Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
com/abstract=1020396
Understanding the Subprime Mortgage Crisis
Yuliya Demyanyk, Otto Van Hemert
∗
This Draft: December 5, 2008
First Draft: October 9, 2007
Abstract
Using loanlevel data, we analyze the quality of subprime mortgage loans by adjusting their performance
for diﬀerences in borrower characteristics, loan characteristics, and macroeconomic conditions. We ﬁnd
that the quality of loans deteriorated for six consecutive years before the crisis and that securitizers
were, to some extent, aware of it. We provide evidence that the rise and fall of the subprime mortgage
market follows a classic lending boombust scenario, in which unsustainable growth leads to the collapse
of the market. Problems could have been detected long before the crisis, but they were masked by high
house price appreciation between 2003 and 2005.
∗
Demyanyk: Banking Supervision and Regulation, Federal Reserve Bank of St. Louis, P.O. Box 442, St. Louis, MO
63166, Yuliya.Demyanyk@stls.frb.org. Van Hemert: Department of Finance, Stern School of Business, New York University,
44 W. 4th Street, New York, NY 10012, ovanheme@stern.nyu.edu. The authors would like to thank Cliﬀ Asness, Joost
Driessen, William Emmons, Emre Ergungor, Scott Frame, Xavier Gabaix, Dwight Jaﬀee, Ralph Koijen, Andreas Lehnert,
Andrew Leventis, Chris Mayer, Andrew Meyer, Toby Moskowitz, Lasse Pedersen, Robert Rasche, Matt Richardson, Stefano
Risa, Bent Sorensen, Matthew Spiegel, Stijn Van Nieuwerburgh, James Vickery, Jeﬀ Wurgler, anonymous referees, and
seminar participants at the Federal Reserve Bank of St. Louis; the Florida Atlantic University; the International Monetary
Fund; the second New York Fed—Princeton liquidity conference; Lehman Brothers; the BaruchColumbiaStern real estate
conference; NYU Stern Research Day; Capula Investment Management; AQR Capital Management,; the Conference on the
Subprime Crisis and Economic Outlook in 2008 at Lehman Brothers; Freddie Mac; Federal Deposit and Insurance Corporation
(FDIC); U.S. Securities and Exchange Comission (SEC); Oﬃce of Federal Housing Enterprise Oversight (OFHEO); Board of
Governors of the Federal Reserve System; Carnegie Mellon University; Baruch; University of British Columbia, University of
Amsterdam; the 44th Annual Conference on Bank Structure and Competition at the Federal Reserve Bank of Chicago; the
Federal Reserve Research and Policy Activities; Sixth Colloquium on Derivatives, RiskReturn and Subprime, Lucca, Italy;
and the Federal Reserve Bank of Cleveland; The views expressed are those of the authors and do not necessarily reﬂect the
oﬃcial positions of the Federal Reserve Bank of St. Louis or the Federal Reserve System.
Electronic copy available at: http://ssrn.com/abstract=1020396 Electronic copy available at: http://ssrn.com/abstract=1020396
1 Introduction
The subprime mortgage crisis of 2007 was characterized by an unusually large fraction of subprime mort
gages originated in 2006 and 2007 becoming delinquent or in foreclosure only months later. The crisis
spurred massive media attention; many diﬀerent explanations of the crisis have been proﬀered. The goal
of this paper is to answer the question: “What do the data tell us about the possible causes of the cri
sis?” To this end we use a loanlevel database containing information on about half of all U.S. subprime
mortgages originated between 2001 and 2007.
The relatively poor performance of vintage 2006 and 2007 loans is illustrated in Figure 1 (left panel).
At every mortgage loan age, loans originated in 2006 and 2007 show a much higher delinquency rate than
loans originated in earlier years at the same ages.
Figure 1: Actual and Adjusted Delinquency Rate
The ﬁgure shows the age pattern in the actual (left panel) and adjusted (right panel) delinquency rate for the diﬀerent vintage years.
The delinquency rate is deﬁned as the cumulative fraction of loans that were past due 60 or more days, in foreclosure, realestate owned,
or defaulted, at or before a given age. The adjusted delinquency rate is obtained by adjusting the actual rate for yearbyyear variation in
FICO scores, loantovalue ratios, debttoincome ratios, missing debttoincome ratio dummies, cashout reﬁnancing dummies, owner
occupation dummies, documentation levels, percentage of loans with prepayment penalties, mortgage rates, margins, composition of
mortgage contract types, origination amounts, MSA house price appreciation since origination, change in state unemployment rate since
origination, and neighborhood median income.
0
5
10
15
20
25
30
35
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Actual Delinquency Rate (%)
0
5
10
15
20
25
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Adjusted Delinquency Rate (%)
We document that the poor performance of the vintage 2006 and 2007 loans was not conﬁned to a
particular segment of the subprime mortgage market. For example, ﬁxedrate, hybrid, purchasemoney,
cashout reﬁnancing, lowdocumentation, and fulldocumentation loans originated in 2006 and 2007 all
1
Electronic copy available at: http://ssrn.com/abstract=1020396
showed substantially higher delinquency rates than loans made the prior ﬁve years. This contradicts a
widely held belief that the subprime mortgage crisis was mostly conﬁned to hybrid or lowdocumentation
mortgages.
We explore to what extent the subprime mortgage crisis can be attributed to diﬀerent loan charac
teristics, borrower characteristics, macroeconomic conditions, and vintage (origination) year eﬀects. The
most important macroeconomic factor is subsequent house price appreciation, measured as the MSAlevel
house price change between the time of origination and the time of loan performance evaluation. For
the empirical analysis, we run a proportional odds duration model with the probability of (ﬁrsttime)
delinquency a function of these factors and loan age.
We ﬁnd that loan and borrower characteristics are very important in terms of explaining the cross
section of loan performance. However, because these characteristics were not suﬃciently diﬀerent in 2006
and 2007 compared with the prior ﬁve years, they cannot explain the unusually weak performance of
vintage 2006 and 2007 loans. For example, a onestandarddeviation increase in the debttoincome ratio
raises the likelihood (the odds ratio) of a current loan turning delinquent in a given month by as much as a
factor of 1.14. However, because the average debttoincome ratio was just 0.2 standard deviations higher
in 2006 than its level in previous years, it contributes very little to the inferior performance of vintage
2006 loans. The only variable in the considered proportional odds model that contributed substantially
to the crisis is the low subsequent house price appreciation for vintage 2006 and 2007 loans, which can
explain about a factor of 1.24 and 1.39, respectively, higherthanaverage likelihood for a current loan to
turn delinquent.
1
Due to geographical heterogeneity in house price changes, some areas have experienced
largerthanaverage house price declines and therefore have a larger explained increase in delinquency and
foreclosure rates.
2
The coeﬃcients of the vintage dummy variables, included as covariates in the proportional odds model,
measure the quality of loans, adjusted for diﬀerences in observed loan characteristics, borrower charac
teristics, and macroeconomic circumstances. In Figure 1 (right panel) we plot the adjusted delinquency
rates, which are obtained by using the estimated coeﬃcients for the vintage dummies and imposing the
requirement that the average actual and average adjusted delinquency rates are equal for any given age.
As shown in Figure 1 (right panel), the adjusted delinquency rates have been steadily rising for the
1
Other papers that research the relationship between house prices and mortgage ﬁnancing include Genesove and Mayer
(1997), Genesove and Mayer (2001), and Brunnermeier and Julliard (2007).
2
Also, house price appreciation may diﬀer in cities versus rural areas. See for example Glaeser and Gyourko (2005) and
Gyourko and Sinai (2006).
2
past seven years. In other words, loan quality—adjusted for observed characteristics and macroeconomic
circumstances—deteriorated monotonically between 2001 and 2007. Interestingly, 2001 was among the
worst vintage years in terms of actual delinquency rates, but is in fact the best vintage year in terms of
the adjusted rates. High interest rates, low average FICO credit scores, and low house price appreciation
created the “perfect storm” in 2001, resulting in a high actual delinquency rate; after adjusting for these
unfavorable circumstances, however, the adjusted delinquency rates are low.
In addition to the monotonic deterioration of loan quality, we show that over time the average combined
loantovalue ratio increased, the fraction of low documentation loans increased, and the subprimeprime
rate spread decreased. The rapid rise and subsequent fall of the subprime mortgage market is therefore
reminiscent of a classic lending boombust scenario.
3
The origin of the subprime lending boom has often
been attributed to the increased demand for socalled privatelabel mortgagebacked securities (MBSs)
by both domestic and foreign investors. Our database does not allow us to directly test this hypothesis,
but an increase in demand for subprime MBSs is consistent with our ﬁnding of lower spreads and higher
volume. Mian and Suﬁ (2008) ﬁnd evidence consistent with this view that increased demand for MBSs
spurred the lending boom.
The proportional odds model used to estimate the adjusted delinquency rates assumes that the co
variate coeﬃcients are constant over time. We test the validity of this assumption for all variables and
ﬁnd that it is the most strongly rejected for the loantovalue (LTV) ratio. HighLTV borrowers in 2006
and 2007 were riskier than those in 2001 in terms of the probability of delinquency, for given values of the
other explanatory variables. Were securitizers aware of the increasing riskiness of highLTV borrowers?
4
To answer this question, we analyze the relationship between the mortgage rate and LTV ratio (along with
the other loan and borrower characteristics). We perform a crosssectional ordinary least squares (OLS)
regression, with the mortgage rate as the dependent variable, for each quarter from 2001Q1 to 2007Q2 for
both ﬁxedrate mortgages and 2/28 hybrid mortgages. Figure 2 shows that the coeﬃcient on the ﬁrstlien
LTV variable, scaled by the standard deviation of the ﬁrstlien LTV ratio, has been increasing over time.
We thus ﬁnd evidence that securitizers were aware of the increasing riskiness of highLTV borrowers, and
3
Berger and Udell (2004) discuss the empirical stylized fact that during a monetary expansion lending volume typically
increases and underwriting standards loosen. Loan performance is the worst for those loans underwritten toward the end
of the cycle. Demirg¨ u¸ cKunt and Detragiache (2002) and Gourinchas, Valdes, and Landerretche (2001) ﬁnd that lending
booms raise the probability of a banking crisis. Dell’Ariccia and Marquez (2006) show in a theoretical model that a change
in information asymmetry across banks might cause a lending boom that features lower standards and lower proﬁts. Ruckes
(2004) shows that low screening activity may lead to intense price competition and lower standards.
4
For loans that are securitized (as are all loans in our database), the securitizer eﬀectively dictates the mortgage rate
charged by the originator.
3
adjusted mortgage rates accordingly.
Figure 2: Sensitivity of Mortgage Rate to FirstLien LoantoValue Ratio
The ﬁgure shows the eﬀect of the ﬁrstlien loantovalue ratio on the mortgage rate for ﬁrstlien ﬁxedrate and 2/28 hybrid mortgages.
The eﬀect is measured as the regression coeﬃcient on the ﬁrstlien loantovalue ratio (scaled by the standard deviation) in an ordinary
least squares regression with the mortgage rate as the dependent variable and the FICO score, ﬁrstlien loantovalue ratio, secondlien
loantovalue ratio, debttoincome ratio, missing debttoincome ratio dummy, cashout reﬁnancing dummy, owneroccupation dummy,
prepayment penalty dummy, origination amount, term of the mortgage, prepayment term, and margin (only applicable to 2/28 hybrid)
as independent variables. Each point corresponds to a separate regression, with a minimum of 18,784 observations.
0
.1
.2
.3
.4
.5
S
c
a
l
e
d
R
e
g
r
e
s
s
i
o
n
C
o
e
f
f
i
c
i
e
n
t
(
%
)
2001 2002 2003 2004 2005 2006 2007
Year
FRM
2/28 Hybrid
We show that our main results are robust to analyzing mortgage contract types separately, focusing
on foreclosures rather than delinquencies, and specifying the empirical model in numerous diﬀerent ways,
like allowing for interaction eﬀects between diﬀerent loan and borrower characteristics. The latter includes
taking into account risklayering—the origination of loans that are risky in several dimensions, such as
the combination of a high LTV ratio and a low FICO score.
As an extension, we estimate our proportional odds model using data just through yearend 2005
and again obtain the continual deterioration of loan quality from 2001 onward. This means that the
seeds for the crisis were sown long before 2007, but detecting them was complicated by high house price
appreciation between 2003 and 2005—appreciation that masked the true riskiness of subprime mortgages.
In another extension, we ﬁnd an increased probability of delinquency for loans originated in low and
moderateincome areas, deﬁned as areas with median income below 80 percent of the larger Metropolitan
Statistical Area median income. This points toward a negative byproduct of the 1977 Community
Reinvestment Act and Government Sponsored Enterprises housing goals, which seek to stimulate loan
4
origination in low and moderateincome areas.
There is a large literature on the determinants of mortgage delinquencies and foreclosures, dating
back to at least Von Furstenberg and Green (1974). Recent contributions include Cutts and Van Order
(2005) and PenningtonCross and Chomsisengphet (2007).
5
Other papers analyzing the subprime crisis
include Gerardi, Shapiro, and Willen (2008), Mian and Suﬁ (2008), DellAriccia, Igan, and Laeven (2008),
and Keys, Mukherjee, Seru, and Vig (2008). Our paper makes several novel contributions. First, we
quantify how much diﬀerent determinants have contributed to the observed high delinquency rates for
vintage 2006 and 2007 loans, which led up to the 2007 subprime mortgage crisis. Our data enables us
to show that the eﬀect of diﬀerent loanlevel characteristics as well as low house price appreciation was
quantitatively too small to explain the poor performance of 2006 and 2007 vintage loans. Second, we
uncover a downward trend in loan quality, determined as loan performance adjusted for diﬀerences in
loan and borrower characteristics and macroeconomic circumstances. We further show that there was a
deterioration of lending standards and a decrease in the subprimeprime mortgage rate spread during the
2001–2007 period. Together these results provide evidence that the rise and fall of the subprime mortgage
market follows a classic lending boombust scenario, in which unsustainable growth leads to the collapse
of the market. Third, we show that the continual deterioration of loan quality could have been detected
long before the crisis by means of a simple statistical exercise. Fourth, securitizers were, to some extent,
aware of this deterioration over time, as evidenced by changing determinants of mortgage rates. Fifth,
we detect an increased likelihood of delinquency in low and middleincome areas, after controlling for
diﬀerences in neighborhood incomes and other loan, borrower, and macroeconomic factors. This empirical
ﬁnding seems to suggest that the housing goals of the Community Reinvestment Act and/or Government
Sponsored Enterprises—those intended to increase lending in low and middleincome areas—might have
created a negative byproduct, that is associated with higher loan delinquencies.
The structure of this paper is as follows. In Section 2 we show the descriptive statistics for the subprime
mortgages in our database. In Section 3 we discuss the empirical strategy we employ. In Section 4 we
present the baselinecase results and in Section 5 we discuss extensions and robustness checks. In Section
6 we demonstrate the increasing riskiness of highLTV borrowers, and the extent to which securitizers
were aware of this risk. In Section 7 we analyze the subprimeprime rate spread and in Section 8 we
conclude. We provide several additional robustness checks in the appendices.
5
Deng, Quigley, and Van Order (2000) discuss the simultaneity of the mortgage prepayment and default option. Campbell
and Cocco (2003) and Van Hemert (2007) discuss mortgage choice over the life cycle.
5
2 Descriptive Analysis
In this paper we use the First American CoreLogic LoanPerformance (henceforth: LoanPerformance)
database, as of June 2008, which includes loanlevel data on about 85 percent of all securitized subprime
mortgages; (more than half of the U.S. subprime mortgage market).
6
There is no consensus on the exact deﬁnition of a subprime mortgage loan. The term subprime can be
used to describe certain characteristics of the borrower (e.g., a FICO credit score less than 620),
7
lender
(e.g., specialization in highcost loans),
8
security of which the loan can become a part (e.g., high projected
default rate for the pool of underlying loans), or mortgage contract type (e.g., no money down and no
documentation provided, or a 2/28 hybrid). The common element across deﬁnitions of a subprime loan
is a high default risk. In this paper, subprime loans are those underlying subprime securities. We do not
include less risky AltA mortgage loans in our analysis. We focus on ﬁrstlien loans and consider the 2001
through 2008 sample period.
9
We ﬁrst outline the main characteristics of the loans in our database at origination. Second, we discuss
the delinquency rates of these loans for various segments of the subprime mortgage market.
2.1 Loan Characteristics at Origination
Table 1 provides the descriptive statistics for the subprime mortgage loans in our database that were
originated between 2001 and 2007. In the ﬁrst block of Table 1 we see that the annual number of
originated loans increased by a factor of four between 2001 and 2006 and the average loan size almost
doubled over those ﬁve years. The total dollar amount originated in 2001 was $57 billion, while in 2006 it
was $375 billion. In 2007, in the wake of the subprime mortgage crisis, the dollar amount originated fell
sharply to $69 billion, and was primarily originated in the ﬁrst half of 2007.
In the second block of Table 1, we split the pool of mortgages into four main mortgage contract types.
6
Mortgage Market Statistical Annual (2007) reports securitization shares of subprime mortgages each year from 2001 to
2006 equal to 54, 63, 61, 76, 76, and 75 percent respectively.
7
The Board of Governors of the Federal Reserve System, The Oﬃce of the Controller of the Currency,
the Federal Deposit Insurance Corporation, and the Oﬃce of Thrift Supervision use this deﬁnition. See e.g.
http://www.fdic.gov/news/news/press/2001/pr0901a.html
8
The U.S. Department of Housing and Urban Development uses HMDA data and interviews lenders to identify subprime
lenders among them. There are, however, some subprime lenders making prime loans and some prime lenders originating
subprime loans.
9
Since the ﬁrst version of this paper in October 2007, LoanPerformance has responded to the request by trustees’ clients
to reclassify some of its subprime loans to AltA status. While it is not clear to us whether the pre or postreclassiﬁcation
subprime data are the most appropriate for research purposes, we checked that our results are robust to the reclassiﬁcation.
In this version we focus on the postclassiﬁcation data.
6
Table 1: Loan Characteristics at Origination for Diﬀerent Vintages
Descriptive statistics for the ﬁrstlien subprime loans in the LoanPerformance database.
2001 2002 2003 2004 2005 2006 2007
Size
Number of Loans (*1000) 452 737 1, 258 1, 911 2, 274 1, 772 316
Average Loan Size (*$1000) 126 145 164 180 200 212 220
Mortgage Type
FRM (%) 33.2 29.0 33.6 23.8 18.6 19.9 27.5
ARM (%) 0.4 0.4 0.3 0.3 0.4 0.4 0.2
Hybrid (%) 59.9 68.2 65.3 75.8 76.8 54.5 43.8
Balloon (%) 6.5 2.5 0.8 0.2 4.2 25.2 28.5
Loan Purpose
Purchase (%) 29.7 29.3 30.1 35.8 41.3 42.4 29.6
Reﬁnancing (cash out) (%) 58.4 57.4 57.7 56.5 52.4 51.4 59.0
Reﬁnancing (no cash out) (%) 11.2 12.9 11.8 7.7 6.3 6.2 11.4
Variable Means
FICO Score 601.2 608.9 618.1 618.3 620.9 618.1 613.2
Combined LoantoValue Ratio (%) 79.4 80.1 82.0 83.6 84.9 85.9 82.8
DebttoIncome Ratio (%) 38.0 38.5 38.9 39.4 40.2 41.1 41.4
Missing DebttoIncome Ratio Dummy (%) 34.7 37.5 29.3 26.5 31.2 19.7 30.9
Investor Dummy (%) 8.2 8.1 8.1 8.3 8.3 8.2 8.2
Documentation Dummy (%) 76.5 70.4 67.8 66.4 63.4 62.3 66.7
Prepayment Penalty Dummy (%) 75.9 75.3 74.0 73.1 72.5 71.0 70.2
Mortgage Rate (%) 9.7 8.7 7.7 7.3 7.5 8.4 8.6
Margin for ARM and Hybrid Mortgage Loans (%) 6.4 6.6 6.3 6.1 5.9 6.1 6.0
7
Most numerous are the hybrid mortgages, accounting for more than half of all subprime loans in our
data set originated between 2001 and 2007. A hybrid mortgage carries a ﬁxed rate for an initial period
(typically 2 or 3 years) and then the rate resets to a reference rate (often the 6month LIBOR) plus a
margin. The ﬁxedrate mortgage contract became less popular in the subprime market over time and
accounted for just 20 percent of the total number of loans in 2006. In contrast, in the prime mortgage
market, most mortgage loans were of the ﬁxedrate type during this period.
10
In 2007, as the subprime
mortgage crisis hit, the popularity of FRMs rose to 28 percent. The proportion of balloon mortgage
contracts jumped substantially in 2006, and accounted for 25 percent of the total number of mortgages
originated that year. A balloon mortgage does not fully amortize over the term of the loan and therefore
requires a large ﬁnal (balloon) payment. Less than 1 percent of the mortgages originated over the sample
period were adjustablerate (nonhybrid) mortgages.
In the third block of Table 1, we report the purpose of the mortgage loans. In about 30 to 40 percent
of cases, the purpose was to ﬁnance the purchase of a house. Approximately 55 percent of our subprime
mortgage loans were originated to extract cash, by reﬁnancing an existing mortgage loan into a larger new
mortgage loan. The share of loans originated in order to reﬁnance with no cash extraction was relatively
small.
In the ﬁnal block of Table 1, we report the mean values for the loan and borrower characteristics that
we will use in the statistical analysis (see Table 2 for a deﬁnition of these variables). The average FICO
credit score rose 20 points between 2001 and 2005. The combined loantovalue (CLTV) ratio, which
measures the value of alllien loans divided by the value of the house, slightly increased over 2001–2006,
primarily because of the increased popularity of secondlien and thirdlien loans. The (backend) debt
toincome ratio (if provided) and the fraction of loans with a prepayment penalty were fairly constant.
For about a third of the loans in our database, no debttoincome ratio was provided (the reported value
in those cases is zero); this is captured by the missing debttoincome ratio dummy variable. The share
of loans with full documentation fell considerably over the sample period, from 77 percent in 2001 to 67
percent in 2007. The mean mortgage rate fell from 2001 to 2004 and rebounded after that, consistent
with movements in both the 1year and 10year Treasury yields over the same period. Finally, the margin
(over a reference rate) for adjustablerate and hybrid mortgages stayed rather constant over time.
10
For example Koijen, Van Hemert, and Van Nieuwerburgh (2007) show that the fraction of conventional, singlefamily,
fully amortizing, purchasemoney loans reported by the Federal Housing Financing Board in its Monthly Interest Rate Survey
that are of the ﬁxedrate type ﬂuctuated between 60 and 90 percent from 2001 to 2006. Vickery (2007) shows that empirical
mortgage choice is aﬀected by the eligibility of the mortgage loan to be purchased by Fannie Mae and Freddie Mac.
8
We do not report summary statistics on the loan source, such as whether a mortgage broker interme
diated, as the broad classiﬁcation used in the database rendered this variable less informative.
2.2 Performance of Loans by Market Segments
We deﬁne a loan to be delinquent if payments on the loan are 60 or more days late, or the loan is reported
as in foreclosure, real estate owned, or in default. We denote the ratio of the number of vintage k loans
experiencing a ﬁrsttime delinquency at age s over the number of vintage k loans with no ﬁrsttime
delinquency for age < s by
˜
P
k
s
. We compute the actual (cumulative) delinquency rate for vintage k at
age t as the fraction of loans experiencing a delinquency at or before age t
Actual
k
t
= 1 −
t
¸
s=1
1 −
˜
P
k
s
(1)
We deﬁne the average actual delinquency rate as
Actual
t
= 1 −
t
¸
s=1
1 −
¯
P
s
, where (2)
¯
P
s
=
1
7
2007
¸
i=2001
˜
P
k
s
(3)
In Figure 1 (left panel) we show that for the subprime mortgage market as a whole, vintage 2006 and
2007 loans stand out in terms of high delinquency rates. In Figure 3, we again plot the age pattern in
the delinquency rate for vintages 2001 through 2007 and split the subprime mortgage market into various
segments. As the ﬁgure shows, the poor performance of the 2006 and 2007 vintages is not conﬁned to a
particular segment of the subprime market, but rather reﬂects a (subprime) marketwide phenomenon.
In the six panels of Figure 3 we see that for hybrid, ﬁxedrate, purchasemoney, cashout reﬁnancing,
lowdocumentation, and fulldocumentation mortgage loans, the 2006 and 2007 vintages show the highest
delinquency rate pattern. In general, vintage 2001 loans come next in terms of high delinquency rates,
and vintage 2003 loans have the lowest delinquency rates. Notice that the scale of the vertical axis diﬀers
across the panels. The delinquency rates for the ﬁxedrate mortgages (FRMs) are lower than those for
hybrid mortgages but exhibit a remarkably similar pattern across vintage years.
In Figure 4 we plot the delinquency rates of all outstanding mortgages. Notice that the fraction of
FRMs that are delinquent remained fairly constant from 2005Q1 to 2007Q2. Delinquency rates in this
9
Figure 3: Actual Delinquency Rate for Segments of the Subprime Mortgage Market
The ﬁgure shows the age pattern in the delinquency rate for diﬀerent segments. The delinquency rate is deﬁned as the cumulative
fraction of loans that were past due 60 or more days, in foreclosure, realestate owned, or defaulted, at or before a given age.
0
5
10
15
20
25
30
35
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Hybrid Mortgage Loans
0
5
10
15
20
25
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Fixed−Rate Mortgage Loans
0
5
10
15
20
25
30
35
40
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Purchase−Money Mortgage Loans
0
5
10
15
20
25
30
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Cash−Out Refinancing Mortgage Loans
0
5
10
15
20
25
30
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Full−Documentation Mortgage Loans
0
5
10
15
20
25
30
35
40
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Low−or−No−Doc Mortgage Loans
10
ﬁgure are deﬁned as the fraction of loans delinquent at any given time, not cumulative. These rates are
consistent with those used in an August 2007 speech by the Chairman of the Federal Reserve System
(Bernanke (2007)), who said “For subprime mortgages with ﬁxed rather than variable rates, for example,
serious delinquencies have been fairly stable.” It is important, though, to realize that this result is driven
by an aging eﬀect of the FRM pool, caused by a decrease in the popularity of FRMs from 2001 to 2006
(see Table 1). In other words, FRMs originated in 2006 in fact performed unusually poorly (Figure 3,
upperright panel), but if one plots the delinquency rate of outstanding FRMs over time (Figure 4, left
panel), the weaker performance of vintage 2006 loans is masked by the aging of the overall FRM pool.
Figure 4: Actual Delinquency Rates of Outstanding Mortgages
The Figure shows the actual delinquency rates of all outstanding FRMs and hybrids from January 2000 through June 2008.
0
5
10
15
20
25
30
35
40
2000 2001 2002 2003 2004 2005 2006 2007 2008
Year
FRM
Hybrid
Actual Delinquency Rate (%)
3 Statistical Model Speciﬁcation
The focus of our paper is on the performance of subprime mortgage loans in the ﬁrst 17 months after
origination, for which we already have data for the vintages of particular interest: 2006 and 2007. Given
this focus on young loans, we include delinquency—the earliest stage of payment problems—in our non
performance measure. Delinquency is an intermediate stage for a loan in trouble; the loan may eventually
cure or terminate with a prepayment or default.
Our paper is related to the vast literature on empirical mortgage termination. Termination occurs
either through a prepayment or a default. An analysis of mortgage termination lends itself naturally to
11
duration (i.e., survival) models, with prepayment and default as competing reasons for termination. Im
portant contributions to this literature include Deng (1997), Ambrose and Capone (2000), Deng, Quigley,
and Van Order (2000), Calhoun and Deng (2002), PenningtonCross (2003), Deng, Pavlov, and Yang
(2005), Clapp, Deng, and An (2006), and PenningtonCross and Chomsisengphet (2007).
We apply the duration model methodology to the intermediate status of delinquency by deﬁning non
survival as “having ever been 60 days delinquent or worse,” which includes formerly delinquent loans that
are prepaid or cured. Transition from survival to nonsurvival will occur when a loan becomes 60 or more
days delinquent or defaults for the ﬁrst time. As a robustness check, we used “being currently 60 or more
days delinquent or in default” as the nonperformance measure, ran a standard logit regression, and found
qualitatively similar results for the eﬀect of explanatory variables and the eﬀect of vintage year dummies
(unreported results).
3.1 Empirical Model Speciﬁcation
We are interested in the number of months (duration) until a loan becomes at least 60 days delinquent
or defaults for the ﬁrst time. Denoting this time by T, we deﬁne the probability that at age t loan i
with covariate values x
i,t
becomes delinquent for the ﬁrst time, conditional on not having been delinquent
before, as
P
i,t
= Pr {T = tT ≥ t, x
i,t
} (4)
Because the monthly choice whether to make a mortgage payment is discrete, we use a proportional
odds model, the discretetime analogue to the popular proportional hazard model:
log
P
i,t
1 − P
i,t
= α
t
+ β
x
i,t
(5)
where α
t
is an agedependent constant and β is a vector of coeﬃcients. The name “proportional odds”
arises from the fact that the vector of coeﬃcients, β, does not have an age subscript and thus the log odds
are proportional to the covariate values at any age.
3.2 Estimation
The proportional odds model, Equation 5, is typically estimated for the full panel at once using either
partial likelihood (see Cox (1972)) or maximum likelihood methodologies. The small sample properties for
12
these two estimation methods are potentially diﬀerent, but for a sample size of 10, 000 loans the methods
already provide very similar estimates for β. For larger sample sizes the computational burden of the
partial likelihood function quickly becomes unmanageable. This is a result of heavily tied data in our
discrete time setup, where the term “tied” refers to loans experiencing ﬁrsttime delinquency at the exact
same age. We therefore estimate the proportional odds model using maximum likelihood, which has the
added advantage that it provides estimates of the loan age eﬀect, α
t
. We use a random sample of 1 million
loans for this exercise.
We use the PROC LOGISTIC procedure in SAS for the maximum likelihood estimation. This method
is able to handle both left censoring (loans entering the sample at a later age) and right censoring (loans
leaving the sample prematurely, not due to a prepayment or default), using the noninformative censoring
assumption.
11
In order to generate unbiased delinquency rate plots (as in Figure 1, right panel), we
classify prepaid loans as nondelinquent and noncensored, because we know for sure that they will never
experience a ﬁrsttime delinquency.
Because we include vintage year dummies as covariates we have to restrict the maximum age considered
in our analysis to 17 months, the latest age for which we have an observation for the 2007 origination
year; the maximum likelihood estimation requires each covariate, including the vintage 2007 dummy, have
some dispersion in the covariate values for each age.
12
For the adjusted delinquency rate plots viewed at
the end of 2005 and 2006 (Figure 5), we restrict the analysis to a maximum age of 11 months.
3.3 Reported Output
The AgeEﬀect statistic is deﬁned as the proportional odds ratio for ﬁrsttime delinquency at a particular
age t for the average (over the full sample) vector of covariate values at age t, ¯ x
t
:
AgeEﬀect
t
= exp (α
t
+ β
¯ x
t
) (6)
The Marginal statistic is deﬁned as the log proportional odds ratio associated with a one standard
11
Loans securitized several months after origination are not observed in our data between the origination date and the
securitization date; therefore, they are left censored. In addition, if the securitizer goes out of business we stop observing
their loans and therefore they are right censored.
12
For more information on this, see page 126 of Allison (2007).
13
deviation increase in variable j, σ
j
:
Marginal
j
= β
j
σ
j
= log
exp (α
t
+ β
x
i,t
+ β
j
σ
j
)
exp (α
t
+ β
x
i,t
)
(7)
The advantage of taking the log in Equation 7 is that the eﬀect of an increase of σ
j
in covariate j is minus
the eﬀect of a decrease of σ
j
, and thus the absolute eﬀect is invariant to the chosen direction of change.
The Deviation statistic measures the diﬀerence between the mean value of a variable in a particular
vintage year and the mean value of that variable measured over the entire sample, expressed in the number
of standard deviations of the variable. For example, for vintage 2001 and variable j it is the diﬀerence
between the mean value for variable j in 2001, x01
j
, and the mean value over all vintages, ¯ x
j
, expressed
in the number of standard deviations:
Deviation
j
=
x01
j
− ¯ x
j
σ
j
(8)
The Contribution statistic measures the deviation of the (average) log proportional odds of ﬁrsttime
delinquency in a particular vintage year from the (average) log proportional odds of ﬁrsttime delinquency
over the entire sample that can be explained by a particular variable. For example for vintage 2001 and
variable j we have:
Contribution
j
= β
j
x01
j
− ¯ x
j
= log
exp
α
t
+ β
x
i,t
+ β
j
x01
j
− ¯ x
j
exp (α
t
+ β
x
i,t
)
(9)
= Marginal
j
∗ Deviation
j
As a straightforward generalization of Equation 9, the combined contribution of two variables is simply
the sum of the individual contributions. This property will be used for reporting the total contribution
of all covariates in Table 3.
The probability of experiencing a ﬁrsttime delinquency at or before age = t is given by
Pr {T ≤ tx
t
, x
t−1
, ..} = 1 −
t
¸
s=1
(1 − P
s
) (10)
To visualize the magnitude of the vintage year eﬀect, we evaluate the above expression for the value of
14
P
s
that satisﬁes
log
P
s
1 − P
s
= log
¯
P
s
1 −
¯
P
s
+ D
k
−
¯
D (11)
where D
k
is the estimated coeﬃcient for for the vintage year k dummy variable and
¯
D is the average
estimated vintage dummy variable. This expression uses the proportional odds property for explana
tory variables, including vintage year dummies, illustrated in Equation 5. Combining Equations 10 and
Equation 11 we obtain the adjusted delinquency rate for vintage year k at age t
Adjusted
k
t
= 1 −
t
¸
s=1
1
1 +
¯
P
s
1−
¯
P
s
exp
¸
D
k
−
¯
D
¸
(12)
Notice that for an average vintage year, D
k
=
¯
D, Equation 12 simpliﬁes to
Adjusted
t
= 1 −
t
¸
s=1
1 −
¯
P
s
= Actual
t
(13)
4 Empirical Results for the BaselineCase Speciﬁcation
In this section we investigate to what extent the proportional odds model model can explain the high levels
of delinquencies for the vintage 2006 and 2007 mortgage loans in our database. All results in this section
are based on a random sample of one million ﬁrstlien subprime mortgage loans, originated between 2001
and 2007.
4.1 Variable Deﬁnitions
Table 2 provides the deﬁnitions of the variables (covariates) included in the baselinecase speciﬁcation of
the proportional odds model.
The borrower and loan characteristics we use in the analysis are: the FICO credit score; the combined
loantovalue ratio; the value of the debttoincome ratio (when provided); a dummy variable indicating
whether the debttoincome ratio was missing (reported as zero); a dummy variable indicating whether the
loan was a cashout reﬁnancing; a dummy variable indicating whether the borrower was an investor (as
opposed to an owneroccupier); a dummy variable indicating whether full documentation was provided; a
dummy variable indicating whether there is a prepayment penalty on a loan; the (initial) mortgage rate;
15
T
a
b
l
e
2
:
B
a
s
e
l
i
n
e

