The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
FEDERAL RESERVE BANK of ATLANTA
Reducing Foreclosures: No Easy Answers
Christopher Foote, Kristopher Gerardi,
Lorenz Goette, and Paul Willen
Working Paper 2009-15
May 2009
WORKING PAPER SERIES
FEDERAL RESERVE BANK o f ATLANTA
WORKING PAPER SERIES
Reducing Foreclosures: No Easy Answers
Christopher Foote, Kristopher Gerardi,
Lorenz Goette, and Paul Willen
Working Paper 2009-15
May 2009
Abstract: This paper takes a skeptical look at a leading argument about what is causing the foreclosure
crisis and what should be done to stop it. We use an economic model to focus on two key decisions: the
borrower’s choice to default on a mortgage and the lender’s subsequent choice whether to renegotiate or
modify the loan. The theoretical model and econometric analysis illustrate that unaffordable loans,
defined as those with high mortgage payments relative to income at origination, are unlikely to be the
main reason that borrowers decide to default. In addition, this paper provides theoretical results and
empirical evidence supporting the hypothesis that the efficiency of foreclosure for investors is a more
plausible explanation for the low number of modifications to date than contract frictions related to
securitization agreements between servicers and investors. While investors might be foreclosing when it
would be socially efficient to modify, there is little evidence to suggest they are acting against their own
interests when they do so. An important implication of our analysis is that policies designed to reduce
foreclosures should focus on ameliorating the immediate effects of job loss and other adverse life events
rather than modifying loans to make them more affordable on a long-term basis.
JEL classification: D11, D12, G21
Key words: mortgage, foreclosure, modification, securitization
The authors gratefully acknowledge Andreas Fuster for both first-rate research assistance and excellent comments and
suggestions. They also thank Daron Acemoglu, Larry Cordell, Jeff Fuhrer, Eileen Mauskopf, Jim Nason, Hui Shan, and Dan
Immergluck for helpful comments. The views expressed here are the authors’ and not necessarily those of the Federal Reserve
Banks of Atlanta or Boston or the Federal Reserve System. Any remaining errors are the authors’ responsibility.
Please address questions regarding content to Christopher Foote, Research Department, Federal Reserve Bank of Boston, P.O.
Box 55882, Boston, MA 02205, 617-973-3077, chris.foote@bos.frb.org; Kristopher Gerardi, Research Department, Federal Reserve
Bank of Atlanta, 1000 Peachtree Street, N.E., Atlanta, GA 30309-4470, 404-498-8561, kristopher.gerardi@atl.frb.org; Lorenz
Goette, University of Geneva, Department of Economics, 40 Bd du Pont d’Arve, Geneva 1211, Switzerland, 617-973-2189,
lorenz.goette@unige.ch; or Paul Willen, Research Department, Federal Reserve Bank of Boston, P.O. Box 55882, Boston, MA
02205, 617-973-3149, paul.willen@bos.frb.org.
Federal Reserve Bank of Atlanta working papers, including revised versions, are available on the Atlanta Fed’s Web site at
www.frbatlanta.org. Click “Publications” and then “Working Papers.” Use the WebScriber Service (at www.frbatlanta.org) to
receive e-mail notifications about new papers.
1
Introduction
One of the most important challenges now facing U.S. policymakers stems from the tide
of foreclosures that now engulfs the country. There is no shortage of suggestions for how
to attack the problem. One of the most influential strands of thought contends that the
crisis can be attenuated by changing the terms of “unaffordable” mortgages. It is thought
that modifying mortgages is not just good for borrowers in danger of losing their homes but
also beneficial for lenders, who will recover more from modifications than they would from
foreclosures. Proponents of this view, however, worry that without government intervention,
this win-win outcome will not occur. Their concern is that the securitization of mortgages
has given rise to contract frictions that prevent lenders and their agents (loan servicers)
from carrying out modifications that would benefit both borrowers and lenders.
In this paper, we take a skeptical look at this argument. Using both a theoretical
model and some loan-level data, we investigate two economic decisions, the borrower’s
decision to default on a mortgage and the lender’s choice between offering a loan modification
and foreclosing on a delinquent loan. We first study the “affordability” of a mortgage,
typically measured by the DTI ratio, which is the size of the monthly payment relative to the
borrower’s gross income.1 We find that the DTI ratio at the time of origination is not a strong
predictor of future mortgage default. A simple theoretical model explains this result. While
a higher monthly payment makes default more likely, other factors, such as the level of house
prices, expectations of future house price growth and intertemporal variation in household
income, matter as well. Movements in all of these factors have increased the probability of
default in recent years, so a large increase in foreclosures is not surprising. Ultimately, the
importance of affordability at origination is an empirical question and the data show scant
evidence of its importance. We estimate that a 10-percentage-point increase in the DTI
ratio increases the probability of a 90-day-delinquency by 7 to 11 percent, depending on
the borrower.2 By contrast, an 1-percentage-point increase in the unemployment rate raises
this probability by 10-20 percent, while a 10-percentage-point fall in house prices raises it
by more than half.
1
DTI ratio stands for “debt-to-income” ratio. A more appropriate name for this ratio is probably
“payment-to-income” ratio, but we use the more familiar terminology. Throughout this paper, we define
DTI as the ratio of mortgage-related payments to income, rather than all debt payments; this is sometimes
called the “front end” DTI.
2
As explained below, these estimates emerge from a duration model of delinquency that are based on
instantaneous hazard rates. So, the statement that an 10-percentage-point increase in DTI increases the
probability of 90-day delinquency by 7 percent means that the DTI increase multiplies the instantaneous
delinquency hazard by 1.07, not that the DTI increase raises the probability of delinquency by 7 percentage
points.
1
The fact that origination DTI explains so few foreclosures should not surprise economists,
given the mountain of economic research on the sources and magnitude of income variation
among U.S. residents. The substantial degree of churning in the labor market, combined
with the trial-and-error path that workers typically follow to find good job matches, suggests
that income today is an imperfect predictor of income tomorrow. Consequently, a mortgage
that is affordable at origination may be substantially less so later on, and vice versa.
We then address the question of why mortgage servicers, who manage loans on behalf
of investors in mortgage-backed securities, have been unwilling to make mass loan modifications. The evidence that a foreclosure loses money for the lender seems compelling. The
servicer typically resells a foreclosed house for much less than the outstanding balance on the
mortgage, in part because borrowers who lose their homes have little incentive to maintain
them during the foreclosure process.3 This would seem to imply that the ultimate owners
of a securitized mortgage, the investors, lose money when a foreclosure occurs. Estimates of
the total gains to investors from modifying rather than foreclosing can run to $180 billion,
more than 1 percent of GDP. It is natural to wonder why investors are leaving so many
$500 bills on the sidewalk. While contract frictions are one possible explanation, another is
that the gains from loan modifications are in reality much smaller or even nonexistent from
the investor’s point of view.
We provide evidence in favor of the latter explanation. First, the typical calculation
purporting to show that an investor loses money when a foreclosure occurs does not capture
all relevant aspects of the problem. Investors also lose money when they modify mortgages
for borrowers who would have repaid anyway, especially if modifications are done en masse,
as proponents insist they should be. Moreover, the calculation ignores the possibility that
borrowers with modified loans will default again later, usually for the same reason they
defaulted in the first place. These two problems are empirically meaningful and can easily
explain why servicers eschew modification in favor of foreclosure.
Turning to the data, we find that the evidence of contract frictions is weak, at least if
these frictions result from the securitization of the loan. Securitization agreements generally
instruct the servicer to behave “as if” it owned the loan in its own portfolio, and the data
are consistent with that principle. Using a dataset that includes both securitized and nonsecuritized loans, we show that these two types of loans are modified at about the same
rate. While there is room for further empirical work on this issue, these results minimize
the likely importance of contract-related frictions in the modification decision. Even though
3
An even more important reason that lenders rarely recover the full balance of the mortgage is that the
borrower owed more on the home than the home was worth. Below, we show that negative equity is a
necessary condition for foreclosure; people rarely lose their homes when they enjoy positive equity.
2
it may be in society’s interest to make modifications (because of the large externalities from
foreclosure), it may not be in the lender’s interest to do so, whether or not this lender is an
investor in a mortgage-backed security or a portfolio lender.4
Our skepticism about the arguments discussed above is not meant to suggest that government has no role in reducing foreclosures. Nor are we arguing that the crisis is completely
unrelated to looser lending standards, which saddled borrowers with high-DTI mortgages,
or interest rates that reset to higher levels a few years into the loans.5 Rather, we argue
that foreclosure-prevention policy should focus on the most important source of defaults. In
the data, this source appears to be the interaction of falling prices and adverse life events,
not mortgages with high-DTI ratios or otherwise relaxed risk characteristics.
The remainder of this paper is organized as follows. Section 2 outlines a simple model
of the default decision that helps organize ideas about potential sources of the foreclosure
crisis. Section 3 shows that, as would be implied by the simple model, the affordability of a
mortgage at origination as measured by DTI is not a strong predictor of mortgage default,
especially compared with other variables that reflect income volatility and falling house
prices in a fundamental way. Section 4 adapts the model to encompass the decision of the
lender to offer a modification, and then provides evidence that securitization contracts are
not unduly preventing modifications. Section 5 concludes with some lessons for foreclosurereduction policy that are suggested by our results.
2
Affordability and Foreclosure: Theory
One of the most commonly cited causes of the current foreclosure crisis is the mass
origination of unaffordable or unsustainable mortgages. Ellen Harnick, the senior policy
counsel for the Center for Responsible Lending, characterized the crisis this way when she
recently testified before Congress:
The flood of foreclosures we see today goes beyond the typical foreclosures of
years past, which were precipitated by catastrophic and unforeseen events such as
job loss, divorce, illness, or death. The current crisis originated in losses triggered
by the unsustainability of the mortgage itself, even without any changes in the
families’ situation, and even where the family qualified for, but was not offered,
4
A foreclosure imposes externalities on society when, for example, a deteriorating foreclosed home drives
down house prices for the entire surrounding neighborhood.
5
For a discussion of the role of looser lending standards, see Mian and Sufi (2009) and Dell’Ariccia, Igan,
and Laeven (2009).
3
a loan that would have been sustainable.6
The claim that the foreclosure crisis results from unaffordable or unsustainable loans has
been endorsed by a number of influential policy analysts.7 But the concept of “unaffordability” is rarely defined precisely. To economists, something is unaffordable if it is unattainable
under any circumstances, even temporarily. For example, an economist might say: “For me,
the penthouse apartment at the Time Warner Center in New York is unaffordable ($50
million when finished in 2004).” But a non-economist might say, “For me, the dry-aged
ribeye at Whole Foods ($19.99 a pound) is unaffordable.” The problem is that, for most
Americans, a regular diet of ribeye steaks is attainable; a consumption bundle that includes
two pounds of ribeye every night is not impossible for most families. They do not choose this
bundle because of relative prices: the tradeoff between the ribeye and other consumption
is unappealing (for example, the family might prefer a new car). In this case, economists,
if they were being precise, would say that the ribeye was “affordable” but “too expensive.”
Along the same lines, economists might argue that an unaffordable mortgage is one that
is really too expensive, in the sense that the benefits that come with making payments on
the mortgage no longer outweigh the opportunity costs of doing so. In the next subsection,
we build a simple model of these benefits and costs in order to evaluate what makes a borrower decide that a mortgage is unaffordable and thus to default on it. In describing this
model, we will use the common usage definition of “affordable,” though we really mean “too
expensive.”
2.1
A simple model
Assume a two-period world (t = 1, 2), with two possible future states, good and bad.
The good state occurs with probability αG , while the bad state occurs with probability αB
(where αB = 1 − αG ). In the first period, the value of the home is P1 with a nominal
mortgage balance of M1 . In this period, the borrower decides between making the mortgage
payment, a fraction m of the mortgage balance M1 , and staying in the home, or stopping
payment and defaulting. Because this is a two-period model, we assume that in the second
6
Harnick (2009), p. 5.
A recent report from the Congressional Oversight Panel of the Troubled Asset Recovery Program (hereafter denoted COP) states that “[t]he underlying problem in the foreclosure crisis is that many Americans
have unaffordable mortgages” (COP report, p. 16). The report adds that the unaffordability problem arises
from five major factors: (1) the fact that many mortgages were designed to be refinanced and cannot be
repaid on their original terms, (2) the extension of credit to less creditworthy borrowers for whom homeownership was inappropriate, (3) fraud on the part of brokers, lenders, and borrowers, (4) the steering of
borrowers who could qualify for lower cost mortgages into higher priced (typically subprime) mortgages,
and (5) the recent economic recession.
7
4
period the borrower either sells the home or defaults on the mortgage. If the good state
occurs, the price of the house in the second period is P2G , while if the bad state occurs, the
price is P2B . We will assume that P2B < M2 , where M2 is the remaining nominal mortgage
balance in the second period.
