Download Non-Owner-Occupancy Misrepresentation and Loan Default

20130224-IA
Internet Appendix for
“Asset Quality Misrepresentation by Financial Intermediaries:
Evidence from RMBS Market”1
TOMASZ PISKORSKI, AMIT SERU, and JAMES WITKIN
This Internet Appendix (IA) contains a more detailed discussion on data construction and
several robustness tests not included in the main text of the paper. It also sketches out a simple
example that is discussed in the paper.
1
Citation format: Piskorski, Tomasz, Amit Seru, and James Witkin, Internet Appendix for
“Asset Quality Misrepresentation by Financial Intermediaries: Evidence from RMBS Market,”
Journal of Finance, doi: 10.1111/jofi.12271. :. Please note: Wiley-Blackwell is not responsible for the
content or functionality of any supporting information supplied by the authors. Any errors (other
than missing material) should be directed to the authors of the article.
I. Data
A. Verifying the Quality of the BlackBox-Equifax Merge
As we mention in Section II of the published article, our primary data set consists of a
merge between (i) loan-level mortgage data collected by BlackBox Logic and (ii) borrower-level
credit report information collected by Equifax. The merge is performed by Equifax using a
proprietary merge algorithm utilizing more than 25 variables. Equifax is one of the three largest
credit bureaus in the U.S. (along with Experian and TransUnion) and collects data on more than
400 million consumers across the world. As such, it is one of the few potential providers for such
detailed consumer credit data, and the firm has built a reputation based on data management. It
provides a variety of analytical tools, from fraud detection to portfolio analytics, to a diverse
group of client companies spanning multiple industry sectors. Given Equifax’s background as a
data services company with over $1.5 billion in annual revenue, it has the expertise to perform
the merge between the BlackBox and Equifax data sets. Nevertheless, as our measures of
underwriting misrepresentation rely on identifying discrepancies between these two data sets, we
independently verify the quality of the merge.
Equifax reports a merge confidence measure that ranges from 0 to 0.90 (ranging from
low to high confidence). The majority of loans fall in the high-confidence buckets, between 0.8
and 0.81 or between 0.89 and 0.90 (see Table IA.I). To further verify the accuracy of the merge,
we compare the dynamic loan payment history reported to BlackBox to that reported to Equifax.
While BlackBox provides loan-level data, the Equifax data are reported at the consumer level
with statistics such as delinquencies either aggregated by debt product (first liens, HELOCs,
credit cards, etc.) or separated by account size and product (largest first lien, second-largest first
lien, etc.). We restrict our analysis to consumers with only one first mortgage in Equifax so that
we can compare the correct payment records.2 Additionally, we limit our analysis to observations
for which the loan’s status in BlackBox is either current or 30, 60, or 90+ days delinquent, as it
can be difficult to correctly map other BlackBox statuses such as bankruptcy or prepayment to a
“correct” status in Equifax. We report the results in Table IA.II.
We find that for the entire sample, the BlackBox and Equifax payment statuses exactly
match 93.8% of the time and are within 30 days of each other 98.0% of the time. We include the
second measure to account for potential lags in reporting across the two databases. We observe
that status-matching rates increase monotonically with Equifax’s confidence measure, with only
0.3% of the highest-confidence loans having significantly different BlackBox and Equifax
delinquency statuses. It is worth noting that among loans with a confidence measure of 0.89 or
greater, about 16% of loans do not have matching BlackBox and Equifax zip codes. 3 This
2
Consider a borrower that has two first mortgages: we would observe the delinquency status of both mortgages, but
we would have to make assumptions regarding which status corresponded to the matched BlackBox loan.
3
This group of loans includes mortgages reported to investors as for non-owner-occupied properties.
2
reinforces the view that Equifax uses unique identifiers in performing its merge, given the
difficulty of performing a merge across millions of loans without requiring a match on zip code.
Based on these findings, we conclude that the most reliable subsample of the merged data
contains loans with a confidence of 0.89 or higher (44.4% of the sample). For the analysis of
misreported second liens, this results in a sample of 854,959 mortgages, down from 1,741,606
loans originated over the 2005 to 2007 period that are reported as having no second liens in
BlackBox (a 51.41% reduction). For the analysis of owner-occupancy misreporting (see Section
II for discussion of owner-occupancy misrepresentation), this results in a sample of 1,563,223
mortgages, down from 3,549,858 loans originated over the 2005 to 2007 period that are reported
to BlackBox as owner-occupied (a 55.96% reduction).