C
a
s
e
V
a
r
i
a
b
l
e
D
e
ﬁ
n
i
t
i
o
n
s
T
h
i
s
t
a
b
l
e
p
r
e
s
e
n
t
s
d
e
ﬁ
n
i
t
i
o
n
s
o
f
t
h
e
b
a
s
e
l
i
n
e

c
a
s
e
v
a
r
i
a
b
l
e
s
(
c
o
v
a
r
i
a
t
e
s
)
u
s
e
d
i
n
t
h
e
p
r
o
p
o
r
t
i
o
n
a
l
o
d
d
s
d
u
r
a
t
i
o
n
m
o
d
e
l
.
T
h
e
ﬁ
r
s
t
t
w
o
v
a
r
i
a
b
l
e
s
a
r
e
u
s
e
d
a
s
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
s
.
T
h
e
o
t
h
e
r
v
a
r
i
a
b
l
e
s
a
r
e
u
s
e
d
a
s
i
n
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
s
.
W
e
r
e
p
o
r
t
t
h
e
e
x
p
e
c
t
e
d
s
i
g
n
f
o
r
t
h
e
i
n
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
s
i
n
p
a
r
e
n
t
h
e
s
e
s
a
n
d
s
o
m
e
t
i
m
e
s
p
r
o
v
i
d
e
a
b
r
i
e
f
m
o
t
i
v
a
t
i
o
n
.
V
a
r
i
a
b
l
e
(
E
x
p
e
c
t
e
d
S
i
g
n
)
E
x
p
l
a
n
a
t
i
o
n
F
I
C
O
S
c
o
r
e
(

)
F
a
i
r
,
I
s
a
a
c
a
n
d
C
o
m
p
a
n
y
(
F
I
C
O
)
c
r
e
d
i
t
s
c
o
r
e
a
t
o
r
i
g
i
n
a
t
i
o
n
.
C
o
m
b
i
n
e
d
L
o
a
n

t
o

V
a
l
u
e
R
a
t
i
o
(
+
)
C
o
m
b
i
n
e
d
v
a
l
u
e
o
f
a
l
l
l
i
e
n
s
d
i
v
i
d
e
d
b
y
t
h
e
v
a
l
u
e
o
f
t
h
e
h
o
u
s
e
a
t
o
r
i
g
i
n
a
t
i
o
n
.
A
h
i
g
h
e
r
c
o
m
b
i
n
e
d
l
o
a
n

t
o

v
a
l
u
e
r
a
t
i
o
m
a
k
e
s
d
e
f
a
u
l
t
m
o
r
e
a
t
t
r
a
c
t
i
v
e
.
D
e
b
t

t
o

I
n
c
o
m
e
R
a
t
i
o
(
+
)
B
a
c
k

e
n
d
d
e
b
t

t
o

i
n
c
o
m
e
r
a
t
i
o
,
d
e
ﬁ
n
e
d
b
y
t
h
e
t
o
t
a
l
m
o
n
t
h
l
y
d
e
b
t
p
a
y
m
e
n
t
s
d
i
v
i
d
e
d
b
y
t
h
e
g
r
o
s
s
m
o
n
t
h
l
y
i
n
c
o
m
e
,
a
t
o
r
i
g
i
n
a
t
i
o
n
.
A
h
i
g
h
e
r
d
e
b
t

t
o

i
n
c
o
m
e
r
a
t
i
o
m
a
k
e
s
i
t
h
a
r
d
e
r
t
o
m
a
k
e
t
h
e
m
o
n
t
h
l
y
m
o
r
t
g
a
g
e
p
a
y
m
e
n
t
.
M
i
s
s
i
n
g
D
e
b
t

t
o

I
n
c
o
m
e
D
u
m
m
y
(
+
)
E
q
u
a
l
s
o
n
e
i
f
t
h
e
b
a
c
k

e
n
d
d
e
b
t

t
o

i
n
c
o
m
e
r
a
t
i
o
i
s
m
i
s
s
i
n
g
a
n
d
z
e
r
o
i
f
p
r
o
v
i
d
e
d
.
W
e
e
x
p
e
c
t
t
h
e
l
a
c
k
o
f
d
e
b
t

t
o

i
n
c
o
m
e
i
n
f
o
r
m
a
t
i
o
n
t
o
b
e
a
n
e
g
a
t
i
v
e
s
i
g
n
a
l
o
n
b
o
r
r
o
w
e
r
q
u
a
l
i
t
y
.
C
a
s
h

O
u
t
D
u
m
m
y
(

)
E
q
u
a
l
s
o
n
e
i
f
t
h
e
m
o
r
t
g
a
g
e
l
o
a
n
i
s
a
c
a
s
h

o
u
t
r
e
ﬁ
n
a
n
c
i
n
g
l
o
a
n
.
P
e
n
n
i
n
g
t
o
n

C
r
o
s
s
a
n
d
C
h
o
m
s
i
s
e
n
g
p
h
e
t
(
2
0
0
7
)
s
h
o
w
t
h
a
t
t
h
e
m
o
s
t
c
o
m
m
o
n
r
e
a
s
o
n
s
t
o
i
n
i
t
i
a
t
e
a
c
a
s
h

o
u
t
r
e
ﬁ
n
a
n
c
i
n
g
a
r
e
t
o
c
o
n
s
o
l
i
d
a
t
e
d
e
b
t
a
n
d
t
o
i
m
p
r
o
v
e
p
r
o
p
e
r
t
y
.
I
n
v
e
s
t
o
r
D
u
m
m
y
(
+
)
E
q
u
a
l
s
o
n
e
i
f
t
h
e
b
o
r
r
o
w
e
r
i
s
a
n
i
n
v
e
s
t
o
r
a
n
d
d
o
e
s
n
o
t
o
w
n
e
r

o
c
c
u
p
y
t
h
e
p
r
o
p
e
r
t
y
.
D
o
c
u
m
e
n
t
a
t
i
o
n
D
u
m
m
y
(

)
E
q
u
a
l
s
o
n
e
i
f
f
u
l
l
d
o
c
u
m
e
n
t
a
t
i
o
n
o
n
t
h
e
l
o
a
n
i
s
p
r
o
v
i
d
e
d
a
n
d
z
e
r
o
o
t
h
e
r
w
i
s
e
.
W
e
e
x
p
e
c
t
f
u
l
l
d
o
c
u
m
e
n
t
a
t
i
o
n
t
o
b
e
a
p
o
s
i
t
i
v
e
s
i
g
n
a
l
o
n
b
o
r
r
o
w
e
r
q
u
a
l
i
t
y
.
P
r
e
p
a
y
m
e
n
t
P
e
n
a
l
t
y
D
u
m
m
y
(
+
)
E
q
u
a
l
s
o
n
e
i
f
t
h
e
r
e
i
s
a
p
r
e
p
a
y
m
e
n
t
p
e
n
a
l
t
y
a
n
d
z
e
r
o
o
t
h
e
r
w
i
s
e
.
W
e
e
x
p
e
c
t
t
h
a
t
a
p
r
e
p
a
y
m
e
n
t
p
e
n
a
l
t
y
m
a
k
e
s
r
e
ﬁ
n
a
n
c
i
n
g
l
e
s
s
a
t
t
r
a
c
t
i
v
e
.
M
o
r
t
g
a
g
e
R
a
t
e
(
+
)
I
n
i
t
i
a
l
i
n
t
e
r
e
s
t
r
a
t
e
a
s
o
f
t
h
e
ﬁ
r
s
t
p
a
y
m
e
n
t
d
a
t
e
.
A
h
i
g
h
e
r
i
n
t
e
r
e
s
t
r
a
t
e
m
a
k
e
s
i
t
h
a
r
d
e
r
t
o
m
a
k
e
t
h
e
m
o
n
t
h
l
y
m
o
r
t
g
a
g
e
p
a
y
m
e
n
t
.
M
a
r
g
i
n
(
+
)
M
a
r
g
i
n
f
o
r
a
n
a
d
j
u
s
t
a
b
l
e

r
a
t
e
o
r
h
y
b
r
i
d
m
o
r
t
g
a
g
e
o
v
e
r
a
n
i
n
d
e
x
i
n
t
e
r
e
s
t
r
a
t
e
,
a
p
p
l
i
c
a
b
l
e
a
f
t
e
r
t
h
e
ﬁ
r
s
t
i
n
t
e
r
e
s
t
r
a
t
e
r
e
s
e
t
.
A
h
i
g
h
e
r
m
a
r
g
i
n
m
a
k
e
s
i
t
h
a
r
d
e
r
t
o
m
a
k
e
t
h
e
m
o
n
t
h
l
y
m
o
r
t
g
a
g
e
p
a
y
m
e
n
t
.
P
r
o
d
u
c
t
T
y
p
e
D
u
m
m
i
e
s
(
+
)
W
e
c
o
n
s
i
d
e
r
f
o
u
r
p
r
o
d
u
c
t
t
y
p
e
s
:
F
R
M
s
,
H
y
b
r
i
d
s
,
A
R
M
s
,
a
n
d
B
a
l
l
o
o
n
s
.
W
e
i
n
c
l
u
d
e
a
d
u
m
m
y
v
a
r
i
a
b
l
e
f
o
r
t
h
e
l
a
t
t
e
r
t
h
r
e
e
t
y
p
e
s
,
w
h
i
c
h
t
h
e
r
e
f
o
r
e
h
a
v
e
t
h
e
i
n
t
e
r
p
r
e
t
a
t
i
o
n
o
f
t
h
e
p
r
o
b
a
b
i
l
i
t
y
o
f
d
e
l
i
n
q
u
e
n
c
y
r
e
l
a
t
i
v
e
t
o
F
R
M
.
B
e
c
a
u
s
e
w
e
e
x
p
e
c
t
t
h
e
F
R
M
t
o
b
e
c
h
o
s
e
n
b
y
m
o
r
e
r
i
s
k

a
v
e
r
s
e
a
n
d
p
r
u
d
e
n
t
b
o
r
r
o
w
e
r
s
,
w
e
e
x
p
e
c
t
p
o
s
i
t
i
v
e
s
i
g
n
s
f
o
r
a
l
l
t
h
r
e
e
p
r
o
d
u
c
t
t
y
p
e
d
u
m
m
i
e
s
.
O
r
i
g
i
n
a
t
i
o
n
A
m
o
u
n
t
(
?
)
S
i
z
e
o
f
t
h
e
m
o
r
t
g
a
g
e
l
o
a
n
.
W
e
h
a
v
e
n
o
c
l
e
a
r
p
r
i
o
r
o
n
t
h
e
e
ﬀ
e
c
t
o
f
t
h
e
o
r
i
g
i
n
a
t
i
o
n
a
m
o
u
n
t
o
n
t
h
e
p
r
o
b
a
b
i
l
i
t
y
o
f
d
e
l
i
n
q
u
e
n
c
y
,
h
o
l
d
i
n
g
c
o
n
s
t
a
n
t
t
h
e
l
o
a
n

t
o

v
a
l
u
e
a
n
d
d
e
b
t

t
o

i
n
c
o
m
e
r
a
t
i
o
.
H
o
u
s
e
P
r
i
c
e
A
p
p
r
e
c
i
a
t
i
o
n
(

)
M
S
A

l
e
v
e
l
h
o
u
s
e
p
r
i
c
e
a
p
p
r
e
c
i
a
t
i
o
n
f
r
o
m
t
h
e
t
i
m
e
o
f
l
o
a
n
o
r
i
g
i
n
a
t
i
o
n
,
r
e
p
o
r
t
e
d
b
y
t
h
e
O
ﬃ
c
e
o
f
F
e
d
e
r
a
l
H
o
u
s
i
n
g
E
n
t
e
r
p
r
i
s
e
O
v
e
r
s
i
g
h
t
(
O
F
H
E
O
)
.
H
i
g
h
e
r
h
o
u
s
i
n
g
e
q
u
i
t
y
l
e
a
d
s
t
o
b
e
t
t
e
r
o
p
p
o
r
t
u
n
i
t
i
e
s
t
o
r
e
ﬁ
n
a
n
c
e
t
h
e
m
o
r
t
g
a
g
e
l
o
a
n
.
C
h
a
n
g
e
U
n
e
m
p
l
o
y
m
e
n
t
R
a
t
e
(
+
)
S
t
a
t
e

l
e
v
e
l
c
h
a
n
g
e
i
n
t
h
e
u
n
e
m
p
l
o
y
m
e
n
t
r
a
t
e
f
r
o
m
t
h
e
t
i
m
e
o
f
l
o
a
n
o
r
i
g
i
n
a
t
i
o
n
,
r
e
p
o
r
t
e
d
b
y
t
h
e
B
u
r
e
a
u
o
f
E
c
o
n
o
m
i
c
A
n
a
l
y
s
i
s
.
A
n
i
n
c
r
e
a
s
e
i
n
t
h
e
s
t
a
t
e
u
n
e
m
p
l
o
y
m
e
n
t
r
a
t
e
i
n
c
r
e
a
s
e
s
t
h
e
p
r
o
b
a
b
i
l
i
t
y
a
h
o
m
e
o
w
n
e
r
l
o
s
t
h
i
s
j
o
b
,
w
h
i
c
h
i
n
c
r
e
a
s
e
s
t
h
e
p
r
o
b
a
b
i
l
i
t
y
o
f
ﬁ
n
a
n
c
i
a
l
p
r
o
b
l
e
m
s
.
N
e
i
g
h
b
o
r
h
o
o
d
I
n
c
o
m
e
(

