihood
that a loan will default in its first 12 months. This allows us to
quantify which factors make default more or less likely and to examine
how the sensitivity to default varies over time and across prime and
subprime loans.
Econometric model
Mortgages can have multiple sources of risk--for example, low
credit quality, high loan-to-value ratios, and contract interest rates
that reset shortly after origination. To take into account these and
other factors that might influence default rates, we estimate a number
of multivariate regression models that allow us to examine the effect of
varying one risk factor while holding others fixed.
The analysis sample includes loans that do not default and are
observed for 12 months after origination and loans that default (become
60 days or more past due) within 12 months of origination. We drop
nondefaulting loans that we do not observe for at least 12 months from
the sample. In effect we are dropping loans for one of three reasons:
The loan was transferred to a different mortgage servicer, the loan was
refinanced in its first 12 months, or we did not have complete data for
the loan. For prime and subprime loans originated in 2004--06, between
13 percent and 16 percent of loans were eliminated for one of these
reasons. For loans originated in 2007, the fraction of loans eliminated
fell to 7.5 percent of subprime loans and 8.6 percent of prime loans.
Among loans that were eliminated, the most common reason was
refinancing. (20) On the one hand, this is a concern for the analysis
because loans that refinance within 12 months of origination may differ
systematically from other loans. The most striking difference that we
observe is that the loans that refinance "early" tend to be in
areas that experienced higher-than-average home price growth. This
suggests that we may be dropping some potentially risky loans from the
analysis, since the areas that saw the greatest home price growth were
often the ones that saw the greatest eventual declines in home prices.
It is also important to keep in mind that some of the new loans on these
properties are probably included in the analysis, since the new loan may
have met the criteria for staying in the sample. On the other hand,
keeping early refinanced and transferred loans in the sample would
understate the share of actual defaults, since by definition these loans
are current for the duration of their (short) presence in the sample.
Our goal is to evaluate the relative strength of associations
between loan default and observable borrower, loan, and macroeconomic
characteristics in different market segments and different years. To
that end, we estimate the following regression:
1) Prob [(default within 12 months).sub.ijk] = [PHI]([[beta].sub.1]
[Loan.sub.ijk], [[beta].sub.2][Borrower.sub.ijk],
[[beta].sub.3][Econ.sub.jk], [[beta].sub.4][D.sub.k]).
The dependent variable is an indicator of whether a loan to
borrower i, originated in an MSAj in state k defaulted within the first
12 months. Default is defined as being 60 days or more past due. We
model this probability as a function of loan and borrower
characteristics, MSA-level economic variables (unemployment, home price
appreciation, and income), and a set of state dummy variables
([D.sub.k]) that capture aspects of the economic and regulatory
environment that vary at the state level. We estimate the model as a
standard maximum likelihood probit with state fixed effects. (21)
To retain maximum flexibility in evaluating the importance of
covariates for prime and subprime defaults, we carry out separate
estimations of equation 1 for prime and subprime loans. To achieve
similar flexibility over time, we further subdivide each of the prime
and subprime samples by year of origination (2004 through 2007).
The economic variables include both the realized growth in the FHFA
HPI and the average realized unemployment rate. Both of these variables
are measured at the MSA level, and both are computed over the 12 months
after loan origination. Consequently, they match the period over which
we are tracking loan performance. In contrast to all of the other
regressors, this information clearly would not be available to the
analyst at the time of loan origination. We can think of the model
described in equation 1 as the sort of analysis one would be able to do
for 2004 loans at the end of 2005. At this point, one would be able to
observe what happened to home prices and unemployment rates over the
same period. The same exercise can be performed for loans originated in
2005 at the end of 2006, for loans originated in 2006 at the end of
2007, and so on.
This is a different exercise than trying to forecast whether or not
a loan will default based on its characteristics at the time of its
origination. Instead, this framework allows us to explore whether the
abrupt reversal in home price appreciation contributed much to the
explosion in defaults on loans originated in 2006 and 2007. As shown in
figure 2 (p. 20) and table 1 (p. 23), growth in home price rates varies
enormously over the four years of our sample period. The 2004 figure of
13.44 percent home price growth for prime loans represents the average
realized 12-month HPI growth rate for loans originated in
January-December of 2004. As such, it averages 12-month home price
appreciation over two years (2004 and 2005) for a nationally
representative sample of prime mortgages. By 2006, these growth rates
fall below 2 percent, and then turn negative in 2007. The realized price
appreciation (and depreciation) of homes financed through subprime loans
shown in table 1 (p. 23) is remarkably similar to the values of homes
financed through prime loans. Subprime mortgage defaults have been
associated with parts of the country where home prices grew very fast
and then declined even more rapidly (for example, California, Florida,
and Arizona). On average, however, subprime and prime mortgages appear
to have been made in similar locations, so we do not observe large
differences in home price growth across the two loan categories. This
means that our analysis examines how different market segments responded
to fairly similar shocks to home values. In contrast with HPI growth,
unemployment rates showed little variation over time or across prime and
subprime loan groups.
Results
The results of the estimation are summarized in table 2. The first
four columns of data depict estimates for prime loans originated in each
of the four sample years, and the next four columns contain the
estimates for subprime loans. The juxtaposition of the data for the two
market segments allows us to easily compare the importance of certain
factors. The table presents estimates of the marginal effects of the
explanatory variables, rather than the coefficients themselves. The
marginal effects tell us how a one-unit change in each explanatory
variable changes the probability that a loan defaults in its first 12
months, holding fixed the impact of the other explanatory variables. For
dummy variables, the marginal effects show the change in the probability
of default when the variable in question goes from zero to one.
The defaults of both prime and subprime loans are strongly
associated with a number of key loan and borrower characteristics. These
include the FICO score, the LTV, and the interest rate at origination.
These variables are strongly statistically significant in virtually
every estimation year for each loan type. For instance, higher FICO
scores are strongly associated with lower default probabilities. For
prime loans, an increase of 100 points in the FICO score in 2004 is
associated with about a 120-basis-point decrease in default likelihood
(the estimated marginal effect of-0.00012, in the first column, sixth
row of table 2, multiplied by 100). The same result is obtained for
2005. The point estimates of marginal effects for 2006 and 2007 increase
about twofold for prime loans, but so does the baseline sample default
rate. For subprime loans, the estimated marginal effects are a full
order of magnitude higher, implying that the same improvement in FICO
scores generates a greater decline in subprime defaults, at least in
absolute terms.
Similarly, higher LTV values have a strong positive association
with defaults for both loan types originated in 2005, 2006, and 2007.
For subprime loans, a rise in LTV generates a stronger absolute increase
in loan defaults. It must be noted that the effect of the leverage on
the likelihood of default may be understated by the LTV measure that we
have. A better measure of how leveraged a borrower is on a given
property would be the combined loan-to-value ratio (CLTV). The CLTV
takes into account second-lien loans on the property in computing the
ratio of indebtedness to the value of the underlying collateral. This
variable is not available in the LPS data, however. If the practice of
obtaining such "piggyback loans" is more prevalent in the
subprime market, then the estimated coefficient for LTV for subprime
loans may be biased downward.