Next, we examine how restricting attention to loans with a high Equifax merge
confidence impacts our sample. Panel A of Table IA.III reports statistics for loans for which
BlackBox reports that the CLTV and LTV are equal. The first two columns report these statistics
without any restrictions, while the last two columns report the statistics for the high-quality
merged sample. Again, imposing these restrictions leads to a sample with slightly higher-quality
observables. FICO scores and origination balances increase, while interest rates and CLTV
decrease. Figure IA.1 displays kernel density plots comparing the merged sample with a sample
of loans with no reported second liens in BlackBox that do not meet the Equifax merge
confidence restrictions. Again, the shapes of the distributions are quite similar, demonstrating
that these two samples are also not meaningfully different.
In Panel B of Table IA.III we display summary statistics for loans reported as owneroccupied in the BlackBox data set to which we apply our second measure of asset quality
misrepresentation. The first two columns report these statistics for all loans in the BlackBox data
set that are reported as owner-occupied. The last two columns report statistics for the subsample
of these loans limited to mortgages with the highest Equifax merge confidence (0.89 or higher).
On average, loans in the high-quality merged sample have somewhat higher-quality observable
characteristics. In particular, these loans have lower interest rates, lower CLTV ratios, and higher
FICO scores and origination balances. Figure IA.2 displays kernel density plots comparing this
high-quality merged sample to the sample of BlackBox loans with reported owner-occupied
status that do not meet the merge confidence restriction. We see that despite the differences in
mean values, the two samples have similar underlying distributions of important observable
characteristics.
Finally, we conduct a placebo test to assess whether there is a relationship between
incorrectly merged records and subsequent loan performance. To identify incorrectly merged
records, we use the origination loan balances that Equifax reports each month for the borrower’s
two largest active first mortgages. We compare these two amounts over the loan’s first six
observations to the origination loan balance reported by BlackBox to construct Balance
Mismatch, which takes the value of one if neither Equifax balance is ever within 2% of the
3
BlackBox origination balance, and zero otherwise.4 We note that the balance of the securitized
first mortgage is unlikely to be misreported to investors because servicers verify and report on a
monthly basis outstanding loan amount and payments to the securitization trust; hence, such
records may indicate incorrectly merged loans across the two databases. Because only the two
largest mortgage balances are reported in the Equifax database, we remove borrowers that have
three or more first mortgages reported in their credit file. The reason is that we are not sure if the
loan in BlackBox would be part of the two largest mortgage balances reported in the Equifax
database.
We find that 2.01% of the loans have mismatched origination loan balances. To examine
the relationship between default likelihood and incorrectly merged records, we run an OLS
regression in which the dependent variable is the same 90-day delinquency dummy variable used
in Table V of the published article. The right-hand variables include Balance Mismatch, the two
previous misrepresentation measures, origination cohort fixed effects, and the same vector of
controls used in prior loan-level regressions. Column (2) clusters standard errors by state level.
Table IA.IV reports the results. While we observe a small positive coefficient on Balance
Mismatch, the effect is economically small relative to what we find for our primary variables.
Moreover, the effect is not statistically significant when standard errors are clustered at the state
level. This suggests that there is no economically meaningful relationship between the potential
merge accuracy of the two databases and subsequent loan performance. Importantly, the
inclusion of Balance Mismatch does not reduce the economic or statistical significance of the
misrepresentation variables. This evidence provides further support that our methodology to
identifies actual misrepresentations other than incorrectly merged records.
B. Constructing Underwriter-Level Measures
In this subsection we discuss construction of the underwriter-level variables used in Table
X of the published article. The measure of the relative importance of the RMBS underwriting
business is constructed by dividing the aggregate dollar value of nonagency RMBS underwritten
in 2005 by the underwriter’s total assets as of the end of 2005, as reported by Compustat. This
variable has a mean of 0.243 and a standard deviation of 0.334. Years of experience in
underwriting subprime deals is a deal-level variable capturing the underwriter’s experience in the
subprime MBS market. To construct this measure, we take each pool’s securitization year and
subtract the year in which the deal’s underwriter entered the subprime MBS market, using
ABSNet’s universe of subprime MBS deals from 1999 onward. This variable has a mean of 5.98
and a standard deviation of 1.45.
The commercial bank underwriter variable, which accounts for differences between
commercial and investment banks, is a dummy that takes a value of one if the underwriter’s
charter identifies it as a commercial bank and zero otherwise. The underwriter bonus-to-salary
4
Results are quantitatively similar regardless of how we define Balance Mismatch with respect to both the sixmonth time window and the strictness of the balance match (within 2%, 5%, $100, etc.).
4
ratio is computed by dividing the cash bonus of each underwriter’s five highest-paid executives
by their base salary in 2005. These data comes from Computstat’s Execucomp database where
available, and otherwise are constructed using firms’ public financial reporting. This variable has
a mean of 14.98 and a standard deviation of 15.53. The Risk Management Index (RMI), based on
Ellul and Yerramili (2013), is computed as the first principal component of the following six risk
management variables: Credit Risk Officer (CRO) is present, CRO is an executive, CRO is a
top-five executive, CRO compensation centrality, risk committee experience, and active risk
committee.