)
Z
i
p

c
o
d
e

l
e
v
e
l
m
e
d
i
a
n
i
n
c
o
m
e
i
n
1
9
9
9
f
r
o
m
t
h
e
U
.
S
.
C
e
n
s
u
s
B
u
r
e
a
u
2
0
0
0
.
T
h
e
b
e
t
t
e
r
t
h
e
n
e
i
g
h
b
o
r
h
o
o
d
,
a
s
p
r
o
x
i
e
d
b
y
t
h
e
m
e
d
i
a
n
i
n
c
o
m
e
,
t
h
e
m
o
r
e
m
o
t
i
v
a
t
e
d
a
b
o
r
r
o
w
e
r
m
a
y
b
e
t
o
s
t
a
y
c
u
r
r
e
n
t
o
n
a
m
o
r
t
g
a
g
e
.
16
and the margin for adjustablerate and hybrid loans.
13
In addition, we use three macro variables in the baselinecase speciﬁcation. First, we construct a
variable that measures house price appreciation from the time of origination until the time we evaluate
whether a loan is delinquent. To this end we use metropolitan statistical area (MSA) level house price
indexes from the Oﬃce of Federal Housing Enterprise Oversight (OFHEO) and match loans with MSAs
by using the zip code provided by LoanPerformance.
14
Second, we include the statelevel change in the
unemployment rate from the time of loan origination until the time of performance evaluation, reported by
the Bureau of Economic Analysis. An increase in the state unemployment rate increases the probability
a homeowner lost his or her job, which increases the probability of ﬁnancial problems. Third, we use a
measure for the quality of the neighborhood: zipcodelevel median household income in 1999. The data
are from U.S. Census Bureau 2000 and are collected every 10 years. The better the neighborhood the
more motivated a borrower may be to stay current on a mortgage. In Table 2 we report the expected sign
for the regression coeﬃcient on each of the explanatory variables in parentheses.
4.2 Determinants of Delinquency
In Tables 3, 4 and 5 we present the determinants of delinquency using the proportional odds methodology.
The tables report the output of a single estimation; due to limited page size, the output is spread over
three separate tables. The ﬁrst column of Table 3 lists the covariates included (other than the vintage and
age dummies which are reported in Tables 4 and 5). Column two documents the marginal eﬀect of the
covariates (Equation 7). All marginal eﬀects have the expected sign, as shown in Table 2. Except for the
hybrid and ARM dummies, all variables are statistically signiﬁcant at the 1% level. The four explanatory
variables with the largest (absolute) marginal eﬀects and thus the most important for explaining cross
sectional diﬀerences in loan performance, are the FICO score, the combined loantovalue ratio, the
mortgage rate, and house price appreciation. According to the estimates, for example, a one standard
deviation increase in the FICO score decreases the log odds of ﬁrsttime delinquency by 47.94 percent;
or, equivalently, changes the odds by a factor exp(−0.4794) = 0.6192. The product type has a relatively
small eﬀect on the performance of a loan, beyond what is explained by other characteristics. In Figure
13
We also studied speciﬁcations that included loan purpose, reported in Table 1, and housing outlook, deﬁned as the house
price accumulation in the year prior to the loan origination. These variables were not signiﬁcant and did not materially
change the regression coeﬃcients on the other variables.
14
Estimating house price appreciation on the MSAlevel, as opposed to the individual property level introduces a potential
measurement error of this variable. To the best of our knowledge, there is no data available to estimate the size of this
measurement error or to evaluate its impact on the results.
17
3 we show that FRMs experience a much lower delinquency rate than hybrid mortgages, which therefore
must be driven by borrowers with better characteristics selecting into FRMs.
15
The contributions of each covariate to explaining diﬀerent delinquency rates for each vintage year are
given in columns 3 through 9 of Table 3. The very high delinquency rate for vintage 2001 loans can
be explained in large part by a near perfect storm of unfavorable lending and economic conditions: low
FICO scores, high mortgage rates, relatively low house price appreciation, and low (negative) change in
unemployment, all contributing to a higher probability of delinquency. In total, the diﬀerent covariates
contributed to a 51.71 percent increase in the log odds of delinquency, compared to a situation with average
covariate values. This is slightly higher than the 45.06 percent for vintage 2006 and slightly smaller than
the 56.09 percent for vintage 2007—years for which low house price appreciation and high mortgage rates
increased the probability of delinquency. For vintage years 2003 and 2005, high house price appreciation
contributed to a reduced probability of delinquency compared to a situation with average covariate values.
Therefore, we can say that high house price appreciation between 2003 and 2005 masked the true riskiness
of subprime mortgages for these vintage years.
16
By construction, the weightedaverage contribution of a variable over 2001–2007 is zero, with weights
equal to the number of originated loans in a particular vintage year. Because the number of loans
originated diﬀers across vintages years, the equalweighted average contribution is not zero, hence rows
for the contribution variable in Table 3 do not add up to zero.
As shown in Table 4, the coeﬃcients of the vintage dummy variables increase every year, which demon
strates that loan quality deteriorated after adjusting for the other covariates included in the speciﬁcation.
This deterioration is illustrated by the adjusted delinquency rates depicted in Figure 1 (right panel), which
is computed using Equation 12. This picture is in sharp contrast with that obtained from actual rates,
where 2003 was the year with the lowest delinquency rates, and 2001 was the year with the thirdhighest
rates (see Figure 1, left panel). We test whether the diﬀerences between every subsequent vintage year
dummy coeﬃcients are statistically signiﬁcant. We ﬁnd that the yearly change (increase) in the dummy
variables are statistically signiﬁcant at the 1% conﬁdence level, except for the small increase from 2006
to 2007. The vintage dummy coeﬃcients reported in Table 4 are in the same units as the contribution of
the diﬀerent variables presented in Table 3. For example, comparing vintages 2001 and 2007, the total
15
Consistent with this ﬁnding, LaCourLittle (2007) shows that individual credit characteristics are important for mortgage
product choice.
16
Shiller (2007) argues that house prices were too high compared to fundamentals in this period and refers to the house
price boom as a classic speculative bubble largely driven by an extravagant expectation for future house price appreciation.
18
T
a
b
l
e
3
:
D
e
t
e
r
m
i
n
a
n
t
s
o
f
D
e
l
i
n
q
u
e
n
c
y
,
V
a
r
i
a
b
l
e
s
O
t
h
e
r
T
h
a
n
V
i
n
t
a
g
e
a
n
d
A
g
e
D
u
m
m
y
V
a
r
i
a
b
l
e
s
T
h
e
t
a
b
l
e
s
h
o
w
s
t
h
e
o
u
t
p
u
t
o
f
t
h
e
p
r
o
p
o
r
t
i
o
n
a
l
o
d
d
s
d
u
r
a
t
i
o
n
m
o
d
e
l
.
T
h
e
ﬁ
r
s
t
c
o
l
u
m
n
r
e
p
o
r
t
s
t
h
e
c
o
v
a
r
i
a
t
e
s
i
n
c
l
u
d
e
d
,
o
t
h
e
r
t
h
a
n
t
h
e
v
i
n
t
a
g
e
a
n
d
a
g
e
d
u
m
m
i
e
s
w
h
i
c
h
a
r
e
r
e
p
o
r
t
e
d
s
e
p
a
r
a
t
e
l
y
i
n
T
a
b
l
e
4
.
T
h
e
s
e
c
o
n
d
c
o
l
u
m
n
r
e
p
o
r
t
s
t
h
e
m
a
r
g
i
n
a
l
e
ﬀ
e
c
t
,
d
e
ﬁ
n
e
d
i
n
E
q
u
a
t
i
o
n
7
.
A
“
∗
”
i
n
d
i
c
a
t
e
s
s
t
a
t
i
s
t
i
c
a
l
s
i
g
n
i
ﬁ
c
a
n
c
e
a
t
t
h
e
1
%
l
e
v
e
l
.
C
o
l
u
m
n
s
t
h
r
e
e
t
h
r
o
u
g
h
n
i
n
e
d
e
t
a
i
l
t
h
e
c
o
n
t
r
i
b
u
t
i
o
n
o
f
a
v
a
r
i
a
b
l
e
t
o
e
x
p
l
a
i
n
a
d
i
ﬀ
e
r
e
n
t
p
r
o
b
a
b
i
l
i
t
y
o
f
d
e
l
i
n
q
u
e
n
c
y
i
n
2
0
0
1
–
2
0
0
7
(
E
q
u
a
t
i
o
n
9
)
.
E
x
p
l
a
n
a
t
o
r
y
V
a
r
i
a
b
l
e
M
a
r
g
i
n
a
l
E
ﬀ
e
c
t
,
%
C
o
n
t
r
i
b
u
t
i
o
n
,
%
2
0
0
1
–
2
0
0
7
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
2
0
0
7
F
I
C
O
S
c
o
r
e
−
4
7
.
9
4
∗
1
3
.
7
4
6
.
7
4
−
0
.
3
6
−
1
.
0
0
−
3
.
0
9
−
1
.
0
4
3
.
0
0
C
o
m
b
i
n
e
d
L
o
a
n

t
o

V
a
l
u
e
R
a
t
i
o
2
4
.
0
2
∗
−
7
.
4
4
−
5
.
9
2
−
2
.
5
2
0
.
2
3
2
.
2
0
3
.
2
9
−
1
.
4
3
D
e
b
t

t
o

I
n
c
o
m
e
R
a
t
i
o
1
2
.
8
6
∗
−
2
.
7
6
−
2
.
7
3
−
0
.
6
6
0
.
3
1
−
0
.
6
3
2
.
8
6
0
.
0
4
M
i
s
s
i
n
g
D
e
b
t

t
o

I
n
c
o
m
e
D
u
m
m
y
1
2
.
5
9
∗
2
.
1
7
2
.
3
5
0
.
2
4
−
0
.
5
1
0
.
8
5
−
2
.
4
0
0
.
6
7
C
a
s
h