The first key insight of the model is that if equity is positive, the borrower will never
default on the house. Selling dominates foreclosure when equity is positive because the
borrower has to move out either way and the former strategy yields cash while the latter
does not. Exactly what constitutes positive equity is a bit tricky empirically. Borrowers
have to pay closing costs to sell the house and may be forced to accept a lower price if they
sell in a hurry. Thus, the balance of the mortgage may be slightly less than the nominal
value of the home, but with these extra expenses factored into the equation, the borrower
may not have positive equity to extract.
The empirical evidence on the role of negative equity in causing foreclosures is overwhelming and incontrovertible. Household-level studies show that the foreclosure hazard
for homeowners with positive equity is extremely small but rises rapidly as equity approaches
and falls below zero. This estimated relationship holds both over time and across localities,
as well as within localities and time-periods, suggesting that it cannot result from the effect
of foreclosures on local-level house prices.8
Because default does not occur if P1 ≥ M1 , we focus on the case where M1 > P1 . The
decision for the borrower is whether or not to make the periodic mortgage payment mM1 .
The cost of making the payment is the payment amount, net of the rent that the borrower
would have to pay for shelter in the event of default. The benefit to the borrower includes
the option in the next period to sell the house at a profit in the good state where P2 > M2 ,
or the option to default in the bad state and lose nothing. We assume that the decision to
default costs the borrower some amount Λ next period, which can be interpreted as some
combination of guilt, shame, and reduced access to future credit. Under these conditions,
we can collapse the default decision into the following inequality:9
αG (P2G − M2 ) + Λ
Default ⇔
> 1 + r.
mM1 − rent1
(1)
The basic point here is that a borrower views the mortgage payment (or more precisely the
excess of the mortgage payment over his rent) as an investment in a security that pays off
in the next period as long as the value of the house exceeds the strike price, which is the
8
See Sherlund (2008), Danis and Pennington-Cross (2005), and Deng, Quigley, and Order (2000) for
default regressions. See Gerardi, Shapiro, and Willen (2007) for an exhaustive discussion of the identification
issues in the study of house prices and foreclosure.
9
For details on a very similar model, see Foote, Gerardi, and Willen (2008).
5
outstanding balance on the mortgage. If the return on the investment exceeds the alternative
investment, here assumed to be the riskless rate, then the borrower stays in the home. If
instead the return falls short, then the borrower decides that the riskless asset is a better
investment and defaults.
Thus far, income appears to play no role in the default decision. In this sense, our
model follows the traditional option-theoretic analyses of the mortgage default decision, in
which the mortgage is viewed as a security priced by arbitrage, and household income is
irrelevant.10
The problem with the model described above is that it gives no role to individual heterogeneity, except potentially through differences in Λ. According to the model, all borrowers
living in similar houses with similar mortgages should default at roughly the same time.
Yet, in the data, we observe enormous heterogeneity in default behavior across otherwise
similar households. Moreover, there is a pattern to this heterogeneity: households that
suffer income disruptions default much more often than households that do not; younger
homeowners default more often; and households with few financial resources default more
often.
To address these limits, we make two small changes to the model. If we assume that
housing is a normal good, households that suffer permanent reductions in income will prefer
less housing, and thus their alternative rent payment will fall. So we allow rent to vary
by individual household, denoting it renti . But, more significantly, we introduce borrowing
constraints. Borrowing constraints mean that the relevant interest rate is no longer “the”
riskless rate but the household’s shadow riskless rate. Under the assumption of log utility
and exponential discounting, this rate equals:
1 + ri = (1 + δi )
−1
−1
ci,1
E
ci,2
where ci,t is consumption of household i at time t and δi is a household-specific discount
rate. Then we can re-write equation (1) as:
Default ⇔
αG (P2G − M2 ) + Λ
> 1 + ri .
mM1 − renti
(2)
This model can shed light on the question of what really constitutes an unaffordable
mortgage. A mortgage is unaffordable if the marginal rate of transformation between current and future consumption implied by the mortgage falls short of the marginal rate of
10
See Kau, Keenan, and Kim (1994), for example.
6
substitution. What makes a mortgage “unaffordable,” that is, too expensive?
1. Low house price appreciation. A higher probability of price appreciation (higher
αG ) increases the expected return to staying in the house. In this sense, our treatment
is similar to the standard user cost calculation in the literature, whereby increased
house price appreciation lowers the cost of owning a home.11
2. High monthly payments. All else equal, higher m makes the mortgage less attractive. This is consistent with the views expressed in the quote that opened in
this section: Many families, for one reason or another, took on mortgages with high
payments that are likely to dissuade them from keeping their mortgage current. Typically, the burden of a mortgage’s payments at origination is measured by the DTI
ratio. Thus, analysts who believe that this type of unaffordability is at the heart of
the crisis often support proposals designed to lower DTI ratios on a long-term basis.
3. Permanent and transitory shocks to income. Permanent shocks lower renti .
Also, if the borrower is constrained, then a transitory shock that leads to a lower
level of income will lead to high consumption growth and thus a high shadow riskless
rate, which makes staying less attractive. The quote that opens this chapter expresses
the view that income shocks were important drivers of foreclosure in the past, but that
these shocks are less important today. However, if income shocks are in fact the most
important source of distress in the housing market, then a policy that grants troubled
borrowers substantial but temporary assistance is warranted. Temporary assistance
may not help borrowers facing permanent income shocks, but it would help borrowers
undergoing transitory setbacks.
4. Low financial wealth. A borrower with little financial wealth is more likely to be
constrained and thus more likely to have a high shadow riskless rate.
2.2
Monthly payments, income, and affordability
Once we recognize the role that unforecastable income shocks can play in foreclosure,
we can further divide the concept of affordability into what we will call ex ante and ex
post affordability. A loan is ex post unaffordable if the borrower decides to default on it.
A loan is ex ante unaffordable if the probability that it will become ex post unaffordable
exceeds some threshold. To decide whether a loan is ex ante affordable, an underwriter or
policymaker needs to forecast the evolution of stochastic variables like income, payments,
11
See Poterba (1984) and, more recently, Himmelberg, Mayer, and Sinai (2005).
7
and house prices, and then choose some threshold probability of ex post unaffordability.
In this section, to clearly convey our points, we consider an extreme model, in which ex
post affordability depends entirely on the ratio of monthly payments to income, the DTI
ratio. Thus, our forecasting model will involve only the required monthly payment and the
borrower’s income.
To forecast income, we follow the macro literature and assume that changes to the
logarithm of a borrower’s labor income yt consist of a predictable drift term αt , a transitory
(and idiosyncratic) shock εt , and a permanent shock ηt :
yt = αt + yt−1 + εt + ηt .
We use estimates from Gourinchas and Parker (2002) for the process for the “average person”
in their sample and assume that the borrower is 30 years old.
For the monthly payments, we assume that either they are constant, or they follow
the typical path of a 2/28 adjustable-rate mortgage (ARM). A 2/28 ARM is a common
subprime mortgage that has a fixed payment for the first two years, after which the payment
is determined by the so-called fully indexed rate, typically hundreds of basis points over the
six-month London interbank offered rate (Libor).12 We assume that the initial rate is 8.5
percent (the average initial rate for a sample of 2/28 ARMs originated in 2005) and that the
first adjustment occurred in 2007, when the six-month Libor was 5.25 percent. A spread
over Libor of 600 basis points was typical during this period and would imply a fully indexed
rate of 11.25 percent, which generates a payment increase of roughly one-third. We focus
on the 2/28 ARMs because they were, by far, the most common type of subprime loan and
have accounted for a hugely disproportionate share of delinquencies and foreclosures in the
last two years. Other loans, like option ARMs, allow for negative amortization and have far
higher payment shocks at reset, but were rarely marketed to subprime borrowers, and thus,
have not accounted for a large share of problem loans so far.
Table 1 shows some basic results. The first key finding is that the threshold for ex post
affordability must be much higher than the threshold for ex ante affordability. If one sets
them equal, then about 70 percent of borrowers will end up with unaffordable mortgages
at some point in the first three years, even without resets. This is important because it
means that one cannot decide on ex ante affordability by using some a priori idea of what
is a reasonable amount to spend on housing. In other words, if spending one third of one’s
income on housing is considered too much (as low-income housing studies often claim), then
one has to set the ex ante criterion well below 33 percent of income.
12
This spread is determined by the risk characteristics of the borrower.
8
The second finding is that resets are of only limited importance. Many commentators
have put the resets at the heart of the crisis, but the simulations illustrate that it is difficult
to support this claim. The payment escalation story is relevant if we assume that there is
no income risk and that the initial DTI is also the threshold for ex post DTI. Then loans
with resets become unaffordable 100 percent of the time and loans without resets never
become unaffordable. But adding income risk essentially ruins this story. If the initial DTI
is also the threshold for ex post DTI, then, with income risk, about 70 percent of the loans
will become unaffordable even without the reset. The reset only raises that figure to about
80 percent. If, on the other hand, we set the ex post affordability threshold well above the
initial DTI, then the resets are not large enough to cause ex post affordability problems.
The only scenario in which the reset makes a significant, quantitative impact is when we set
the initial DTI very low and the threshold for ex post affordability very high. In this case
the likelihood of default roughly doubles with resets.
The third finding is that setting the right initial DTI can help reduce foreclosures if the
ex post affordability criterion is sufficiently high, but this finding is very sensitive to the
assumption about income volatility. The first column of Panel C shows that if the ex post
criterion is 50 percent, then loans with 31 percent DTI at origination become unaffordable
only about 16 percent of the time, whereas those with 50 percent DTI do so roughly 70
percent of the time. The problem here is that the troubled borrowers who obtain subprime
loans or who need help right now are unlikely to have the baseline parameters from Gourinchas and Parker (2000). If we assume that they have a standard deviation of transitory
shocks twice as large as average, then column 4 shows that the benefits of low DTI are
much smaller.Going from 38 percent DTI to 31 percent DTI only lowers the number of
borrowers who will face ex post unaffordability by 30 percent from 54 percent to 38 percent.
Put another way, if our goal is “sustainable” mortgages, neither 31 percent nor 38 percent
would fit that definition.
3
Affordability and Foreclosure: Evidence
In this section we perform an empirical analysis of the potential determinants of default
identified in the previous section, including falling house prices, labor income shocks, and
high DTI ratios. Because a loan that is prepaid is no longer at risk of default, we also
investigate prepayments in a competing risks framework.
9
3.1
Data
The data used in this paper come from loan-level records, compiled by LPS Applied
Analytics, Inc., from large loan-servicing organizations.13 This dataset has fields for key
variables set at the time of each loan’s origination, including the amount of the loan, the
appraised value and location of the property that secures the loan, whether the loan is
classified as prime or subprime, whether the loan is a first or second lien, and whether the
loan is held in portfolio or has been packaged into a mortgage-backed security (MBS). We
can also observe a host of interest-rate variables, such as whether the loan is fixed-rate
or adjustable-rate and the manner in which the interest rate changes in the latter case.
Additionally, the performance of each loan can be monitored over time. For each month
in which a given loan is in the data, we know its outstanding balance, the current interest
rate, and the borrower’s payment status (that is, current, 30-, 60-, or 90-days delinquent, in
foreclosure, etc.). We also know whether a loan ended in payment, prepayment, or default.
As of December 2008, the LPS dataset covered nearly 60 percent of active residential
mortgages in the United States, representing about 29 million loans with a total outstanding
balance of nearly $6.5 trillion.14 Nine of the top 10 servicers in the U.S. are present in our
data, including Bank of America/Countrywide and Wells Fargo. Cordell, Watson, and
Thomson (2008) write that because the LPS data come from large servicers (who now
dominate the servicing market), the unconditional credit quality of the average loan in
the LPS data is probably lower than that of a randomly sampled U.S. mortgage, because
smaller servicers are more prevalent in the prime market. However, when assessing the
representativeness of the LPS data, it is important to note that we can tell whether a
loan in the data is prime or subprime.15 Additionally, we usually have access to other
variables reflecting risk, including the borrower’s credit (that is, FICO) score, loan-to-value
at origination, etc. This allows us to condition on several factors affecting loan quality.
One of the strengths of the LPS dataset is that it is one of the few loan-level databases
that include both conforming prime loans and subprime loans. Table 2 lists the numbers
of prime and subprime loans in the data, disaggregated by the investors for whom the
servicers are processing payments and the seniority of the mortgage (first lien, second lien,
13
The dataset was originally created by a company called McDash Analytics; LPS acquired McDash
in mid-2008. Among housing researchers, the dataset is still generally called the “McDash data.” The
description of the LPS dataset in this section draws heavily from Cordell, Watson, and Thomson (2008).
The dataset was purchased in late 2008 by a consortium that included the Board of Governors of the Federal
Reserve System and eight regional Federal Reserve Banks.
14
Because of the size of the data (about 600 gigabytes), we never took possession of it when performing
our analysis. Instead we downloaded random samples of various size from the servers of the Federal Reserve
Bank of Kansas City.
15
Subprime loans are defined by the servicers themselves as loans with a grade of either “B” or “C.”