5
Table IA.I
Distribution of Equifax Merge Confidence
The table shows the distribution of loans in the BlackBox data set across Equifax’s merge confidence level. The
highest merge confidence of a given loan is 0.90.
Merge Confidence
Number
of Loans
[0.00,0.80)
3.6%
[0.80,0.81)
41.8%
[0.81,0.89)
10.2%
Highest Confidence Available:
[0.89,0.90)
44.4%
Total
100%
Table IA.II
Delinquency Status Matching by Equifax Merge Confidence
The table shows the fraction of observations with delinquency statuses that match or are within 30 days of each
other in BlackBox and Equifax. The sample is restricted to borrowers with only one first mortgage record in Equifax
and to observations for which the loan’s status in BlackBox is either current or 30, 60, or 90 days delinquent. A
“Status Match” is defined as an exact match between the reported BlackBox and Equifax delinquency statuses.
“Status within 30 Days” is defined as two statuses that are less than 60 days apart, for example, 90 days delinquent
in BlackBox and 60 days delinquent in Equifax.
Merge Confidence
Fraction Payment Status
Match
Fraction Payment Status
Within 30 Days
[0.00,0.80)
0.841
0.924
[0.80,0.81)
0.904
0.964
[0.81,0.89)
0.942
0.991
[0.89,0.90)
0.976
0.997
6
Table IA.III
Comparison of Descriptive Statistics
The table presents a comparison of summary statistics for key variables. The first two columns of each panel present
statistics for all loans in a given sample, and the second two columns restrict attention to loans that have the highest
merge confidence. Panel A presents these statistics for loans reported as having no second lien to the RMBS trustee,
while Panel B presents these statistics for mortgages reported as for owner-occupied properties.
Interest Rate
FICO
Balance
CLTV
Purchase
No Cash Out Refi
Cash Out Refi
Low or No Doc.
ARM
Option ARM
Number of Loans
Interest Rate
FICO
Balance
CLTV
Purchase
No Cash Out Refi
Cash Out Refi
Low or No Doc.
ARM
Option ARM
Number of Loans
Panel A: Sample of Loans Reported as Having No Second Liens
Full Sample
Merged Sample
Mean
SD
Mean
SD
6.925
2.236
6.458
2.424
652.8
76.48
663.5
76.90
277.6
225.3
293.1
238.9
79.79
10.03
78.93
9.955
0.366
0.482
0.361
0.480
0.117
0.321
0.121
0.326
0.506
0.500
0.507
0.500
0.472
0.499
0.514
0.500
0.597
0.491
0.475
0.499
0.113
0.317
0.169
0.375
1,741,606
854,959
Panel B: Sample of Loans Reported as for Owner-Occupied Properties
Full Sample
Merged Sample
SD
Mean
SD
6.990
1.947
6.605
2.024
662.6
74.95
680.2
73.20
291.0
254.2
324.6
248.3
84.33
11.44
83.72
11.71
0.483
0.500
0.496
0.500
0.149
0.356
0.136
0.343
0.360
0.480
0.358
0.480
0.419
0.493
0.463
0.499
0.606
0.489
0.527
0.499
0.0991
0.299
0.103
0.304
3,549,858
1,563,223
7
Table IA.IV
Balance Mismatch Placebo Table
The table presents OLS estimates from regressions in which the dependent variable takes the value of one if the loan
defaults (90+ days delinquent) in the first two years since origination and zero otherwise. Balance Mismatch takes a
value of one if the origination loan balance reported by BlackBox is not within 2% of any first-mortgage origination
loan balance reported to Equifax during the loan’s first six months, and zero otherwise. Misreported Second takes a
value of one if the loan is characterized as having a second-lien misrepresentation and zero otherwise. Misreported
Non-Owner-Occupant takes a value of one if the loan is characterized as having a non-owner-occupancy
misrepresentation and zero otherwise. Other Controls include origination variables such as FICO score, interest rate,
and LTV ratio. Squared and cubed terms for FICO and CLTV are included to account for potential nonlinear effects.
The estimates are in percentage terms; standard errors are in parentheses. * p < 0.10, ** p < 0.05, and *** p < 0.01.