At first glance, the interest rate at origination is similar to LTV
and FICO score in having a strong statistical and economic effect on
both prime and subprime loan defaults in each origination year. What
stands out is the sheer magnitude of the estimated effects. However, one
must be cautious in interpreting hypothetical marginal effects of the
interest rate. While LTV and FICO score cover fairly wide ranges for
both prime and subprime loans, interest rate values are more tightly
distributed. (22) For example, the standard deviation of interest rates
on prime loans across all years of our sample is 81 basis points: A one
standard deviation increase in the interest rate for prime loans would
raise the average rate from 6.25 percent to 7.06 percent. The equivalent
one standard deviation increase in interest rates for subprime loans
would raise the average rate from 7.93 percent to 9.24 percent. If we
see two loans with otherwise identical characteristics but one has a
higher interest rate, a likely explanation is that the lender has
additional information about the credit quality of the borrower and is
charging a higher interest rate to take into account additional risk
factors, over and above those that are captured by the borrower's
FICO score.
There are also a number of notable differences between the prime
and subprime samples. Perhaps the most interesting finding is the
different sensitivity of defaults to changes in home prices. For
subprime loans, defaults are much lower when home price growth is higher
for three out of the four sample years. This relationship is
particularly striking for 2006 loan originations, many of which
experienced home price declines over their first 12 months. For prime
loans, 2006 is the only year of origination in which changes in home
prices are significantly correlated with loan defaults. These results
suggest that, relative to subprime defaults, prime defaults have a
weaker relationship with home prices, once key borrower and loan
characteristics (LTV, FICO score, and so on) are taken into account.
The contrast between prime and subprime loans is even sharper for
the debt-to-income ratio and loan margin rate. The DTI is widely
considered to be one of the main determinants of loan affordability,
since it relates household monthly income to debt service flows. The DTI
for prime loans is not significantly correlated with defaults, except
for loans originated in 2007, but it matters consistently for subprime
loans. The absence of any measurable effects of DTI even on defaults of
prime loans originated in 2006 can be interpreted as a sign of the
resilience of prime borrowers who experienced significant changes in the
prices of their homes.
The loan margin rate is one of the key terms in an ARM contract. It
defines the spread to a reference rate (usually the London interbank
offered rate, or Libor). At reset, the ARM's interest rate goes up
to the sum of Libor and the loan margin. The margin is set by the
lender, and is often thought to capture additional aspects of a
borrower's creditworthiness. This is consistent with the fact that
the margin rate is, on average, substantially higher for subprime
borrowers (see table l, p. 23). We find that this variable has no
association with defaults among prime loans, with the exception of loans
originated in 2006. In contrast, defaults on subprime loans originated
in every year except 2007 are significantly higher for loans with higher
margin rates, all else being equal. This suggests that, for the subprime
borrower, the margin rate contains additional information on borrower
quality not reflected in FICO scores and other loan characteristics. It
is also interesting that ARMs with introductory periods of three years
or less--the most common mortgage contract in the subprime market--have
the same correlation with subprime defaults as fixed-rate mortgages do.
Put differently, once loan and borrower characteristics are accounted
for, the choice of an ARM with a short introductory period is not
associated with higher subprime defaults.
Several other results merit comment. For prime loans, being
securitized within one month of origination (as opposed to remaining in
the lender's portfolio) is associated with lower defaults for loans
made in 2004, 2005, and 2006. This does not necessarily mean that the
securitization process has been successful in identifying loans of
higher quality. Since nearly all loans in the sample are securitized
over the 12-month default horizon (see table 1, p. 23), the difference
in defaults probably captures differences between fast-to-securitize and
slow-to-securitize originators, rather than differences between
securitized loans and those held on portfolio. We find similar results
for securitized subprime loans made in 2005, 2006, and 2007. Subprime
loans also have extremely high rates of eventual securitization, and the
relationship between subprime default and securitization can be
interpreted in the same way. This hints at the possibility that
originators with business models focused on securitization are better at
screening loan quality. These originators would have more to lose if
their reputations were damaged by weak ex post performance of the loans
they originated.
Finally, we note that purchase loans, as opposed to refinance
loans, are consistently associated with higher defaults in nearly all
sample years, in both the prime and subprime market segments. This may
seem surprising, since borrowers who extract equity from their homes in
cash-out refinancings may be particularly vulnerable to economic shocks
and experience higher defaults as a result. However, not all cash-out
refinancing is done by borrowers who need to finance current
consumption. Since a prerequisite for any cash-out transaction is the
availability of positive home equity, these borrowers, as a group, may
have a greater incentive to avoid defaults to maintain this option for
the future. In addition, purchase loans include first-time home buyers
who, in retrospect, were buying houses at the peak of the market and had
little time to build up equity to offset subsequent price declines.
Comparisons across years and across loan types
Since table 2 contains regression estimates from multiple
nonoverlapping samples, the comparison of the relative importance of the
explanatory variables can be tricky. The distribution of loan
characteristics varies from year to year and across prime and subprime
loans. In addition, the baseline rates of actual defaults are quite
different across samples. Because of this, one cannot simply compare two
point estimates and conclude that a bigger one indicates a stronger
correspondence with defaults.
To compare the relative importance of the explanatory variables
across the samples, we conduct the following exercise. For each
independent variable, we change its value for each observation by a
specified increment. Then, we compute the predicted sample default rate
using estimated coefficients for each year of origination and loan type.
We compare the new predicted default to the original one. The difference
between the original prediction and the new one tells us the marginal
contribution of that variable to the overall default rate. (23) We
compare these figures across years and across prime loans (table 3,
panel A) and subprime loans (table 3, panel B). For example, for 2004
prime loans we increase all FICO scores by 50 points, predict a new
default rate, and compare it to the original default rate. The
difference is -1.16 percentage points, or a 53 percent decrease in the
likelihood of default for loans originated in 2004 (fourth column,
second full row of table 3, panel A). For brevity, we look at just six
key explanatory variables: HPI growth, FICO score, LTV, DTI, interest
rate at origination, and loan margin rate. (24) Table 3 also reports the
mean of the relevant variable, its standard deviation, and the absolute
change that we impose. We tried to keep the magnitude of the absolute
changes reasonably close to the standard deviations.
An increase of 10 percentage points in home price appreciation (HPI
growth) substantially lowers default probabilities (first full row of
each panel in table 3). This effect is more consistent for subprime
loans originated in various years, where it translates to decreases of
between 10 percent and 18 percent relative to the baseline default rate
in 2004, 2005, and 2006. For prime loans, the 10-percentage-point
increase in the HPI has a big effect only for loans originated in 2006,
where the estimates imply that defaults would have been 1.78 percentage
points, or 42 percent, lower. The effect of FICO score stands out. A
50-point uniform increase in FICO scores (second full row of each panel)
is associated with a 41 percent to 53 percent decline in predicted
default rates relative to the baseline for prime loans, and a 20 percent
to 34 percent relative decline for subprime loans. The average marginal
effects of the LTV are greater (in a relative sense) for prime loans
than for subprime loans. (25) Finally, higher interest rates at
origination appear to generate incredible increases in defaults for both
market segments. For instance, a 1 percentage point increase in interest
rates translates into a jump in defaults on 2007 prime loans of more
than 3 percentage points--a rise of 66 percent relative to the actual
default rate. Increasing everyone's interest rates by 1 percentage
point is equivalent to a substantial deterioration in the quality of the
borrower pool, and thus translates into much higher predicted defaults.