O
u
t
D
u
m
m
y
−
1
2
.
7
3
∗
−
0
.
9
8
−
0
.
6
0
−
0
.
6
9
−
0
.
3
5
0
.
5
9
0
.
7
8
−
0
.
9
3
I
n
v
e
s
t
o
r
D
u
m
m
y
3
.
8
6
∗
0
.
0
4
−
0
.
0
3
−
0
.
0
1
0
.
0
2
0
.
0
0
−
0
.
0
2
−
0
.
0
3
D
o
c
u
m
e
n
t
a
t
i
o
n
D
u
m
m
y
−
1
3
.
7
9
∗
−
3
.
4
2
−
1
.
4
9
−
0
.
5
4
0
.
0
1
0
.
7
8
0
.
9
0
−
0
.
5
2
P
r
e
p
a
y
m
e
n
t
P
e
n
a
l
t
y
D
u
m
m
y
6
.
1
3
∗
0
.
5
0
0
.
3
7
0
.
1
8
0
.
0
5
−
0
.
0
6
−
0
.
3
2
−
0
.
4
6
M
o
r
t
g
a
g
e
R
a
t
e
2
9
.
2
1
∗
3
9
.
5
1
1
6
.
8
6
−
4
.
7
8
−
1
3
.
0
0
−
8
.
2
5
1
1
.
3
4
1
3
.
6
8
M
a
r
g
i
n
1
1
.
8
4
∗
−
2
.
0
7
0
.
1
9
−
1
.
8
1
0
.
3
2
1
.
1
0
0
.
3
2
−
2
.
4
3
H
y
b
r
i
d
D
u
m
m
y
1
.
8
1
−
0
.
2
3
0
.
0
3
−
0
.
1
2
0
.
2
9
0
.
3
4
−
0
.
5
3
−
0
.
9
5
A
R
M
D
u
m
m
y
0
.
3
6
0
.
0
0
0
.
0
0
−
0
.
0
1
0
.
0
0
0
.
0
0
0
.
0
0
0
.
0
0
B
a
l
l
o
o
n
D
u
m
m
y
3
.
7
0
∗
−
0
.
1
6
−
0
.
7
1
−
0
.
9
4
−
1
.
0
2
−
0
.
4
8
2
.
4
8
2
.
9
5
O
r
i
g
i
n
a
t
i
o
n
A
m
o
u
n
t
1
5
.
9
1
∗
−
7
.
3
1
−
5
.
0
3
−
2
.
6
4
−
0
.
6
2
2
.
0
5
3
.
3
1
4
.
0
8
H
o
u
s
e
P
r
i
c
e
A
p
p
r
e
c
i
a
t
i
o
n
−
2
9
.
9
6
∗
7
.
2
4
5
.
3
2
−
1
0
.
5
2
−
1
6
.
4
0
−
2
.
2
4
2
1
.
3
8
3
2
.
8
6
C
h
a
n
g
e
U
n
e
m
p
l
o
y
m
e
n
t
7
.
6
9
∗
1
2
.
7
6
3
.
0
3
−
1
.
8
5
−
2
.
4
0
−
2
.
3
1
2
.
2
5
5
.
1
8
N
e
i
g
h
b
o
r
h
o
o
d
I
n
c
o
m
e
−
8
.
8
1
∗
0
.
1
1
−
0
.
4
0
−
0
.
5
3
−
0
.
1
1
0
.
1
1
0
.
4
8
0
.
3
8
T
o
t
a
l
−
5
1
.
7
1
1
7
.
9
9
−
2
7
.
5
5
−
3
4
.
1
9
−
9
.
0
4
4
5
.
0
6
5
6
.
0
9
19
Table 4: Determinants of Delinquency, Vintage Dummy Variables
The table shows the output of the proportional odds duration model. The ﬁrst column reports the vintage dummies included. The
values for the other covariates are outlined in Tables 3 and 5. The second column shows the estimated coeﬃcients. The vintage 2001
dummy was not included as covariate; hence the coeﬃcients on the other covariates are relative to the 2001 vintage and we report a
zero value for the coeﬃcient of 2001. A “∗” indicates statistical signiﬁcance at the 1% level.
Explanatory Variable Estimate, %
Vintage 2001 0.00
Vintage 2002 5.12
∗
Vintage 2003 23.64
∗
Vintage 2004 49.71
∗
Vintage 2005 57.93
∗
Vintage 2006 64.65
∗
Vintage 2007 66.69
∗
Table 5: Determinants of Delinquency, Age Dummy Variables
The table shows the output of the proportional odds duration model. Each column reports the corresponding odds ratio, AgeEﬀect,
estimated for each age dummy variable based on Equation 6. The values for the other covariates are reported in Tables 3 and 4.
Age, Months 3 4 5 6 7 8 9 10
AgeEﬀect (*100) 0.04 0.31 0.55 0.67 0.72 0.72 0.73 0.72
Age, Months 11 12 13 14 15 16 17 –
AgeEﬀect (*100) 0.73 0.72 0.73 0.71 0.70 0.69 0.67 –
20
contribution of the (nonvintage dummy) covariates increased by 4.38 percent (from 51.71 percent to
56.09 percent). This pales in comparison to the 66.69 percent increase in the vintage dummy variable
over 2001–2007.
To illustrate the eﬀect of age on the conditional probability of ﬁrsttime delinquency, in Table 5 we
report the odds statistic deﬁned in Equation 6. We see that the odds of ﬁrsttime delinquency peaks
around age of 7–13 months.
Next we study the following question: Based on information available at the end of 2005, was the
dramatic deterioration of loan quality since 2001 already apparent? Notice that we cannot answer this
question by simply inspecting vintages 2001 through 2005 in Figure 1 (right panel), because the com
putation of the adjusted delinquency rate for, say, vintage 2001 loans, makes use of a regression model
estimated using data from 2001 through 2008. Hence, we reestimate the proportional odds model under
lying Figure 1 (right panel) making use of only 2001–2005 data. The resulting age pattern in adjusted
delinquency rates is plotted in Figure 5 (left panel). We again obtain the result that the adjusted delin
quency rate rose monotonically from 2001 onward. We therefore conclude that the dramatic deterioration
of loan quality in this decade should have been apparent by the end of 2005. Figure 5 (right panel) depicts
the situation when we use data available at the end of 2006. Again, the deterioration is clearly visible.
17
The ﬁnding of a continual decline in loan quality also occurs when analyzing foreclosure rates (Ap
pendix A), and analyzing hybrid mortgages and FRMs separately (Appendix B). Moreover, the main
result documenting the monotonic rise in adjusted delinquency rates is found based on numerous alterna
tive model speciﬁcations discussed in Section 5.
5 Empirical Results for Alternative Speciﬁcations
In this Section we explore various alternative model speciﬁcations.
5.1 Diﬀerent Loan and Borrower Characteristics
We explore numerous alternative loan and borrower characteristics as covariates for robustness. First, we
consider as covariates those of the baseline case presented in Table 3, plus the 10 interaction and quadratic
terms that can be constructed from the four most important variables: the FICO score, the CLTV ratio,
17
One reason why investors did not massively start to avoid or short subprimerelated securities is that the timing of the
subprime market downturn may have been hard to predict. Moreover, a short position is associated with a high cost of carry
(Feldstein (2007)).
21
Figure 5: Adjusted Delinquency Rate, Viewed at the End of 2005 and 2006
The ﬁgure shows the adjusted delinquency rate using data available at the end of 2005 (left panel) and 2006 (right panel). The
delinquency rate is deﬁned as the cumulative fraction of loans that were past due 60 or more days, in foreclosure, realestate owned, or
defaulted, at or before a given age. The adjusted delinquency rate is obtained by adjusting the actual rate for yearbyyear variation in
FICO scores, loantovalue ratios, debttoincome ratios, missing debttoincome ratio dummies, cashout reﬁnancing dummies, owner
occupation dummies, documentation levels, percentage of loans with prepayment penalties, mortgage rates, margins, composition of
mortgage contract types, origination amounts, MSA house price appreciation since origination, change in state unemployment rate since
origination, and neighborhood median income.
0
2
4
6
8
10
12
14
16
3 4 5 6 7 8 9 10 11
Loan Age (Months)
2005
2004
2003
2002
2001
Adjusted Delinquency Rate (%), End of 2005
0
2
4
6
8
10
12
14
16
3 4 5 6 7 8 9 10 11
Loan Age (Months)
2006
2005
2004
2003
2002
2001
Adjusted Delinquency Rate (%), End of 2006
22
the mortgage rate, and subsequent house price appreciation. Allowing for these additional terms, we take
into account the eﬀect of risklayering—such as, for example, the eﬀect of a combination of a borrower’s
low FICO score and a high CLTV ratio—on the probability of delinquency. It is in this case not a priori
clear what the sign on the FICOCLTV interaction variable should be. A negative sign would mean that
a low FICO and a high CLTV reinforce each other and give rise to a predicted delinquency probability
that is higher than that without interaction eﬀect. A positive sign could be explained by lenders who
originate a low FICO and high CLTV loan only if they have positive private information on the loan or
borrower quality. We ﬁnd that the coeﬃcient on the FICOCLTV interaction term is close to zero and
insigniﬁcant.
More certain is the sign we expect on the HPACLTV variable. Low house price appreciation is
expected to especially give rise to a higher delinquency probability for a high CLTV ratio, because the
borrower is closer to a situation with negative equity in the house (combined value of the mortgage loans
larger than the market value of the house). Consistent with this intuition, we ﬁnd a negative and signiﬁcant
(at the 1% level) coeﬃcient on this interaction term. The vintage dummy variables are still increasing
every year. Inclusion of the 10 interaction covariates does not substantially increase the overall ﬁt of the
model, as measured by the log likelihood ratio.
Second, we included a dummy for the presence of a secondlien loan. We ﬁnd that the coeﬃcient
is positive and statistically signiﬁcant and that it inherits some of the statistical power of the CLTV
variable. The coeﬃcients of the other covariates are virtually unchanged. Inclusion of the dummy does
not substantially increase the overall ﬁt of the model, as measured by the log likelihood ratio.
Third, we considered as an additional covariate a dummy variable taking the value one whenever
the CLTV equals 80 percent. With this variable, we are aiming to control for silent seconds, referring
to a situation where an investor takes out a secondlien loan not reported in our database typically
in combination with an 80 percent ﬁrstlien loan. This dummy variable is statistically signiﬁcant but
economically not very large and moreover hardly improved the overall ﬁt.
Fourth, we excluded the loans where the debttoincome ratio is missing from the sample to make sure
the measurement error associated with this variable does not lead to a signiﬁcant bias in the results. The
estimates based on the smaller subsample, in which the debttoincome variable has nonzero reported
values, are statistically and economically similar to those based on the entire sample of loans.
Fifth, we performed several robustness checks regarding the reset rate for hybrid mortgages. The
23
initial mortgage rate for hybrid mortgages is potentially lower during the initial ﬁxedrate period. For
99.57 percent of mortgages in our database, the duration of the initial ﬁxed rate period is 24 months or
more; the most common ﬁxed period is 24 months, followed by 36 months. Since we focus on delinquency
in the ﬁrst 17 months, we expect the initial mortgage rate to be an important covariate. However, because
households may have factored in the rise in the mortgage rate before the actual reset date, we include the
postreset margin as an additional covariate in the baseline case speciﬁcation. The margin is in excess of
a reference rate, which in 99.34 percent of the cases is the 6month LIBOR rate. As a robustness check
we performed the analysis for FRMs and Hybrid mortgages separately, with FRMs not being subject to
resets, and in both cases obtain our main result that adjusted delinquency rates rose monotonically over
2001–2007 (see Figure 9).
We performed two additional robustness checks. First, instead of the margin we included a covariate
deﬁned as the margin plus the 6month LIBOR interest rate at origination (obtained from Bloomberg).
Second, instead of the margin we included a covariate deﬁned as the margin plus the 6month LIBOR
interest rate at origination minus the initial mortgage rate. This captures the potential change in interest
rate at the time of reset, based on the 6month LIBOR rate at origination. For FRMs the values of the
new covariates are set to zero. The marginal eﬀect for the two new covariates is 12.27 percent and 7.83
percent respectively; the marginal eﬀects for the margin covariates in the baseline case are 11.84 percent,
as reported in Table 3. Among the other covariates, only the coeﬃcient for the initial rate is slightly
aﬀected. The marginal eﬀect of the initial rate is 29.58 percent and 33.60 percent for the two alternative
speciﬁcations respectively, compared to a marginal eﬀect of 29.21 percent in the baseline case, as reported
in Table 3. The log likelihood of the diﬀerent speciﬁcations is virtually identical.
5.2 Local housebyhouse return volatility
In this section we study local housebyhouse return volatility as as additional covariate. We obtained the
data from the Oﬃce of Federal Housing Enterprise Oversight (OFHEO) at the state level.
18
OFHEO uses
a repeatedsales methodology to compute house price appreciation in a geographical unit (like a state or
MSA) and, as a byproduct, obtains an estimate of the variation of the return around the mean in the
geographical unit.
19
18
We would like to thank OFHEO for providing us with the data. MSAlevel data exist but was not available for public
release.
19
For more details on this procedure, see the oﬃcial OFHEO documentation by Calhoun (1996). Also De Jong, Driessen,
and Van Hemert (2008) explain the interpretation and computation of the volatility parameters.
24
The housebyhouse volatility is 8.05 percent, annualized and averaged over the full panel (50 states plus
the District of Columbia in the 2001Q12008Q2 sample period). Most of the variation of the volatility is
in the crosssection: the standard deviation of the volatility estimate across states (averaged over 2001Q1
2008Q2) is 0.75 percent. The standard deviation of the volatility estimate over time (averaged over the
50 states plus the District of Columbia) is only 0.18 percent.
Ambrose, LaCourLittle, and Huszar (2005) use the volatility estimate of OFHEO to compute the
probability of negative equity. In the special case that the equity in the house is exactly zero, based on
house price appreciation in the geographical area, there is a 50 percent probability the household actually
has negative equity based on the (unobserved) house price appreciation of the individual house, assuming
a symmetric (e.g., normal) distribution for the housebyhouse deviation from the MSA mean. When the
equity is positive based on house price appreciation in the geographical area, the higher the housebyhouse
return volatility, the higher the probability an individual house has negative equity. When the equity is
negative, the situation is reversed.
While it is not just the probability of having negative equity for an individual house that matters,
but in general the whole probability distribution, it does seem intuitive that the eﬀect of local houseby
house return volatility may interact with HPA. Hence, in the empirical implementation we add both the
housebyhouse return volatility by itself and the interaction of this volatility with HPA as covariates.
Consistent with Ambrose, LaCourLittle, and Huszar (2005), the estimated coeﬃcient on the interaction
term is positive; when house price appreciation is high in the geographical area, high volatility increases
the probability of having negative equity for a speciﬁc house. The eﬀect is statistically signiﬁcant at
the 1% level, but has a small Chisquared statistic compared to the baseline case covariates and has a
negligible impact on the coeﬃcients of the other covariates.
5.3 CRA and GSE Housing Goals
In this section we explore additional covariates related to the Community Reinvestment Act (CRA) and
Government Sponsored Enterprises (GSEs) housing goals. The 1977 Community Reinvestment Act (CRA)
is a United States federal law designed to encourage commercial banks and savings associations to meet
the needs of borrowers in all segments of their communities, including low and moderateincome (LMI)
households and neighborhoods. The CRA does not list speciﬁc criteria for evaluating the performance
of ﬁnancial institutions, but indicates that the evaluation process should accommodate the situation and
25
context of each individual institution. An institution’s CRA compliance record is, for example, taken
into account when applying for deposit facilities.
20
LMI neighborhoods have a median income level
that is less than 80 percent of the median income of a broader geographic area, e.g., the MSA for urban
neighborhoods.
21
With a similar objective of helping poor and underserved individuals and neighborhoods,
the Congress in 1992 established an aﬀordable housing mission for Fannie Mae and Freddie Mac by
directing the Department of Housing and Urban Developement (HUD) to create speciﬁc mortgage purchase
goals for these Government Sponsored Enterprises (GSEs). The goals are primarily deﬁned in terms of (i)
household income relative to median MSA income (for urban neighborhoods), (ii) neighborhood median
income relative to MSA median income, and (iii) minority concentration in the neighborhood.
22
To isolate the eﬀect of CRA and the GSE housing goals from pure neighborhood eﬀects, we study
the eﬀect of LMI neighborhood status on the probability of delinquency, controlling for the baseline
speciﬁcation covariates, including the neighborhood median income.
We either include an LMI dummy that equals one if the median neighborhood income is less than 80
percent of the MSA median income, or we add the MSA median income. We use zipcode level median
household income and median MSA household income from the U.S. Census Bureau to compute the LMI
dummy.
23
The LMI dummy or MSA median income level pick up the diﬀerence in the probability of delinquency
for loans in MSAs with diﬀerent median incomes, but in neighborhoods with the same median income and
with the same loan and borrower characteristics, giving rise to diﬀerent incentives for CRAcomplying
institutions and the GSEs. We interpret any eﬀect from the LMI dummy or MSA median income to come
from the CRA and GSE housing goals, and thus assume that there is no direct eﬀect of LMI status or
MSA median income on borrower behavior after controlling for neighborhood median income.
The results are presented in Table 6. Speciﬁcation I corresponds to the baseline case of Table 3 and
is included as a benchmark. Of the baseline case covariates, we only report the neighborhood income to
preserve space. In speciﬁcation II we add the MSA median income to the baseline case covariates. A higher
20
See http://www.federalreserve.gov/dcca/cra/
21
See Laderman (2004) for a further discussion.
22
See http://www.huduser.org/Datasets/GSE/gse2006.pdf Table 1 for the speciﬁc goals for 20002006.
23
The precise identiﬁcation of CRA and GSE targeted neighborhoods is not feasible in our analysis for several reasons: (i)
the location of subprime lenders is not reported in the LoanPerformance data, thus we cannot identify the CRA targeted
community; and (ii) most frequently, CRA and GSE targeted communities are deﬁned using census tractlevel measures (such
as income, poverty level, and minority concentration). LoanPerformance data identiﬁes each property location in zip codes,
which do not directly match with census tracts. Therefore, in our analysis we identify neighborhoods that can potentially
fall under CRA and GSE housing goals using either the LMI dummy variable or MSAlevel median household income.
26
MSA median income increases the possibility a loan was made to advance CRA or GSE housing goals.
The positive and statistically signiﬁcant (at the 1% level) marginal eﬀect reported in the table eﬀect thus
implies that CRA and GSE housing goal eligibility is associated with a higher probability of delinquency,
ceteris paribus. In speciﬁcation III we add the LMI dummy to the baseline case covariates. We ﬁnd a
positive and signiﬁcant marginal eﬀect, again implying that CRA and GSE housing goal eligibility raises
the probability of delinquency. Finally in speciﬁcation IV we interact the LMI dummy with vintage
dummies to see how the LMI eﬀect changed over time. For all vintage years the marginal eﬀect is positive
and statistically signiﬁcant at the 1% level, except for 2003, which is positive and statistically signiﬁcant
at the 5% level. There is no clear trend over time.
Table 6: Low and ModerateIncome Neighborhood Eﬀect
This table reports marginal eﬀects for diﬀerent measures of neighborhood (median household) income relative to MSA (median household)
income. In all four speciﬁcations we include the baseline case variables used in Table 3, of which we only report neighborhood income
to preserve space. Speciﬁcation I is the baseline case with no variable capturing the neighborhood income relative to the MSA income.
Speciﬁcation II includes MSA income. Speciﬁcation III includes a dummy with value one if neighborhood income is less than 80 percent
of MSA income; i.e., if it is a low and moderateincome (LMI) neighborhood. Speciﬁcation IV includes interaction variables of the
LMI dummy with vintage year dummies. The marginal eﬀect for the interaction terms is computed as the product of the estimated
coeﬃcient and the full sample standard deviations of LMI. A “∗” indicates statistical signiﬁcance at the 1% level.
I II III IV
Neighborhood Income −8.81%* −9.10%∗ −6.95%∗ −6.94%∗
MSA Income − 1.07%∗ − −
LMI Dummy − − 2.82%∗ −
LMI 2001 − − − 6.26%∗
LMI 2002 − − − 4.52%∗
LMI 2003 − − − 1.75%
LMI 2004 − − − 3.64%∗
LMI 2005 − − − 2.67%∗
LMI 2006 − − − 1.79%∗
LMI 2007 − − − 3.79%∗
27
6 NonStationarity of the LoantoValue Eﬀect
The proportional odds model used in Section 4 assumes that the covariate coeﬃcients are constant over
time. That is, the eﬀect of a unit change in a given covariate on the probability of ﬁrsttime delinquency
is the same in, for example, 2006, as it is in 2001, holding constant the values of the other covariates. We
test the validity of this assumption for all variables in our analysis and ﬁnd that the strongest rejection of a
constant coeﬃcient is for the CLTV ratio. In this section we ﬁrst discuss this ﬁnding and then turn to the
question of whether lenders were aware of the nonstationarity of the loantovalue eﬀect, by investigating
the relationship between the loantovalue ratio and mortgage rates over time.
We estimate the proportional odds model with CLTV interacted with the seven vintage dummy vari
ables, instead of just CLTV without interaction, as we have in the baseline case speciﬁcation (Section 4).
We compute the marginal eﬀect for a particular loan vintage as the product of the estimated coeﬃcient
for the associated interaction variable and the standard deviation of CLTV for all vintages together, which
totals 0.10, 0.11, 0.14, 0.15, 0.22, 0.35, 0.34 for years 2001 to 2007, respectively. All seven coeﬃcients are
highly statistically signiﬁcant at the 1% level. Hence, the CLTV is an increasingly important determinant
of delinquency over time.
We examine whether lenders were aware that high LTV ratios were increasingly associated with riskier
borrowers. The combined LTV ratio rather than the ﬁrstlien LTV ratio is believed to be the main
determinant of delinquency because it is the burden of all the debt together that may trigger ﬁnancial
problems for the borrower. In contrast, the ﬁrstlien LTV is the more important determinant of the
mortgage rate on a ﬁrstlien mortgage, because it captures the dollar amount at stake for the ﬁrst
lien lender.
24
For this reason, we test whether the sensitivity of the lender’s interest rate to the ﬁrst
lien LTV ratio changed over time. We perform a crosssectional OLS regression with the mortgage
rate as the dependent variable, and loan characteristics, including the ﬁrstlien LTV and secondlien
LTV (CLTV minus ﬁrstlien LTV), as independent variables.
25
We perform one such regression for
each calendar quarter in our sample period. We can only expect to get accurate results when using
relatively homogeneous groups of loans, and therefore consider fully amortizing FRM and 2/28 hybrid
loans separately. Together these two contract types account for more than half of all mortgage loans in
24
This is conﬁrmed by our empirical results. To conserve space the results are not reported.
25
Speciﬁcally, we use the FICO score, ﬁrstlien loantovalue ratio, secondlien loantovalue ratio, debttoincome ratio, a
dummy for a missing debttoincome ratio, a cashout reﬁnancing dummy, a dummy for owner occupation, documentation
dummy, prepayment penalty dummy, margin, origination amount, term of the mortgage, and prepayment term as the right
handside variables.
28
our database. Each crosssectional regression is based on a minimum of 18,784 observations.
Figure 2 shows the regression coeﬃcient on the ﬁrstlien LTV ratio for each quarter from 2001Q1
through 2007Q2.
26
We scaled the coeﬃcients by the standard deviation of the ﬁrstlien LTV ratio, and
they can therefore be interpreted as the changes in the mortgage rates when the ﬁrstlien LTV ratios are
increased by one standard deviation. In the fourth quarter of 2006, a onestandarddeviation increase in the
ﬁrstlien LTV ratio corresponded to about a 30basispoint increase in the mortgage rate for 2/28 hybrids
and about a 40basispoint increase for FRMs, keeping constant other loan characteristics. In contrast,
in the ﬁrst quarter of 2001, the corresponding rate increase was 10 and 16 basis points, respectively.
This provides evidence that lenders were to some extent aware of high LTV ratios being increasingly
associated with risky borrowers.
27
In Appendix C we show that this result is robust to allowing for a
nonlinear relationship between the mortgage rate and the ﬁrstlien LTV ratio. Finally, notice that the
eﬀect of a onestandarddeviation increase in the ﬁrstlien LTV ratio on the 2/28 mortgage rate increased
substantially in the wake of the subprime mortgage crisis: from 30 basis points in 2007Q1 to 42 basis
points in 2007Q2.
7 SubprimePrime Rate Spread
In general, interest rates on subprime mortgages are higher than on prime mortgages to compensate the
lender for the (additional) default risk associated with subprime loans. In this section we analyze the
time series of the subprimeprime rate spread, both with and without adjustment for changes in loan
and borrower characteristics. We focus on FRMs for this exercise. For hybrid mortgages the subprime
prime comparison is more complicated because (i) both the initial (teaser) rate and the margin should be
factored in, and (ii) we don’t have good data on the prime initial rates and margins.
In Figure 6 we show the actual subprimeprime rate spread, deﬁned in Equation (15) below. The
subprime rate for this exercise is calculated as the average across individual loans initial mortgage rate
for each calendar month (the data source is LoanPerformance); the prime rate is the contract rate on
FRMs reported by the Federal Housing Finance Board (FHFB) in its Monthly Interest Rate Survey.
28
The subprimeprime spread—the diﬀerence between the average subprime and prime rates—decreased
26
Our data extends to 2007Q3, but due to a near shutdown of the securitized subprime mortgage market we lack statistical
power in this quarter.
27
The eﬀects of other loan characteristics on mortgage rates have been much more stable over time, as unreported results
suggest.
28
Available at http://www.fhfb.gov/GetFile.aspx?FileID=6416.
29
substantially over time, with the largest decline between 2001 and 2004, which coincides with the most
rapid growth in the number of loans originated (see Table 1). In Figure 6 we also plot the yield spread
between 10year BBB and AAA corporate bonds, which we obtained from Standard and Poor’s Global
Fixed Income Research. Compared to the corporate BBBAAA yield spread, the actual subprimeprime
rate spread declined much more and more steadily, hence the decline cannot just be attributed to a change
in the overall level of risk aversion.
Figure 6: FRM Rate Spread and Corporate Bond Yield Spread
The ﬁgure shows the FRM subprimeprime rate spread and the yield spread between 10year BBB and AAA corporate bonds.
0
.5
1
1.5
2
2.5
3
3.5
2001 2002 2003 2004 2005 2006 2007
Year
Subprime−Prime Spread (%)
BBB−AAA Spread (%)
We perform a crosssectional OLS regression with the loanlevel spread as the dependent variable and
the prime rate and various subprime loan and borrower characteristics as the explanatory variables, using
data from 2001 through 2006.
29
spread
it
= β
0
+ β
1
prime
t
+ β
2
characteristics
it
+ error
it
, (14)
spread
it
= subprime
it
− prime
t
(15)
Notice that the β
1
prime
t
term corrects for the fact that the spread is aﬀected by the prime rate itself,
and thus changes over the business cycle, because a higher prime rate increases the default probability
29
The explanatory factors in the regression are the FICO credit score, a dummy variable that equals one if full docu