(1)
0.709***
(0.192)
(2)
0.709
(0.443)
Misreported Second
8.214***
(0.16)
8.214***
(0.859)
Misreported Non-Owner-Occupant
7.485***
(0.124)
7.485***
(0.54)
Other Controls
Yes
Yes
Half-Year Origination Cohorts
Yes
Yes
No
1487121
14.44
2.010
0.155
Yes
1487121
14.44
2.010
0.155
Balance Mismatch
SEs Clustered by State
Number of Loans
Percent 90 DPD
Percent Balance Mismatch
R-Squared
8
Panel A: Origination CLTV Ratios
Panel B: Origination FICO Scores
Panel C: Origination Interest Rates
Panel D: Origination Loan Balance
Figure IA.1. Comparability of loans with no second-lien reported by merge restrictions – kernel density of observables. The figure shows the
kernel density plots of loan origination CLTV ratios, FICO credit scores, interest rates, and loan balances for loans reported as not having any second liens in the
BlackBox data set. The group of loans meeting the Equifax merge restrictions is represented by the solid line, and the group that fails these restrictions is
represented by the dashed line.
9
Panel A: Origination CLTV Ratios
Panel B: Origination FICO Scores
Panel C: Origination Interest Rates
Panel D: Origination Loan Balance
Figure IA.2. Comparability of loans reporting owner-occupied status by merge restrictions – kernel density of observables. The figure shows the kernel
density plots of loan origination CLTV, FICO credit scores, interest rates, and loan balances for loans reported as owner-occupied in the BlackBox data set. The
group of loans meeting the Equifax merge restrictions is represented by the solid line, and the group that fails these restrictions is represented by the dashed line.
10
II. Non-Owner-Occupancy Misrepresentation
A. Construction of Non-Owner-Occupancy Misrepresentation Measure
In this section we consider a second measure of asset quality misrepresentation. This
measure concerns the occupancy status of the property backing the loan. We identify a loan as
misreported if the disclosure to investors shows that the loan is collateralized by owner-occupied
property, when in fact the credit bureau data show that the property was owned by borrowers
with a different primary residence (i.e., is an investment property or a second home).
Prospectuses also make statements related to asset quality on this dimension. 5
To construct this measure we start with our base sample and consider only loans reported
as owner-occupied properties to investors at the time of origination. This gives us a sample of
1,563,223 loans. We then take advantage of the mailing address zip code reported to Equifax to
construct our measure. The identifying assumption here is that a zip code reported to a credit
bureau, such as Equifax, should generally match the property address zip code reported to the
trustee of the RMBS for owner-occupied homes. To allow for reporting delays, we follow a very
conservative approach and compare the Equifax zip code reported each month over the first year
of the loan’s life to the property zip code reported to investors. If none of these 12 Equifax zip
codes matches with the BlackBox data set, we classify such a loan as having a Misreported NonOwner-Occupant.
In addition to ensuring that the match between the two data sets, done by the credit
bureau, is of high quality we have to address another issue regarding our second measure of asset
quality misrepresentation. In particular, even if the match between the two databases is perfect,
only our first measure – Misreported Second-Lien – allows us to directly identify asset
misrepresentation. For the second measure, which is based on occupancy status, the
misrepresentation is inferred since we do not observe occupancy status in the Equifax database.
Instead, we infer whether a property is owner-occupied if the mailing zip code reported in the
credit bureau data matches the property zip code disclosed to investors. Consequently, we need
to verify that the measure based on occupancy status is reasonable.
Again, our measure appears to be reasonable for several reasons. First, among loans
reported as for non-owner-occupied properties, the majority (about two-thirds) have a mailing
address zip code in the Equifax database that does not match the property address in the
BlackBox database.6 Similarly, a majority (about 70%) of reported owner-occupants residing in
the same zip code have only one first mortgage on file in the Equifax data. These features
For instance, the prospectus for Series 2006-FF15 states: “The prospectus supplement will disclose the aggregate
principal balance of Mortgage Loans secured by Mortgaged Properties that are owner-occupied.” The prospectus
supplements also usually describe the frequency of loans for second homes or investor properties. For example, the
supplement for the Wells Fargo MBS Series 2007–8 states that “Approximately 0.02% of the mortgage loans in the
mortgage pool are expected to be secured by investor properties.”
6 About a third of properties whose owner occupancy status is reported to a trustee of the RMBS as being nonowner-occupied likely have both properties in their zip code of residence – a fact we corroborated to be potentially
true based on our conversations with several industry practitioners.
5
11
underscore our measure of owner-occupancy misrepresentation. In addition, a majority (more
than 60%) of misreported owner-occupants have multiple first mortgages on their credit file,
indicating that they own multiple properties and are likely to be real estate investors.7
Second, ex post delinquencies of loans we identify as misrepresented on the dimension of
owner occupancy are significantly higher when compared with otherwise similar loans. Finally,
we do not detect owner-occupancy misrepresentations for cases in which the loan is truly made
to a non-owner-occupant who resides in the same zip code as the property backing the loan, and
we discard cases in which the zip code of the property differs from that in the credit bureau data
for a period less than 12 consecutive months. It is worth noting that, for the above reasons, we
will likely classify loans that may be misrepresented as not having been misrepresented. This
should bias against finding that misrepresented loans have higher ex post delinquencies relative
to otherwise identical loans.