10
etc.). About 33 percent of the mortgages in the dataset are held in the securities of Fannie
Mae, with another 22 percent held in Freddie Mac securities. Around 18 percent of the
loans are held in “private securitized” pools; these are the loans that are also covered by
the well-known LoanPerformance dataset.16 A little less than 10 percent of the loans in the
LPS data are held in the portfolio of the servicer itself.
While the LPS dataset now covers more than half of the U.S. mortgage market, coverage
was not as extensive in earlier years. The LPS dataset has grown over time as new servicers
have been added, with a substantial spread in coverage of the market in 2005 (when most of
our samples begin). Whenever a new servicer is added to the dataset, that servicer’s existing
portfolio is incorporated into the dataset. Future loans from that servicer are added a month
or two after the loans close. This pattern has the potential to introduce unrepresentative
loans into the data, because loans that stay active for many years (and thus are likely to be
added when their servicers enter the LPS data) are a nonrandom sample of all loans. One
way to ameliorate potential problems of left-censoring is to analyze only those loans that
enter the data within the year that the loans were originated.17 A separate issue is the fact
that not all servicers collected the exact same variables, so the preponderance of missing
data changes over time. Unfortunately, DTI is recorded for only about half the loans in the
sample, as shown in Table 3. On one hand, this is disheartening, because an analysis of
DTI is a prime goal of this section. On the other hand, the sample is sufficiently large that
we do not want for observations. Moreover, the fact that DTI is so spottily recorded —
especially in comparison to the FICO score — indicates that investors and servicers place
little weight on it when valuing loans. This is, of course, what the model of section 2 would
predict. A final concern about the LPS data is that we do not know whether there are other
loans on the property that secures any given loan. Thus, given some path of local house
prices, we are able to construct an ongoing loan-to-value ratio for any loan in the dataset,
but we cannot construct a combined loan-to-value ratio for the borrower on that loan. We
are therefore unable to calculate precise estimates of total home equity.18
16
The dataset from LoanPerformance FirstAmerican Corp. includes loans that were securitized outside
of the government-sponsored agencies, Fannie Mae and Freddie Mac. It therefore includes loans that are
subprime, Alt-A, and non-conforming (that is, jumbo loans). The coverage of private securitized loans is
broader in the LoanPerformance data than it is in LPS, as LoanPerformance has about 90 percent of the
private-label market.
17
Most loans in our sample were included in the data one or two months after origination.
18
For a borrower with only one mortgage, the loan-to-value ratio on his single mortgage will, of course,
be his total loan-to-value figure. However, we are unable to know whether any particular borrower in the
data has more than one mortgage.
11
3.2
Affordability and origination DTI: Results from duration models
To learn how different risk characteristics and macroeconomic variables affect loan outcomes, we run Cox proportional hazard models for both defaults and prepayments.19 In
this context, the proportional hazard model assumes that there are common baseline hazard
functions that are shared by all loans in the data. The model allows for regressors that can
shift this hazard up or down in a multiplicative fashion. The specific type of proportional
hazard model that we estimate, the Cox model, makes no assumption about the functional
form of the baseline hazard. Rather, the Cox model essentially “backs out” the baseline
hazard after taking account of the effects of covariates. The baseline hazards for both potential outcomes (default and prepayment) are likely to be different across the two types
of loans (prime and subprime), so we estimate four separate Cox models in all. We define
default as the loan’s first 90-day delinquency, and our main estimation period runs from
2005 through 2008. In this section, we use a random 5 percent sample of the LPS data.
The results of these models should not be interpreted as causal effects. If we see that
borrowers with low loan-to-value ratios (LTVs) default less often (and we will), we cannot tell
whether this arises because of something about the loan or something about the borrowers
likely to choose low-LTV mortgages. Even so, a finding that DTI at origination is not a
very strong predictor of default would undermine the claim that unaffordable mortgages are
a more important cause of default than income shocks and falling prices.
Table 4 presents summary statistics of the loan-level characteristics that are included
in the proportional hazard models. The average DTI at origination for prime loans in our
sample is 35.1 percent, while the mean DTI for subprime loans is about 5 percentage points
higher. Subprime loans also have generally higher LTVs and lower FICO scores. Figure
1 provides some additional detail about these risk characteristics by presenting the entire
distributions of DTIs, LTVs, and FICO scores. While the distribution of prime DTIs is
somewhat symmetric, the distribution of DTIs for subprime loans is strongly skewed, with
a peak near 50 percent. Another interesting feature of the data emerges in the bottom row
of panels, which presents LTVs. For both prime and subprime loans, the modal LTV is 80
percent, with additional bunching at multiplies of five lying between 80 and 100. Recall that
in the LPS data, an LTV of 80 percent does not necessarily correspond to 20 percent equity.
This is because the borrower may have used a second mortgage to purchase the home (or
may have taken out a second mortgage as part of a refinance). Unfortunately, there is no
way to match loans to the same borrower in the LPS dataset, nor is there a flag to denote
19
For details about hazard models, see Kiefer (1988).
12
whether any given loan is the only lien on the property. The large number of 80-percent
LTVs, however, strongly suggests that these loans were accompanied by second mortgages.
Thus, in our empirical analysis, we include a dummy variable that denotes whether the
particular loan has an LTV of exactly 80 percent.20
In addition to loan-specific characteristics, the Cox models also include the cumulative
changes in statewide house prices and county-level unemployment rates that have occurred
since the loan was originated.21 Figures 2 and 3 present the distributions for these data;
unlike the figures for DTI, FICO, and LTV, each loan in the sample contributes a number
of monthly observations to each of these two figures. Figure 2 shows that the distribution
of price changes is skewed toward positive changes. In part, this reflects the large number
of loans originated in the early years of the sample (2005–2006), when house prices were
rising. In our empirical work we allow positive price changes to have different effects than
negative price changes.22
Finally, we also include a number of interactions among risk characteristics and macro
variables. These interactions play an important role, given the strong functional form assumption embedded in the proportional hazard model. Denote h(t|xj ) as the hazard rate for
either a default or a prepayment, conditional on a vector of covariates xj . The proportional
hazard assumption is
h(t|xj ) = h0 (t) exp(xj βx ),
where h0 (t) is the shared baseline hazard and βx represent coefficient estimates. Because
exp(β1 x1 + β2 x2 ) equals exp(β1 x1 ) exp(β2 x2 ), there is in a sense a multiplicative interaction
“built in” to the proportional hazard assumption. Entering various interactions directly
ensures that interactions implied by the estimated model are not simply consequences of the
functional form assumption. Of course, as with any regression, the presence of interactions
makes interpretation of the level coefficients more difficult, because the level coefficients will
now measure marginal effects at zero values of the other variables. Hence, we subtract 80
from the loan’s LTV before entering this variable in the regressions. In this way, a value of
zero in the transformed variable will correspond to the most common value of LTV in the
data. We transform DTI by subtracting 35 for prime loans and 40 for subprime loans, and
20
For ease of interpretation, we define this variable to equal one if the borrower does not have an LTV of
80 percent.
21
Obviously, county-level house prices would be preferable to state-level prices, but high-quality, disaggregated data on house prices are not widely available. Our state-level house prices come from the Federal
Housing Financing Authority (formerly the OFHEO price index).
22
Because of the importance of negative equity in default, the difference between a price increase of 10
percent and an increase of 20 percent may be much less consequential for a loan’s outcome than whether
the house price declines by 10 or 20 percent. However, recall that we cannot figure total equity in the house,
because we do not observe all mortgages.
13
we transform FICO by subtracting 700 for prime loans and 600 for subprime loans.
Figure 4 graphs the baseline default hazards for both prime and subprime loans. The
subprime default hazard (dotted line) is much higher than the hazard for prime loans (note
the different vertical scales on the figure). There is an increase in the subprime default
hazard shortly after 24 months, a time when many loans reset to a higher interest rate.
At first blush, this feature of the subprime default hazard would appear to lend support
to oft-made claims that unaffordable resets caused the subprime crisis. Recall, however,
that a hazard rate measures the instantaneous probability of an event occurring at time t
among all subjects in the risk pool at time t − 1. While the default hazard shows that the
default probability rises shortly after 24 months, the subprime prepayment hazard, graphed
in Figure 5, shows that prepayments also spiked at the same time. The surge in prepayments
means that the relevant pool of at-risk mortgages is shrinking, so that the absolute number
of subprime mortgages that default shortly after the reset is rising to a much smaller extent
than the hazard rate seems to imply. Thus, our results are not inconsistent with other
research that shows that most subprime borrowers who defaulted did so well before their
reset date.23
Table 5 presents the coefficients from the Cox models. The model for prime defaults
(first column) generates a significantly positive coefficient for the DTI ratio: .0105, with a
state-clustered standard error of .0009. When working with proportional hazard models, it
is common to report results in terms of “hazard ratios,” exp(βj ), the multiplicative shift
in the baseline hazard engendered by a unit change in the regressor of interest. The DTI
coefficient in the prime default regression generates a hazard ratio of exp(0.0105) ≈ 1.0105,
indicating that a one-percentage-point increase in DTI shifts the default hazard up by 1.05
percent.24 While statistically significant, the effect is small as a practical matter. Recall that
Table 4 showed that the standard deviation of DTI in the prime sample is 13.8 percentage
points, so a one-standard-deviation increase in DTI for prime borrowers results in a hazard
ratio of exp(13.8 · 0.0105) ≈ 1.156. This effect can be compared to the effect of decreasing a
borrower’s FICO score by one standard deviation. The FICO coefficient in the first column
(–.0124) has about the same absolute value as the DTI coefficient, but the standard deviation
in FICO scores is much greater (61.6 points). Thus, a one-standard-deviation drop in the
FICO score results in a hazard ratio of exp(−61.6 · −0.0124) ≈ 2.147.
Other coefficients in the first column also have reasonable signs and magnitudes. More
defaults are to be expected among loans with high LTVs as well as loans with LTVs that
23
See Sherlund (2008), Mayer and Pence (2008), and Foote, Gerardi, Goette, and Willen (2008).
Because of the way we transformed our variables, this marginal effect corresponds to a prime borrower
with a 700 FICO score, a DTI of 35, and an LTV of 80 percent.
24
14
are exactly 80 percent (and which thus suggest the presence of a second mortgage). The
unemployment rate enters the regression with a large coefficient (.2068), so that a onepercentage-point increase in the unemployment rate results in a hazard ratio of about 1.23.
House-price changes also enter significantly, though there is little evidence for different
coefficients based on the direction of the price change (both the positive-change and negativechange coefficients are close to –.058).25 These estimates indicate that a 10-percentage-point
increase in housing prices shifts the hazard down by about 44 percent. When evaluating the
effect of these macroeconomic coefficients on defaults, it is important to recall the earlier
qualifications about identification. An exogenous increase in delinquencies may increase
housing-related unemployment and cause housing prices to fall. Nevertheless, it is gratifying
to see that the results of the model are consistent with other work that shows a direct
causal effect of prices on default in ways that are immune to the reverse-causation argument
(Gerardi, Shapiro, and Willen (2007)).
The second column of the table presents the estimates from the subprime default model.26
As in the prime column, all of the individual-level risk characteristics enter the model
significantly. And, as before, movements in FICO scores have a more potent effect on default
than movements in DTI, though the difference is not as extreme. For subprime borrowers, a
one standard-deviation increase in DTI results in a hazard ratio of exp(.0072∗11.1) = 1.083.
This percentage change is smaller than the corresponding shift for prime mortgages, but
recall that the baseline default hazard for subprime mortgages is also much higher. In any
case, for subprime loans, the effect of raising DTI by one standard deviation is still smaller
than the effect of lowering FICO by one standard deviation, shifting the baseline hazard up
by about 21 percent rather than 8.3 percent.
Two additional results from the default regressions are consistent with the idea that
idiosyncratic income risk is an important determinant of mortgage outcomes. First, among
subprime borrowers, the effect of DTI on the likelihood of default is smaller for borrowers
with high FICO scores. The coefficient on the interaction of FICO and DTI in the second
column is significantly negative (–.000055, with a standard error of .000017). Thus, for a
subprime borrower with a 700 FICO score, the total marginal effect of an increase in DTI
on his default probability is only .0017, an effect that is insignificantly different from zero.27
The fact that high-FICO borrowers in the subprime pool are better able to tolerate high
25
Negative price changes are entered as a negative numbers, not as absolute values.
The level coefficients for LTV, FICO, and DTI now correspond to marginal effects for a subprime
borrower with a 600 FICO score, an LTV of 80 percent, and a 40 percent DTI.
27
To see this, note that a 700 FICO score corresponds to a score of 100 in our transformed FICO metric
for subprime borrowers. Thus, the relevant DTI coefficient for a 700-FICO borrower is the level coefficient
on DTI (.0072) plus 100 times the interaction of DTI and FICO (–.000055). This sum approximately equals
.0017.
26
15
DTIs suggests that these borrowers may have been able to make good predictions of their
future incomes and of the likely variation in these incomes. These borrowers may have
desired high-DTI mortgages that were unattractive to prime lenders, so they entered the
subprime pool. A second set of results pointing to the importance of income volatility is the
coefficients on the unemployment–FICO interactions. These coefficients are significantly
negative in both the prime and subprime regressions, indicating that the ARMs of high
FICO borrowers are generally hurt more severely, in percentage terms, by increases in the
aggregate unemployment rate. If idiosyncratic income variation among high-FICO borrowers is relatively low, then it is perhaps not surprising that their mortgages are relatively
more sensitive to aggregate fluctuations.