As mentioned earlier, the DTI and the margin rate do not have strong
associations with prime mortgage defaults. In contrast, higher values of
these variables consistently indicate higher default rates for subprime
mortgages. However, the economic magnitude of marginal effects of DTI
and the margin rate on defaults (fourth and sixth full rows of each
panel) is somewhat muted.
What if?
In this section, we use the estimates discussed previously to do
two things. First, we examine how much (or how little) of the increase
in mortgage defaults from 2004 through 2007 can be explained by changes
in the characteristics of loans and borrowers, as opposed to changes in
the responsiveness of defaults to those characteristics. Next, we
examine how forecasts of prime and subprime mortgage defaults vary with
different assumptions about the future path of home prices.
Predicted versus actual defaults
The descriptive regressions in the previous section provide
insights into the factors that are associated with realized defaults for
different types of loans originated in different years. An open question
is how useful these relationships could have been in forecasting the
defaults of future loans. To address this, we conduct the following
thought experiment. For each set of estimates, we compute predicted
defaults using observed loan, borrower, and economic characteristics
from other origination years. For instance, we take the relationship
between borrower characteristics and loan characteristics that we
estimate using data from prime loans originated in 2004 (the marginal
effects reported in the first column of table 2, pp. 28-29) and see what
it would imply for defaults for prime loans originated in 2007. In other
words, we pick a particular year and fix the relationship (that is, the
estimated coefficients) between defaults and characteristics, but let
the characteristics vary as they actually did in the data. This exercise
lets us see to what extent higher defaults on loans originated in 2007
can be explained by changes in characteristics alone. We show these
results in table 4, where this exercise is carried out separately for
prime loans (panels A and B) and subprime loans (panels C and D). Panels
A and C of the table show the predicted default rate for loans
originated in each of the sample years (rows) using the coefficients
estimated with data from each of the sample years (columns). The numbers
in bold that run diagonally through the panels are predictions that use
characteristics and coefficients from the same year. A useful way to
summarize the results is to look at predictions below and above the
diagonal bold entries. Those below the diagonal bold entries are
forecasts of future defaults using historical models (for example,
relationships between characteristics and defaults from 2004 and
characteristics from 2007). Those above the diagonal run this exercise
in reverse--they apply recent model estimates to loans from earlier
years (for example, relationships between characteristics and defaults
from 2007 and characteristics from 2004). These two groups of
predictions are strikingly different in terms of their predictions
relative to the actual realized default rates.
Perhaps the easiest way to see this is in panels B and D of table
4, which express predicted defaults as a percentage of realized
defaults. The diagonal forecasts in those two panels are close to 100
percent. This is a feature of the estimation procedure. However,
forward-looking forecasts--those below the diagonal bold entries--are
nearly always less than 100 percent for both prime and subprime loans.
(26) This means that predictions based on relationships between
characteristics and defaults from 2004 and 2005 uniformly underpredict
defaults in 2006 and 2007. The underprediction is more dramatic for
prime loans, where less than 75 percent of the realized defaults in 2006
and 2007 are accounted for by the "old" relationships between
characteristics and defaults. Even for subprime loans, the shortfall is
substantial, suggesting increased defaults cannot be accounted for by
changes in loan characteristics alone. The relationship between
observable loan and economic characteristics and defaults appears to
have changed for loans originated after 2005. The sharp rise in the
default rate cannot be explained just by looser underwriting standards
or by changes in the composition of loan contracts.
In contrast, the backward-looking forecasts--those above the
diagonal bold entries--are typically greater than 100 percent. This
means that the world described by defaults observed in 2006 and 2007
would have resulted in defaults higher than observed in 2004 and 2005.
This overprediction holds uniformly for subprime mortgages, but not for
the predictions based on the 2006 model of prime defaults. Moreover, the
overprediction is particularly severe for the 2007-based model
coefficients, again suggesting a structural difference in the
determinants of loan defaults that occurred after the rapid reversal in
home price appreciation.
What role do home prices play?
We turn our attention now to the role of home prices. We know that
home prices were increasing very rapidly in 2004 and 2005 and began to
fall quite dramatically beginning in 2006. A closer look at the
potential impact of this shock may help to illuminate why defaults of
both prime and subprime mortgages increased so much. We are also
interested in refining the discussion by being very clear about what
information would have been available to analysts at different points in
time. This will allow us to gauge the extent to which market
participants were "surprised" by the performance of prime and
subprime loans originated in 2006 and 2007.
The results in table 4 give the impression that an analyst would be
able to predict 2006 loan defaults, using 2005 model estimates. In
reality, this would not have been possible because the model of 2005
defaults could only be estimated in full in December 2006 when loans
made in December 2005 had been observed for a full 12 months. Moreover,
some of the key variables--notably, future growth in the HPI--are not of
course available at the time a loan is made. To create estimates of
default predictions that take into account only the available
information, we conduct a series of experiments that are reported in
table 5.
In panels A and B of table 5, the first row presents predicted
defaults for 2006, using coefficients from estimates of default for
loans made in 2004. The information necessary to do this exercise would
not have been available at the beginning of 2006. In panels A and B of
table 5, the second row does the same for 2007 defaults, using
coefficients from estimates of default for loans made in 2005. Each of
the scenarios uses the same coefficients (2004 or 2005), but differs in
assumptions about HPI growth. Scenario I assumes that the analyst can
perfectly predict future home prices (this simply restates the
appropriate value from table 4). Scenario II assumes that an analyst
forecasts that the MSA-specific HPI growth in a given year will change
by exactly as much as it did in the most recent observable 12-month
period. In other words, for a particular MSA, HPI growth in 2006 will
look just like it did in 2005, and this growth in 2007 will look just
like it did in 2006. In scenario HI, the hypothetical analyst becomes
very pessimistic and assumes that house prices in 2006 and 2007 will not
grow at all. Panel C of table 5 summarizes the average MSA-level HPI
growth rates assumed in each scenario for prime and subprime loans.
The results for prime loans (table 5, panel A) suggest that varying
assumptions about HPI growth has little effect on predicted default
rates. Whether one uses past experience to extrapolate future home price
growth or arbitrarily sets the growth rate to zero, the model
substantially underpredicts the actual default rate. The historical
experience in the prime market for loans originated in 2004 and 2005
suggested that there was essentially no relationship between home price
appreciation and loan defaults. Using the experience of mortgages
originated during this period together with any assumption about HPI
growth would not have helped analysts forecast the spike in prime loan
defaults. What would have been helpful would have been if an analyst
could have foreseen that prime mortgages might respond to home prices
the way subprime mortgages did.