B. The Extent of Non-Owner-Occupancy Misrepresentation and Its Relation with Loan
Performance and Mortgage Rates
Table IA.V displays summary statistics for loans with reported owner-occupied status. As
can be seen, 6.42% of mortgages reported as financing owner-occupied properties were given to
borrowers with a different primary residence, and thus were misrepresented based on our method.
Stated differently, 27% of loans obtained by non-owner-occupants misreported their true purpose.
As in the case of the second-lien misrepresentation, the prevalence of owner-occupancy
misrepresentation is higher for loans used to purchase the properties. Among the purchase loans
with reported owner-occupied status, about 9.6% misrepresent owner occupancy compared with
only 3.1% for refinance mortgages. There are a number of reasons why misrepresentation on this
dimension could be more prevalent for purchase loans. For example, it may be harder to
misrepresent the occupancy status for refinancing loans, especially if borrowers are involved, if
the borrower does not reside in the property that is his or her primary residence.
Overall, as there is some overlap in both types of misrepresentation across mortgages, our
measures imply that about 9.1% of loans have misreported second-lien or misreported owneroccupancy status (or both). This estimate would be about 12.2% were we to include HELOCs
when inferring misreported second liens.
There is significant regional variation in the prevalence of owner-occupancy
misrepresentations across the U.S. Areas where owner-occupancy misrepresentation is most
prevalent correspond to markets that were booming leading up to the crisis. This is consistent
with Chinco and Mayer (2012), who show significant activity of non-owner-occupants acquiring
residential real estate in these regional markets for speculation prior to the recent housing crisis.
Our measure indicates that a significant portion of the purchases made by non-owner-occupants
in these regions was financed with mortgages that misrepresented their true purpose. For
7
This is also true for reported non-owner-occupants, as about 80% have multiple first mortgages on the credit file.
12
example, in our data about 22.5% of loans in Florida were reported to be non-owner-occupant
loans. At the same time, according to our measure, about 9.5% of mortgaged properties in
Florida reported as owner-occupied were actually non-owner-occupied properties. This implies
that about a third of loans obtained by non-owner-occupants in Florida misreported their true
purpose. Rural states such as Nebraska and South Dakota have some of the lowest fractions of
loans that misrepresent owner occupancy status (2.5% and 2.8%, respectively). This lower
incidence of misrepresentation is most likely tied to the less pronounced presence of non-owneroccupant borrowers in these regional markets.
Table IA.VI shows that loans with misrepresentation on owner-occupancy status are
more likely to have higher CLTVs. Moreover, these loans often tend to be option ARMs8 and
have low or no additional documentation provided. We note that the relation between
observables and asset misrepresentation also seems economically meaningful. For instance, the
probability of asset misrepresentation based on the owner-occupancy status of a loan increases
by more than 7% (more than 100% relative to the mean) if the loan contract is an option ARM.
Finally, as before, misrepresentation of owner-occupancy status is much less common among
refinance loans. Broadly, the features that are positively related to misrepresentation of
occupancy status are typically associated with purchase mortgages of the Alt-A type of credit
risk.9
Although the addition of origination cohort controls does not seriously impact the sign or
significance of the regression coefficients, the fixed effects themselves illustrate an interesting
trend. Column (2) of Table IA.VI indicates that the average propensity of owner-occupancy
misrepresentation peaks in the first half of 2007 and then drops sharply thereafter. These timeseries patterns are consistent with models such as those of Povel, Singh, and Winton (2007) and
Hartman-Glaser (2013), who predict that misrepresentations are most likely to occur in relatively
good times and can grow over time. Our findings also suggest that the tightening of credit
standards just before the collapse of the nonagency market may have led to increased scrutiny.
Overall, misreported loans have characteristics that are broadly expected of similar loans
with truthful reporting. As we show below, even after accounting for characteristics reported to
investors, there are significant differences in the performance of misrepresented loans on the
dimension of owner-occupancy relative to similar loans with no misrepresentation. This indicates
that reported characteristics of mortgages, including interest rates on mortgages, may not have
fully accounted for the riskiness induced by misrepresentation.
8
Unlike more traditional fixed-rate mortgages (FRMs) or ARMs, an option ARM is an adjustable-rate mortgage that
lets borrowers pay only the interest portion of the debt or even less than that, with the loan balance growing above
the amount initially borrowed up to a certain limit (see Piskorski and Tchistyi (2010)).
9 The Alt-A risk category consists of loans that, for various reasons (e.g., low documentation), loan originators
consider riskier than prime loans but less risky than subprime mortgages.