Results from the prepayment regressions are presented in the third and fourth columns
of Table 5. Prime borrowers tend to refinance somewhat more quickly out of high-DTI
mortgages, while DTI has an insignificant effect on subprime prepayment. Of particular note
in both regressions is the strong effect that house prices have on prepayment. The coefficients
on all price terms are positive, indicating that higher prices encourage prepayment and lower
prices reduce it. The effect of price declines on subprime refinancing is particularly strong.
Figure 6 puts the pieces together by simulating the number of monthly defaults under
various assumptions about loan characteristics, house prices, and unemployment. To do this,
we first shift the baseline hazards for both default and prepayment to be consistent with the
assumptions and the coefficient estimates from the model. We then calculate what these
adjusted hazards would imply for the size of an initial risk set of 100 loans.28 Multiplying
the risk set in a given month times the hazard of either defaults or prepayments gives the
total number of the 100 original loans that are expected to default or prepay in that month.
Panel A of Figure 6 presents the data for prime defaults. The solid line assumes a baseline
case of no changes in house prices or unemployment along with the baseline DTI value (35
percent for prime loans). The dashed line just above it assumes that DTI is 45 rather than
35. As one would expect from the modest size of the coefficient in the first column of Table
5, increasing DTI has a modest effect on monthly defaults. The next lines return DTI to 35
but either raise the unemployment rate by 2 percentage points or reduce housing prices by
10 percent. These assumptions have a much larger positive effect on prime defaults than the
assumption of higher DTI. Falling house prices also strongly discourage prime prepayments,
as shown in Panel B.
The bottom two panels of Figure 6 present the results for subprime loans. In Panel C, we
28
For example, if both the default and prepayment hazards have been adjusted upwards by the implied
assumptions on covariates and coefficient estimates, then the risk set will be whittled more quickly away by
defaults and prepayments.
16
see a small uptick in defaults between 24 and 30 months, presumably due to the interest-rate
resets on subprime 2/28 mortgages. This increase, however, is smaller than the bulge in the
baseline hazard at about this time, because the risk set has been significantly reduced by
prepayments. Panel C also shows the nearly imperceptible effect of higher DTI. Here, the
experiment is raising DTI from the baseline subprime value of 40 percent to 50 percent. As
with prime defaults, the effect of this increase is small relative to the effect of unemployment
and house prices. Finally, Panel D shows that falling house prices have particularly severe
effects on the prepayments of subprime loans.
The patterns displayed in Figure 6 are consistent with a large role for income volatility
in mortgage defaults discussed in section 2. Higher unemployment rates increase defaults,
as more people are likely to lose jobs and become liquidity constrained during recessions.
Falling housing prices also raise defaults, because they increase the likelihood that a homeowner who receives a negative income shock will also have negative equity, and will thus
be unable to sell his home for enough to repay the mortgage. This interaction of income
shocks and falling prices is sometimes called the “double-trigger” model of default, because
it claims that defaults occur when two things happen at the same time: the borrower suffers
some adverse life event while he also has negative equity in his home.
3.3
Affordability and falling prices: Quantifying “walk-away” defaults
The previous subsection showed that high levels of origination DTI are not predictive
of high default rates, especially in comparison to variables like FICO scores and features
of the macroeconomic environment like falling house prices and rising unemployment. Our
preferred interpretation of this pattern is that falling prices lead to negative equity, which
can lead to default and foreclosure when a borrower receives a large negative income shock.
However, as the model of section 2 shows, housing prices have a direct effect on the affordability of a home that does not involve income volatility. A lower probability of future price
appreciation (lower αG ) raises the user cost of owning a home and makes default more likely.
If there is no hope that the price of the house will ever recover to exceed the outstanding
balance on the mortgage, the borrower may engage in “ruthless default” and simply walk
away from the home. Kau, Keenan, and Kim (1994) show that optimal ruthless default
takes place at a negative-equity threshold that is well below zero, due to the option value
of waiting to see whether the house price recovers.29 Once the default threshold has been
29
The presence of this option value explains why negative equity is a necessary but not sufficient condition
for default.
17
reached, however, default remains optimal if no new information arrives.
Of course, we cannot observe the expectations of individual homeowners to see whether
their defaults coincide with extremely gloomy forecasts of future house prices. However, we
can exploit a particular feature of the ruthless default model to get a rough upper bound
on how many people are walking away from their homes. If the ruthless default model is
a good characterization of the data, then delinquent borrowers should simply stop making
payments, never to resume again. There is no reason for a ruthless defaulter to change
his mind and start making payments once more (unless his expectation of future house
prices suddenly improves). On the other hand, if income volatility is interacting with falling
prices to produce double-trigger defaults, then we should see delinquent borrowers cycling
through various stages of delinquency as various shocks to their incomes are realized and
they struggle to keep their homes. In the LPS data, we observe each borrower’s monthly
delinquency status so we can compare the number of “direct defaults” to the number of
“protracted defaults.” The fraction of 90-day delinquencies that arise via direct defaults
will be an upper bound on the importance of walk-away defaults, because some people
may have suffered particularly severe declines in income and had to stop making payments
abruptly, even though they wanted to keep their homes.
To set the stage for this analysis, we first present so-called “roll rates,” which measure the
likelihood that a borrower in one stage of delinquency will transition into another. Figure 7
graphs these rates for borrowers who start a month in different delinquency stages.30 Panel
A considers people who begin a month in current status. Since January 2001, about 1 to
2 percent of current borrowers have become 30 days delinquent each month. Interestingly,
the number of people rolling from current to 30 days delinquent has only recently exceeded
the levels of the 2001 recession, even though foreclosures have been far higher than they
were then. Another interesting pattern in this panel is that the current-to-30-day roll rate
was low in 2004 and 2005, when many supposedly unaffordable mortgages were originated.
Panel B considers borrowers who begin the month 30 days late. A fairly constant 40 percent
of these borrowers make their next payment to remain 30 days late the next month. Until
2007, about 40 percent of borrowers who were 30 days late made two payments to become
current again, with the remaining 20 percent failing to make a payment at all and thereby
becoming 60 days late. In the past few months, however, more persons who were 30 days
late are rolling into 60-day status, considered the start of serious delinquency. Panel C shows
that the fraction of 60-day delinquencies that roll into 90-day status has risen sharply over
the past two years, with corresponding declines in the fractions of borrowers making two or
30
As was the case with the duration models, the roll rates are based on a random 5 percent sample of the
LPS data.
18
three payments. Yet the fraction of 60-day delinquencies making one payment to remain 60
days late has remained fairly constant. Finally, Panel D analyzes borrowers who begin the
month 90 days late. This is a somewhat absorbing state, because there is no formal 120-day
status.
The main takeaway from Figure 7 is that many people who are delinquent have no
desire to stay that way. Many people who are seriously delinquent come up with two or
three payments in an attempt to climb out of the status, or manage one payment so as not
to slide further down. Still, these graphs do not answer the precise question of how many
people who become 90 days delinquent simply stopped making payments. We define this
type of direct default as a 90-day delinquency that satisfies three requirements:
• The borrower is current for three consecutive months, then registers a 30-day, a 60-day,
and a 90-day delinquency in succession during the next three months;
• The borrower had never been seriously delinquent before this six-month stretch;
• The borrower never becomes current or rolls down to 30-day or 60-day status after
this stretch.
Panel A of Table 6 lists the fraction of direct defaults for the entire United States,
starting in 2003. These rates differ by the year that the mortgage is originated and the
year in which the default occurred. Among all 2003–2008 mortgages that defaulted in 2008,
fewer than half, 41.6 percent, were direct defaults. This percentage was higher for loans
made at the height of the housing boom, as 44.6 percent of 2005 mortgages defaulting in
2008 were direct defaults. This is consistent with the idea that mortgages likely to have
the largest amounts of negative equity are the most likely to ruthlessly default. But among
these mortgages, fewer than half simply stopped making payments, and even this fraction is
an upper bound on the true fraction of ruthless defaults.31 Panel B Table 6 uses data from
four states that have had particularly severe price declines and thus are more likely to have
ruthless defaulters.32 As we would expect, the share of direct defaulters is higher in these
states, reaching 55.1 percent in 2008. The 2008 fraction of direct defaults in the remaining
47 states (including DC) is less than one-third, as seen in Panel C.
To sum up, falling house prices are no doubt causing some people to ruthlessly default.
But the data indicate that ruthless defaults are not the biggest part of the foreclosure
problem. For the nation as a whole, less than 40 percent of homeowners who had their first
90-day delinquency in 2008 stopped making payments abruptly. Because this figure is an
31
It is also important to point out that right-censoring may be inflating these numbers a little, since some
of the borrowers who we identify as direct defaulters in the last 3 months of the data, may make a mortgage
payment in the future.
32
The states are Arizona, California, Nevada, and Florida.
19
upper bound on the fraction of ruthless defaults, it suggests ruthless default is not the main
reason why falling house prices have caused so many foreclosures.
4
Foreclosure and Renegotiation
A distressing feature of the ongoing foreclosure crisis is the seeming inability of the
private market to stop it. A lender typically suffers a large loss when it (or its agent)
forecloses on a house. On the surface, it would appear that the lender would be better off
modifying any delinquent loan in the borrower’s favor and taking a small loss, as opposed
to refusing a modification, foreclosing on the mortgage, and suffering a large loss. Lender
behavior is especially perplexing if high DTI ratios are causing the crisis. Surely making
the mortgage affordable by reducing a borrower’s DTI to 38 or 31 percent is preferable to
foreclosure for the lender as well as the borrower. Given this apparent puzzle, a number of
analysts have argued that the securitization of mortgages into trusts with diffuse ownership
is preventing “win-win” modifications from taking place. In this section, we provide an
alternative explanation for why modifications are rare. We then consult the LPS dataset
and the historical record to see how the different explanations square with the data.
4.1
The renegotiation-failure theory
Lenders often take large losses on foreclosed homes, which are typically sold for much less
than the outstanding balances of the defaulted mortgages. Conversely, the modifications
offered to borrowers are generally modest. A study by White (2009) provides the following
data:
The average loss for the 21,000 first mortgages liquidated in November was
$145,000, representing an average loss of 55 percent of the amount due. Losses on
second lien mortgages were close to 100 percent. In comparison, for the modified
loans with some amount of principal or interest written off, the average loss
recognized was $23,610. This seven-to-one difference between foreclosure losses
and modification write-offs is striking, and lies at the heart of the failure of the
voluntary mortgage modification program. Particularly for foreclosed loans with
losses above the 57 percent average, some of which approach 100 percent, the
decisions of servicers to foreclose is mystifying.... At a minimum, there is room
for servicers to be more generous in writing down debt for the loans they are
modifying, while still recovering far more than from foreclosures in the depressed
20
real estate market of late 2008.33
To explain the small number of concessions and the large number of foreclosures, many
analysts blame institutional factors related to the collection of mortgages into mortgagebacked securities (MBS). Such loans are owned by trusts on behalf of a large number of
individual investors, rather than by a single entity (such as a local bank). White’s quote
mentions the decisions of loan servicers, who are responsible for funneling mortgage payments to these MBS investors and performing various other tasks related to securitized
mortgages.34 Most importantly, when a borrower falls behind on his mortgage, it is the
servicer who decides whether a loan modification or a foreclosure is more appropriate.
Analysts who blame securitization for the low number of modifications argue that the
incentives of the servicers have become decoupled from those of investors, who ultimately
bear the losses entailed in foreclosure. We label this claim the renegotiation-failure theory.
Securitization can potentially limit modifications in at least two ways. First, servicers can
be hamstrung by restrictive agreements they signed with investors at the origination of the
mortgage trust, well before the crisis hit.35 The actions of a servicer working for a trust are
governed by so-called Pooling and Servicing Agreements (PSAs). Among other things, these
agreements specify the latitude that servicers have when deciding between modification and
foreclosure. As a general rule, PSAs allow servicers to make modifications, but only in cases
where default is likely and where the benefit of a modification over foreclosure can be shown
with a net-present-value (NPV) calculation. Second, proponents of the renegotiation-failure
theory claim that servicers are afraid that they will be sued by one tranche of investors
in the MBS if they make modifications, even if these modifications benefit the investors in
the trust as a whole. Because different tranches of investors have different claims to the
payment streams from the MBS, a modification may alter these streams in a way that will
benefit one tranche at the expense of another. One might think that the PSAs would have
foreseen this possibility, but some analysts claim that the PSAs were not written with an
eye to the current foreclosure crisis. Thus, it is claimed that there is enough ambiguity in
the PSAs to make servicers wary of getting caught up in “tranche warfare,” so servicers are
thought to follow the path of least resistance and foreclose on delinquent borrowers.36
33
White (2009), pp. 14–15.