Even when home price appreciation was relatively high during 2004
and 2005, default rates among subprime borrowers were quite sensitive to
home prices. Indeed, the results for the subprime loans, shown in table
5, panel B, suggest that assuming zero price growth would have gone a
long way in closing the gap between forecasts based on simple
extrapolation of past growth and actual defaults. This is especially
true for loans originated during 2006, which proved to be the pivotal
year for home prices and loan performance. For those loans,
extrapolation of past trends using coefficients from estimates of
default for loans made in 2004 predicted default rates of 17.5 percent.
With the assumption of zero growth, the same model produced default
rates of 20.3 percent, much closer to the actual rate of 24 percent.
Even relying on the somewhat stale coefficient estimates, it appears to
have been possible to forecast a sharp deterioration in default rates on
subprime loans with fairly mild HPI growth assumptions. However, the
same cannot be said for prime loans.
Conclusion
We have analyzed the default experience of prime and subprime loans
originated over the period 2004--07. Similar to other studies, we
document some decline in underwriting standards during this period for
both prime and subprime loans. We also find that characteristics such as
the loan-to-value ratio, FICO score, and interest rate at origination
are important predictors of defaults for both prime and subprime loans.
However, changes in loan and borrower characteristics are not enough to
have predicted the incredible increase we have seen in prime and
subprime mortgage defaults. While changes in borrower and loan
characteristics can get us closer to observed default rates for subprime
loans than they can for prime loans, for both market segments there were
other factors at work.
Home prices play a very important role in determining mortgage
outcomes; this became particularly evident for subprime loans by the end
of 2005. For prime loans, it is only when we analyze data through the
end of 2007 (that is, evaluate the performance of loans originated in
2006) that we are able to document this sensitivity. Even very
pessimistic assumptions about the future path of home prices would not
have been enough to substantially improve contemporaneous forecasts of
prime mortgage defaults for loans made in 2006 and 2007. In hindsight,
of course, it appears self-evident that the relationships between HPI
growth and defaults on prime loans might be different in periods with
declining home prices. However, recognizing this in real time would not
have been possible using the available data from the recent past. It
could, perhaps, have been done by analyzing data that included earlier
episodes of substantial regional home price declines.
REFERENCES
Bajari, Patrick, Sean Chu, and Minjung Park, 2008, "An
empirical model of subprime mortgage default from 2000 to 2007,"
University of Minnesota, manuscript.
Calomiris, Charles, 2008, "The subprime turmoil: What's
old, what's new, and what's next," Columbia Business
School, working paper.
Chomsisengphet, Souphala, and Anthony Pennington-Cross, 2006,
"The evolution of the subprime mortgage market," Review,
Federal Reserve Bank of St. Louis, Vol. 88, No. l, January/February, pp.
31-56.
Dell'Ariccia, Giovanni, Deniz Igan, and Luc Laeven, 2008,
"Credit booms and lending standards: Evidence from the subprime
mortgage market," Center for Economic Policy Research, discussion
paper, No. DP6683, February.
Demyanyk, Yuliya, and Otto Van Hemert, 2009, "Understanding
the subprime mortgage crisis," Review of Financial Studies,
forthcoming.
DiMartino, Danielle, and John V. Duca, 2007, "The rise and
fall of subprime mortgages," Economic Letter, Federal Reserve Bank
of Dallas. Vol. 2, No. 11, November.
Foote, Christopher L., Kristopher S. Gerardi, Lorenz Goette, and
Paul S. Willen, 2008, "Just the facts: An initial analysis of
subprime's role in the housing crisis," Journal of Housing
Economics, Vol. 17, No. 4, December, pp. 291-305.
Gerardi, Kristopher S., Andreas Lehnert, Shane M. Sherlund, and
Paul S. Willen, 2008, "Making sense of the subprime crisis,"
Brookings Papers on Economic Activity, Vol. 2008, No. 2, pp. 69-159.
Gerardi, Kristopher S., Adam Hale Shapiro, and Paul S. Willen,
2008, "Subprime outcomes: Risky mortgages, homeownership
experiences, and foreclosures," Federal Reserve Bank of Boston,
working paper, No. 07-15, revised May 4, 2008.
International Monetary Fund, 2008, Global Financial Stability
Report: Containing Systemic Risks and Restoring Financial Soundness,
World Economic and Financial Surveys, Washington, DC, April, available
at www.imf.org/external/pubs/ft/gfsr/2008/01/pdf/ text.pdf.
Keys, Benjamin J., Tanmoy K. Mukherjee, Amit Seru, and Vikrant Vig,
2010, "Did securitization lead to lax screening? Evidence from
subprime loans," Quarterly Journal of Economics, Vol. 125, No. 1,
forthcoming.
Mian, Atif R., and Amir Sufi, 2009, "The consequences of
mortgage credit expansion: Evidence from the U.S. mortgage default
crisis," Quarterly Journal of Economics, Vol. 124, No. 4,
forthcoming.
NOTES
(1) These numbers are based on authors' calculations using
data from Lender Processing Services (LPS) Applied Analytics, described
in detail later in the article.
(2) This figure is equal to 75 percent of total home mortgage debt
outstanding of $11,1212 billion in the third quarter of 2008, reported
in table L.2, line 11 of the Board of Governors of the Federal Reserve
System's Z. I release, dated March 12, 2009, available at
www.federalreserve.gov/releases/z1/.
(3) We emphasize that these are estimates of the direct costs from
mortgage defaults, not the cost to society. These estimates use data
from the Mortgage Bankers Association on the fraction of prime and
subprime mortgages that are past due or in foreclosure as of the end of
the third quarter of 2008. For prime mortgages, 4 34 percent are 30 days
past due. 2.11 percent of prime mortgages are 60 days or more past due,
and 1.58 percent are in foreclosure. For subprime mortgages, the
analogous figures are 20.03 percent, 11.47 percent, and 12.55 percent.
See wwwmortgagebankers.org/NewsandMedia/ PressCenter/66626.htm We assume
that 70 percent of subprime mortgages that are 60 days or more past due
will eventually go into foreclosure and that 40 percent of subprime
loans that are 30 days past due will become seriously delinquent For
prime mortgages, we assume that 25 percent of loans that are 30 days
past due become seriously delinquent and that 50 percent of seriously
delinquent loans eventually foreclose. For both prime and subprime
mortgages, we assume that lenders lose 50 percent of the outstanding
value of the loan in foreclosure. These estimates are of course very
sensitive to the assumptions. If we assume that a higher fraction of
past due subprime loans eventually default compared with prime loans,
the difference in the loss amounts will be larger. Note also that these
estimates do not include any mark-to-market losses on securities
associated with the underlying mortgages.
(4) The full official name for Fannie Mae is the Federal National
Mortgage Association. The full official name for Freddie Mac is the
Federal Home Loan Mortgage Corporation.
(5) Note that we are abstracting from the potential role of private
mortgage insurers here.
(6) In our pessimistic forecasts, home prices stay flat in 2006 and
2007. Although such forecasts must have looked gloomy in 2005, the
actual experience for home prices turned out to be quite a bit worse,
particularly in 2007.