13
As in the case of second-lien misrepresentation, we examine the relation between
subsequent loan performance and misrepresentation of non-owner-occupancy status using the
following loan-level specification:
(IA.1)
𝑌𝑖 = 𝛼 + 𝛽𝑋𝑖 + 𝛾𝛿𝑖 + 𝜀𝑖 ,
where the dependent variable is a dummy that takes a value of one if the mortgage goes 90 days
past due on payments during the first two years since origination and zero otherwise. The main
explanatory variable is the dummy Misreported Non-Owner-Occupant (𝛿𝑖 ), which takes the
value of one if loan i has misrepresentation on owner occupancy as per our method and zero
otherwise.
As we see from Table IA.VII, conditional on all the other observables, loans with
misrepresented owner occupancy are associated with a higher likelihood of subsequent
delinquency. This estimate is also economically significant: a loan with misrepresented borrower
occupancy status has about a 9.4% higher likelihood of default, compared with loans with similar
characteristics but with truthfully reported owner-occupancy status of the borrower. This
estimate implies that loans with misreported owner-occupancy status have more than a 60%
higher default rate relative to the mean default rate of owner-occupants.
The positive association between Misreported Non-Owner-Occupant and the delinquency
of the loan holds across all origination cohorts, and as in the case of misrepresented second liens,
the absolute effect grows over time. In terms of economic magnitudes, the effect is smallest for
loans originated in 2005. However, even in 2005, the estimate suggests an increase in absolute
delinquency of 4.04% for loans misrepresenting owner occupancy, implying a relative increase
of more than 56% when compared with the mean default rate of 7.15% for loans with truthfully
reported owner-occupancy status of borrowers.
So far we show that mortgage defaults on loans with misreported owner-occupancy status
are significantly higher than similar loans truthfully disclosed as having owner-occupant
borrowers. As in the case of second-lien misrepresentation, we now run another comparison. In
Panel B of Table IA.VII, we also examine whether loans with misreported owner occupancy
perform differently than loans with truthfully reported non-owner-occupant borrowers. To do so,
we expand the sample to include loans backed by properties reported as either second homes or
investor properties. We then estimate regressions similar to the above specification where the
dependent variable is the same delinquency dummy variable used earlier. The two variables of
interest are Misreported Non-Owner-Occupant and Reported Non-Owner-Occupant, a dummy
that takes a value of one if the loan is reported as secured by a second home/investor property.
The results show that Reported Non-Owner-Occupant has a positive and significant
coefficient in the default regression, implying an approximate 3.5% absolute increase in the
default rate for these loans (about 23% relative increase in the default rate compared with loans
with owner-occupant borrowers). This result is in line with prior work that finds that, all else
equal, loans with non-owner-occupancy status are more likely to default (see Mayer, Pence, and
14
Sherlund (2009) and Haughwout et al. (2011)). Such borrowers are likely to default more quickly
on homes that are not their primary residence, for reasons such as weaker neighborhood and
social ties.
More interestingly, Misreported Non-Owner-Occupant has a positive and much larger
coefficient in the default regression, implying an approximately 9.5% absolute increase in the
default rate for these loans (about a 64% increase in the default rate relative to loans with owneroccupant borrowers). These results show that loans with misreported non-owner-occupancy
status perform worse not only with respect to loans of owner-occupants, but also relative to loans
of non-owner-occupants that are truthfully reported to investors. Thus, misrepresentation on the
dimension of owner occupancy indicates borrowers with significantly lower-quality
unobservable characteristics, even compared with truthfully reported non-owner-occupants.
We also investigate how the impact of borrower occupancy status on default varies over
the loan’s life. Figure IA.3 plots the resulting cumulative delinquency rates over time for
properties with different borrower occupancy status, holding all other observables at the overall
sample mean. Regardless of the age of the loan, those with misreported non-owner-occupants
consistently default at higher levels than loans with truthfully reported non-owner-occupants and
loans with truthfully reported owner-occupants.
Next we investigate whether non-owner-occupancy misrepresentation is related to the
mortgage interest rates charged by lenders. To do so, we estimate specifications for the mortgage
interest rate with Misreported Non-Owner-Occupant and Reported Non-Owner-Occupant. In
Table IA.VIII we find that Reported Non-Owner-Occupant has a positive and significant
coefficient. This result is in line with the literature that finds that loans with non-owneroccupancy status are perceived to be riskier than those for owner-occupied properties. More
interestingly, however, the coefficient on Misreported Non-Owner-Occupant is also positive and
significant. This suggests that, to some extent, asset misrepresentations on this margin were
captured in interest rates charged by lenders. Note, however, that the increase in interest rates for
misreported non-owner-occupants is smaller than that for truthfully reported ones (by around 13
bps), even though we show that loans with misreported non-owner-occupants default
significantly more than loans of truthfully reported non-owner-occupants.10
Overall, the evidence suggests that lenders were at least partly aware of the higher risk of
misrepresented loans, because they charged higher interest rates on these loans. However, the
interest rate markups on misrepresented loans are significantly smaller relative to loans with
similar or even lower default risk – that is, mortgages truthfully reported to have a second lien or
10
To give an alternative perspective on these results, we also analyze the distribution of variation in interest rates
across loans grouped by second-lien status and owner-occupancy status, controlling for other risk characteristics.