Mortgages held in the portfolio of a single financial institution are normally serviced by that institution.
35
For example, the authors of the COP report write that “[r]estrictions on mortgage servicers’ ability to
modify loans are an obstacle that has contributed to foreclosure that destroys value for homeowners and
investors alike” (p. 50.)
36
The authors of the COP report write that “[s]ervicers may also be reluctant to engage in more active
loan modification efforts because of litigation risk” (p. 46).
34
21
A central implication of this theory is that securitization and the related frictions embedded in the contracts between investors and servicers are preventing modifications that
would make even the lender better off. As Eggert (2007) states:
The complex webs that securitization weaves can be a trap and leave no one, not
even those who own the loans, able effectively to save borrowers from foreclosure.
With the loan sliced and tranched into so many separate interests, the different
claimants with their antagonistic rights may find it difficult to provide borrowers
with the necessary loan modifications, whether they want to or not (p. 292).37
4.2
Reasons to doubt the renegotiation-failure theory
There are, however, reasons to doubt the renegotiation-failure theory. First, there is
little evidence on the extent to which PSAs have limited modifications in practice.38 A 2007
study by Credit Suisse of approximately 30 PSAs concluded that fewer than 10 percent of
them completely ruled out modifications. About 40 percent of the PSAs allowed modifications, but with some restrictions. These restrictions included a limit on the percentage
of mortgages in the pool that could be modified without permission from the trustee of
the mortgage-backed security (often 5 percent), and/or a floor for the mortgage rate that
could be applied in the event of a modification that entailed a reduction in the borrower’s
interest rate. The remainder of PSAs contained no restrictions. It is unlikely that even
PSAs with 5-percent caps are preventing modification to any significant degree. The Congressional Oversight Panel for the Troubled Asset Recovery Program examined a number of
securitized pools with 5-percent caps and found that none had yet approached this cap.39
Moreover, one can make a case that the typical PSA actually compels the servicer to make
modifications if these modifications are in the best interests of the investor. According to
Cordell, Dynan, Lehnert, Liang, and Mauskopf (2008), “While investors seem somewhat
concerned about servicer capacity, they do not convey widespread concern that servicers
are relying overmuch on foreclosures relative to modifications.” In fact, investors opposed
additional incentives for modifications:
Investors with whom we spoke were not enthusiastic about an idea to reimburse
37
Other policy analysts have adopted a similar view. For example, the COP writes in its recent report
that “A series of impediments now block the negotiations that would bring together can-pay homeowners
with investors who hold their mortgages .... Because of these impediments, foreclosures that injure both
the investor and homeowner continue to mount” (COP report, p. 2).
38
For a discussion of the role of PSAs in reducing modifications, see Cordell, Dynan, Lehnert, Liang, and
Mauskopf (2008), which also discusses the incentives faced by servicers more generally.
39
COP report (p. 44).
22
servicers for expenses of loss mitigation. In their view, such payments could
lead to more modifications than warranted by the NPV calculations. They also
felt that the PSA adequately specified that modifications that maximized NPV
should be undertaken. A typical response from an investor was, Why should I
pay servicers for doing something that I already paid them to do?40
Regarding the fear of lawsuits, no servicer has yet been sued for making too many loan
modifications. There has been a well-publicized lawsuit filed by a group of investors against
a servicer doing modifications, but the details of this suit should not make other servicers
wary about making modifications.41 Moreover, Hunt (2009) studied a number of subprime
securitization contracts and found not only that outright bans on modifications were rare,
but also that most contracts allowing modifications essentially instructed the servicer to
behave as if it were the single owner of the loan:
The most common rules [in making modifications] are that the servicer must
follow generally applicable servicing standards, service the loans in the interest
of the certificate holders and/or the trust, and service the loans as it would
service loans held for its own portfolio. Notably, these conditions taken together
can be read as attempting to cause the loans to be serviced as if they had not
been securitized. (p. 8, insertion added)
While there can be substantial disagreement about the importance of any particular
institutional impediment to loan modification, perhaps the most compelling reason to be
skeptical about the renegotiation-failure theory is the sheer size of the losses it implies. We
can use White’s figures quoted above to come up with a back-of-the-envelope calculation for
the total losses that follow from the renegotiation-failure theory. One figure often cited for
the total number of foreclosures that can be prevented with modifications is 1.5 million.42
For a dollar figure, we can multiply this number of preventable foreclosures by the $120,000
that White claims is lost by investors for each foreclosure performed.43 This results in a
total deadweight loss of $180 billion.
40
Cordell, Dynan, Lehnert, Liang, and Mauskopf (2008), p. 19.
Specifically, an MBS investor has sued two large servicers, Countrywide and Bank of America, for
promising to make mass modifications as part of a settlement that Countrywide and Bank of America
struck with the government in a predatory lending case. The key argument by the investor in this lawsuit
was that the modifications were done not because they were profitable for the investors, but rather to
settle a predatory lending lawsuit, which the plaintiffs of that lawsuit claimed was the responsibility of
Countrywide, in its capacity as the originator of the troubled loans.
42
This figure comes from FDIC Chairman Sheila Bair. For details see “Sheila Bair’s Mortgage Miracle,”
Wall Street Journal, December 3, 2008.
43
White (2008)
41
23
Losses of this size may be hard to square with economic theory, as Eric Maskin recently
pointed out in a letter to the New York Times. Maskin wrote his letter in response to an
earlier op-ed that had claimed the government has a role in facilitating loan modifications,
specifically mass write-downs of principal balances.44 According to Maskin: “If, as claimed,
such write-downs are truly ‘win-win’ moves — allowing borrowers to keep their homes and
giving mortgage holders a higher return than foreclosure — they may not need the government’s assistance.” The writers of the original op-ed column had claimed that servicers now
have an undue incentive to foreclose rather than modify loans. Maskin pointed out that if
this were the case, then
mortgage holders themselves have strong motivation to renegotiate those contracts, so that the servicers’ incentives are corrected. That would be a win-winwin move (for mortgage holders, servicers and borrowers), and to complete their
argument, the writers must show why it won’t happen.
Economists will recognize the reasoning in Maskin’s critique. The Coase Theorem implies
that economically efficient decisions will be made as long as property rights are well-defined
and transactions costs are not of first-order importance. Under these conditions, it does
not matter that servicers are not the ones who suffer the $180 billion losses entailed in
foreclosure, or even that existing PSAs might unduly limit modifications. The party that
suffers the potential losses — the investors — has an incentive to make side payments or
to change contractual arrangements so as to prevent these massive losses from occurring.
To take this reasoning one step further, if one class of investors has more to gain from
modification than another class stands to lose, the first class has an incentive to strike deals
with (or buy out) the second class. Consequently, to be consistent with the Coase Theorem,
the renegotiation-failure theory must also assert that the transactions costs implied by
securitization are large enough to derail these efficiency-enhancing arrangements, at the
cost to lenders of $180 billion.
4.3
A theory of loan modifications
There is another way to explain the low number of modifications that does not rely on
enormous transactions costs and yet is consistent with the Coase Theorem. It is simply
that most potential modifications are negative-NPV transactions from the standpoint of
investors. In other words, when all the relevant costs and benefits are considered, servicers
44
The op-ed to which Maskin responded is Geanakoplos and Koniak (2008). Maskin’s letter appeared on
March 7, 2009.
24
may already be acting in the best interests of the investors when they foreclose.45
To start with, modifications do not always prevent foreclosures, especially when defaults
are of the double-trigger variety. Consider a borrower who has lost his job. No permanent
modification can make the house affordable if the borrower has no income. Lenders often
offer “forbearance” in these cases, whereby the borrower pays sharply reduced payments
for a time. The borrower is then obligated to make up these arrears, with interest, later
on. Lenders may be reluctant to offer forbearance for any length of time if they are unsure
when the borrower will find a new job (and at what wage). When the value of the house
that collateralizes the loan is falling, and when all parties know that the house has probably
become unaffordable to the borrower, then the servicer may simply decide to take a loss
now by foreclosing, rather than risk an even larger loss down the road.46
The possibility that borrowers will re-default on their loans reduces the benefits of loan
modifications and thereby makes them less likely to occur. There are also reasons to think
that costs of modifications are higher than many housing analysts recognize. These analysts
typically ignore the costs of modifications that are made to borrowers who would have
repaid their loans anyway. Consider a lender facing a troubled borrower who is requesting
a modification. If the lender fails to modify the loan and the borrower defaults, the lender
will lose because (as White points out above), the cost of modifying the loan falls far short
of the cost of foreclosing. We will call this loss “Type I error.” However, Type I error is
only part of the story, as the lender faces another potential problem. If, unbeknownst to
the lender, the borrower requesting the modified loan will not default in the absence of a
modification, then the lender will lose the money he would have received according to the
original terms of the loan. We call this situation “Type II error.” For a modification to
make economic sense from the lender’s perspective, Type I error must exceed Type II error.
More formally, we can follow Foote, Gerardi, and Willen (2008), who consider a lender
with a borrower who owes m on a house currently worth pH dollars. This borrower will
default with probability α0 , in which case the lender recovers pH less λ dollars in foreclosure
costs. A modification lowers the value of the loan to m∗ < m and the probability of
foreclosure to α1 < α0 . Note that we do not assume that modification guarantees full
repayment of the mortgage — there is some probability of re-default when α1 > 0. Some
45
Note that because of externalities from foreclosures, modifications may be in society’s interests even if
they are not investors’ interests.
46
We have been told that there is a macabre saying in the servicing industry: “The first loss is the best
loss.”
25
simple arithmetic shows that renegotiation occurs when:
Renegotiation ⇔
(α0 − α1 )
| {z }
Reduction in
foreclosure prob.
× (m∗ − (pH − λ)) >
|
{z
}
Reduced
loss
(1 − α0 )
| {z }
Pct. repay without
mitigation
×
(m − m∗ )
| {z }
. (3)
Reduced value of
the mortgage
The first term corresponds to the Type I error — if a foreclosure is prevented, the lender
recovers m∗ rather than pH − λ. The second term corresponds to the Type II error —
borrowers who would have repaid in full, but take advantage of principal reduction to
reduce their debt burden.
The following reformulation of equation (3) is instructive:
m − m∗ <
α0 − α1
[m − (pH − λ)].
1 − α1
(4)
The right-hand side is the maximum possible concession the lender can profitably make.
To understand this, consider some simple examples. If we set α0 , the probability of default
without a modification, equal to 1, then equation (5) becomes
m∗ > pH − λ.
This is the case that White (2009) has in mind when he writes, “Particularly for foreclosed
loans with losses above the 57 percent average, some of which approach 100 percent, the
decisions of servicers to foreclose is mystifying.”47 In White’s extreme example of 100 percent
loss given default, even a modification that reduces the probability of default from 1 to
anything even infinitesimally less than one, and in which the lender recovers infinitesimally
more than 0, makes economic sense.
However, even a little uncertainty about whether the borrower will default invalidates
the above logic. If we assume modification ensures that the loan will repay with certainty
(α1 = 0), then equation (5) becomes:
m − m∗ < α0 [m − (pH − λ)].
(5)
It is easy to see in this equation exactly how the math works against modification. Suppose
the expected loss is 57 percent and the likelihood of default is 50 percent, then the lender
can only reduce the value of the loan by 28.5 percent.
How big are Type I and Type II errors in practice? Results in Gerardi and Willen
47
White (2009), p. 15.
26
(2009) show that for most categories of homeowners in Massachusetts, Type II is large
relative to Type I error: even with major stresses, most homeowners will not default on
their mortgages. The authors find that concessionary modifications make sense only for
multi-family properties purchased with subprime mortgages.
Equation (3) clearly illustrates that the observation that a foreclosure, on the surface,
seems to lead to greater monetary losses than an apparently reasonable modification is not
prima facie evidence of inefficiency. Such foreclosures may well be ex ante efficient, when the
issue of moral hazard is factored into the equation. This type of moral hazard explains why
mortgage investors are not unduly concerned about too few modifications being performed,
and why, to date, there have been no lawsuits against servicers encouraging them to do
more modifications.
4.4
Statistical evidence on loan modifications
The LPS data allow us to perform an econometric test of the renegotiation-failure theory,
because these data contain information on the ultimate holder (investor) of the residential
mortgages. Specifically, we are able to tell whether a mortgage is held on the balance sheet
of a financial institution, securitized by a government sponsored enterprise (GSE) such as
Freddie Mac (FHLMC) or Fannie Mae (FNMA), or securitized by a non-agency, private
institution. With this information, combined with information that allows us to identify
modified loans, we are able to compare the relative modification frequency between loans
held in portfolio and loans that are securitized. If institutional constraints inherent in the
securitization process are preventing profitable modifications, then we expect to see in the
data relatively few modifications among securitized loans, as compared with loans held in
portfolio.