(7) The servicers included in the data set are those that
participate in the HOPE NOW alliance (www.hopenow.com/members.
html#mortgage). This alliance includes some of the country's
largest home lenders--Bank of America, Citibank, JPMorgan Chase, and
Wells Fargo.
(8) Alt-A loans are a middle category of loans--riskier than prime
and less risky than subprime. They are generally made to borrowers with
good credit ratings, but the loans have characteristics that make them
ineligible to be sold to the GSEs--for example, limited documentation of
the income or assets of the borrower or higher loanto-value ratios than
those specified by GSE limits.
(9) Note that the HMDA data do not represent the entire universe of
home mortgages originated in a given year either. The HMDA data include
all mortgages originated by lenders that have a home or branch office in
a metropolitan statistical area and exceed exemption thresholds on the
size and the number of home purchase or refinancing loans made in a
calendar year. For depository institutions, the threshold on asset size
is adjusted annually on the basis of changes in the Consumer Price Index
for Urban Wage Earners and Clerical Workers (CPI-W). For 2008 loan
reporting, it was set at $39 million. Also, for depository institutions,
the threshold for the number of loans is one per year. For nondepository
institutions, the threshold on asset size is set at $10 million, and the
threshold for the number of loans is 100 per year. In our comparison of
the LPS data with the HMDA data, we have dropped LPS loans made in zip
codes outside of an MSA. However, loans made outside of an MSA may still
be included in the HMDA data if the lender is based in an MSA
(10) Note that these numbers include loans made outside of an MSA.
(11) B loans might arguably be considered near prime, but for
convenience, we group both B and C loans together as subprime in this
article.
(12) As Bajari, Chu, and Park (2008) emphasize, an important
feature of the FICO score is that it measures a borrower's
creditworthiness prior to taking out the mortgage. FICO scores range
between 300 and 850. Typically, a FICO score above 800 is considered
very good, while a score below 620 is considered poor. As reported on
the Fair Isaac Corporation website (www.myfico.com), borrowers with FICO
scores above 760 are able to take out 30-year fixed-rate mortgages with
interest rates that are 160 basis points lower, on average, than those
available for borrowers with scores in the 620-639 range.
(13) If we repeat the analysis using alternative outcome variables
and different time periods (in default after 18 months, in foreclosure,
30 days or more past due, and so on), the results are very similar.
(14) As part of the Housing and Economic Recovery Act of 2008
(HERA), the Federal Housing Finance Regulatory Reform Act of 2008
established a single regulator, the FHFA, for GSEs involved in the home
mortgage market, namely, Fannie Mae, Freddie Mac, and the 12 Federal
Home Loan Banks. The FHFA was formed by a merger of the Office of
Federal Housing Enterprise Oversight, the Federal Housing Finance Board
(FHFB), and the U.S. Department of Housing and Urban Development's
government-sponsored enterprise mission team (see www.fhfa.gov for
additional details).
(15) Note that we are looking at a relatively short period, and
other researchers document changes in underwriting criteria that
occurred prior to 2004 (see, for example, Gerardi et al., 2008).
(16) Such mortgages are known as "hybrid ARMs." They are
also commonly identified as "2/28" and "3/27" loans,
referring to 30-year ARMs that reset after two and three years,
respectively.
(17) The maximum dollar value for loans that can be securitized by
GSEs (the conforming loan limit) is set annually by the Federal Housing
Finance Agency. Prior to 2008, the conforming loan limit was set at the
same level throughout most of the country. (In Alaska and Hawaii, the
limit is equal to 150 percent of the limit for the other states.) The
Housing and Economic Recovery Act of 2008 allowed the limit to increase
to 115 percent of local median housing prices, not to exceed 150 percent
of the standard loan limit of $417,000. This provision affected a number
of counties in certain high-cost areas.
(18) As of December 2008, the LPS data are estimated to cover about
18 percent of the total value of loans held on portfolio.
(19) In September 2007, when the private securitization market had
all but shut down, the GSEs were encouraged by members of Congress to
expand their portfolios to support the market; see the correspondence
from James B. Lockhart, the director of the OFHEO, to Senator Charles E
Schumer (D-NY), at www fhfa.
gov/webfiles/2298/Schumerletter810attachment.pdf
(20) Early refinancings were more common among loans originated in
2004 and 2007. Consequently, early refinancings accounted for the vast
majority of loans dropped from the sample--about two-thirds of all
dropped prime loans originated in 2004 and 2007 (and more than 90
percent of all dropped subprime loans) By comparison, among the loans
eliminated from the sample, just over 50 percent of the prime loans
originated in 2005 and 2006 (and 78 percent of the subprime loans) were
dropped for this reason.
(21) In effect, we are taking a slice through the two panels in
figure 1 (p. 19) at 12 months. An alternative modeling approach would be
to estimate loan-level time to default as a function of a similar array
of characteristics, using a proportional hazard model. This approach is
chosen by Demyanyk and Van Hemert (2009) and Gerardi et al. (2008),
among others. As there are few time-varying covariates, the results from
a straightforward probit model are likely to be qualitatively similar.
(22) Keep in mind that, for simplicity, the analysis uses the
actual interest rate at loan origination and not the difference between
this rate and some reference risk-free rate.
(23) This exercise amounts to computing the average of marginal
effects for individual loans, instead of the marginal effect at the
mean, which is obtained by multiplying a hypothetical change in an
explanatory variable by its regression coefficient
(24) The loan margin is increased for only ARMs, since fixed-rate
loans by definition have a zero margin under all circumstances
Similarly, we incremented DTI for only those loans that had nonmissing
DTI values.
(25) As discussed earlier, this may be due to our inability to
account for piggyback loans.
(26) Forecasts of 2007 loan defaults using 2006 model coefficients
are the only exceptions.
Gene Amromin is a senior financial economist in the Financial
Markets Group and Anna L. Paulson is a senior financial economist in the
Economic Research Department at the Federal Reserve Bank of Chicago. The
authors are grateful to Leslie McGranahan for very helpful feedback and
to Edward Zhong and Arpit Gupta for excellent research assistance.