mentation was provided, a dummy variable that equals one if prepayment penalty is present, origination amount, value of
debttoincome ratio, a dummy variable that equals one if debttoincome was not provided, a dummy variable that equals
one if loan is a reﬁnancing, a dummy variable that equals one if a borrower is an investor, loantovalue ratio based on a
ﬁrstlien, and loantovalue ratios based on a second, third, etc. liens if applicable.
30
on subprime loans for a given spread. In Figure 7 we plot the prediction error, averaged per origination
month t, along with a ﬁtted linear trend.
Figure 7: Prediction Error in the SubprimePrime Rate Spread
The ﬁgure shows the prediction error in the subprimeprime rate spread, determined in a regression of the spread on the prime rate
and the following loan and borrower characteristics: the FICO credit score, a dummy variable that equals one if full documentation
was provided, a dummy variable that equals one if a prepayment penalty is present, origination amount, value of debttoincome ratio,
a dummy variable that equals one if debttoincome was not provided, a dummy variable that equals one if the loan is a reﬁnancing,
a dummy variable that equals one if a borrower is an investor, the loantovalue ratio based on a ﬁrst lien, and the loantovalue ratio
based on a second, third, etc. liens if applicable.
−.5
−.25
0
.25
.5
.75
2001 2002 2003 2004 2005 2006 2007
Year
Prediction Error (%)
Fitted Trend
The downward trend in Figure 7 indicates that the subprimeprime spread, after adjusting for diﬀer
ences in observed loan and borrower characteristics, declined between 2001 to 2007. In Figure 1 (right
panel) we showed that loan quality, obtained by adjusting loan performance for diﬀerences in loan and
borrower characteristics and subsequent house price appreciation, deteriorated over the period, and thus
the (adjusted) riskiness of loans rose. Therefore, on a perunitofrisk basis, the subprimeprime mortgage
spread decreased even more than the level of the spread.
8 Concluding Remarks
The subprime mortgage market experienced explosive growth between 2001 and 2006. Angell and Rowley
(2006) and Kiﬀ and Mills (2007), among others, argue that this was facilitated by the development of
socalled privatelabel mortgage backed securities (they do not carry any kind of credit risk protection by
the Government Sponsored Enterprises). Investors in search of higher yields kept increasing their demand
31
for privatelabel mortgagebacked securities, which also led to sharp increases in the subprime share of
the mortgage market (from around 8 percent in 2001 to 20 percent in 2006) and in the securitized share
of the subprime mortgage market (from 54 percent in 2001 to 75 percent in 2006).
In this paper we show that during the dramatic growth of the subprime (securitized) mortgage market,
the quality of the market deteriorated dramatically. We measure loan quality as the performance of loans,
adjusted for diﬀerences in borrower characteristics (such as the credit score, a level of indebtedness, an
ability to provide documentation), loan characteristics (such as a product type, an amortization term, a
loan amount, an mortgage interest rate), and macroeconomic conditions (such as house price appreciation,
level of neighborhood income and change in unemployment).
The decline in loan quality was monotonic, but not equally spread among diﬀerent types of borrowers.
Over time, highLTV borrowers became increasingly risky (their adjusted performance worsened more)
compared to lowLTV borrowers. Securitizers seem to have been aware of this particular pattern in the
relative riskiness of borrowers: We show that over time mortgage rates became more sensitive to the LTV
ratio of borrowers. In 2001, for example, a borrower with a one standard deviation aboveaverage LTV
ratio paid a 10 basis point premium compared to an average LTV borrower. By 2006, in contrast, the
premium paid by the high LTV borrower was around 30 basis points.
In principal, the subprimeprime mortgage rate spread (subprime markup) should account for the
default risk of subprime loans. For the rapid growth of the subprime mortgage market to have been
sustainable, the increase in the overall riskiness of subprime loans should have been accompanied by an
increase in the subprime markup. In this paper we show that this was not the case: The subprime mark
up—adjusted and not adjusted for changes in diﬀerences in borrower and loan characteristics—declined
over time. With the beneﬁt of hindsight we now know that indeed this situation was not sustainable,
and the subprime mortgage market crashed in 2007. In many respects, the subprime market experienced
a classic lending boombust scenario with rapid market growth, loosening underwriting standards, dete
riorating loan performance, and decreasing risk premiums.
30
Argentina in 1980, Chile in 1982, Sweden,
Norway, and Finland in 1992, Mexico in 1994, Thailand, Indonesia, and Korea in 1997 all experienced the
culmination of a boombust scenario, albeit in diﬀerent economic settings.
Were problems in the subprime mortgage market apparent before the actual crisis erupted in 2007?
30
A more detailed discussion, theory, and empirical evidence on such episodes is available in Dell’Ariccia and Marquez
(2006), Demirg¨ u¸ cKunt and Detragiache (2002), Gourinchas, Valdes, and Landerretche (2001), and Kamisky and Reinhart
(1999), among many others.
32
Our answer is yes, at least by the end of 2005. Using the data available only at the end of 2005, we
show that the monotonic degradation of the subprime market was already apparent. Loan quality had
been worsening for ﬁve years in a row at that point. Rapid appreciation in housing prices masked the
deterioration in the subprime mortgage market and thus the true riskiness of subprime mortgage loans.
When housing prices stopped climbing, the risk in the market became apparent.
33
References
Allison, Paul, 2007, Survival Analysis Using SAS: A Practical Guide (SAS Institute, 10th Printing, Cary, NC,
USA).
Ambrose, Brent, and Charles Capone, 2000, The hazard rates of ﬁrst and second defaults, Journal of Real Estate
Finance and Economics 20, 275–293.
Ambrose, Brent, Michael LaCourLittle, and Zsuzsa Huszar, 2005, A note on hybrid mortgages, Real Estate Eco
nomics 33, 765–782.
Angell, Cynthia, and Clare D. Rowley, 2006, Breaking new ground in U.S. mortgage lending, FDIC: Outlook
Summer 2006 Federal Deposit Insurance Corporation.
Berger, Allen, and Gregory Udell, 2004, The institutional memory hypothesis and the procyclicality of bank lending
behavior, Journal of Financial Intermediation 12, 458–495.
Bernanke, Ben, 2007, Housing, housing ﬁnance, and monetary policy, Speech Federal Reserve Bank of Kansas City’s
Economic Symposium.
Brunnermeier, Markus, and Christian Julliard, 2007, Money illusion and housing frenzies, Review of Financial
Studies forthcoming.
Calhoun, Charles, 1996, Ofheo house price indexes: HPI technical description, Oﬃcial OFHEO document.
Calhoun, Charles, and Yongheng Deng, 2002, A dynamic analysis of ﬁxed and adjustablerate mortgage termina
tions, Journal of Real Estate Finance and Economics 24, 9–33.
Campbell, John Y., and Jo˜ao F. Cocco, 2003, Household risk management and optimal mortgage choice, Quarterly
Journal of Economics 118, 1449–1494.
Clapp, John, Yongheng Deng, and Xudong An, 2006, Unobserved heterogeneity in models of competing mortgage
termination risks, Real Estate Economics 34, 243–273.
Cox, David, 1972, Regression models and life tables, Journal of the Royal Statistical Society B34, 187–220.
Cutts, Amy Crews, and Robert A. Van Order, 2005, On the economics of subprime lending, The Journal of Real
Estate Finance and Economics 30, 167–196.
De Jong, Frank, Joost Driessen, and Otto Van Hemert, 2008, Hedging house price risk: Portfolio choice with housing
futures, Working paper.
DellAriccia, Giovanni, Deniz Igan, and Luc Laeven, 2008, Credit booms and lending standards: Evidence from the
subprime mortgage market, Working paper.
34
Dell’Ariccia, Giovanni, and Robert Marquez, 2006, Lending booms and lending standards, Journal of Finance 61,
2511–2546.
Demirg¨ u¸ cKunt, Asli, and Enrica Detragiache, 2002, Does deposit insurance increase banking system stability? An
empirical investigation, Journal of Monetary Economics 49, 1373–1406.
Deng, Yongheng, 1997, Mortgage termination: An empirical hazard model with stochastic term structure, Journal
of Real Estate Finance and Economics 14, 309–331.
Deng, Yongheng, Andrey Pavlov, and Lihong Yang, 2005, Spatial heterogeneity in mortgage terminations by reﬁ
nance, sale and default, Real Estate Economics 33, 739–764.
Deng, Yongheng, John M. Quigley, and Robert A. Van Order, 2000, Mortgage terminations, heterogeneity and the
exercise of mortgage options, Econometrica 68, 275–308.
Feldstein, Martin S., 2007, Housing, credit markets and the business cycle, NBER working paper 13471.
Genesove, David, and Christopher Mayer, 1997, Equity and time to sale in the real estate market, American
Economic Review 87, 255–269.
Genesove, David, and Christopher Mayer, 2001, Loss aversion and seller behavior: Evidence from the housing
market, Quarterly Journal of Economics 116, 1233–1260.
Gerardi, Kristopher, Adam Shapiro, and Paul Willen, 2008, Subprime outcomes: Risky mortgages, homeownership
experiences, and foreclosures, Working paper.
Glaeser, Edward, and Joseph Gyourko, 2005, Urban decline and durable housing, Journal of Political Economy 113,
345–375.
Gourinchas, PierreOliver, Rodrigo Valdes, and Oscar Landerretche, 2001, Lending booms: Latin America and the
world, Economia 1, 47–99.
Gyourko, Joseph, Mayer Christopher, and Todd Sinai, 2006, Superstar cities, NBER working paper 12355.
Kamisky, Graciela L., and Carmen M. Reinhart, 1999, The twin crises: The causes of banking and balanceof
payment problems, American Economic Review 89, 473–500.
Keys, Benjamin J., Tanmoy Mukherjee, Amit Seru, and Vikrant Vig, 2008, Did securitization lead to lax screening?
Evidence from subprime loans, Working paper.
Kiﬀ, John, and Paul Mills, 2007, Money for nothing and checks for free: Recent developments in U.S. subprime
mortgage markets, IMF working paper.
Koijen, Ralph, Otto Van Hemert, and Stijn Van Nieuwerburgh, 2007, Mortgage timing, NBER working paper 13361.
35
LaCourLittle, Michael, 2007, The home purchase mortgage preferences of lowandmoderate income households,
Real Estate Economics forthcoming.
Laderman, Liz, 2004, Has the cra increased lending for lowincome home purchases? Federal reserve of san francisco
economic letter, 200416.
Mian, Atif, and Amir Suﬁ, 2008, The consequences of mortgage credit expansion: Evidence from the 2007 mortgage
default crisis, Working paper.
PenningtonCross, Anthony, 2003, Credit history and the performance of prime and nonprime mortgages, Journal
of Real Estate Finance and Economics 27, 279–301.
PenningtonCross, Anthony, and Souphala Chomsisengphet, 2007, Subprime reﬁnancing: Equity extraction and
mortgage termination, Real Estate Economics 35, 233–263.
Ruckes, Martin, 2004, Bank competition and credit standards, review of ﬁnancial studies, Review of Financial
Studies 17, 1073–1102.
Shiller, Robert J., 2007, Understanding recent trends in house prices and home ownership, NBER working paper
13553.
Van Hemert, Otto, 2007, Household interest rate risk management, Working paper.
Vickery, James, 2007, Interest rates and consumer choice in the residential mortgage market, Working paper.
Von Furstenberg, George M., and R. Jeﬀery Green, 1974, Home mortgage delinquencies: A cohort analysis, Journal
of Finance 29, 1545–1548.
36
A Foreclosure Rates
In this Appendix we show the continual deterioration of adjusted loan performance using foreclosure, instead of delinquency,
as a measure of loan performance. Foreclosure is deﬁned as a loan being in foreclosure, realestate owned, or in default. In
Figure 8 we present actual (left panel) and adjusted (right panel) foreclosure rates. The actual foreclosure rate for loans age
six months and younger is close to zero, in contrast to the actual delinquency rate at this age (presented in Figure 1 (left
panel)). For older aged loans the actual foreclosure rate is roughly speaking twice as low as the actual delinquency rate.
Similar to the actual delinquency rates (Figure 1 (left panel), the actual foreclosure rates (Figure 8, left panel) are highest
for 2007, 2006, and 2001 and lowest for 2003 and 2004.
Using foreclosure instead of delinquency as a measure for nonperformance, the adjusted foreclosure rates every vintage
year starting in2002, as can be seen from Figure 8, right panel. The adjusted foreclosure rates of vintage 2001 loans are
between vintages 2002 and 2003. The change for each year is statistically signiﬁcant at the 1% conﬁdence level.
Figure 8: Actual and Adjusted Foreclosure Rates
The ﬁgure shows the age pattern in the actual (left panel) and adjusted (right panel) foreclosure rate for the diﬀerent vintage years. The
foreclosure rate is deﬁned as the cumulative fraction of loans that were in foreclosure, realestate owned, or defaulted, at or before a given
age. The adjusted foreclosure rate is obtained by adjusting the actual rate for yearbyyear variation in FICO scores, loantovalue ratios,
debttoincome ratios, missing debttoincome ratio dummies, cashout reﬁnancing dummies, owneroccupation dummies, documentation
levels, percentage of loans with prepayment penalties, mortgage rates, margins, composition of mortgage contract types, origination
amounts, MSA house price appreciation since origination, change in state unemployment rate since origination, and neighborhood
median income.
0
2
4
6
8
10
12
14
16
18
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Actual Foreclosure Rate (%)
0
2
4
6
8
10
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Adjusted Foreclosure Rate (%)
B Adjusted Delinquency Rate for Hybrids and FRMs Separately
In this Appendix we show that the continual deterioration of adjusted loan performance over the 2001–2007 period also
occurs when estimating a separate proportional odds model for the main contract types, as opposed to the baseline case in
the main text where we perform a single estimation for all loans together, but include contract type dummies in the regression
37
speciﬁcation. Figure 9 shows the adjusted delinquency rate for the two main contract types: 2/28 hybrids and FRMs. For
both contract types, the adjusted delinquency rates have increased monotonically over time. Except for a level diﬀerence,
the age pattern for the diﬀerent vintage years looks very much the same for the two contract types.
Figure 9: Adjusted Delinquency Rates for Hybrids and FRMs Separately
The ﬁgure shows the adjusted delinquency rates based on hybrid mortgages (left panel) and FRMs (right panel) separately. The
delinquency rate is deﬁned as the cumulative fraction of loans that were past due 60 or more days, in foreclosure, realestate owned, or
defaulted, at or before a given age. The adjusted delinquency rate is obtained by adjusting the actual rate for yearbyyear variation in
FICO scores, loantovalue ratios, debttoincome ratios, missing debttoincome ratio dummies, cashout reﬁnancing dummies, owner
occupation dummies, documentation levels, percentage of loans with prepayment penalties, mortgage rates, margins, composition of
mortgage contract types, origination amounts, MSA house price appreciation since origination, change in state unemployment rate since
origination, and neighborhood median income.
0
5
10
15
20
25
30
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Hybrid Mortgage Loans
0
5
10
15
20
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Loan Age (Months)
2007
2006
2005
2004
2003
2002
2001
Fixed−Rate Mortgage Loans
C NonLinearity in the Sensitivity of the Mortgage Rate to the LTV
In Figure 2 we plotted the sensitivity of the ﬁxedrate and 2/28 hybrid mortgage rates to the ﬁrstlien LTV ratio. The
sensitivity is deﬁned as the regression coeﬃcient on the ﬁrstlien LTV (scaled by the standard deviation) in a regression
with the mortgage rate as dependent variable and the ﬁrstlien LTV, the secondlien LTV, and the other loan and borrower
characteristics listed in Section 6, as independent variables.
In this appendix we study the robustness of this result to adding the square of the ﬁrstlien LTV and the square of the
secondlien LTV as independent variables, therefore allowing for a nonlinear functional form. In Figure 10 we report the
resulting scaled marginal eﬀect of the ﬁrstlien LTV for ﬁxedrate and 2/28 hybrid mortgages evaluated at a ﬁrstlien LTV
of 80 percent (left panel) and 90 percent (right panel). Without nonlinear terms the marginal eﬀect is simply given by
the regression coeﬃcient. This is what we plotted in Figure 2. With the quadratic terms, the marginal eﬀect is given by
β
LTV
+ 2β
LTV
2X, where the βs are the regression coeﬃcients and X is the ﬁrstlien LTV ratio at which the marginal eﬀect
is evaluated.
38
Figure 10: Sensitivity of Mortgage Rate to FirstLien LTV Ratio Allowing for NonLinearity
The ﬁgure shows the scaled marginal eﬀect of the ﬁrstlien loantovalue (LTV) ratio on the mortgage rate for ﬁrstlien ﬁxedrate and
2/28 hybrid mortgages, evaluated at a ﬁrstlien LTV of 80 percent (left panel) and 90 percent (right panel). The eﬀect is determined using
an OLS regression with the interest rate as dependent variable and the FICO score, ﬁrstlien LTV (and the square), secondlien LTV
(and the square), debttoincome ratio, missing debttoincome ratio dummy, cashout reﬁnancing dummy, owneroccupation dummy,
prepayment penalty dummy, origination amount, term of the mortgage, prepayment term, and margin as independent variables.
0
.2
.4
.6
.8
2001 2002 2003 2004 2005 2006 2007
Year
FRM
2/28 Hybrid
Scaled Marginal Effect of First−Lien LTV = 80% (%)
0
.2
.4
.6
.8
1
1.2
2001 2002 2003 2004 2005 2006 2007
Year
FRM
2/28 Hybrid
Scaled Marginal Effect of First−Lien LTV = 90% (%)
As shown in Figure 10, the marginal eﬀect is rising over time, consistent with the baseline case results presented in Figure
2. Moreover, we ﬁnd that there is a statistically and economically signiﬁcant nonlinear eﬀect of the ﬁrstlien LTV on the
mortgage rate. Comparing the left and right panels in Figure 10, the higher the ﬁrstlien LTV ratio, the more sensitive is the
mortgage rate to changes in the ﬁrstlien LTV. The largest diﬀerence between the results based on speciﬁcations with and
without nonlinearity is observed for 2/28 hybrid mortgages in 2007 at a ﬁrstlien LTV of 90 percent (right panel). The scaled
marginal eﬀect increases by 27 basis points over the course of 3 months in 2007 when the model allows for nonlinearity. In
contrast, in the case without nonlinearity, as in Figure 2, the increase in the scaled marginal eﬀect is only 13 basis points.
39
1
Introduction
The subprime mortgage crisis of 2007 was characterized by an unusually large fraction of subprime mortgages originated in 2006 and 2007 becoming delinquent or in foreclosure only months later. The crisis spurred massive media attention; many diﬀerent explanations of the crisis have been proﬀered. The goal of this paper is to answer the question: “What do the data tell us about the possible causes of the crisis?” To this end we use a loanlevel database containing information on about half of all U.S. subprime mortgages originated between 2001 and 2007. The relatively poor performance of vintage 2006 and 2007 loans is illustrated in Figure 1 (left panel). At every mortgage loan age, loans originated in 2006 and 2007 show a much higher delinquency rate than loans originated in earlier years at the same ages. Figure 1: Actual and Adjusted Delinquency Rate
The ﬁgure shows the age pattern in the actual (left panel) and adjusted (right panel) delinquency rate for the diﬀerent vintage years. The delinquency rate is deﬁned as the cumulative fraction of loans that were past due 60 or more days, in foreclosure, realestate owned, or defaulted, at or before a given age. The adjusted delinquency rate is obtained by adjusting the actual rate for yearbyyear variation in FICO scores, loantovalue ratios, debttoincome ratios, missing debttoincome ratio dummies, cashout reﬁnancing dummies, owneroccupation dummies, documentation levels, percentage of loans with prepayment penalties, mortgage rates, margins, composition of mortgage contract types, origination amounts, MSA house price appreciation since origination, change in state unemployment rate since origination, and neighborhood median income.
Actual Delinquency Rate (%) 35 30 25 20 15 10 5 0 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) 5 0 3 4 5
2007 2006 2005 2004 2003 2002 2001
Adjusted Delinquency Rate (%) 25 20 15 10
2007 2006 2005 2004 2003 2002 2001
6
7
8 9 10 11 12 13 14 15 16 17 Loan Age (Months)
We document that the poor performance of the vintage 2006 and 2007 loans was not conﬁned to a particular segment of the subprime mortgage market. For example, ﬁxedrate, hybrid, purchasemoney, cashout reﬁnancing, lowdocumentation, and fulldocumentation loans originated in 2006 and 2007 all
1
Electronic copy available at: http://ssrn.com/abstract=1020396
showed substantially higher delinquency rates than loans made the prior ﬁve years. This contradicts a widely held belief that the subprime mortgage crisis was mostly conﬁned to hybrid or lowdocumentation mortgages. We explore to what extent the subprime mortgage crisis can be attributed to diﬀerent loan characteristics, borrower characteristics, macroeconomic conditions, and vintage (origination) year eﬀects. The most important macroeconomic factor is subsequent house price appreciation, measured as the MSAlevel house price change between the time of origination and the time of loan performance evaluation. For the empirical analysis, we run a proportional odds duration model with the probability of (ﬁrsttime) delinquency a function of these factors and loan age. We ﬁnd that loan and borrower characteristics are very important in terms of explaining the crosssection of loan performance. However, because these characteristics were not suﬃciently diﬀerent in 2006 and 2007 compared with the prior ﬁve years, they cannot explain the unusually weak performance of vintage 2006 and 2007 loans. For example, a onestandarddeviation increase in the debttoincome ratio raises the likelihood (the odds ratio) of a current loan turning delinquent in a given month by as much as a factor of 1.14. However, because the average debttoincome ratio was just 0.2 standard deviations higher in 2006 than its level in previous years, it contributes very little to the inferior performance of vintage 2006 loans. The only variable in the considered proportional odds model that contributed substantially to the crisis is the low subsequent house price appreciation for vintage 2006 and 2007 loans, which can explain about a factor of 1.24 and 1.39, respectively, higherthanaverage likelihood for a current loan to turn delinquent.1 Due to geographical heterogeneity in house price changes, some areas have experienced largerthanaverage house price declines and therefore have a larger explained increase in delinquency and foreclosure rates.2 The coeﬃcients of the vintage dummy variables, included as covariates in the proportional odds model, measure the quality of loans, adjusted for diﬀerences in observed loan characteristics, borrower characteristics, and macroeconomic circumstances. In Figure 1 (right panel) we plot the adjusted delinquency rates, which are obtained by using the estimated coeﬃcients for the vintage dummies and imposing the requirement that the average actual and average adjusted delinquency rates are equal for any given age. As shown in Figure 1 (right panel), the adjusted delinquency rates have been steadily rising for the
Other papers that research the relationship between house prices and mortgage ﬁnancing include Genesove and Mayer (1997), Genesove and Mayer (2001), and Brunnermeier and Julliard (2007). 2 Also, house price appreciation may diﬀer in cities versus rural areas. See for example Glaeser and Gyourko (2005) and Gyourko and Sinai (2006).
1
2
Electronic copy available at: http://ssrn.com/abstract=1020396
past seven years. In other words, loan quality—adjusted for observed characteristics and macroeconomic circumstances—deteriorated monotonically between 2001 and 2007. Interestingly, 2001 was among the worst vintage years in terms of actual delinquency rates, but is in fact the best vintage year in terms of the adjusted rates. High interest rates, low average FICO credit scores, and low house price appreciation created the “perfect storm” in 2001, resulting in a high actual delinquency rate; after adjusting for these unfavorable circumstances, however, the adjusted delinquency rates are low. In addition to the monotonic deterioration of loan quality, we show that over time the average combined loantovalue ratio increased, the fraction of low documentation loans increased, and the subprimeprime rate spread decreased. The rapid rise and subsequent fall of the subprime mortgage market is therefore reminiscent of a classic lending boombust scenario.3 The origin of the subprime lending boom has often been attributed to the increased demand for socalled privatelabel mortgagebacked securities (MBSs) by both domestic and foreign investors. Our database does not allow us to directly test this hypothesis, but an increase in demand for subprime MBSs is consistent with our ﬁnding of lower spreads and higher volume. Mian and Suﬁ (2008) ﬁnd evidence consistent with this view that increased demand for MBSs spurred the lending boom. The proportional odds model used to estimate the adjusted delinquency rates assumes that the covariate coeﬃcients are constant over time. We test the validity of this assumption for all variables and ﬁnd that it is the most strongly rejected for the loantovalue (LTV) ratio. HighLTV borrowers in 2006 and 2007 were riskier than those in 2001 in terms of the probability of delinquency, for given values of the other explanatory variables. Were securitizers aware of the increasing riskiness of highLTV borrowers?4 To answer this question, we analyze the relationship between the mortgage rate and LTV ratio (along with the other loan and borrower characteristics). We perform a crosssectional ordinary least squares (OLS) regression, with the mortgage rate as the dependent variable, for each quarter from 2001Q1 to 2007Q2 for both ﬁxedrate mortgages and 2/28 hybrid mortgages. Figure 2 shows that the coeﬃcient on the ﬁrstlien LTV variable, scaled by the standard deviation of the ﬁrstlien LTV ratio, has been increasing over time. We thus ﬁnd evidence that securitizers were aware of the increasing riskiness of highLTV borrowers, and
Berger and Udell (2004) discuss the empirical stylized fact that during a monetary expansion lending volume typically increases and underwriting standards loosen. Loan performance is the worst for those loans underwritten toward the end of the cycle. Demirg¨cKunt and Detragiache (2002) and Gourinchas, Valdes, and Landerretche (2001) ﬁnd that lending u¸ booms raise the probability of a banking crisis. Dell’Ariccia and Marquez (2006) show in a theoretical model that a change in information asymmetry across banks might cause a lending boom that features lower standards and lower proﬁts. Ruckes (2004) shows that low screening activity may lead to intense price competition and lower standards. 4 For loans that are securitized (as are all loans in our database), the securitizer eﬀectively dictates the mortgage rate charged by the originator.