The LPS dataset does not include direct information on loan modifications. However,
it does contain updated loan terms at a monthly frequency, with which we are able to
identify loan modifications indirectly (and imperfectly).48 With these data we label a loan
as modified if there is a change in its terms that was not stipulated by the initial terms of the
contract. These changes include interest-rate reductions, principal-balance reductions, and
term extensions. We can also identify principal-balance and mortgage-payment increases
that reflect the addition of arrears to the balance of a loan.49
48
The Office of Thrift Supervision (OTS) and Office of the Comptroller of Currency (OCC) used very
similar data from LPS to analyze the outcomes of recent mortgage modification programs (OCC and OTS
Mortgage Metrics Report, Third Quarter 2008) In their report, they used supplementary data directly from
large mortgage servicers that included the identification of loans in the LPS data that had been modified.
While we do not have access to those data, our findings are fairly consistent with theirs.
49
There are two potential mistakes we can make in this exercise. First, we may falsely identify modifica-
27
Table 7 reports the number of modifications made by quarter from the first quarter of
2007 through the last quarter of 2008, disaggregated by the type of modification made.
Each of the numbers in the table is a multiple of 10 because we used a 10 percent random
sample and scaled up the numbers we found. The first column simply reports the total
number of loan modifications performed and shows that they have become more common
as the housing market has weakened. By our calculations, there appear to be more than
seven times as many modifications performed in the fourth quarter of 2008 as in the first
quarter of 2007.
In addition to the rapid growth in loan modifications, the composition of modifications
has changed over time. This can be seen in the remaining columns of Table 7, which list
the incidence of modifications of different types.50 A somewhat surprising finding is that
most modifications entailed increases in the principal balance of a mortgage. Such increases
are likely due to the addition of arrears to the outstanding mortgage balance for delinquent
borrowers, and they often increase the monthly mortgage payment by a nontrivial amount.
Table 7 shows that while the absolute numbers of balance-increasing modifications are still
rising, they are falling as a percentage of total modifications. In the last few quarters in our
data, interest-rate reductions, which necessarily involve a decrease in the mortgage payment,
have become more frequent, rising to more than 25 percent of all modifications performed
in 2008:Q4. Adelino, Gerardi, and Willen (2009) provide further information regarding
the behavior of monthly mortgage payments for loans that have undergone a modification.
The authors find that until the fourth quarter of 2008, modifications involving payment
increases were more common than those involving payment decreases. In addition, they find
that the average and median magnitudes of payment decreases have recently increased from
approximately 10–14 percent in the period between 2007:Q1 and 2008:Q2, to approximately
20 percent in the final two quarters of 2008. Based on the logic from our simple framework
above, it is likely that these will have more success than modifications involving increases
in the payment and/or balance.
Figure 8 contains some evidence from the LPS data to support this claim. The figure
tions (“false positives”) because of measurement error in the data (for example, a mistake in the updated
balance or interest rate) or some endogenous behavior on the part of the borrower (for example, a borrower
making extra principal payments). Second, we may miss modifications (“false negatives”) because our algorithm for finding modifications is incomplete. In this section we are more concerned with false positives
than with false negatives, so we use a conservative set of criteria. See Adelino, Gerardi, and Willen (2009)
for a detailed explanation of the exact algorithm used to identify modified loans in the LPS data.
50
In many cases a mortgage will experience multiple types of modifications at the same time. For example,
we see cases in the data in which the interest rate is decreased and at the same time the term of the loan is
extended. Thus, the percentages in Table 7 are not calculated with respect to the number of loans modified,
but rather with respect to the number of modifications performed.
28
contains Kaplan-Meier non-parametric, survival estimates (also known as the product limit
estimator) of the transition from modification to default.51 The figure considers a loan to
be in default when it becomes 90 days delinquent (approximately three missed payments).
The figure shows that modifications involving a decrease in the monthly payment are far
more successful than those involving an increase in the payment. For example, after one
year, the probability that a modified loan involving a payment increase becomes 90 days
delinquent is approximately 69 percent. In contrast, a modified loan involving a payment
decrease has a probability of becoming 90 days delinquent of approximately 52 percent.52 Of
course, it should be noted that the underlying data in Figure 8 come predominantly from
loan modifications that took place in 2007 and early-to-mid 2008, while the majority of
modifications in the LPS data occurred in the last two quarters of 2008. The Kaplan-Meier
estimator does account for right-censoring, but in order to draw more conclusive inferences
we will need to observe more data on these recent modifications. Another noteworthy
observation from Table 7 is that the incidence of principal reductions is extremely low
in our data. This is likely due to two factors. First, the LPS data under-represent the
subprime mortgage market.53 A few servicers that focus almost exclusively on subprime
mortgages have recently begun modification programs that involve principal reduction.54
In addition, from a theoretical perspective, principal reduction plans suffer from the severe
incomplete-information problem noted earlier. Balance reductions are appealing to both
borrowers in danger of default and those who are not. As a result, lenders have a strong
incentive to provide modifications only to those borrowers who are most likely to default.
Adelino, Gerardi, and Willen (2009) provide evidence to support this claim, as they show
that modified loans in the LPS dataset are characterized by high leverage, high initial debt51
The Kaplan-Meier estimate of the survival function for delinquency is given by:
Sˆt =
Y n i − mi
,
ni
t <t
(6)
i
where S(t) is the probability that a borrower will not default through time t, di corresponds to the number
of loans that default at time ti , while ni corresponds to the number of loans that are “at-risk” of default at
time ti , or in other words the number of loans that are still active and that have not defaulted before time
ti .
52
The pattern is similar if we assume the more stringent definition of default, corresponding to the
situation in which foreclosure proceedings are initiated by the holder of the mortgage. In this case, modified
loans involving a payment increase have a probability of experiencing a foreclosure of about 34 percent,
while modifications involving a payment decrease have an associated probability of about 17 percent.
53
The majority of subprime mortgages are securitized by non-agency firms, and for the period of interest,
the LPS dataset includes approximately 35 percent of mortgages securitized by non-agency corporations.
54
An October report by Credit Suisse notes that Ocwen Loan Servicing, LLC and Litton Loan Servicing LP were the only subprime servicers that had performed a nontrivial number of principal reduction
modifications. Neither of these servicers contribute to the LPS dataset.
29
to-income ratios, and low initial credit scores. These are the loans that are most likely to
default without a modification (that is, loans where α0 is high).
Table 8 contains modification statistics broken down by the holder of the mortgage. We
distinguish between mortgages held in portfolio, mortgages securitized by a GSE such as
Fannie Mae or Freddie Mac, and mortgages securitized by a private entity. For each quarter
of 2008, we calculate the percentage of loans outstanding at the beginning of each quarter
that were modified at some point in that quarter. Each panel in the table corresponds
to a different sample of mortgages. Panel A corresponds to all types of mortgages in the
data. Panel B corresponds to both subprime and Alt-A mortgages.55 Finally, each panel in
the table is disaggregated into three parts, corresponding to different denominators used in
calculating the percentages. The first part uses all loans outstanding at the beginning of the
respective quarter, the second part uses all loans that are 30 days delinquent at the start of
the respective quarter, and the third part uses all loans that are 60 days delinquent at the
start of the respective quarter. By limiting the sample to delinquent loans, we are partially
controlling for differences in credit quality between loans held in portfolio and loans that are
securitized. This control turns out to be important. In both of the panels, and in almost all
quarters, modifications for privately securitized loans are more frequent than for portfolio
loans when the relevant universe is the full sample of loans. However, privately securitized
loans are generally riskier than other loans, so this discrepancy may simply reflect the fact
that more privately securitized loans are in danger of foreclosure and are thus, candidates
for modification. When we narrow the focus to delinquent loans, the results become more
balanced. Portfolio loans have a slightly higher incidence of modification compared with
privately securitized in Panel A, while modifications are more common among portfolio
loans in many instances in Panel B (except in the fourth quarter of 2008).
There are at least two patterns of note in Table 8. First, while delinquent loans held
in portfolio appear to be modified more frequently than privately securitized mortgages
(except for subprime and Alt-A mortgages, as defined in the LPS data), the discrepancy
is not as large as it is often made out to be in policy circles and in media reports. For
the sample of all 30-day delinquent loans (Panel A) held in portfolio, 6.81 percent were
modified in the third quarter of 2008 and 8.55 percent in the fourth quarter of 2008. In
comparison, 6.28 percent and 6.23 percent of privately securitized mortgages were modified
in the third quarter and fourth quarter of 2008, respectively. We see similar, although
slightly larger discrepancies for 60-day delinquent loans, but in many instances the sign
changes for subprime and Alt-A loans (Panel B). The second take-away from the table
55
The definition of subprime and Alt-A comes directly from the servicers that contribute to the LPS
dataset. There is no additional distinction between subprime and Alt-A in the LPS dataset.
30
is that the GSEs appear to have been much more reluctant to modify loans, with the
exception of Freddie Mac in the third and fourth quarters of 2008.56 While the summary
statistics presented above suggest that the incidence of modification does not seem to be
greatly impeded by the process of securitization, there are a variety of factors that could be
contributing to the variation in Table 8, including substantial differences in characteristics
between portfolio-held loans and securitized loans. In addition, there may be significant lags
between the time when a loan becomes delinquent and the point when it is modified that are
not captured in Table 8. For example, if it were the case that the percentages of modified
loans were the same, but portfolio-held loans were modified more quickly than privately
securitized loans, Table 8 would show more portfolio-held loans being modified (since the
slower, privately securitized modifications would not be picked up in the table). For this
reason, a slightly more formal analysis is necessary, in which other observable differences
between securitized and portfolio loans are controlled for, and in which the timing issues
as well as right-censoring are also taken into account. Censoring is an especially important
problem, as there are currently many delinquent loans outstanding that are, or will soon
be, good candidates for modification, as the housing market continues to decline.
Figure 9 displays Kaplan-Meier estimates of the survival function with respect to the
transition from delinquency to modification, broken down by the holder of the mortgage.
While the Kaplan-Meier estimator does not control for other observable differences in mortgage characteristics, it does account for censoring and the timing issues discussed above.57
The figure contains two plots. The first plot displays estimates of the survival function
corresponding to the transition from 30 days delinquency (one mortgage payment behind)
to modification of all mortgages originated after 2004 in the LPS dataset, while the second
plot uses only data from subprime/Alt-A mortgages in the LPS data originated after 2004.
There are a few notable patterns contained in Figure 9. First, looking at the universe of
all mortgages, privately securitized loans and GNMA loans are more likely to have been
modified than loans held in portfolio and FNMA loans over a fairly long horizon. Conditional on 30-day delinquency, a privately securitized loan has a 15 percent probability of
being modified after two years, and a 26 percent probability after three years, compared
56
Piskorski, Seru, and Vig (2009) find differences in the performance of seriously delinquent securitized
and non-securitized loans. They attribute this difference to a difference in the willingness of servicers to
modify securitized vs. non-securitized mortgages. However, they do not attempt to identify modifications
directly, as we do. Thus, our findings are at odds with their interpretation of the data. See Adelino, Gerardi,
and Willen (2009) for details.
57
Adelino, Gerardi, and Willen (2009) estimate Cox proportional hazard models of the transition from
delinquency to modification, in which differences in observable loan and borrower characteristics are controlled for, and find results that support the patterns in Figure 9.
31
with 11 percent and 16 percent for loans held in portfolio, respectively.58 Over a shorter
horizon, (less than one year), there is very little difference across different types of loans
when conditioning on 30-day delinquency. The patterns are slightly different for the sample of subprime/Alt-A loans, as the incidence of modification is virtually the same over all
horizons for portfolio-held and privately securitized loans.59
4.5
Historical evidence on loan modifications
In addition to comparing securitized vs. non-securitized loans today, we can evaluate
claims about contract-related frictions by looking at the historical record. It is often claimed
that renegotiation was frequent in the past, before securitized mortgages were common. For
example, a report from the Congressional Oversight Panel for the Troubled Asset Recovery
Program states that
For decades, lenders in this circumstance could negotiate with can-pay borrowers
to maximize the value of the loan for the lender (100 percent of the market value)
and for the homeowner (a sustainable mortgage that lets the family stay in the
home). Because the lender held the mortgage and bore all the loss if the family
couldn’t pay, it had every incentive to work something out if a repayment was
possible.60
Other authors, including Zingales (2008) and Geanakoplos and Koniak (2008), have also
claimed that renegotiation used to be common, but we know of no historical studies that
verify this claim. There are, however, reasons to be skeptical. First, foreclosures were quite
common in the past. Between 1929 and 1936, lenders carried out 1.8 million foreclosures
in the United States. To put that number in perspective, keep in mind that the number
of occupied dwellings more than quadrupled from 22.9 million in 1930 to 105 million in
2000.61 In addition, increases in credit and increases in owner-occupancy have resulted
in the number of owner-occupied, mortgaged homes rising from 4.8 million in 1940 to 39
58
These probabilities increase substantially for loans that become 60 days delinquent, but the relative
patterns are similar. Conditional on 60-day delinquency, a privately securitized loan has a 27 percent
probability of being modified after two years, and a 40 percent probability after three years, compared with
23 percent and 32 percent for loans held in portfolio, respectively. See Adelino, Gerardi, and Willen (2009)
for these plots.