TABLE 1
Mortgage characteristics
Prime mortgages
2004 2005
Default (%) first 12 months 2.43 2.39
Default (% first 18 months 3.90 3.74
Default (% first 21 months 5.11 4.91
HPI growth (%), 12 months since origination 13.44 9.10
Unemployment rate (%), 12 months since origination 5.15 4.70
Median annual income in zip code ($) 50,065 49,486
Origination amount ($) 173,702 200,383
FICO score 710 715
LTV (%) 75.92 74.89
DTI, if nonmissing (%) 35.95 37.87
DTI missing (% of loans) 52.80 32.10
Interest rate at origination (%) 5.60 6.00
Margin rate for ARMS (%) 2.30 2.40
Share (%) of loans that are:
ARMS 26.45 26.04
Reset > 3 years 14.52 13.32
Reset [less than or equal to] 3 years 11.93 12.71
Prepayment penalty 2.67 9.82
Purchase loans 44.89 50.12
Refinancing loans 40.51 41.92
Cash-out refinancing loans 12.19 20.65
Refinancing, no cash-out 6.69 1.93
Refinancing, unknown cash-out 21.63 19.35
Investment property loans 4.90 7.31
Conforming loans 60.68 66.50
1 month since origination (% of loans)
Loans sold to GSE 31.10 34.75
Loans sold to private securitizer 18.20 27.63
Loans held on portfolio 50.44 37.61
12 months since origination (% of loans)
Loans sold to GSE 74.17 70.72
Loans sold to private securitizer 19.08 23.73
Loans held on portfolio 6.75 5.55
Number of loans in the sample 11,604 18,388
Prime mortgages
2006 2007
Default (%) first 12 months 4.33 4.93
Default (% first 18 months 7.67 6.86
Default (% first 21 months 10.51 6.40
HPI growth (%), 12 months since origination 1.94 -4.19
Unemployment rate (%), 12 months since origination 4.45 4.80
Median annual income in zip code ($) 48,417 48,221
Origination amount ($) 211,052 205,881
FICO score 708 706
LTV (%) 75.99 77.75
DTI, if nonmissing (%) 37.25 38.74
DTI missing (% of loans) 27.60 20.80
Interest rate at origination (%) 6.70 6.50
Margin rate for ARMS (%) 2.90 2.70
Share (%) of loans that are:
ARMS 23.16 12.93
Reset > 3 years 12.11 10.38
Reset [less than or equal to] 3 years 11.05 2.55
Prepayment penalty 10.91 5.56
Purchase loans 53.33 49.68
Refinancing loans 40.70 45.44
Cash-out refinancing loans 20.85 20.97
Refinancing, no cash-out 1.26 2.14
Refinancing, unknown cash-out 18.59 22.32
Investment property loans 7.72 7.15
Conforming loans 66.18 57.82
1 month since origination (% of loans)
Loans sold to GSE 34.42 45.76
Loans sold to private securitizer 28.25 12.84
Loans held on portfolio 37.32 40.66
12 months since origination (% of loans)
Loans sold to GSE 72.30 82.83
Loans sold to private securitizer 23.08 10.56
Loans held on portfolio 4.56 6.40
Number of loans in the sample 15,992 15,039
Subprime mortgage
2004 2005
Default (%) first 12 months 11.19 16.22
Default (% first 18 months 15.92 23.35
Default (% first 21 months 23.35 31.72
HPI growth (%), 12 months since origination 13.99 9.70
Unemployment rate (%), 12 months since origination 5.28 4.83
Median annual income in zip code ($) 45,980 44,965
Origination amount ($) 167,742 172,316
FICO score 617 611
LTV (%) 79.63 80.69
DTI, if nonmissing (%) 39.55 38.35
DTI missing (% of loans) 41.00 30.90
Interest rate at origination (%) 7.10 7.50
Margin rate for ARMS (%) 5.20 5.40
Share (%) of loans that are:
ARMS 73.31 69.49
Reset > 3 years 1.05 0.96
Reset [less than or equal to] 3 years 72.26 68.53
Prepayment penalty 70.98 75.42
Purchase loans 41.12 43.47
Refinancing loans 53.83 53.65
Cash-out refinancing loans 35.03 42.95
Refinancing, no cash-out 0.27 0.65
Refinancing, unknown cash-out 18.53 10.05
Investment property loans 2.82 3.82
Conforming loans 28.89 24.47
1 month since origination (% of loans)
Loans sold to GSE 3.34 4.22
Loans sold to private securitizer 53.65 66.51
Loans held on portfolio 43.01 29.27
12 months since origination (% of loans)
Loans sold to GSE 4.09 5.76
Loans sold to private securitizer 90.92 91.89
Loans held on portfolio 4.99 2.35
Number of loans in the sample 6,889 20,778
Subprime mortgage
2006 2007
Default (%) first 12 months 23.79 25.48
Default (% first 18 months 34.91 33.87
Default (% first 21 months 43.75 32.15
HPI growth (%), 12 months since origination 1.52 -3.94
Unemployment rate (%), 12 months since origination 4.55 4.81
Median annual income in zip code ($) 43,790 43,817
Origination amount ($) 179,003 172,667
FICO score 607 597
LTV (%) 80.40 80.56
DTI, if nonmissing (%) 39.78 40.72
DTI missing (% of loans) 27.20 8.00
Interest rate at origination (%) 8.50 8.40
Margin rate for ARMS (%) 5.50 5.30
Share (%) of loans that are:
ARMS 61.78 38.92
Reset > 3 years 1.93 6.63
Reset [less than or equal to] 3 years 59.85 32.28
Prepayment penalty 73.70 48.52
Purchase loans 40.21 29.46
Refinancing loans 57.34 70.04
Cash-out refinancing loans 46.59 57.47
Refinancing, no cash-out 0.80 0.16
Refinancing, unknown cash-out 9.95 12.41
Investment property loans 4.24 3.17
Conforming loans 23.67 13.40
1 month since origination (% of loans)
Loans sold to GSE 5.96 32.52
Loans sold to private securitizer 64.07 42.60
Loans held on portfolio 29.96 24.88
12 months since origination (% of loans)
Loans sold to GSE 8.63 40.12
Loans sold to private securitizer 88.59 55.06
Loans held on portfolio 2.78 4.81
Number of loans in the sample 18,189 8,562
Notes: All reported statistics are subject to rounding. For
definitions of the variables, see box 1 on p. 22.
Source: Authors' calculations based on data from Lender
Processing Services (LPS) Applied Analytics.