3
3
origination amount. focusing on foreclosures rather than delinquencies.784 observations. Scaled Regression Coefficient (%) . In another extension.and moderateincome areas. This means that the seeds for the crisis were sown long before 2007. with a minimum of 18.adjusted mortgage rates accordingly. cashout reﬁnancing dummy. The eﬀect is measured as the regression coeﬃcient on the ﬁrstlien loantovalue ratio (scaled by the standard deviation) in an ordinary least squares regression with the mortgage rate as the dependent variable and the FICO score.4 . secondlien loantovalue ratio. deﬁned as areas with median income below 80 percent of the larger Metropolitan Statistical Area median income.3 .2 . term of the mortgage. such as the combination of a high LTV ratio and a low FICO score.1 0 2001 FRM 2/28 Hybrid 2002 2003 2004 Year 2005 2006 2007 We show that our main results are robust to analyzing mortgage contract types separately. owneroccupation dummy. but detecting them was complicated by high house price appreciation between 2003 and 2005—appreciation that masked the true riskiness of subprime mortgages. and margin (only applicable to 2/28 hybrid) as independent variables. like allowing for interaction eﬀects between diﬀerent loan and borrower characteristics. we estimate our proportional odds model using data just through yearend 2005 and again obtain the continual deterioration of loan quality from 2001 onward. which seek to stimulate loan 4 . Figure 2: Sensitivity of Mortgage Rate to FirstLien LoantoValue Ratio The ﬁgure shows the eﬀect of the ﬁrstlien loantovalue ratio on the mortgage rate for ﬁrstlien ﬁxedrate and 2/28 hybrid mortgages. debttoincome ratio. prepayment term. and specifying the empirical model in numerous diﬀerent ways. missing debttoincome ratio dummy. ﬁrstlien loantovalue ratio. The latter includes taking into account risklayering—the origination of loans that are risky in several dimensions. we ﬁnd an increased probability of delinquency for loans originated in low. This points toward a negative byproduct of the 1977 Community Reinvestment Act and Government Sponsored Enterprises housing goals. prepayment penalty dummy. Each point corresponds to a separate regression.5 . As an extension.
DellAriccia. and macroeconomic factors. dating back to at least Von Furstenberg and Green (1974). we uncover a downward trend in loan quality. First. and Willen (2008). borrower. securitizers were. determined as loan performance adjusted for diﬀerences in loan and borrower characteristics and macroeconomic circumstances. In Section 7 we analyze the subprimeprime rate spread and in Section 8 we conclude. Recent contributions include Cutts and Van Order (2005) and PenningtonCross and Chomsisengphet (2007). There is a large literature on the determinants of mortgage delinquencies and foreclosures. Shapiro. and Laeven (2008). In Section 2 we show the descriptive statistics for the subprime mortgages in our database. as evidenced by changing determinants of mortgage rates. we detect an increased likelihood of delinquency in low. Campbell and Cocco (2003) and Van Hemert (2007) discuss mortgage choice over the life cycle.and moderateincome areas. 5 Deng. which led up to the 2007 subprime mortgage crisis. Fourth. The structure of this paper is as follows. Quigley. Second. Our paper makes several novel contributions. Mukherjee.and middleincome areas—might have created a negative byproduct.origination in low. after controlling for diﬀerences in neighborhood incomes and other loan. 5 . Together these results provide evidence that the rise and fall of the subprime mortgage market follows a classic lending boombust scenario. in which unsustainable growth leads to the collapse of the market. This empirical ﬁnding seems to suggest that the housing goals of the Community Reinvestment Act and/or Government Sponsored Enterprises—those intended to increase lending in low. we quantify how much diﬀerent determinants have contributed to the observed high delinquency rates for vintage 2006 and 2007 loans.and middleincome areas. In Section 6 we demonstrate the increasing riskiness of highLTV borrowers. In Section 4 we present the baselinecase results and in Section 5 we discuss extensions and robustness checks. We further show that there was a deterioration of lending standards and a decrease in the subprimeprime mortgage rate spread during the 2001–2007 period. and Vig (2008). Fifth. In Section 3 we discuss the empirical strategy we employ. Third. we show that the continual deterioration of loan quality could have been detected long before the crisis by means of a simple statistical exercise. and the extent to which securitizers were aware of this risk. aware of this deterioration over time.5 Other papers analyzing the subprime crisis include Gerardi. that is associated with higher loan delinquencies. and Van Order (2000) discuss the simultaneity of the mortgage prepayment and default option. and Keys. Igan. We provide several additional robustness checks in the appendices. Our data enables us to show that the eﬀect of diﬀerent loanlevel characteristics as well as low house price appreciation was quantitatively too small to explain the poor performance of 2006 and 2007 vintage loans. to some extent. Seru. Mian and Suﬁ (2008).
(more than half of the U.fdic. we discuss the delinquency rates of these loans for various segments of the subprime mortgage market. and the Oﬃce of Thrift Supervision use this deﬁnition. 61. which includes loanlevel data on about 85 percent of all securitized subprime mortgages. and 75 percent respectively.9 We ﬁrst outline the main characteristics of the loans in our database at origination. See e. or a 2/28 hybrid). In the second block of Table 1. specialization in highcost loans). The term subprime can be used to describe certain characteristics of the borrower (e. 63. While it is not clear to us whether the pre. subprime loans are those underlying subprime securities. 76. The Oﬃce of the Controller of the Currency. however.7 lender (e. The total dollar amount originated in 2001 was $57 billion. 7 The Board of Governors of the Federal Reserve System. 9 Since the ﬁrst version of this paper in October 2007.. some subprime lenders making prime loans and some prime lenders originating subprime loans. a FICO credit score less than 620). as of June 2008.2 Descriptive Analysis In this paper we use the First American CoreLogic LoanPerformance (henceforth: LoanPerformance) database.g. high projected default rate for the pool of underlying loans). while in 2006 it was $375 billion. in the wake of the subprime mortgage crisis.. In this paper.g. or mortgage contract type (e. We focus on ﬁrstlien loans and consider the 2001 through 2008 sample period.gov/news/news/press/2001/pr0901a.g. LoanPerformance has responded to the request by trustees’ clients to reclassify some of its subprime loans to AltA status.8 security of which the loan can become a part (e. Second.or postreclassiﬁcation subprime data are the most appropriate for research purposes.html 8 The U. We do not include less risky AltA mortgage loans in our analysis.S. The common element across deﬁnitions of a subprime loan is a high default risk. no money down and no documentation provided.6 There is no consensus on the exact deﬁnition of a subprime mortgage loan. In 2007. 76. Department of Housing and Urban Development uses HMDA data and interviews lenders to identify subprime lenders among them. the dollar amount originated fell sharply to $69 billion..g. the Federal Deposit Insurance Corporation.S.1 Loan Characteristics at Origination Table 1 provides the descriptive statistics for the subprime mortgage loans in our database that were originated between 2001 and 2007. In this version we focus on the postclassiﬁcation data. we checked that our results are robust to the reclassiﬁcation. subprime mortgage market). 6 Mortgage Market Statistical Annual (2007) reports securitization shares of subprime mortgages each year from 2001 to 2006 equal to 54. http://www. 6 .. we split the pool of mortgages into four main mortgage contract types.g. 2. In the ﬁrst block of Table 1 we see that the annual number of originated loans increased by a factor of four between 2001 and 2006 and the average loan size almost doubled over those ﬁve years. and was primarily originated in the ﬁrst half of 2007. There are.
2 42.2 41.3 180 23.1 19.1 70.9 6.9 212 19.7 6.1 200 18.5 29.7 6.3 620.2 8.1 82.3 8.4 6.6 39.7 618.5 37.0 34.5 8.2 43.3 0.4 6.2 Mortgage Type 27.6 0.0 0.3 63.4 72.6 6.3 83.7 8.8 74.9 41.3 66.4 73.4 12.4 59.5 7.8 41.8 30. 2001 2002 2003 2004 2005 2006 2007 Size Number of Loans (*1000) Average Loan Size (*$1000) FRM (%) ARM (%) Hybrid (%) Balloon (%) Purchase (%) Reﬁnancing (cash out) (%) Reﬁnancing (no cash out) (%) FICO Score Combined LoantoValue Ratio (%) DebttoIncome Ratio (%) Missing DebttoIncome Ratio Dummy (%) Investor Dummy (%) Documentation Dummy (%) Prepayment Penalty Dummy (%) Mortgage Rate (%) Margin for ARM and Hybrid Mortgage Loans (%) 452 126 33.2 2.5 8.9 0.2 0.9 8.5 5.0 7. 274 1.1 57.3 65.2 737 1.4 80.3 71. 258 1.6 59.8 0.3 57.2 85.3 75.4 51.9 79.7 70.8 618.3 6.7 58.0 7 .7 11.0 8.9 40.6 0.1 38.4 68.3 52.4 30.9 84.2 35.2 31.4 6.9 164 33.5 29.1 67.2 608.9 29.5 0.5 29.Table 1: Loan Characteristics at Origination for Diﬀerent Vintages Descriptive statistics for the ﬁrstlien subprime loans in the LoanPerformance database.6 618.0 38. 911 2.5 75.8 56.0 11.9 9.1 613.1 82.2 66.2 8.2 76.7 8.2 62.5 7.4 316 220 Loan Purpose Variable Means 601.4 11.1 7.8 28.4 26.4 76. 772 145 29.4 75.4 54.5 25.8 0.4 38.8 4.3 8.7 6.
Less than 1 percent of the mortgages originated over the sample period were adjustablerate (nonhybrid) mortgages. 10 8 . consistent with movements in both the 1year and 10year Treasury yields over the same period.Most numerous are the hybrid mortgages. For about a third of the loans in our database. the popularity of FRMs rose to 28 percent. In contrast. The proportion of balloon mortgage contracts jumped substantially in 2006. The share of loans with full documentation fell considerably over the sample period. accounting for more than half of all subprime loans in our data set originated between 2001 and 2007. The ﬁxedrate mortgage contract became less popular in the subprime market over time and accounted for just 20 percent of the total number of loans in 2006. by reﬁnancing an existing mortgage loan into a larger new mortgage loan. The average FICO credit score rose 20 points between 2001 and 2005. singlefamily. The combined loantovalue (CLTV) ratio. In the third block of Table 1. For example Koijen. purchasemoney loans reported by the Federal Housing Financing Board in its Monthly Interest Rate Survey that are of the ﬁxedrate type ﬂuctuated between 60 and 90 percent from 2001 to 2006. and accounted for 25 percent of the total number of mortgages originated that year. as the subprime mortgage crisis hit. and Van Nieuwerburgh (2007) show that the fraction of conventional. the purpose was to ﬁnance the purchase of a house. A hybrid mortgage carries a ﬁxed rate for an initial period (typically 2 or 3 years) and then the rate resets to a reference rate (often the 6month LIBOR) plus a margin. Finally. we report the purpose of the mortgage loans. Approximately 55 percent of our subprime mortgage loans were originated to extract cash. the margin (over a reference rate) for adjustablerate and hybrid mortgages stayed rather constant over time. in the prime mortgage market. most mortgage loans were of the ﬁxedrate type during this period. A balloon mortgage does not fully amortize over the term of the loan and therefore requires a large ﬁnal (balloon) payment. In about 30 to 40 percent of cases. fully amortizing. The (backend) debttoincome ratio (if provided) and the fraction of loans with a prepayment penalty were fairly constant. In the ﬁnal block of Table 1. slightly increased over 2001–2006. this is captured by the missing debttoincome ratio dummy variable. Van Hemert. we report the mean values for the loan and borrower characteristics that we will use in the statistical analysis (see Table 2 for a deﬁnition of these variables). which measures the value of alllien loans divided by the value of the house.10 In 2007. The share of loans originated in order to reﬁnance with no cash extraction was relatively small. from 77 percent in 2001 to 67 percent in 2007. The mean mortgage rate fell from 2001 to 2004 and rebounded after that. primarily because of the increased popularity of secondlien and thirdlien loans. Vickery (2007) shows that empirical mortgage choice is aﬀected by the eligibility of the mortgage loan to be purchased by Fannie Mae and Freddie Mac. no debttoincome ratio was provided (the reported value in those cases is zero).
we again plot the age pattern in the delinquency rate for vintages 2001 through 2007 and split the subprime mortgage market into various segments. the poor performance of the 2006 and 2007 vintages is not conﬁned to a particular segment of the subprime market. As the ﬁgure shows. lowdocumentation. Delinquency rates in this 9 . and fulldocumentation mortgage loans. cashout reﬁnancing. In general. such as whether a mortgage broker intermediated. or in default.We do not report summary statistics on the loan source. We denote the ratio of the number of vintage k loans experiencing a ﬁrsttime delinquency at age s over the number of vintage k loans with no ﬁrsttime ˜k delinquency for age < s by Ps . purchasemoney. real estate owned. where ˜k Ps (2) (3) i=2001 In Figure 1 (left panel) we show that for the subprime mortgage market as a whole. In Figure 3. the 2006 and 2007 vintages show the highest delinquency rate pattern. or the loan is reported as in foreclosure. vintage 2006 and 2007 loans stand out in terms of high delinquency rates. In Figure 4 we plot the delinquency rates of all outstanding mortgages. vintage 2001 loans come next in terms of high delinquency rates. In the six panels of Figure 3 we see that for hybrid. but rather reﬂects a (subprime) marketwide phenomenon. The delinquency rates for the ﬁxedrate mortgages (FRMs) are lower than those for hybrid mortgages but exhibit a remarkably similar pattern across vintage years. 2. We compute the actual (cumulative) delinquency rate for vintage k at age t as the fraction of loans experiencing a delinquency at or before age t t k Actualt = 1 − s=1 ˜k 1 − Ps (1) We deﬁne the average actual delinquency rate as t Actualt = 1 − ¯ Ps = 1 7 s=1 2007 ¯ 1 − Ps .2 Performance of Loans by Market Segments We deﬁne a loan to be delinquent if payments on the loan are 60 or more days late. as the broad classiﬁcation used in the database rendered this variable less informative. Notice that the scale of the vertical axis diﬀers across the panels. ﬁxedrate. and vintage 2003 loans have the lowest delinquency rates. Notice that the fraction of FRMs that are delinquent remained fairly constant from 2005Q1 to 2007Q2.
in foreclosure. The delinquency rate is deﬁned as the cumulative fraction of loans that were past due 60 or more days. at or before a given age. Hybrid Mortgage Loans 35 30 25 20 15 10 5 0 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) 5 0 3 4 5 6 2007 2006 2005 2004 2003 2002 2001 Fixed−Rate Mortgage Loans 25 20 15 10 2007 2006 2005 2004 2003 2002 2001 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) Purchase−Money Mortgage Loans 40 35 30 25 20 15 10 5 0 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) 5 0 3 4 2007 2006 2005 2004 2003 2002 2001 Cash−Out Refinancing Mortgage Loans 30 25 20 15 10 2007 2006 2005 2004 2003 2002 2001 5 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) Full−Documentation Mortgage Loans 30 25 20 15 10 5 0 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) 2007 2006 2005 2004 2003 2002 2001 Low−or−No−Doc Mortgage Loans 40 35 30 25 20 15 10 5 0 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) 2007 2006 2005 2004 2003 2002 2001 10 . or defaulted.Figure 3: Actual Delinquency Rate for Segments of the Subprime Mortgage Market The ﬁgure shows the age pattern in the delinquency rate for diﬀerent segments. realestate owned.
caused by a decrease in the popularity of FRMs from 2001 to 2006 (see Table 1). who said “For subprime mortgages with ﬁxed rather than variable rates. FRMs originated in 2006 in fact performed unusually poorly (Figure 3. Figure 4: Actual Delinquency Rates of Outstanding Mortgages The Figure shows the actual delinquency rates of all outstanding FRMs and hybrids from January 2000 through June 2008. serious delinquencies have been fairly stable. though. for example. the weaker performance of vintage 2006 loans is masked by the aging of the overall FRM pool. to realize that this result is driven by an aging eﬀect of the FRM pool. Termination occurs either through a prepayment or a default. upperright panel). we include delinquency—the earliest stage of payment problems—in our nonperformance measure.ﬁgure are deﬁned as the fraction of loans delinquent at any given time. Given this focus on young loans. In other words. but if one plots the delinquency rate of outstanding FRMs over time (Figure 4. Our paper is related to the vast literature on empirical mortgage termination. for which we already have data for the vintages of particular interest: 2006 and 2007. Delinquency is an intermediate stage for a loan in trouble. not cumulative. These rates are consistent with those used in an August 2007 speech by the Chairman of the Federal Reserve System (Bernanke (2007)). Actual Delinquency Rate (%) 40 35 30 25 20 15 10 5 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 Year FRM Hybrid 3 Statistical Model Speciﬁcation The focus of our paper is on the performance of subprime mortgage loans in the ﬁrst 17 months after origination. An analysis of mortgage termination lends itself naturally to 11 .” It is important. the loan may eventually cure or terminate with a prepayment or default. left panel).
t log = αt + β xi.duration (i. and An (2006). The name “proportional odds” arises from the fact that the vector of coeﬃcients. we deﬁne the probability that at age t loan i with covariate values xi.2 Estimation The proportional odds model. Important contributions to this literature include Deng (1997). does not have an age subscript and thus the log odds are proportional to the covariate values at any age. ran a standard logit regression. survival) models.t (5) where αt is an agedependent constant and β is a vector of coeﬃcients. PenningtonCross (2003). and found qualitatively similar results for the eﬀect of explanatory variables and the eﬀect of vintage year dummies (unreported results).t 1 − Pi. is typically estimated for the full panel at once using either partial likelihood (see Cox (1972)) or maximum likelihood methodologies. Denoting this time by T . We apply the duration model methodology to the intermediate status of delinquency by deﬁning nonsurvival as “having ever been 60 days delinquent or worse.. with prepayment and default as competing reasons for termination. The small sample properties for 12 . Ambrose and Capone (2000). Calhoun and Deng (2002). β.t becomes delinquent for the ﬁrst time. Deng.” which includes formerly delinquent loans that are prepaid or cured. we use a proportional odds model.t } (4) Because the monthly choice whether to make a mortgage payment is discrete. the discretetime analogue to the popular proportional hazard model: Pi. Transition from survival to nonsurvival will occur when a loan becomes 60 or more days delinquent or defaults for the ﬁrst time. 3.e. conditional on not having been delinquent before. 3. Deng. Pavlov. and PenningtonCross and Chomsisengphet (2007). and Van Order (2000). and Yang (2005). Deng. Equation 5. As a robustness check.1 Empirical Model Speciﬁcation We are interested in the number of months (duration) until a loan becomes at least 60 days delinquent or defaults for the ﬁrst time. we used “being currently 60 or more days delinquent or in default” as the nonperformance measure.t = Pr {T = tT ≥ t. Clapp. as Pi. xi. Quigley.
but for a sample size of 10. including the vintage 2007 dummy. For larger sample sizes the computational burden of the partial likelihood function quickly becomes unmanageable. they are left censored. αt . because we know for sure that they will never experience a ﬁrsttime delinquency. if the securitizer goes out of business we stop observing their loans and therefore they are right censored. This is a result of heavily tied data in our discrete time setup.3 Reported Output The AgeEﬀect statistic is deﬁned as the proportional odds ratio for ﬁrsttime delinquency at a particular age t for the average (over the full sample) vector of covariate values at age t. see page 126 of Allison (2007). xt : ¯ AgeEﬀectt = exp (αt + β xt ) ¯ (6) The M arginal statistic is deﬁned as the log proportional odds ratio associated with a one standard Loans securitized several months after origination are not observed in our data between the origination date and the securitization date. 000 loans the methods already provide very similar estimates for β. which has the added advantage that it provides estimates of the loan age eﬀect.these two estimation methods are potentially diﬀerent. the maximum likelihood estimation requires each covariate. In addition. 12 For more information on this. we classify prepaid loans as nondelinquent and noncensored. not due to a prepayment or default). We use the PROC LOGISTIC procedure in SAS for the maximum likelihood estimation. This method is able to handle both left censoring (loans entering the sample at a later age) and right censoring (loans leaving the sample prematurely. We use a random sample of 1 million loans for this exercise. where the term “tied” refers to loans experiencing ﬁrsttime delinquency at the exact same age. Because we include vintage year dummies as covariates we have to restrict the maximum age considered in our analysis to 17 months. using the noninformative censoring assumption. the latest age for which we have an observation for the 2007 origination year.12 For the adjusted delinquency rate plots viewed at the end of 2005 and 2006 (Figure 5). have some dispersion in the covariate values for each age. We therefore estimate the proportional odds model using maximum likelihood.11 In order to generate unbiased delinquency rate plots (as in Figure 1. 3. we restrict the analysis to a maximum age of 11 months. right panel). 11 13 . therefore.
x01j . for vintage 2001 and variable j it is the diﬀerence between the mean value for variable j in 2001. σj : M arginalj = βj σj = log exp (αt + β xi.t + βj x01j − xj ¯ exp (αt + β xi. This property will be used for reporting the total contribution of all covariates in Table 3. the combined contribution of two variables is simply the sum of the individual contributions.t + βj σj ) exp (αt + β xi. expressed in the number of standard deviations of the variable. The probability of experiencing a ﬁrsttime delinquency at or before age = t is given by t Pr {T ≤ txt . . we evaluate the above expression for the value of 14 . and the mean value over all vintages. xj .t ) ¯ Contributionj = βj x01j − xj = log (9) = M arginalj ∗ Deviationj As a straightforward generalization of Equation 9. xt−1 .} = 1 − s=1 (1 − Ps ) (10) To visualize the magnitude of the vintage year eﬀect. expressed ¯ in the number of standard deviations: x01j − xj ¯ σj Deviationj = (8) The Contribution statistic measures the deviation of the (average) log proportional odds of ﬁrsttime delinquency in a particular vintage year from the (average) log proportional odds of ﬁrsttime delinquency over the entire sample that can be explained by a particular variable.t ) (7) The advantage of taking the log in Equation 7 is that the eﬀect of an increase of σj in covariate j is minus the eﬀect of a decrease of σj ..deviation increase in variable j. The Deviation statistic measures the diﬀerence between the mean value of a variable in a particular vintage year and the mean value of that variable measured over the entire sample. For example. For example for vintage 2001 and variable j we have: exp αt + β xi. and thus the absolute eﬀect is invariant to the chosen direction of change.
The borrower and loan characteristics we use in the analysis are: the FICO credit score. a dummy variable indicating whether there is a prepayment penalty on a loan. originated between 2001 and 2007. All results in this section are based on a random sample of one million ﬁrstlien subprime mortgage loans. 4. a dummy variable indicating whether full documentation was provided. Dk = D. the (initial) mortgage rate. illustrated in Equation 5. 15 . Combining Equations 10 and Equation 11 we obtain the adjusted delinquency rate for vintage year k at age t t Adjustedk = 1 − t s=1 1 1+ ¯ Ps ¯ 1−Ps ¯ exp Dk − D (12) ¯ Notice that for an average vintage year. a dummy variable indicating whether the loan was a cashout reﬁnancing. Equation 12 simpliﬁes to t Adjustedt = 1 − s=1 ¯ 1 − Ps = Actualt (13) 4 Empirical Results for the BaselineCase Speciﬁcation In this section we investigate to what extent the proportional odds model model can explain the high levels of delinquencies for the vintage 2006 and 2007 mortgage loans in our database.1 Variable Deﬁnitions Table 2 provides the deﬁnitions of the variables (covariates) included in the baselinecase speciﬁcation of the proportional odds model. including vintage year dummies. a dummy variable indicating whether the borrower was an investor (as opposed to an owneroccupier). a dummy variable indicating whether the debttoincome ratio was missing (reported as zero).Ps that satisﬁes log Ps 1 − Ps = log ¯ Ps ¯ 1 − Ps ¯ + Dk − D (11) ¯ where Dk is the estimated coeﬃcient for for the vintage year k dummy variable and D is the average estimated vintage dummy variable. the value of the debttoincome ratio (when provided). the combined loantovalue ratio. This expression uses the proportional odds property for explanatory variables.