59
There are a trivial number of GNMA subprime loans in the data, and thus we drop GNMA from the
graph. In addition, there are only a small number of FNMA and FHLMC subprime loans that are seasoned
beyond two years, and thus we decided to truncate the graph for these types of loans after two years.
60
COP report, p. 2.
61
Source: U.S. Census of Housing, 2000, Table DP-4, and 1950, Part 1, Table J.
32
million in 2000. Thus, an equivalent figure for the current crisis would be between 8.3 and
17 million foreclosures.
Another way to compare foreclosures in the current era with foreclosures during the
Depression is to the look at the performance of vintages of loans. The top panel of Figure
10 shows the fraction of loans foreclosed upon by year of origination for the three principal sources of credit in that period: savings and loan institutions (S&L), life insurance
companies, and commercial banks. The worst vintages were those of the late 1920s, when
approximately 30 percent of loans originated by life insurance companies ended in foreclosure, 20 percent of S&L mortgages ended in foreclosure, and about 15 percent of commercial
bank loans were foreclosed upon. The bottom panel shows the fraction of homeownerships
(not loans) originated each year in Massachusetts from 1988 through 2008 that eventually
ended in foreclosure.62 Since at least some of these foreclosures did not occur on purchase
mortgages, but rather on subsequent refinances, one can view this as an upper bound on a
similar measure using current data. What is clear is that we see far fewer foreclosures than
we did in the 1930s. These statistics are difficult to square with the claim that renegotiation
was more common in the past.
In fact, historical documents do suggest that modifications occurred in the past. The
Home Owners Loan Corporation (HOLC), set up by the federal government in 1933 in
the midst of the Great Depression, would buy loans at a deep discount from lenders and
re-underwrite the borrower into a new mortgage consistent with the borrower’s financial
situation at the time. However, it is important to understand that the economic situation
was extremely poor, as 40 percent of American homeowners were more than 15 months in
arrears. In terms of our model, this made Type I error large and Type II error small.63
Unfortunately, we do not have detailed data on the subsequent mortgages to analyze the
ultimate experiences of HOLC borrowers.
In addition, commercial banks commonly modified loans in this time period. Behrens
(1952) shows that as many as 40 percent of the loans originated in a given year would
be modified at least once, and as many as half of those more than once. However, it is
important to understand that until the 1930s, commercial banks could not make long-term
amortized loans, so renegotiation for term extensions and interest rate changes was common.
According to Behrens, “It should also be observed that the low level of interest rates current
in the 1930s as compared with that prevailing during the 1920s doubtless stimulated a good
many of the loan modifications, primarily for those loans in good standing...”
62
63
See Gerardi, Shapiro, and Willen (2007) for details regarding the Massachusetts data.
See Harriss (1951) for details about HOLC.
33
In general, discussions of foreclosure from contemporary sources in “past decades” never
mention concessionary modification as a strategy for dealing with troubled borrowers. A
book on what we would now call “best practices” in mortgage banking, written in the
mid-1950s, gives a detailed discussion of how to contact delinquent borrowers, but then
recommends turning the problem over to an attorney.64 The author then discusses how to
deal with the sale of a foreclosed property but never suggests that the servicer should make
concessions to help the borrower continue making payments. Even HOLC, to a large extent,
considered mostly non-concessionary modifications and foreclosed on almost 20 percent of
the borrowers to whom it lent.65
Foreclosure has always been a common outcome in mortgage lending, even for the bestintentioned of lenders. The first borrower ever to obtain a loan from a Building and Loan
Society in the U.S. was eventually forced out of his home. A man named Comly Rich took
out a mortgage on April 11, 1831, but “was frequently fined for failure to pay his dues and
interest.” The problems were resolved in what amounts to a foreclosure: both the house and
the mortgage were transferred to another borrower.66
5
Conclusion
In this paper, we have attempted to make two main points. First, while the concept
of mortgage “affordability” is often used in explanations of the current rise in mortgage
defaults, this concept is not helpful if it is not defined precisely. Many people believe
that the affordability of a mortgage is adequately summarized in the DTI at origination.
However, this ratio does not appear to be a strong predictor of default. What really matters
in the default decision is the mortgage payment relative to the borrower’s income in the
present and future, not the borrower’s income in the past. Consequently, the high degree
of volatility in individual incomes means that mortgages that start out with low DTIs can
end in default if housing prices are falling. A second, related point concerns the apparent
unwillingness of loan servicers to turn “bad” (that is, high-DTI) mortgages into “good”
(low-DTI) mortgages. It is true that lenders may lose a great deal of money with each
individual foreclosure, but the loan modifications might have negative NPV if they are
sometimes extended to people who are likely to pay on time anyway. And the benefits of
modifications are uncertain if borrowers have lost their jobs.
What do these findings suggest for foreclosure-reduction policy? One suggestion would
64
Pease and Cherrington (1953).
Harriss (1951).
66
See Bodfish (1931), pp. 66–72.
65
34
be to focus a program on the effects of income volatility, helping people who lose their
jobs get through difficult periods without having to leave their homes. For example, the
government could replace a portion of lost income for a period of one or two years, through
a program of loans or grants to individual homeowners.67 For more permanent and very
large setbacks, a good policy might help homeowners transition to rentership through short
sales or other procedures. In any case, the results of this paper suggest that policies that
encourage moderate, long-term reductions in DTIs face important hurdles in addressing the
current foreclosure crisis.
67
For details of such a plan, see http://bosfed.org/economic/paymentsharingproposal.pdf.
35
References
Adelino, M., K. Gerardi, and P. S. Willen (2009). Renegotiating Home Mortgages: Evidence from the Subprime Mortgage Crisis. Working Paper .
Behrens, C. F. (1952). Commercial Bank Activities in Urban Mortgage Financing. New
York, NY: National Bureau of Economic Research.
Bodfish, H. M. (1931). History of Building and Loan in the United States. Chicago, IL:
United States Building and Loan League.
Congressional Oversight Panel of the Troubled Assets Recovery Program (COP) (2009).
Foreclosure Crisis: Working Toward A Solution.
Cordell, L., K. Dynan, A. Lehnert, N. Liang, and E. Mauskopf (2008). The Incentives of
Mortgage Servicers: Myths and Realities. Finance and Economics Discussion Series,
Federal Reserve Board 46.
Cordell, L., M. Watson, and J. Thomson (2008). McDash Data Warehouse Seminar.
Danis, M. A. and A. N. Pennington-Cross (2005). The Delinquency of Subprime Mortgages. SSRN eLibrary.
Dell’Ariccia, G., D. Igan, and L. Laeven (2009). Credit booms and lending standards:
Evidence from the subprime mortgage market.
Deng, Y., J. Quigley, and R. V. Order (2000). Mortgage terminations, heterogeneity and
the exercise of mortgage options. Econometrica 68 (2), 275–307.
Eggert, K. (2007). Comment on Michael A. Stegman et al.’s “Preventive Servicing Is Good
for Business and Affordable Homeownership Policy”: What Prevents Loan Modifications. Housing Policy Debate 18 (2), 279–297.
Foote, C., K. Gerardi, L. Goette, and P. S. Willen (2008). Just the facts: An initial
analysis of subprime’s role in the housing crisis. Journal of Housing Economics 17 (4),
1–24.
Foote, C., K. Gerardi, and P. Willen (2008). Negative equity and foreclosure: Theory and
evidence. Journal of Urban Economics 6 (2), 234–245.
Geanakoplos, J. and S. Koniak (2008). The best way to help homeowners – and the
economy. Available at http://cowles.econ.yale.edu/gean/crisis/index.htm.
Gerardi, K., A. Shapiro, and P. Willen (2007). Subprime outcomes: Risky mortgages,
homeownership experiences, and foreclosures. Federal Reserve Bank of Boston Working Paper 07-15.
36
Gerardi, K. and P. Willen (2009). Subprime Mortgages, Foreclosures and Urban Neighborhoods. The B.E. Journal of Economic Analysis & Policy 9 (3 (Symposium)).
Gourinchas, P. and J. Parker (2002). Consumption over the life cycle. Econometrica,
47–89.
Harnick, E. (2009). Examining the Making Home Affordable Program, Testimony Before
the U.S. House of Representatives Committee on Financial Services, Subcommittee
on Housing and Community Opportunity.
Harriss, C. L. (1951). History and Policies of the Home Owners’ Loan Corporation. New
York, NY: National Bureau of Economic Research.
Himmelberg, C., C. Mayer, and T. Sinai (2005). Assessing high house prices: bubbles,
fundamentals and misperceptions. Journal of Economic Perspectives 19 (4), 67–92.
Hunt, J. (2009). What do subprime securitization contracts actually say about loan modification. Berkeley Center for Law, Business and the Economy.
Kau, J. B., D. C. Keenan, and T. Kim (1994). Default probabilities for mortgages. Journal
of Urban Economics 35 (3), 278–296.
Kiefer, N. (1988). Economic duration data and hazard functions. Journal of economic
literature, 646–679.
Mayer, C. J. and K. Pence (2008, June). Subprime mortgages: What, where, and to
whom? Working Paper 14083, National Bureau of Economic Research.
Mian, A. R. and A. Sufi (2009). The Consequences of Mortgage Credit Expansion: Evidence from the U.S. Mortgage Default Crisis. Quarterly Journal of Economics forthcoming.
Morton, J. E. (Ed.) (1956). Urban Mortgage Lending: Comparative Markets and Experience. New York, NY: National Bureau of Economic Research.
Pease, R. H. and H. V. Cherrington (Eds.) (1953). Mortgage Banking. New York, NY:
McGraw-Hill.
Piskorski, T., A. Seru, and V. Vig (2009). Securitization and Distressed Loan Renegotiation: Evidence from the Subprime Mortgage Crisis.
Poterba, J. (1984). Tax subsidies to owner-occupied housing: an asset-market approach.
The Quarterly Journal of Economics, 729–752.
Sherlund, S. (2008). The past, present, and future of subprime mortgages. Finance and
Economics Discussion Series 2008-63, Federal Reserve Board.
37
White, A. M. (2008). Rewriting Contracts, Wholesale: Data on Voluntary Mortgage
Modifications from 2007 and 2008 Remittance Reports. Fordham Urban Law Journal forthcoming.
White, A. M. (2009). Deleveraging the American Homeowner: The Failure of 2008 Voluntary Mortgage Contract Modifications. Connecticut Law Review forthcoming.
Zingales, L. (2008). Plan b. Economists’ Voice 5 (6), 4.
38
Table 1: Probability that a loan will become “unaffordable” at least once in the first three years, where
unaffordability is defined as DTI above a certain threshold.
(1)
(2)
std(ηt )
15%
15%
std(εt )
21%
21%
Resets
No
Yes
A. Ex Post Unaffordable Defined as DTI>31%
Initial DTI = 31%
70.1
81.7
B. Ex Post Unaffordable Defined as DTI>38%
31%
45.6
60.5
Initial DTI =
38%
70.3
81.8
C. Ex Post Unaffordable Defined as DTI>50%
31%
16.4
30.6
36.5
51.8
Initial DTI = 38%
50%
69.7
81.4
(3)
0%
0%
Yes
(4)
15%
42%
No
100.0
72.6
0.0
100.0
58.6
74.7
0.0
0.0
100.0
38.3
54.8
72.0
Table 2: Shares of various loans in LPS data, by seniority, subprime status, and investor: December 2008
Investor
First-lien Prime
and Near Prime
Second-lien Prime
and Near Prime
First-lien
Subprime
Second-Lien
Subprime
Other
Total
Panel A: Counts
GSE Securitized:
Fannie Mae
Freddie Mac
Ginnie Mae
Private Securitized
Portfolio
Unknown
Other
Total
9,410,856
6,342,870
4,709,406
4,224,463
2,224,951
121,635
271,696
27,305,877
7,292
2,672
391
208,722
412,691
1,830
4,173
637,771
48,093
7,911
751
486,469
87,843
7,953
122
639,142
130
0 9,466,371
0
15 6,353,468
1
6 4,710,555
121,987
250 5,041,891
11,823 32,267 2,769,575
76
0
131,494
0
0
275,991
134,017 32,538 28,749,345
Panel B: Percentages
GSE Securitized:
Fannie Mae
Freddie Mac
Ginnie Mae
Private Securitized
Portfolio
Unknown
Other
Total
32.73
22.06
16.38
14.69
7.74
0.42
0.95
94.98
0.03
0.01
0.00
0.73
1.44
0.01
0.01
2.22
0.17
0.03
0.00
1.69
0.31
0.03
0.00
2.22
0.00
0.00
0.00
0.42
0.04
0.00
0.00
0.47
0.00
0.00
0.00
0.00
0.11
0.00
0.00
0.11
32.93
22.10
16.38
17.54
9.63
0.46
0.96
100.00
Notes: The investor “Other” category includes local housing authorities, the Federal Home Loan Bank
(FHLB), and GNMA Buyout Loans.