Consistent with our earlier results, loans that truthfully report non-owner-occupant status receive higher interest
rates than those of misreported non-owner-occupants, which in turn have interest rates that are higher than loans that
are truthfully reported as owner-occupants (see Panel B of Figure IA.4).
15
to have a non-owner-occupant. Our results therefore suggest that, relative to prevailing interest
pricing, interest rates on misrepresented mortgages did not fully reflect their higher default risk.
These results are consistent with those in Table IA.VII indicating that misrepresented loans had
higher default rates, even after conditioning on a host of observable characteristics including
interest rates.
C. Where Did Non-Owner-Occupancy Misrepresentation Occur?
The data we use so far do not allow us to directly investigate where in the supply chain of
credit (i.e., borrower, lender, and/or underwriter) the misrepresentations took place. To shed
more light on this question, we again employ the New Century data to investigate owneroccupancy-status misrepresentation. Recall that the advantage of this data set is that it contains
the loan characteristics that were internally recorded by the lender at loan origination. These data
are unlikely to have been manipulated, since New Century entered bankruptcy quite quickly at
the start of the crisis, and an independent trustee sold this data set during the bankruptcy process.
The data allow us to assess whether New Century internal records correctly reflected non-owner
occupancy status for loans that were misrepresented to investors on this dimension.
Table IA.IX indicates that in the merged New Century sample, none of the loans we
identify as having misreported non-owner-occupied status in Equifax was reported as for nonowner-occupied properties in the New Century data. Moreover, virtually all the loans that we
identify as misrepresenting borrower occupancy status in the Equifax data were originated by
brokers selling to New Century (more than 99%)—a much larger proportion relative to the
overall percentage of broker-originated loans by New Century (about 65%).
This evidence suggests that the misrepresentation concerning owner-occupancy status
was made early in the origination process, possibly by the borrower or broker originating the
loan on behalf of New Century. 11 Recall that our previous results show that loans with
misrepresented owner-occupancy status had much higher default rates compared with loans that
correctly reported non-owner occupancy status. Thus, if owner-occupancy misrepresentation is
occurring at the borrower/broker level, the default results suggest that lenders may have been
largely unaware of such misrepresentation and thus faced an adverse selection of borrowers.
We further investigate the question of which party was aware of misrepresentation by
assessing the credit history of borrowers associated with non-owner-occupancy
misrepresentation. In particular, we exploit borrower credit card payment histories provided by
Equifax and create a dummy variable that takes a value of one if the borrower is 60 days past due
on any credit card account in the first two years following mortgage origination. We then
estimate regressions that investigate how credit card defaults relate to our misrepresentation
measures.
11
Berndt et al. (2010) analyze broker compensation using New Century data and find that brokers earned high
profits on loans that turned out to be riskier ex post.
16
Table IA.X displays these results. We find that borrowers obtaining loans with
misreported owner-occupancy status are about 7.6% more likely to default on their credit cards
compared with borrowers who were truthfully reported as owner-occupants. Importantly,
truthfully reported non-owner-occupants have a similar likelihood of default on credit cards as
truthfully reported owner-occupants. In contrast, borrowers associated with second-lien
misrepresentation had somewhat lower credit card defaults compared to borrowers with reported
seconds. These results, taken together with our previous findings, reaffirm our prior result that
financial institutions may have been unaware of the riskiness of borrowers in the case of owneroccupancy misrepresentation.
Finally, we investigate the prevalence of non-owner occupancy misrepresentations across
underwriters. To do so, we use specifications similar to Table IA.VI with Misreported NonOwner Occupant and Misreported Either as dependent variables with underwriter fixed effects in
addition to the control variables. The Misreported Either variable takes the value of one if the
loan has a misreported second lien, a misreported non-owner-occupant borrower, or both.
In Figure IA.4, we plot the estimated level of underwriter misrepresentation, calculated
using the underwriter fixed effects obtained from this regression and fixing other controls at their
means, along with the 95% confidence interval. Panel A presents results for misrepresentation on
the dimension of owner occupancy while Panel B presents results using both second-lien and
owner-occupancy misrepresentation. The omitted category in these figures is Credit Suisse. We
find that 4.43% of loans misrepresent the owner-occupancy status of the borrower, and 5.69%
have either type of misrepresentation.