at origination.Table 2: BaselineCase Variable Deﬁnitions This table presents deﬁnitions of the baselinecase variables (covariates) used in the proportional odds duration model. applicable after the ﬁrst interest rate reset. Equals one if the mortgage loan is a cashout reﬁnancing loan. ARMs. MSAlevel house price appreciation from the time of loan origination. We have no clear prior on the eﬀect of the origination amount on the probability of delinquency. reported by the Bureau of Economic Analysis. We expect that a prepayment penalty makes reﬁnancing less attractive. Equals one if the backend debttoincome ratio is missing and zero if provided. holding constant the loantovalue and debttoincome ratio. Census Bureau 2000. DebttoIncome Ratio (+) Missing DebttoIncome Dummy (+) CashOut Dummy () Investor Dummy (+) Documentation Dummy () 16 Prepayment Penalty Dummy (+) Mortgage Rate (+) Margin (+) Product Type Dummies (+) Origination Amount (?) House Price Appreciation () Change Unemployment Rate (+) Neighborhood Income () . A higher interest rate makes it harder to make the monthly mortgage payment. Backend debttoincome ratio. An increase in the state unemployment rate increases the probability a homeowner lost his job. We report the expected sign for the independent variables in parentheses and sometimes provide a brief motivation. Zipcodelevel median income in 1999 from the U. reported by the Oﬃce of Federal Housing Enterprise Oversight (OFHEO).S. We consider four product types: FRMs. which therefore have the interpretation of the probability of delinquency relative to FRM. We include a dummy variable for the latter three types. we expect positive signs for all three product type dummies. Equals one if the borrower is an investor and does not owneroccupy the property. A higher debttoincome ratio makes it harder to make the monthly mortgage payment. Initial interest rate as of the ﬁrst payment date. Size of the mortgage loan. Higher housing equity leads to better opportunities to reﬁnance the mortgage loan. Margin for an adjustablerate or hybrid mortgage over an index interest rate. Equals one if there is a prepayment penalty and zero otherwise. as proxied by the median income. which increases the probability of ﬁnancial problems. Equals one if full documentation on the loan is provided and zero otherwise. PenningtonCross and Chomsisengphet (2007) show that the most common reasons to initiate a cashout reﬁnancing are to consolidate debt and to improve property. the more motivated a borrower may be to stay current on a mortgage. Hybrids. deﬁned by the total monthly debt payments divided by the gross monthly income. Explanation FICO Score () Combined LoantoValue Ratio (+) Combined value of all liens divided by the value of the house at origination. and Balloons. Statelevel change in the unemployment rate from the time of loan origination. Because we expect the FRM to be chosen by more riskaverse and prudent borrowers. Variable (Expected Sign) Fair. Isaac and Company (FICO) credit score at origination. We expect the lack of debttoincome information to be a negative signal on borrower quality. A higher margin makes it harder to make the monthly mortgage payment. We expect full documentation to be a positive signal on borrower quality. The better the neighborhood. A higher combined loantovalue ratio makes default more attractive. The ﬁrst two variables are used as dependent variables. The other variables are used as independent variables.
2 Determinants of Delinquency In Tables 3.and the margin for adjustablerate and hybrid loans. or. 14 Estimating house price appreciation on the MSAlevel.94 percent.4794) = 0. the combined loantovalue ratio. To the best of our knowledge. and house price appreciation. Census Bureau 2000 and are collected every 10 years. a one standard deviation increase in the FICO score decreases the log odds of ﬁrsttime delinquency by 47. The product type has a relatively small eﬀect on the performance of a loan. the mortgage rate. All marginal eﬀects have the expected sign. reported by the Bureau of Economic Analysis. To this end we use metropolitan statistical area (MSA) level house price indexes from the Oﬃce of Federal Housing Enterprise Oversight (OFHEO) and match loans with MSAs by using the zip code provided by LoanPerformance. deﬁned as the house price accumulation in the year prior to the loan origination. Third. Column two documents the marginal eﬀect of the covariates (Equation 7). The tables report the output of a single estimation. These variables were not signiﬁcant and did not materially change the regression coeﬃcients on the other variables. as shown in Table 2.13 In addition.S. In Figure We also studied speciﬁcations that included loan purpose. we construct a variable that measures house price appreciation from the time of origination until the time we evaluate whether a loan is delinquent. equivalently. due to limited page size. The four explanatory variables with the largest (absolute) marginal eﬀects and thus the most important for explaining crosssectional diﬀerences in loan performance. for example. In Table 2 we report the expected sign for the regression coeﬃcient on each of the explanatory variables in parentheses. 4 and 5 we present the determinants of delinquency using the proportional odds methodology.6192. According to the estimates. we use a measure for the quality of the neighborhood: zipcodelevel median household income in 1999. we include the statelevel change in the unemployment rate from the time of loan origination until the time of performance evaluation. An increase in the state unemployment rate increases the probability a homeowner lost his or her job. are the FICO score. which increases the probability of ﬁnancial problems. First. changes the odds by a factor exp(−0. the output is spread over three separate tables. and housing outlook. The ﬁrst column of Table 3 lists the covariates included (other than the vintage and age dummies which are reported in Tables 4 and 5). there is no data available to estimate the size of this measurement error or to evaluate its impact on the results. Except for the hybrid and ARM dummies. as opposed to the individual property level introduces a potential measurement error of this variable. 4. we use three macro variables in the baselinecase speciﬁcation. The data are from U. all variables are statistically signiﬁcant at the 1% level. The better the neighborhood the more motivated a borrower may be to stay current on a mortgage.14 Second. reported in Table 1. beyond what is explained by other characteristics. 13 17 .
15 18 . high house price appreciation contributed to a reduced probability of delinquency compared to a situation with average covariate values. the total Consistent with this ﬁnding. left panel). relatively low house price appreciation.16 By construction.15 The contributions of each covariate to explaining diﬀerent delinquency rates for each vintage year are given in columns 3 through 9 of Table 3. and 2001 was the year with the thirdhighest rates (see Figure 1. As shown in Table 4. For vintage years 2003 and 2005. the equalweighted average contribution is not zero. except for the small increase from 2006 to 2007. which is computed using Equation 12.71 percent increase in the log odds of delinquency. LaCourLittle (2007) shows that individual credit characteristics are important for mortgage product choice. the coeﬃcients of the vintage dummy variables increase every year.3 we show that FRMs experience a much lower delinquency rate than hybrid mortgages. 16 Shiller (2007) argues that house prices were too high compared to fundamentals in this period and refers to the house price boom as a classic speculative bubble largely driven by an extravagant expectation for future house price appreciation. where 2003 was the year with the lowest delinquency rates. This picture is in sharp contrast with that obtained from actual rates. Therefore. high mortgage rates.09 percent for vintage 2007—years for which low house price appreciation and high mortgage rates increased the probability of delinquency. The very high delinquency rate for vintage 2001 loans can be explained in large part by a near perfect storm of unfavorable lending and economic conditions: low FICO scores. For example. we can say that high house price appreciation between 2003 and 2005 masked the true riskiness of subprime mortgages for these vintage years. The vintage dummy coeﬃcients reported in Table 4 are in the same units as the contribution of the diﬀerent variables presented in Table 3. comparing vintages 2001 and 2007. which demonstrates that loan quality deteriorated after adjusting for the other covariates included in the speciﬁcation. all contributing to a higher probability of delinquency. This deterioration is illustrated by the adjusted delinquency rates depicted in Figure 1 (right panel). compared to a situation with average covariate values.06 percent for vintage 2006 and slightly smaller than the 56. Because the number of loans originated diﬀers across vintages years. hence rows for the contribution variable in Table 3 do not add up to zero. which therefore must be driven by borrowers with better characteristics selecting into FRMs. with weights equal to the number of originated loans in a particular vintage year. This is slightly higher than the 45. and low (negative) change in unemployment. We test whether the diﬀerences between every subsequent vintage year dummy coeﬃcients are statistically signiﬁcant. the weightedaverage contribution of a variable over 2001–2007 is zero. the diﬀerent covariates contributed to a 51. In total. We ﬁnd that the yearly change (increase) in the dummy variables are statistically signiﬁcant at the 1% conﬁdence level.
69∗ −8.51 −0.23 0.34 −0.36 2001 2002 2003 2004 2005 2006 Marginal Eﬀect.00 −8. The ﬁrst column reports the covariates included.07 −0.00 2.20 0.01 −0. Columns three through nine detail the contribution of a variable to explain a diﬀerent probability of delinquency in 2001–2007 (Equation 9).94∗ 24.35 0.37 16.71 3. The second column reports the marginal eﬀect.29 −1.17 2.40 −3.49 0.86∗ −13.19 −9.92 −2.86 5.46 11.05 5.11 17. deﬁned in Equation 7. Variables Other Than Vintage and Age Dummy Variables The table shows the output of the proportional odds duration model.04 13.40 −2.54 0.02 0.55 −34.Table 3: Determinants of Delinquency.94 −2.73 −0.32 −2.63 0.96∗ 7.32 −0.59 0.31 21.64 −1.00 −0.03 7.69 2.79∗ 6.25 0.02 −0.74 −0.32 0.08 32.24 12.00 −0. % Contribution.11 0.42 −1.23 2.90 −0.29 0.93 0.24 −0.66 −7.85 −0.04 −0. other than the vintage and age dummies which are reported separately in Table 4.84∗ 1.38 56.51 0.36 3.40 0.04 0.81∗ − −2.00 3.13∗ 29.43 0.09 −1.38 2.67 0.35 0. % 2007 3.68 0.78 −4.44 −5.02 −0.95 0.86 0.09 FICO Score −1.18 0.48 3.60 −0.00 2.74 6.81 0.81 −0.78 −0.03 −0.31 −5.52 0.86 0.00 39.18 −1.03 0.01 −2.32 −10.50 0.99 −27.10 0.48 2.48 45.31 −0.19 0.53 −0.91∗ −29.02∗ 12.12 −0.05 −0.73∗ 3.25 0.85 −2.04 Combined LoantoValue Ratio DebttoIncome Ratio Missing DebttoIncome Dummy CashOut Dummy Investor Dummy 19 Documentation Dummy Prepayment Penalty Dummy Mortgage Rate Margin Hybrid Dummy ARM Dummy Balloon Dummy Origination Amount House Price Appreciation Change Unemployment Neighborhood Income Total .78 −13.95 4.03 0.52 −16.76 51.59∗ −12.06 0.43 2.06 −0.16 −0.76 −2.40 −2.53 −0.62 1. Explanatory Variable 2001–2007 −47.34 13.01 −0.00 −3.00 −0.71 −7. A “∗” indicates statistical signiﬁcance at the 1% level.00 −1.21∗ 11.24 −2.11 −0.98 −0.70∗ 15.86∗ 12.03 0.52 0.31 −0.
73 0. The ﬁrst column reports the vintage dummies included. A “∗” indicates statistical signiﬁcance at the 1% level. Months 3 4 5 6 7 8 9 10 AgeEﬀect (*100) 0. estimated for each age dummy variable based on Equation 6. The second column shows the estimated coeﬃcients. Vintage Dummy Variables The table shows the output of the proportional odds duration model. The vintage 2001 dummy was not included as covariate.67 0.73 0.73 0.55 0.93∗ 64.70 0. The values for the other covariates are reported in Tables 3 and 4.69 0.00 5.72 0.71 0.64∗ 49. Each column reports the corresponding odds ratio. Explanatory Variable Estimate. AgeEﬀect. hence the coeﬃcients on the other covariates are relative to the 2001 vintage and we report a zero value for the coeﬃcient of 2001.12∗ 23.69∗ Table 5: Determinants of Delinquency. Age.72 0.65∗ 66.31 0. The values for the other covariates are outlined in Tables 3 and 5.67 – 20 .72 0. Age Dummy Variables The table shows the output of the proportional odds duration model. Months 11 12 13 14 15 16 17 – AgeEﬀect (*100) 0.04 0.Table 4: Determinants of Delinquency.72 Age. % Vintage 2001 Vintage 2002 Vintage 2003 Vintage 2004 Vintage 2005 Vintage 2006 Vintage 2007 0.71∗ 57.
Moreover. plus the 10 interaction and quadratic terms that can be constructed from the four most important variables: the FICO score.17 The ﬁnding of a continual decline in loan quality also occurs when analyzing foreclosure rates (Appendix A). in Table 5 we report the odds statistic deﬁned in Equation 6. Moreover. This pales in comparison to the 66.1 Diﬀerent Loan and Borrower Characteristics We explore numerous alternative loan and borrower characteristics as covariates for robustness. the deterioration is clearly visible. We therefore conclude that the dramatic deterioration of loan quality in this decade should have been apparent by the end of 2005. because the computation of the adjusted delinquency rate for. makes use of a regression model estimated using data from 2001 through 2008.69 percent increase in the vintage dummy variable over 2001–2007. we consider as covariates those of the baseline case presented in Table 3.contribution of the (nonvintage dummy) covariates increased by 4. To illustrate the eﬀect of age on the conditional probability of ﬁrsttime delinquency.71 percent to 56. say. and analyzing hybrid mortgages and FRMs separately (Appendix B). the CLTV ratio. Figure 5 (right panel) depicts the situation when we use data available at the end of 2006. vintage 2001 loans. Again.09 percent).38 percent (from 51. First. a short position is associated with a high cost of carry (Feldstein (2007)). 5 Empirical Results for Alternative Speciﬁcations In this Section we explore various alternative model speciﬁcations. was the dramatic deterioration of loan quality since 2001 already apparent? Notice that we cannot answer this question by simply inspecting vintages 2001 through 2005 in Figure 1 (right panel). The resulting age pattern in adjusted delinquency rates is plotted in Figure 5 (left panel). 17 21 . One reason why investors did not massively start to avoid or short subprimerelated securities is that the timing of the subprime market downturn may have been hard to predict. Next we study the following question: Based on information available at the end of 2005. we reestimate the proportional odds model underlying Figure 1 (right panel) making use of only 2001–2005 data. the main result documenting the monotonic rise in adjusted delinquency rates is found based on numerous alternative model speciﬁcations discussed in Section 5. 5. We see that the odds of ﬁrsttime delinquency peaks around age of 7–13 months. Hence. We again obtain the result that the adjusted delinquency rate rose monotonically from 2001 onward.
Viewed at the End of 2005 and 2006 The ﬁgure shows the adjusted delinquency rate using data available at the end of 2005 (left panel) and 2006 (right panel). cashout reﬁnancing dummies. documentation levels. Adjusted Delinquency Rate (%). at or before a given age. owneroccupation dummies. margins. End of 2005 16 14 12 10 8 6 4 2 0 3 4 5 6 7 8 Loan Age (Months) 9 10 11 2005 2004 2003 2002 2001 Adjusted Delinquency Rate (%). missing debttoincome ratio dummies. MSA house price appreciation since origination.Figure 5: Adjusted Delinquency Rate. origination amounts. change in state unemployment rate since origination. The adjusted delinquency rate is obtained by adjusting the actual rate for yearbyyear variation in FICO scores. or defaulted. percentage of loans with prepayment penalties. mortgage rates. debttoincome ratios. loantovalue ratios. realestate owned. End of 2006 16 14 12 10 8 6 4 2 0 3 4 5 6 7 8 Loan Age (Months) 9 10 11 2006 2005 2004 2003 2002 2001 22 . in foreclosure. and neighborhood median income. The delinquency rate is deﬁned as the cumulative fraction of loans that were past due 60 or more days. composition of mortgage contract types.
referring to a situation where an investor takes out a secondlien loan not reported in our database typically in combination with an 80 percent ﬁrstlien loan. We ﬁnd that the coeﬃcient on the FICOCLTV interaction term is close to zero and insigniﬁcant. we take into account the eﬀect of risklayering—such as. Allowing for these additional terms. We ﬁnd that the coeﬃcient is positive and statistically signiﬁcant and that it inherits some of the statistical power of the CLTV variable. are statistically and economically similar to those based on the entire sample of loans. This dummy variable is statistically signiﬁcant but economically not very large and moreover hardly improved the overall ﬁt. Inclusion of the dummy does not substantially increase the overall ﬁt of the model. Third. Fourth. for example. as measured by the log likelihood ratio. we included a dummy for the presence of a secondlien loan. Fifth. The estimates based on the smaller subsample.the mortgage rate. the eﬀect of a combination of a borrower’s low FICO score and a high CLTV ratio—on the probability of delinquency. because the borrower is closer to a situation with negative equity in the house (combined value of the mortgage loans larger than the market value of the house). we are aiming to control for silent seconds. we performed several robustness checks regarding the reset rate for hybrid mortgages. A negative sign would mean that a low FICO and a high CLTV reinforce each other and give rise to a predicted delinquency probability that is higher than that without interaction eﬀect. The 23 . and subsequent house price appreciation. we ﬁnd a negative and signiﬁcant (at the 1% level) coeﬃcient on this interaction term. Consistent with this intuition. With this variable. It is in this case not a priori clear what the sign on the FICOCLTV interaction variable should be. Second. we considered as an additional covariate a dummy variable taking the value one whenever the CLTV equals 80 percent. The vintage dummy variables are still increasing every year. Inclusion of the 10 interaction covariates does not substantially increase the overall ﬁt of the model. we excluded the loans where the debttoincome ratio is missing from the sample to make sure the measurement error associated with this variable does not lead to a signiﬁcant bias in the results. More certain is the sign we expect on the HPACLTV variable. as measured by the log likelihood ratio. A positive sign could be explained by lenders who originate a low FICO and high CLTV loan only if they have positive private information on the loan or borrower quality. Low house price appreciation is expected to especially give rise to a higher delinquency probability for a high CLTV ratio. in which the debttoincome variable has nonzero reported values. The coeﬃcients of the other covariates are virtually unchanged.
First. 19 For more details on this procedure.60 percent for the two alternative speciﬁcations respectively. we include the postreset margin as an additional covariate in the baseline case speciﬁcation. We performed two additional robustness checks. the marginal eﬀects for the margin covariates in the baseline case are 11. For FRMs the values of the new covariates are set to zero. Second. The log likelihood of the diﬀerent speciﬁcations is virtually identical. the duration of the initial ﬁxed rate period is 24 months or more.21 percent in the baseline case. 18 24 . 5. This captures the potential change in interest rate at the time of reset. and in both cases obtain our main result that adjusted delinquency rates rose monotonically over 2001–2007 (see Figure 9). see the oﬃcial OFHEO documentation by Calhoun (1996). The margin is in excess of a reference rate. followed by 36 months. Since we focus on delinquency in the ﬁrst 17 months.84 percent.18 OFHEO uses a repeatedsales methodology to compute house price appreciation in a geographical unit (like a state or MSA) and. instead of the margin we included a covariate deﬁned as the margin plus the 6month LIBOR interest rate at origination (obtained from Bloomberg). compared to a marginal eﬀect of 29. as reported in Table 3. the most common ﬁxed period is 24 months. as a byproduct.34 percent of the cases is the 6month LIBOR rate. which in 99. We obtained the data from the Oﬃce of Federal Housing Enterprise Oversight (OFHEO) at the state level. as reported in Table 3. However. instead of the margin we included a covariate deﬁned as the margin plus the 6month LIBOR interest rate at origination minus the initial mortgage rate. because households may have factored in the rise in the mortgage rate before the actual reset date. we expect the initial mortgage rate to be an important covariate. MSAlevel data exist but was not available for public release. Among the other covariates. and Van Hemert (2008) explain the interpretation and computation of the volatility parameters. only the coeﬃcient for the initial rate is slightly aﬀected.19 We would like to thank OFHEO for providing us with the data.58 percent and 33. The marginal eﬀect for the two new covariates is 12.83 percent respectively.27 percent and 7. based on the 6month LIBOR rate at origination. For 99. As a robustness check we performed the analysis for FRMs and Hybrid mortgages separately. Also De Jong.57 percent of mortgages in our database. with FRMs not being subject to resets. Driessen. obtains an estimate of the variation of the return around the mean in the geographical unit.initial mortgage rate for hybrid mortgages is potentially lower during the initial ﬁxedrate period. The marginal eﬀect of the initial rate is 29.2 Local housebyhouse return volatility In this section we study local housebyhouse return volatility as as additional covariate.
75 percent. annualized and averaged over the full panel (50 states plus the District of Columbia in the 2001Q12008Q2 sample period). 5. The standard deviation of the volatility estimate over time (averaged over the 50 states plus the District of Columbia) is only 0. in the empirical implementation we add both the housebyhouse return volatility by itself and the interaction of this volatility with HPA as covariates. but indicates that the evaluation process should accommodate the situation and 25 . based on house price appreciation in the geographical area. the estimated coeﬃcient on the interaction term is positive.3 CRA and GSE Housing Goals In this section we explore additional covariates related to the Community Reinvestment Act (CRA) and Government Sponsored Enterprises (GSEs) housing goals. When the equity is positive based on house price appreciation in the geographical area. The 1977 Community Reinvestment Act (CRA) is a United States federal law designed to encourage commercial banks and savings associations to meet the needs of borrowers in all segments of their communities. the higher the probability an individual house has negative equity. the higher the housebyhouse return volatility.05 percent. there is a 50 percent probability the household actually has negative equity based on the (unobserved) house price appreciation of the individual house. The CRA does not list speciﬁc criteria for evaluating the performance of ﬁnancial institutions. and Huszar (2005). normal) distribution for the housebyhouse deviation from the MSA mean.The housebyhouse volatility is 8.18 percent.and moderateincome (LMI) households and neighborhoods. LaCourLittle. While it is not just the probability of having negative equity for an individual house that matters. the situation is reversed. and Huszar (2005) use the volatility estimate of OFHEO to compute the probability of negative equity. In the special case that the equity in the house is exactly zero. LaCourLittle. The eﬀect is statistically signiﬁcant at the 1% level. including low.g. Ambrose. Hence. when house price appreciation is high in the geographical area. but has a small Chisquared statistic compared to the baseline case covariates and has a negligible impact on the coeﬃcients of the other covariates. Most of the variation of the volatility is in the crosssection: the standard deviation of the volatility estimate across states (averaged over 2001Q12008Q2) is 0. it does seem intuitive that the eﬀect of local housebyhouse return volatility may interact with HPA.. but in general the whole probability distribution. When the equity is negative. high volatility increases the probability of having negative equity for a speciﬁc house. Consistent with Ambrose. assuming a symmetric (e.
We either include an LMI dummy that equals one if the median neighborhood income is less than 80 percent of the MSA median income.context of each individual institution. LoanPerformance data identiﬁes each property location in zip codes. An institution’s CRA compliance record is. The goals are primarily deﬁned in terms of (i) household income relative to median MSA income (for urban neighborhoods). and (iii) minority concentration in the neighborhood. Census Bureau to compute the LMI dummy. In speciﬁcation II we add the MSA median income to the baseline case covariates.pdf Table 1 for the speciﬁc goals for 20002006. poverty level. 21 20 26 . The results are presented in Table 6. including the neighborhood median income. CRA and GSE targeted communities are deﬁned using census tractlevel measures (such as income.org/Datasets/GSE/gse2006. controlling for the baselinespeciﬁcation covariates. Therefore.23 The LMI dummy or MSA median income level pick up the diﬀerence in the probability of delinquency for loans in MSAs with diﬀerent median incomes. (ii) neighborhood median income relative to MSA median income. 22 See http://www. Of the baseline case covariates. thus we cannot identify the CRA targeted community. for example. e. giving rise to diﬀerent incentives for CRAcomplying institutions and the GSEs. A higher See http://www. We interpret any eﬀect from the LMI dummy or MSA median income to come from the CRA and GSE housing goals. 23 The precise identiﬁcation of CRA and GSE targeted neighborhoods is not feasible in our analysis for several reasons: (i) the location of subprime lenders is not reported in the LoanPerformance data. we only report the neighborhood income to preserve space. the MSA for urban neighborhoods. taken into account when applying for deposit facilities.gov/dcca/cra/ See Laderman (2004) for a further discussion. Speciﬁcation I corresponds to the baseline case of Table 3 and is included as a benchmark.S.huduser. or we add the MSA median income.22 To isolate the eﬀect of CRA and the GSE housing goals from pure neighborhood eﬀects..federalreserve. which do not directly match with census tracts.21 With a similar objective of helping poor and underserved individuals and neighborhoods. we study the eﬀect of LMI neighborhood status on the probability of delinquency. and thus assume that there is no direct eﬀect of LMI status or MSA median income on borrower behavior after controlling for neighborhood median income. and minority concentration). in our analysis we identify neighborhoods that can potentially fall under CRA and GSE housing goals using either the LMI dummy variable or MSAlevel median household income.20 LMI neighborhoods have a median income level that is less than 80 percent of the median income of a broader geographic area.g. and (ii) most frequently. We use zipcode level median household income and median MSA household income from the U. but in neighborhoods with the same median income and with the same loan and borrower characteristics. the Congress in 1992 established an aﬀordable housing mission for Fannie Mae and Freddie Mac by directing the Department of Housing and Urban Developement (HUD) to create speciﬁc mortgage purchase goals for these Government Sponsored Enterprises (GSEs).