39
Table 3: Incidence of Missing DTI Ratios and FICO Scores in LPS data, By Year of Loan Origination
2002
2003
2004
2005
2006
2007
2008
All
88.3
65.1
44.1
40.4
40.3
31.7
42.5
All years
50.1
DTI Ratio
Prime Subprime
88.1
100.0
64.5
90.9
42.7
60.8
40.6
38.8
40.4
39.8
32.1
22.5
42.5
26.2
50.2
48.6
All
22.2
22.2
16.1
15.5
17.4
13.3
12.3
17.4
FICO Score
Prime Subprime
22.4
10.0
22.1
26.2
17.1
4.4
16.5
5.1
17.9
12.9
13.8
1.6
12.3
0.0
17.9
8.7
Table 4: Summary Statistics: Loans Originated from 2005–2008
Prime
DTI Ratio
FICO Score
LTV Ratio
Adjustable Rate Dummy
Number of Loans
Mean
35.1
714.1
73.4
.21
Std Dev
13.8
61.6
18.2
.40
501,317
40
Mean
40.0
609.0
79.2
.56
Subprime
Std Dev
11.1
54.9
12.5
.50
41,132
Figure 1: Loan-Specific Characteristics in LPS Sample
0
0
.01
.01
.02
Density
Density
.03
.02
.04
.05
Panel B: Subprime DTI
.03
Panel A: Prime DTI
20
40
60
80
100
0
20
40
60
80
DTI Ratio
DTI Ratio
Panel C: Prime FICO
Panel D: Subprime FICO
100
0
0
.002
.002
Density
.004
Density
.004
.006
.006
.008
.01
.008
0
400
450
500
550
600
650
FICO Score
700
750
800
850
400
450
500
600
650
FICO Score
700
750
800
850
Panel F: Subprime LTV
0
0
.05
.05
.1
Density
.1
Density
.15
.15
.2
.2
.25
Panel E: Prime LTV
550
0
25
50
75
100
125
LTV Ratio
0
25
50
75
LTV Ratio
41
100
125
0
.02
.04
Density
.06
.08
.1
Figure 2: Cumulative Changes in State-Level House Prices for LPS Loans
−20
0
20
Change in Prices Since Origination
40
0
.2
Density
.4
.6
Figure 3: Cumulative Changes in County-Level Unemployment for LPS Loans
−5
0
Change in Unemployment Since Origination
42
5
0
0
.001
.01
Prime
.002
.02
Subprime
.003
.03
.004
.04
Figure 4: Baseline Default Hazards: Prime and Subprime Loans
0
12
24
Months Since Origination
Prime
36
Subprime
0
0
.002
.005
Prime
.004
.01
Subprime
.006
.015
.02
.008
Figure 5: Baseline Prepayment Hazards: Prime and Subprime Loans
0
12
24
Months Since Origination
Prime
Subprime
43
36
Table 5: Estimates from Cox Proportional Hazard Models
DTI Ratio
FICO Score
LTV Ratio
LTV 6= 80 dummy
Adjustable Rate Dummy
∆ UR
∆ HP >= 0
∆ HP < 0
∆ HP * ∆ UR
∆ HP * LTV
∆ HP * DTI
∆ HP * FICO
∆ UR * FICO
FICO * DTI
DTI * ∆ UR
No. of monthly observations
No. of loans
Prime
Defaults
.0105**
(.0009)
–.0124**
(.0003)
.0308**
(.0021)
–.2973**
(.0453)
.7521**
(.0539)
.2068**
(.0207)
–.0571**
(.0061)
–.0592**
(.0051)
.0061**
(.0009)
–.0001
(.0001)
–.0000
(.0001)
–.0000
(.0000)
.0010**
(.0001)
.0000
(.0000)
–.0008
(.0005)
10,796,387
501,317
Subprime
Defaults
.0072**
(.0012)
–.0035**
(.0003)
.0212**
(.0026)
–.1836*
(.0738)
.5074**
(.0354)
.1007**
(.0156)
–.0516**
(.0071)
–.0451**
(.0049)
.0069**
(.0008)
–.0003**
(.0001)
.0001
(.0001)
–.0000
(.0000)
.0003**
(.0001)
–.000055**
(.000017)
.0003
(.0005)
821,020
41,132
Prime
Prepayments
.0046**
(.0005)
–.0004
(.0002)
–.0108**
(.0010)
.1126**
(.0219)
.6465**
(.0568)
–.0344
(.0210)
.0236**
(.0032)
.0555**
(.0062)
.0015
(.0012)
.0007**
(.0001)
.00010*
(.00004)
–.00012**
(.00002)
.0002
(.0001)
–.0000
(.0000)
–.0005
(.0003)
10,796,387
501,317
Subprime
Prepayments
–.0003
(.0012)
–.0016**
(.0005)
–.0234**
(.0017)
.2447**
(.0286)
.5605**
(.0537)
–.0476
(.0345)
.0384**
(.0043)
.0925**
(.0088)
.0019
(.0016)
.0009**
(.0001)
.0001
(.0001)
.0000
(.0000)
-.0000
(.0000)
–.00004**
(.00001)
.0008
(.0006)
821,020
41,132
Notes: Standard errors are clustered by state. * denotes significance at 5 percent. **
denotes significance at 1 percent. A negative value of a house-price change (HP < 0) is
entered directly in the regression (not as an absolute value.)
44
Figure 6: Model-Generated Monthly Defaults and Prepayments (Per 100 Loans Originated) Under Various Assumptions
Monthly Defaults
0
.2
.4
.6
0
10
20
30
Months Since Origination
Panel B: Prime Prepayments
Monthly Prepayments
0 .2 .4 .6 .8
Panel A: Prime Defaults
40
0
10
20
30
Months Since Origination
40
45
Monthly Defaults
0 .5 1 1.5 2 2.5
Monthly Prepayments
0
.5
1
Panel C: Subprime Defaults
0
10
20
30
Months Since Origination
40
No Chg in DTI, UR, or HP
Raise UR by 2 pct pts
Panel D: Subprime Prepayments
0
10
20
30
Months Since Origination
Raise DTI by 10 pct pts
Reduce HP by 10 pct
40
Panel B: 30 Days Late
.5
Panel A: Current Loans
30
30
.01
Fraction
.2
.3
.4
Fraction
.015 .02
.025
Figure 7: Roll Rates by Initial Delinquency Status
60
2003m1
2005m1
2007m1
2001m1
2005m1
2007m1
60
30
.8
Panel D: 90 Days Late
90
Fraction
.2
.4
2003m1
90
Fraction
.2
.4
.6
.6
Panel C: 60 Days Late
Fore
Cur
2001m1
2003m1
2005m1
2007m1
0
Cur
0
46
2001m1
.1
.005
Cur
2001m1
2003m1
60
30
2005m1 2007m1
Table 6: Direct Defaults as a Share of All Defaults, by Year of Origination and Year of Default
Year of
Year of Default
Origination 2003 2004 2005 2006 2007 2008
Panel A: All States
2003 11.4 10.8 15.3 14.1 18.0 28.1
8.2 12.5 14.5 22.7 34.4
2004
2005
11.2 15.8 31.4 44.6
2006
11.7 25.7 44.0
25.5 39.7
2007
2008
38.1
All Orig. Years 11.4 10.3 13.1 14.5 26.6 41.6
Panel B: AZ, CA, FL, & NV
2003 9.4
5.0
5.4
8.7 20.9 41.0
2004
7.3
6.6 13.7 33.9 49.3
2005
5.0 18.1 43.9 57.7
2006
12.5 34.4 55.8
2007
31.3 53.9
47.1
2008
All Orig. Years 9.4
5.5
6.0 15.4 36.8 55.1
Panel C: 47 Remaining States
2003 11.8 11.9 16.3 14.6 17.6 24.9
8.4 13.4 14.7 19.9 27.3
2004
2005
12.6 15.0 23.0 31.5
2006
11.3 19.8 31.6
2007
22.0 29.6
2008
34.5
All Orig. Years 11.8 11.2 14.1 14.3 20.7 30.4
Total
16.4
20.4
32.1
34.9
37.4
38.1
30.8
16.8
30.5
46.2
46.5
50.6
47.1
44.8
16.4
17.8
22.9
25.0
28.3
34.5
22.7
Table 7: Modification Statistics by Type: 2007:Q1–2008:Q4
# Loans
Modified
2007:Q1
2007:Q2
2007:Q3
2007:Q4
2008:Q1
2008:Q2
2008:Q3
2008:Q4
10,940
14,600
17,720
27,150
36,230
44,750
62,190
74,800
Interest Rate
Principal Balance
Reductions
Reductions
#
(% total)
#
(% total)
600
5.3
700
6.2
820
5.4
550
3.7
770
4.1
810
4.3
2,990
9.7
700
2.3
6,010
13.8
900
2.1
9,050
16.4 1,300
2.4
16,280
20.3
940
1.2
28,630
26.7 1,450
1.4
Principal Balance Term Extensions
Increases
#
(% total)
#
(% total)
8,660
76.4 1,380
12.2
11,630
77.3 2,050
13.6
15,170
81.2 1,940
10.4
22,520
72.8 4,740
15.3
32,100
73.8 4,500
10.3
39,750
72.1 5,030
9.1
56,940
70.9 6,110
7.6
65,960
61.5 11,230
10.5
Notes: These statistics were computed using a 10 percent random sample of the LPS data. Quantities
obtained from the data are multiplied by a factor of 10. The percentages are taken with respect to the total
number of modifications, and not loans modified. Thus, there is double-counting in the sense that some
loans received multiple types of modifications in a given quarter.
47
Table 8: Modification Statistics by Loan Holder
Panel A – All Loan Types
Modification % of
all loans outstanding
30 days delinquent or worse
60 days delinquent or worse
2008:Q1 2008:Q2 2008:Q3 2008:Q4 2008:Q1 2008:Q2 2008:Q3 2008:Q4 2008:Q1 2008:Q2 2008:Q3 2008:Q4
GNMA
0.04
0.04
0.03
0.03
0.39
0.36
0.31
0.27
0.80
0.74
0.64
0.51
0.10
0.06
0.05
0.04
2.32
1.30
0.88
0.61
4.87
2.51
1.58
1.03
FNMA
FHLMC
0.05
0.05
0.16
0.23
1.95
1.75
4.72
5.26
4.56
3.74
9.06
9.30
Private Securitized
0.55
0.84
1.25
1.42
3.45
4.63
6.28
6.23
5.03
6.41
8.49
8.31
0.53
0.65
0.69
1.05
6.31
7.53
6.81
8.55
10.23
11.47
10.32
12.57
Portfolio
Panel B – Subprime/Alt-A Loans (LPS Defintion)
Modification % of
all outstanding loans
30 days delinquent or worse
60 days delinquent or worse
2008:Q1 2008:Q2 2008:Q3 2008:Q4 2008:Q1 2008:Q2 2008:Q3 2008:Q4 2008:Q1 2008:Q2 2008:Q3 2008:Q4
FNMA
0.80
0.42
0.37
0.19
3.42
1.70
1.32
0.56
6.01
3.05
2.24
0.87
0.23
0.12
2.48
1.70
1.30
0.56
9.59
5.35
2.92
1.18
17.86
8.68
FHLMC
1.59
2.58
4.39
4.56
4.41
6.65
10.41
9.46
6.28
9.11
14.13
12.55
Private Securitized
Portfolio
1.41
2.51
3.97
6.93
3.72
6.23
9.95
15.83
5.21
8.57
13.55
21.75
48
Figure 8: Kaplan-Meier Survival Estimates: Transition from Modification to Default
90-Days Delinquent
1
49
Survival Probability
0.8
Payment Decrease
0.6
All Modifications
0.4
Payment Increase
0.2
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Months Since Modification
17
18
19
20
21
22
23
24
Figure 9: Kaplan-Meier Survival Estimates: Transition from Delinquency to Modification
30 Days Delinquent to Modification: All Mortgages
1
Survival Probability
0.9
0.8
0.7
GNMA
FNMA
FHLMC
Privately Securitized
Portfolio
0.6
11
9
7
5
3
1
13
15
17
19
21
23
25
27
29
31
33
35
Month Since First 30-Day Delinquency
30 Days Delinquent to Modification: Subprime/Alt-A Mortgages
1
Survival Probability
0.9
0.8
0.7
0.6
1
3
5
7
9
11
13
15
17
19
21
23
25
Month Since First 30-Day Delinquency
50
27
29
31
33
35
Figure 10: Default Probability by Year.
The top panel reports foreclosures on loans originated in that year. Loans may be purchase or refinance.
Data comes from Morton (1956). The bottom panel reports foreclosures on homes purchased with mortgages
in that year. For these data, we count a loan as foreclosed if there was a foreclosure on that loan or any
subsequent mortgage to that owner. Thus the probabilities in the lower panel are an upper bound on the
probabilities in the top panel. See Gerardi, Shapiro, and Willen (2007) for details.
Foreclosure probabilities between 1920 and 1947 by origination year
30
Life Insurerց
Probability of default in %
25
20
15
ւS&Ls
10
5
Commercial Banksր
0
1920
1925
1930
1935
Foreclosure probabilities between 1989 and 2008 by origination year
30
Probability of default in %
25
20
15
10
5
0
1990
1995
2000
51
2005
@FedFRASER