TABLE 2
Probability of defaulting within 12 months of mortgage origination
Marginal effects (dF/dx)
Prime mortgages
Variables 2004 2005
Estimation sample mean 0.0221 0.0217
of default rate
HPI growth -0.00166 -0.00494
(0.0139) (0.0109)
Unemployment rate -0.0104 0.222 ***
(0.0507) (0.0393)
Median annual income -0.00149 -0.00253
in zip code (0.00428) (0.00388)
Origination amount -0.000122 0.000176
(0.000387) (0.000425)
FICO score -0.00012 *** -0.00012 ***
(1.64e-05) (1.64e-05)
LTV 0.00433 0.0144 ***
(0.00528) (0.00500)
DTI (0 if missing) 0.000502 -0.00160
(0.00409) (0.00350)
DTI missing dummy 0.00248 0.00148
(0.00218) (0.00187)
Interest rate at 0.337 *** 0.257 **
origination (0.103) (0.107)
ARMS with reset > 3 years 0.00151 -0.00366
dummy (0.00685) (0.00228)
ARMS with reset [less than 0.00180 -0.00566 ***
or equal to] 3 years dummy (0.00687) (0.00204)
Margin rate -0.192 0.150
(0 if FIRM) (0.245) (0.118)
Prepayment penalty 0.00286 0.00374
dummy (0.00572) (0.00325)
Cash-out refinancing 0.00274 0.00445 **
dummy (0.00229) (0.00222)
Purchase loan dummy 0.0029 * 0.00222 *
(0.00151) (0.00131)
Investment property loan -0.000422 0.00468
dummy (0.00305) (0.00304)
Conforming loan dummy -0.00290 -0.00667 ***
(0.00190) (0.00184)
GSE-securitized dummy -0.0113 *** -0.00623 ***
(0.00268) (0.00225)
Private-label-securitized -0.00578 ** -0.000475
dummy (0.00269) (0.00207)
Observations 8,887 15,653
R-squared 0.2587 0.2364
Marginal effects (dF/dx)
Prime mortgages
Variables 2006 2007
Estimation sample mean 0.0423 0.0483
of default rate
HPI growth -0.137 *** -0.00356
(0.0294) (0.0243)
Unemployment rate -0.0370 0.131
(0.115) (0.0968)
Median annual income -0.00689 -0.0231 ***
in zip code (0.00772) (0.00880)
Origination amount 0.000830 0.0019 **
(0.000806) (0.000758)
FICO score -0.000262 *** -0.000318 ***
(1.97e-05) (2.26e-05)
LTV 0.0636 *** 0.0814 ***
(0.0109) (0.0123)
DTI (0 if missing) 0.00841 0.0343 ***
(0.00859) (0.00845)
DTI missing dummy 0.00974 ** 0.0119 **
(0.00491) (0.00586)
Interest rate at 1.351 *** 1.653 ***
origination (0.186) (0.218)
ARMS with reset > 3 years -0.00112 0.0358 **
dummy (0.00485) (0.0183)
ARMS with reset [less than -0.014 *** 0.0462
or equal to] 3 years dummy (0.00364) (0.0317)
Margin rate 0.322 ** -0.165
(0 if FIRM) (0.147) (0.340)
Prepayment penalty 0.00757 -0.00390
dummy (0.00467) (0.00476)
Cash-out refinancing 0.000601 -0.00157
dummy (0.00318) (0.00340)
Purchase loan dummy 0.00552 ** 0.00604 **
(0.00244) (0.00295)
Investment property loan 0.000339 0.00159
dummy (0.00378) (0.00486)
Conforming loan dummy -4.45e-05 0.00166
(0.00288) (0.00344)
GSE-securitized dummy -0.019 *** -0.00352
(0.00417) (0.00530)
Private-label-securitized -0.0093 ** -0.000801
dummy (0.00424) (0.00635)
Observations 13,941 12,932
R-squared 0.1997 0.1962
Marginal effects (dF/dx)
Subprime mortgages
Variables 2004 2005
Estimation sample mean 0.1076 0.1572
of default rate
HPI growth -0.183 * -0.168 ***
(0.0981) (0.0500)
Unemployment rate 0.218 0.743 ***
(0.324) (0.239)
Median annual income -0.0398 -0.0572 ***
in zip code (0.0281) (0.0215)
Origination amount 0.0135 *** 0.016 ***
(0.00436) (0.00341)
FICO score -0.000733 *** -0.00122 ***
(9.12e-05) (6.84e-05)
LTV 0.0523 0.082 ***
(0.0368) (0.0255)
DTI (0 if missing) 0.0876 ** 0.143 ***
(0.0399) (0.0302)
DTI missing dummy 0.0265 0.0593 ***
(0.0191) (0.0148)
Interest rate at 2.487 *** 3.092 ***
origination (0.388) (0.282)
ARMS with reset > 3 years 0.0288 -0.0328
dummy (0.0609) (0.0313)
ARMS with reset [less than -0.0345 0.00391
or equal to] 3 years dummy (0.0321) (0.0200)
Margin rate 1.235 ** 0.776 **
(0 if FIRM) (0.521) (0.363)
Prepayment penalty 0.00369 0.0109
dummy (0.00894) (0.00753)
Cash-out refinancing -0.00985 -0.0186 **
dummy (0.0110) (0.00915)
Purchase loan dummy 0.0243 *** 0.0415 ***
(0.00863) (0.00602)
Investment property loan -0.00612 -0.00102
dummy (0.0216) (0.0143)
Conforming loan dummy 0.0154 0.0226 ***
(0.0120) (0.00864)
GSE-securitized dummy 0.0255 -0.0312
(0.0271) (0.0194)
Private-label-securitized 0.00680 -0.086 ***
dummy (0.0183) (0.0148)
Observations 5,825 19,356
R-squared 0.1138 0.0934
Marginal effects (dF/dx)
Subprime mortgages
Variables 2006 2007
Estimation sample mean 0.2399 0.2539
of default rate
HPI growth -0.447 *** -0.0105
(0.0934) (0.121)
Unemployment rate -0.749 ** -0.476
(0.346) (0.503)
Median annual income -0.0942 *** -0.0672
in zip code (0.0299) (0.0412)
Origination amount 0.0247 *** 0.0331 ***
(0.00598) (0.00627)
FICO score -0.00131 *** -0.00116 ***
(9.10e-05) (0.000136)
LTV 0.193 *** 0.193 ***
(0.0330) (0.0477)
DTI (0 if missing) 0.0900 ** 0.106 **
(0.0388) (0.0446)
DTI missing dummy 0.0183 0.000879
(0.0180) (0.0258)
Interest rate at 4.692 *** 5.384 ***
origination (0.345) (0.476)
ARMS with reset > 3 years -0.0956 *** 0.187
dummy (0.0294) (0.133)
ARMS with reset [less than -0.0197 0.203 *
or equal to] 3 years dummy (0.0325) (0.121)
Margin rate 1.841 *** -2.483
(0 if FIRM) (0.568) (2.014)
Prepayment penalty -0.0189 * 0.00180
dummy (0.0111) (0.0159)
Cash-out refinancing -0.0284 ** -0.0117
dummy (0.0122) (0.0164)
Purchase loan dummy 0.0856 *** 0.0729 ***
(0.00815) (0.0132)
Investment property loan -0.00164 0.0446
dummy (0.0163) (0.0301)
Conforming loan dummy 0.0196 * 0.0119
(0.0107) (0.0184)
GSE-securitized dummy -0.138 *** -0.00455
(0.0284) (0.0294)
Private-label-securitized -0.168 *** -0.0619 **
dummy (0.0228) (0.0241)
Observations 17,359 8,349
R-squared 0.0926 0.0745
* Significant at the 10 percent level.
** Significant at the 5 percent level.
*** Significant at the 1 percent level.
Notes: These are probit regressions with state fixed effects.
Standard errors are in parentheses. The securitization status
(GSE or private label) is measured during the first month since
origination. For definitions of the variables, see box 1 on p.
22.
Source: Authors' calculations based on data from Lender
Processing Services (LPS) Applied Analytics.