if it is a low..79%∗ 3. I Neighborhood Income MSA Income LMI Dummy LMI 2001 LMI 2002 LMI 2003 LMI 2004 LMI 2005 LMI 2006 LMI 2007 II III IV −8. A “∗” indicates statistical signiﬁcance at the 1% level. The positive and statistically signiﬁcant (at the 1% level) marginal eﬀect reported in the table eﬀect thus implies that CRA and GSE housing goal eligibility is associated with a higher probability of delinquency.and moderateincome (LMI) neighborhood. i.64%∗ 2.07%∗ − − − − − − − − − 2. The marginal eﬀect for the interaction terms is computed as the product of the estimated coeﬃcient and the full sample standard deviations of LMI.67%∗ 1. Finally in speciﬁcation IV we interact the LMI dummy with vintage dummies to see how the LMI eﬀect changed over time.82%∗ − − − − − − − − − 6.94%∗ − − − − − − − − − 1. Table 6: Low. In speciﬁcation III we add the LMI dummy to the baseline case covariates.and ModerateIncome Neighborhood Eﬀect This table reports marginal eﬀects for diﬀerent measures of neighborhood (median household) income relative to MSA (median household) income. In all four speciﬁcations we include the baseline case variables used in Table 3.95%∗ −6. of which we only report neighborhood income to preserve space. Speciﬁcation I is the baseline case with no variable capturing the neighborhood income relative to the MSA income. Speciﬁcation III includes a dummy with value one if neighborhood income is less than 80 percent of MSA income.75% 3.81%* −9.MSA median income increases the possibility a loan was made to advance CRA or GSE housing goals. Speciﬁcation IV includes interaction variables of the LMI dummy with vintage year dummies.79%∗ 27 . There is no clear trend over time.10%∗ −6.26%∗ 4.52%∗ 1.e. ceteris paribus. We ﬁnd a positive and signiﬁcant marginal eﬀect. For all vintage years the marginal eﬀect is positive and statistically signiﬁcant at the 1% level. which is positive and statistically signiﬁcant at the 5% level. except for 2003. Speciﬁcation II includes MSA income. again implying that CRA and GSE housing goal eligibility raises the probability of delinquency.
The combined LTV ratio rather than the ﬁrstlien LTV ratio is believed to be the main determinant of delinquency because it is the burden of all the debt together that may trigger ﬁnancial problems for the borrower. ﬁrstlien loantovalue ratio. 25 24 28 . the CLTV is an increasingly important determinant of delinquency over time. including the ﬁrstlien LTV and secondlien LTV (CLTV minus ﬁrstlien LTV). That is. We examine whether lenders were aware that high LTV ratios were increasingly associated with riskier borrowers. and loan characteristics. for example. 0. term of the mortgage.15. We estimate the proportional odds model with CLTV interacted with the seven vintage dummy variables. Hence. 0. origination amount.25 We perform one such regression for each calendar quarter in our sample period. the eﬀect of a unit change in a given covariate on the probability of ﬁrsttime delinquency is the same in. a cashout reﬁnancing dummy. instead of just CLTV without interaction.14. 0. as independent variables.34 for years 2001 to 2007. a dummy for owner occupation. documentation dummy. In contrast. as we have in the baseline case speciﬁcation (Section 4). 0. which totals 0. a dummy for a missing debttoincome ratio. To conserve space the results are not reported.22. by investigating the relationship between the loantovalue ratio and mortgage rates over time. as it is in 2001. holding constant the values of the other covariates. 2006. margin. We can only expect to get accurate results when using relatively homogeneous groups of loans.10. secondlien loantovalue ratio. and therefore consider fully amortizing FRM and 2/28 hybrid loans separately. prepayment penalty dummy. we use the FICO score. and prepayment term as the righthandside variables.6 NonStationarity of the LoantoValue Eﬀect The proportional odds model used in Section 4 assumes that the covariate coeﬃcients are constant over time. Together these two contract types account for more than half of all mortgage loans in This is conﬁrmed by our empirical results. respectively. the ﬁrstlien LTV is the more important determinant of the mortgage rate on a ﬁrstlien mortgage. All seven coeﬃcients are highly statistically signiﬁcant at the 1% level. because it captures the dollar amount at stake for the ﬁrstlien lender. debttoincome ratio.24 For this reason. In this section we ﬁrst discuss this ﬁnding and then turn to the question of whether lenders were aware of the nonstationarity of the loantovalue eﬀect. Speciﬁcally. 0. we test whether the sensitivity of the lender’s interest rate to the ﬁrstlien LTV ratio changed over time.35. 0. We compute the marginal eﬀect for a particular loan vintage as the product of the estimated coeﬃcient for the associated interaction variable and the standard deviation of CLTV for all vintages together. We test the validity of this assumption for all variables in our analysis and ﬁnd that the strongest rejection of a constant coeﬃcient is for the CLTV ratio.11. We perform a crosssectional OLS regression with the mortgage rate as the dependent variable.
Figure 2 shows the regression coeﬃcient on the ﬁrstlien LTV ratio for each quarter from 2001Q1 through 2007Q2. and they can therefore be interpreted as the changes in the mortgage rates when the ﬁrstlien LTV ratios are increased by one standard deviation. This provides evidence that lenders were to some extent aware of high LTV ratios being increasingly associated with risky borrowers.fhfb. In this section we analyze the time series of the subprimeprime rate spread. In Figure 6 we show the actual subprimeprime rate spread. respectively. For hybrid mortgages the subprimeprime comparison is more complicated because (i) both the initial (teaser) rate and the margin should be factored in. We focus on FRMs for this exercise.27 In Appendix C we show that this result is robust to allowing for a nonlinear relationship between the mortgage rate and the ﬁrstlien LTV ratio. both with and without adjustment for changes in loan and borrower characteristics. but due to a near shutdown of the securitized subprime mortgage market we lack statistical power in this quarter. as unreported results suggest.our database. 26 29 .gov/GetFile. the corresponding rate increase was 10 and 16 basis points. notice that the eﬀect of a onestandarddeviation increase in the ﬁrstlien LTV ratio on the 2/28 mortgage rate increased substantially in the wake of the subprime mortgage crisis: from 30 basis points in 2007Q1 to 42 basis points in 2007Q2. keeping constant other loan characteristics.28 The subprimeprime spread—the diﬀerence between the average subprime and prime rates—decreased Our data extends to 2007Q3. interest rates on subprime mortgages are higher than on prime mortgages to compensate the lender for the (additional) default risk associated with subprime loans. In the fourth quarter of 2006. 27 The eﬀects of other loan characteristics on mortgage rates have been much more stable over time. 28 Available at http://www. the prime rate is the contract rate on FRMs reported by the Federal Housing Finance Board (FHFB) in its Monthly Interest Rate Survey. Each crosssectional regression is based on a minimum of 18. a onestandarddeviation increase in the ﬁrstlien LTV ratio corresponded to about a 30basispoint increase in the mortgage rate for 2/28 hybrids and about a 40basispoint increase for FRMs.784 observations.aspx?FileID=6416. 7 SubprimePrime Rate Spread In general. in the ﬁrst quarter of 2001. and (ii) we don’t have good data on the prime initial rates and margins. Finally. deﬁned in Equation (15) below. The subprime rate for this exercise is calculated as the average across individual loans initial mortgage rate for each calendar month (the data source is LoanPerformance).26 We scaled the coeﬃcients by the standard deviation of the ﬁrstlien LTV ratio. In contrast.
which coincides with the most rapid growth in the number of loans originated (see Table 1).5 2 1. origination amount. hence the decline cannot just be attributed to a change in the overall level of risk aversion. etc. 3. with the largest decline between 2001 and 2004. a dummy variable that equals one if a borrower is an investor. the actual subprimeprime rate spread declined much more and more steadily. third. a dummy variable that equals one if prepayment penalty is present. In Figure 6 we also plot the yield spread between 10year BBB and AAA corporate bonds. using data from 2001 through 2006. value of debttoincome ratio. a dummy variable that equals one if full documentation was provided.substantially over time. a dummy variable that equals one if loan is a reﬁnancing. which we obtained from Standard and Poor’s Global Fixed Income Research. loantovalue ratio based on a ﬁrstlien. spreadit = subprimeit − primet (14) (15) Notice that the β1 primet term corrects for the fact that the spread is aﬀected by the prime rate itself.29 spreadit = β0 + β1 primet + β2 characteristicsit + errorit . and loantovalue ratios based on a second. and thus changes over the business cycle.5 0 2001 Subprime−Prime Spread (%) BBB−AAA Spread (%) 2002 2003 2004 Year 2005 2006 2007 We perform a crosssectional OLS regression with the loanlevel spread as the dependent variable and the prime rate and various subprime loan and borrower characteristics as the explanatory variables. Compared to the corporate BBBAAA yield spread.5 1 . 29 30 . liens if applicable. Figure 6: FRM Rate Spread and Corporate Bond Yield Spread The ﬁgure shows the FRM subprimeprime rate spread and the yield spread between 10year BBB and AAA corporate bonds.5 3 2. because a higher prime rate increases the default probability The explanatory factors in the regression are the FICO credit score. a dummy variable that equals one if debttoincome was not provided.
In Figure 1 (right panel) we showed that loan quality. along with a ﬁtted linear trend. third. averaged per origination month t. Figure 7: Prediction Error in the SubprimePrime Rate Spread The ﬁgure shows the prediction error in the subprimeprime rate spread. obtained by adjusting loan performance for diﬀerences in loan and borrower characteristics and subsequent house price appreciation. . determined in a regression of the spread on the prime rate and the following loan and borrower characteristics: the FICO credit score.on subprime loans for a given spread. value of debttoincome ratio. a dummy variable that equals one if debttoincome was not provided.75 . a dummy variable that equals one if the loan is a reﬁnancing. origination amount. a dummy variable that equals one if full documentation was provided. after adjusting for diﬀerences in observed loan and borrower characteristics.25 −. deteriorated over the period. liens if applicable. a dummy variable that equals one if a prepayment penalty is present. the loantovalue ratio based on a ﬁrst lien.5 . etc. argue that this was facilitated by the development of socalled privatelabel mortgage backed securities (they do not carry any kind of credit risk protection by the Government Sponsored Enterprises). Investors in search of higher yields kept increasing their demand 31 .25 0 −. 8 Concluding Remarks The subprime mortgage market experienced explosive growth between 2001 and 2006. declined between 2001 to 2007. on a perunitofrisk basis. a dummy variable that equals one if a borrower is an investor. In Figure 7 we plot the prediction error. Therefore. the subprimeprime mortgage spread decreased even more than the level of the spread. Angell and Rowley (2006) and Kiﬀ and Mills (2007). and thus the (adjusted) riskiness of loans rose. and the loantovalue ratio based on a second. among others.5 2001 2002 2003 2004 Year 2005 Prediction Error (%) Fitted Trend 2006 2007 The downward trend in Figure 7 indicates that the subprimeprime spread.
and macroeconomic conditions (such as house price appreciation. the quality of the market deteriorated dramatically. Were problems in the subprime mortgage market apparent before the actual crisis erupted in 2007? A more detailed discussion. 30 32 . for example. and empirical evidence on such episodes is available in Dell’Ariccia and Marquez (2006). Sweden. For the rapid growth of the subprime mortgage market to have been sustainable. In this paper we show that during the dramatic growth of the subprime (securitized) mortgage market. and the subprime mortgage market crashed in 2007. a level of indebtedness. loan characteristics (such as a product type. highLTV borrowers became increasingly risky (their adjusted performance worsened more) compared to lowLTV borrowers. theory.for privatelabel mortgagebacked securities. the increase in the overall riskiness of subprime loans should have been accompanied by an increase in the subprime markup. In 2001. Over time. The decline in loan quality was monotonic. and Landerretche (2001). Indonesia. an ability to provide documentation). the subprimeprime mortgage rate spread (subprime markup) should account for the default risk of subprime loans. among many others. Chile in 1982. Demirg¨cKunt and Detragiache (2002). In many respects.30 Argentina in 1980. deteriorating loan performance. In this paper we show that this was not the case: The subprime markup—adjusted and not adjusted for changes in diﬀerences in borrower and loan characteristics—declined over time. but not equally spread among diﬀerent types of borrowers. the subprime market experienced a classic lending boombust scenario with rapid market growth. albeit in diﬀerent economic settings. an amortization term. Valdes. adjusted for diﬀerences in borrower characteristics (such as the credit score. Gourinchas. a loan amount. and Finland in 1992. Securitizers seem to have been aware of this particular pattern in the relative riskiness of borrowers: We show that over time mortgage rates became more sensitive to the LTV ratio of borrowers. and decreasing risk premiums. and Kamisky and Reinhart u¸ (1999). which also led to sharp increases in the subprime share of the mortgage market (from around 8 percent in 2001 to 20 percent in 2006) and in the securitized share of the subprime mortgage market (from 54 percent in 2001 to 75 percent in 2006). Thailand. and Korea in 1997 all experienced the culmination of a boombust scenario. an mortgage interest rate). By 2006. Norway. a borrower with a one standard deviation aboveaverage LTV ratio paid a 10 basis point premium compared to an average LTV borrower. In principal. in contrast. the premium paid by the high LTV borrower was around 30 basis points. level of neighborhood income and change in unemployment). We measure loan quality as the performance of loans. With the beneﬁt of hindsight we now know that indeed this situation was not sustainable. Mexico in 1994. loosening underwriting standards.
Rapid appreciation in housing prices masked the deterioration in the subprime mortgage market and thus the true riskiness of subprime mortgage loans. at least by the end of 2005. the risk in the market became apparent. Loan quality had been worsening for ﬁve years in a row at that point. we show that the monotonic degradation of the subprime market was already apparent. Using the data available only at the end of 2005. When housing prices stopped climbing. 33 .Our answer is yes.
housing ﬁnance. Giovanni. 34 . Calhoun. Unobserved heterogeneity in models of competing mortgage termination risks. and Otto Van Hemert. Paul.S. Cutts. NC. Journal of the Royal Statistical Society B34. Rowley. 1449–1494. Deniz Igan. Speech Federal Reserve Bank of Kansas City’s Economic Symposium. Markus. and monetary policy. and Charles Capone. and Clare D. The institutional memory hypothesis and the procyclicality of bank lending behavior. 2007. Oﬃcial OFHEO document. Quarterly a Journal of Economics 118. Brent. 2003. DellAriccia. Ofheo house price indexes: HPI technical description. Ambrose. 2007. 458–495. Real Estate Economics 34. and Jo˜o F. 275–293. Brunnermeier. John Y. Amy Crews. Hedging house price risk: Portfolio choice with housing futures. Survival Analysis Using SAS: A Practical Guide (SAS Institute. Real Estate Economics 33. The Journal of Real Estate Finance and Economics 30. Charles. Breaking new ground in U. Berger. De Jong. 765–782. 2004. Ben. and Yongheng Deng. Review of Financial Studies forthcoming. 1972. Ambrose. Brent. 2008. Clapp. Bernanke. 243–273. Household risk management and optimal mortgage choice. 2008. USA). Frank. 2000. and Christian Julliard. Yongheng Deng. FDIC: Outlook Summer 2006 Federal Deposit Insurance Corporation. Cary. Credit booms and lending standards: Evidence from the subprime mortgage market. Working paper. Journal of Real Estate Finance and Economics 20. Joost Driessen. On the economics of subprime lending. and Zsuzsa Huszar. 167–196.. 2005. and Xudong An. and Gregory Udell. Working paper. Michael LaCourLittle. John. Cox. Van Order. Cynthia. Charles. mortgage lending. 2005. 9–33. 2006. 1996. Allen. 187–220. 2002. Calhoun. Journal of Real Estate Finance and Economics 24. 2006. 2007. Money illusion and housing frenzies. Journal of Financial Intermediation 12. The hazard rates of ﬁrst and second defaults. Housing. 10th Printing. Regression models and life tables. Cocco. A dynamic analysis of ﬁxed.References Allison. A note on hybrid mortgages.and adjustablerate mortgage terminations. Campbell. and Robert A. Angell. and Luc Laeven. David.
2000. 1997. Journal of Finance 61. 2008. Martin S. Joseph. and Paul Willen. Superstar cities. Spatial heterogeneity in mortgage terminations by reﬁnance. Did securitization lead to lax screening? Evidence from subprime loans. homeownership experiences. Working paper. Loss aversion and seller behavior: Evidence from the housing market. NBER working paper 13471. Gyourko. Kiﬀ. 1373–1406. 275–308. Lending booms and lending standards. and Christopher Mayer. IMF working paper. Journal of Real Estate Finance and Economics 14. and Oscar Landerretche. 1233–1260. Gourinchas. Money for nothing and checks for free: Recent developments in U. Benjamin J. and foreclosures. Deng. Genesove. and Todd Sinai. Economia 1. Feldstein. 2005. Real Estate Economics 33. Quarterly Journal of Economics 116. Lending booms: Latin America and the world. Rodrigo Valdes. Asli. 1997. David. and Lihong Yang. sale and default. Andrey Pavlov. Keys. 2002. and Enrica Detragiache. and Paul Mills. 473–500. 2007. NBER working paper 12355. 1999. 47–99. Tanmoy Mukherjee. Econometrica 68. Reinhart. Mortgage timing. Graciela L. Mayer Christopher. Giovanni. Amit Seru. Equity and time to sale in the real estate market. Otto Van Hemert. Mortgage termination: An empirical hazard model with stochastic term structure.Dell’Ariccia. Ralph. credit markets and the business cycle. 2005. Yongheng. Journal of Political Economy 113. Deng. 2001. 2006. 345–375. Glaeser. Housing. 2007. subprime mortgage markets. John. Subprime outcomes: Risky mortgages. Urban decline and durable housing. Kristopher. and Stijn Van Nieuwerburgh. Journal of Monetary Economics 49. 309–331. American Economic Review 89. and Joseph Gyourko. Yongheng. Quigley. Adam Shapiro. Demirg¨cKunt. 2006. 35 . and Carmen M. Kamisky. Edward. David. American Economic Review 87. heterogeneity and the exercise of mortgage options.S. Koijen. Yongheng. 2007. 2001.. John M.. Van Order.. 2511–2546. Working paper. 2008. 255–269. and Vikrant Vig. NBER working paper 13361. Deng. and Robert A. Does deposit insurance increase banking system stability? An u¸ empirical investigation. The twin crises: The causes of banking and balanceofpayment problems. PierreOliver. Genesove. and Robert Marquez. Gerardi. and Christopher Mayer. 739–764. Mortgage terminations.
James. Real Estate Economics 35.LaCourLittle. Vickery. The consequences of mortgage credit expansion: Evidence from the 2007 mortgage default crisis. 2004. Journal of Finance 29. 2004. Real Estate Economics forthcoming.. PenningtonCross. Otto. 1974. 1545–1548. Has the cra increased lending for lowincome home purchases? Federal reserve of san francisco economic letter. and R. Laderman. 36 . Credit history and the performance of prime and nonprime mortgages. PenningtonCross. 2007. 1073–1102. Journal of Real Estate Finance and Economics 27. 200416. Von Furstenberg. Liz. Jeﬀery Green. Mian. 233–263. Shiller. Review of Financial Studies 17. 2007. Anthony. NBER working paper 13553. Robert J. Working paper. Ruckes. Anthony. and Amir Suﬁ. and Souphala Chomsisengphet. 2003. 279–301. Subprime reﬁnancing: Equity extraction and mortgage termination. Working paper. 2007. The home purchase mortgage preferences of lowandmoderate income households. Atif. Working paper. 2008.. 2007. 2007. Understanding recent trends in house prices and home ownership. review of ﬁnancial studies. Martin. Michael. Household interest rate risk management. George M. Home mortgage delinquencies: A cohort analysis. Interest rates and consumer choice in the residential mortgage market. Van Hemert. Bank competition and credit standards.
but include contract type dummies in the regression 37 . loantovalue ratios. and neighborhood median income. The actual foreclosure rate for loans age six months and younger is close to zero. Using foreclosure instead of delinquency as a measure for nonperformance. Actual Foreclosure Rate (%) 18 16 14 12 10 8 6 4 2 0 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) 0 3 4 5 4 2 2007 2006 2005 2004 2003 2002 2001 Adjusted Foreclosure Rate (%) 10 8 6 2007 2006 2005 2004 2003 2002 2001 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) B Adjusted Delinquency Rate for Hybrids and FRMs Separately In this Appendix we show that the continual deterioration of adjusted loan performance over the 2001–2007 period also occurs when estimating a separate proportional odds model for the main contract types. as can be seen from Figure 8. MSA house price appreciation since origination. change in state unemployment rate since origination. The foreclosure rate is deﬁned as the cumulative fraction of loans that were in foreclosure. realestate owned. realestate owned. cashout reﬁnancing dummies. owneroccupation dummies. as a measure of loan performance. Figure 8: Actual and Adjusted Foreclosure Rates The ﬁgure shows the age pattern in the actual (left panel) and adjusted (right panel) foreclosure rate for the diﬀerent vintage years. Foreclosure is deﬁned as a loan being in foreclosure. The change for each year is statistically signiﬁcant at the 1% conﬁdence level. documentation levels. or in default. the actual foreclosure rates (Figure 8. right panel.A Foreclosure Rates In this Appendix we show the continual deterioration of adjusted loan performance using foreclosure. 2006. or defaulted. The adjusted foreclosure rate is obtained by adjusting the actual rate for yearbyyear variation in FICO scores. The adjusted foreclosure rates of vintage 2001 loans are between vintages 2002 and 2003. mortgage rates. the adjusted foreclosure rates every vintage year starting in2002. In Figure 8 we present actual (left panel) and adjusted (right panel) foreclosure rates. debttoincome ratios. at or before a given age. instead of delinquency. as opposed to the baseline case in the main text where we perform a single estimation for all loans together. origination amounts. missing debttoincome ratio dummies. Similar to the actual delinquency rates (Figure 1 (left panel). percentage of loans with prepayment penalties. and 2001 and lowest for 2003 and 2004. composition of mortgage contract types. left panel) are highest for 2007. For older aged loans the actual foreclosure rate is roughly speaking twice as low as the actual delinquency rate. in contrast to the actual delinquency rate at this age (presented in Figure 1 (left panel)). margins.
where the βs are the regression coeﬃcients and X is the ﬁrstlien LTV ratio at which the marginal eﬀect is evaluated. Figure 9 shows the adjusted delinquency rate for the two main contract types: 2/28 hybrids and FRMs. and neighborhood median income. the age pattern for the diﬀerent vintage years looks very much the same for the two contract types. as independent variables.speciﬁcation. The delinquency rate is deﬁned as the cumulative fraction of loans that were past due 60 or more days. MSA house price appreciation since origination. missing debttoincome ratio dummies. debttoincome ratios. Figure 9: Adjusted Delinquency Rates for Hybrids and FRMs Separately The ﬁgure shows the adjusted delinquency rates based on hybrid mortgages (left panel) and FRMs (right panel) separately. This is what we plotted in Figure 2. or defaulted. The sensitivity is deﬁned as the regression coeﬃcient on the ﬁrstlien LTV (scaled by the standard deviation) in a regression with the mortgage rate as dependent variable and the ﬁrstlien LTV. mortgage rates. composition of mortgage contract types. the adjusted delinquency rates have increased monotonically over time. The adjusted delinquency rate is obtained by adjusting the actual rate for yearbyyear variation in FICO scores. In this appendix we study the robustness of this result to adding the square of the ﬁrstlien LTV and the square of the secondlien LTV as independent variables. Without nonlinear terms the marginal eﬀect is simply given by the regression coeﬃcient. For both contract types. realestate owned. percentage of loans with prepayment penalties. cashout reﬁnancing dummies. owneroccupation dummies. With the quadratic terms. Hybrid Mortgage Loans 30 25 20 15 10 5 5 0 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) 0 3 4 5 6 2007 2006 2005 2004 2003 2002 2001 Fixed−Rate Mortgage Loans 20 2007 2006 2005 2004 2003 2002 2001 15 10 7 8 9 10 11 12 13 14 15 16 17 Loan Age (Months) C NonLinearity in the Sensitivity of the Mortgage Rate to the LTV In Figure 2 we plotted the sensitivity of the ﬁxedrate and 2/28 hybrid mortgage rates to the ﬁrstlien LTV ratio. In Figure 10 we report the resulting scaled marginal eﬀect of the ﬁrstlien LTV for ﬁxedrate and 2/28 hybrid mortgages evaluated at a ﬁrstlien LTV of 80 percent (left panel) and 90 percent (right panel). Except for a level diﬀerence. origination amounts. 38 . loantovalue ratios. change in state unemployment rate since origination. and the other loan and borrower characteristics listed in Section 6. the marginal eﬀect is given by βLT V + 2βLT V 2 X. at or before a given age. in foreclosure. the secondlien LTV. documentation levels. therefore allowing for a nonlinear functional form. margins.
the higher the ﬁrstlien LTV ratio. in the case without nonlinearity. prepayment penalty dummy. 39 . origination amount. and margin as independent variables. The eﬀect is determined using an OLS regression with the interest rate as dependent variable and the FICO score.2 0 2001 FRM 2/28 Hybrid Scaled Marginal Effect of First−Lien LTV = 90% (%) 1. the more sensitive is the mortgage rate to changes in the ﬁrstlien LTV.2 1 . secondlien LTV (and the square). In contrast.Figure 10: Sensitivity of Mortgage Rate to FirstLien LTV Ratio Allowing for NonLinearity The ﬁgure shows the scaled marginal eﬀect of the ﬁrstlien loantovalue (LTV) ratio on the mortgage rate for ﬁrstlien ﬁxedrate and 2/28 hybrid mortgages. Comparing the left and right panels in Figure 10. Moreover.4 . as in Figure 2. the marginal eﬀect is rising over time. owneroccupation dummy.8 . ﬁrstlien LTV (and the square). missing debttoincome ratio dummy. The scaled marginal eﬀect increases by 27 basis points over the course of 3 months in 2007 when the model allows for nonlinearity.2 0 2001 FRM 2/28 Hybrid 2002 2003 2004 Year 2005 2006 2007 2002 2003 2004 Year 2005 2006 2007 As shown in Figure 10.4 . cashout reﬁnancing dummy. consistent with the baseline case results presented in Figure 2.8 . debttoincome ratio. the increase in the scaled marginal eﬀect is only 13 basis points.6 . Scaled Marginal Effect of First−Lien LTV = 80% (%) . we ﬁnd that there is a statistically and economically signiﬁcant nonlinear eﬀect of the ﬁrstlien LTV on the mortgage rate.6 . prepayment term. The largest diﬀerence between the results based on speciﬁcations with and without nonlinearity is observed for 2/28 hybrid mortgages in 2007 at a ﬁrstlien LTV of 90 percent (right panel). term of the mortgage. evaluated at a ﬁrstlien LTV of 80 percent (left panel) and 90 percent (right panel).
This action might not be possible to undo. Are you sure you want to continue?
Use one of your book credits to continue reading from where you left off, or restart the preview.