TABLE 3
Average marginal effect on mortgage default rates from changes in key
explanatory variables
2004-07
Standard
Variables Mean deviation Change (+)
A. Prime mortgages
Baseline predicted default rate
HPI growth 4.80 10.30 10 ppts
Percent change in default likelihood
FICO score 710 62 50 points
Percent change in default likelihood
LTV 76.10 16.80 10 ppts
Percent change in default likelihood
DTI, if nonmissing 37.70 14.90 10 ppts
Percent change in default likelihood
Interest rate at origination 6.25 0.81 1 ppt
Percent change in default likelihood
Margin rate (0 if FIRM) 2.58 1.14 1 ppt
Percent change in default likelihood
B. Subprime mortgages
Baseline predicted default rate
HPI growth 5.40 9.60 10 ppts
Percent change in default likelihood
FICO score 608 55 50 points
Percent change in default likelihood
LTV 80.40 12.60 10 ppts
Percent change in default likelihood
DTI, if nonmissing 39.40 10.70 10 ppts
Percent change in default likelihood
Interest rate at origination 7.93 1.31 1 ppt
Percent change in default likelihood
Margin rate (0 if FIRM) 5.39 0.72 1 ppt
Percent change in default likelihood
Default first 12 months
Variables 2004 2005
A. Prime mortgages
Baseline predicted default rate 0.0220 0.0217
HPI growth -0.0003 -0.001
Percent change in default likelihood -1 -5
FICO score -0.0116 *** -0.01 ***
Percent change in default likelihood -53 -46
LTV 0.0009 0.0031 ***
Percent change in default likelihood 4 14
DTI, if nonmissing -0.0001 -0.0002
Percent change in default likelihood 0 -1
Interest rate at origination 0.0105 *** 0.0062 **
Percent change in default likelihood 48 29
Margin rate (0 if FIRM) -0.0007 0.0008
Percent change in default likelihood -3 4
B. Subprime mortgages
Baseline predicted default rate 0.1075 0.1571
HPI growth -0.0189 * -0.0163 ***
Percent change in default likelihood -18 -10
FICO score -0.0351 *** -0.0532 ***
Percent change in default likelihood -33 -34
LTV 0.0059 0.0084 ***
Percent change in default likelihood 5 5
DTI, if nonmissing 0.0069 ** 0.0108 ***
Percent change in default likelihood 6 7
Interest rate at origination 0.0302 *** 0.0331 ***
Percent change in default likelihood 28 21
Margin rate (0 if FIRM) 0.0112 ** 0.0059 **
Percent change in default likelihood 10 4
Default first 12 months
Variables 2006 2007
A. Prime mortgages
Baseline predicted default rate 0.0422 0.0482
HPI growth -0.0178 *** -0.0006
Percent change in default likelihood -42 -1
FICO score -0.0172 *** -0.0204 ***
Percent change in default likelihood -41 -42
LTV 0.0113 *** 0.0141 ***
Percent change in default likelihood 27 29
DTI, if nonmissing 0.0009 0.0049 ***
Percent change in default likelihood 2 10
Interest rate at origination 0.027 *** 0.032 ***
Percent change in default likelihood 64 66
Margin rate (0 if FIRM) 0.0018 ** -0.0004
Percent change in default likelihood 4 -1
B. Subprime mortgages
Baseline predicted default rate 0.2396 0.2538
HPI growth -0.0406 *** -0.001
Percent change in default likelihood -17 0
FICO score -0.0581 *** -0.0519 ***
Percent change in default likelihood -24 -20
LTV 0.0189 *** 0.0189 ***
Percent change in default likelihood 8 7
DTI, if nonmissing 0.0066 ** 0.0095 **
Percent change in default likelihood 3 4
Interest rate at origination 0.0471 *** 0.0542 ***
Percent change in default likelihood 20 21
Margin rate (0 if FIRM) 0.0125 *** -0.0101
Percent change in default likelihood 5 -4
* Significant at the 10 percent level.
** Significant at the 5 percent level.
*** Significant at the 1 percent level.
Notes: All values in the first two columns are in percent except
for FICO score. The abbreviation opt indicates percentage point.
For definitions of the variables, see box 1 on p. 22.
Source: Authors' calculations based on data from Lender
Processing Services (LPS) Applied Analytics.
TABLE 4
Predicted probability of defaulting within 12 months of
mortgage origination
Prediction model coefficients from
Mortgages origi- Actual 2004 2005 2006 2007
nated in default
rate
(-----------------percent----------------------)
A. Prime mortgages: Predicted default rate
2004 2.21 2.20 # 2.55 1.47 2.86
2005 2.17 2.04 2.17 # 1.72 3.77
2006 4.23 3.09 3.08 4.22 # 5.62
2007 4.83 3.54 3.49 5.82 4.82 #
B. Prime mortgages: Predicted default rate as percentage of actual
default rate
2004 100 99.89 # 115.63 66.63 129.72
2005 100 93.94 99.77 # 79.15 173.55
2006 100 72.98 72.68 99.73 # 132.71
2007 100 73.32 72.11 120.45 99.77 #
C. Subprime mortgages: Predicted default rate
2004 10.76 10.75 # 13.72 13.26 16.62
2005 15.72 13.56 15.71 # 16.78 19.93
2006 23.99 19.92 20.68 23.96 # 24.95
2007 25.39 22.27 23.06 25.49 25.38 #
D. Subprime mortgages: Predicted default rate as percentage of actual
default rate
2004 100 99.83 # 127.48 123.21 154.41
2005 100 86.31 99.98 # 106.80 126.78
2006 100 83.04 86.20 99.89 # 104.01
2007 100 87.70 90.80 100.40 99.97 #
Notes: Average predicted default rate for mortgages, using
estimated coefficients from different mortgage origination years.
Bold numbers are predictions that use characteristics and
coefficients from the same year. See the text for further
details.
Source: Authors' calculations based on data from Lender
Processing Services (LPS) Applied Analytics.
Note: Bold numbers are predictions that use characteristics and
coefficients from the same year. See the text for further details
is indicated with #.
TABLE 5
Mortgage default rate forecasts under different assumptions for the
future path of home prices
Predictions
Model year for mortgages Actual
coefficients originated in default rate Scenario I
(-----------percent------------)
A. Default rate forecasts for prime mortgages
2004 2006 4.23 3.09
2005 2007 4.83 3.49
B. Default rate forecasts for subprime mortgages
2004 2006 23.99 19.92
2005 2007 25.39 23.06
Prime mortgages Subprime mortgages
Mortgages origi- Scenario I Scenario II Scenario I
nated in
(-----------percent------------)
C. Average HPI growth rates assumed in each scenario
2004 13.44 10.63 13.99
2005 9.10 14.30 9.70
2006 1.94 9.57 1.52
2007 -4.19 2.65 -3.94
Model year
coefficients Scenario II Scenario III
(-----------percent------------)
A. Default rate forecasts for prime mortgages
2004 3.03 3.07
2005 3.39 3.43
B. Default rate forecasts for subprime mortgages
2004 17.47 20.32
2005 21.51 22.18
All mortgages
Mortgages origi- Scenario II Scenario III
nated in
(-----------percent------------)
C. Average HPI growth rates assumed in each scenario
2004 12.89 0.00
2005 15.00 0.00
2006 9.87 0.00
2007 3.29 0.00
Notes: The forecast horizon is 12 months since mortgage
origination. HPI means House Price Index from the Federal Housing
Finance Agency. Scenario I is perfect foresight; it uses actual
realized HPI growth. Scenario II is a simple extrapolation; it
uses the HPI growth rate from the preceding 12 months. Scenario
III assumes zero HPI growth for the forecasting period. See the
text for further details.
Source: Authors' calculations based on data from Lender
Processing Services (LPS) Applied Analytics.