The figure shows significant heterogeneity across underwriters in the propensity to
securitize misrepresented loans. In particular, both panels of Figure IA.4 indicate sizeable
variation across underwriters in the share of their loans having either type of misrepresentation.
Importantly, all underwriters have misrepresentation levels significantly greater than zero, with
pools underwritten by Countrywide and Lehman being the worst in terms of misrepresentations.
Overall, although there is substantial heterogeneity across underwriters, a significant
degree of misrepresentation exists across all underwriters, including the most reputable financial
institutions. The result of a significant extent of asset quality misrepresentation on a dimension
other than second-lien reporting confirms that our estimates regarding second-lien
misrepresentation establish a lower bound for the overall amount of misrepresentation in the
RMBS market.
17
References
Berndt, Antje, Burton Hollifield, and Patrik Sandås, 2010, The role of mortgage brokers in the
subprime crisis, NBER Working paper 16175.
Chinco, Alexander, and Chris Mayer, 2012, Distant speculators and asset bubbles in the housing
market, Working paper, [affiliation].
Ellul, Andrew, and Vijay Yerramilli, 2013, Stronger risk controls, lower risk: Evidence from U.S.
bank holding companies, Journal of Finance 68, 1757-1803.
Hartman-Glaser, Barney, 2013, Reputation and signaling in asset sales, Working paper, SSRN.
Haughwout, Andrew, Donghoon Lee, Joseph Tracy, and Wilbert van der Klaauw, 2011, Real
estate investors, the leverage cycle, and the housing market crisis, Federal Reserve Bank
of New York Staff Reports, 514.
Mayer, Christopher, Karen Pence, and Shane Sherlund, 2009, The rise in mortgage defaults,
Journal of Economic Perspectives 23, 27-50.
Piskorski, Tomasz, and Alexei Tchistyi, 2010, Optimal mortgage design, Review of Financial
Studies 23, 3098-3140.
Povel, Paul, Rajdeep Singh, and Andrew Winton, 2007, Booms, busts, and fraud, Review of
Financial Studies 20, 1219-1254.
18
Table IA.V
Descriptive Statistics and Percent of Loans with Misreported
Non-Owner-Occupants
The table presents summary statistics of key variables for mortgages reported as owner-occupied to the RMBS
trustee (loans reported as such in the BlackBox dataset). The sample consists of these loans merged with a high level
of confidence with the credit bureau data. Interest Rate is the loan interest rate at origination in percentage terms.
FICO is the borrower’s FICO credit score at loan origination. Balance is the initial loan balance (in thousands of
dollars). CLTV is the loan’s origination combined loan-to-value ratio in percentage terms reported to investors. No
Cash Out Refi and Cash Out Refi are dummies that take a value of one if the loan purpose was a no cash-out
refinancing or cash-out refinancing, respectively, and zero otherwise. Low or No Doc. is a dummy that takes a value
of one if the loan was originated with no or limited documentation, and is zero otherwise. ARM and Option ARM are
dummies that take a value of one if the loan type was an ARM or option ARM, respectively, and are zero otherwise.
Interest Rate
FICO
Balance
CLTV
Purchase
No Cash Out Refi
Cash Out Refi
Low or No Doc.
ARM
Mean
6.60
680.
324.6
83.72
0.49
0.13
0.35
0.46
0.52
SD
2.02
73.20
248.3
11.71
0.50
0.34
0.48
0.49
0.49
Option ARM
0.10
0.30
Number of Loans
1,563,223
Percent of Loans for Misreported
Non-Owner-Occupants
6.42
Percent of Fully Documented Loans for
Misreported Non-Owner-Occupants
4.79
Percent of Purchase Loans for Misreported
Non-Owner-Occupants
9.67
Percent of Refinance Loans for Misreported
Non-Owner-Occupants
3.12
19
Table IA.VI
Non-Owner-Occupancy Misrepresentation and Loan Characteristics
This table presents OLS estimates from regressions in which the dependent variable takes a value of one if the loan
is Misreported Non-Owner Occupant and zero otherwise. All the controls represent the values reported to investors
(the corresponding values in the BlackBox database). Column (2) adds the fixed effects corresponding to the loan
origination date, with 2005 being the omitted category. Column (3) adds fixed effects corresponding to the location
of the property backing the loan. Column (4) adds fixed effects capturing the identity of the underwriter that sold the
loan. Column (5) reports standard errors clustered at the state level. All estimates are in percentage terms; standard
errors are in parentheses. * p < 0.10, ** p < 0.05, and *** p < 0.01.
(1)
0.00907***
(0.000369)
(2)
0.00902***
(0.000371)
(3)
0.00915***
(0.000372)
(4)
0.0103***
(0.000381)
(5)
0.0103***
(0.00159)
0.00262***
(0.0000872)
0.00256***
(0.0000891)
0.00133***
(0.0000948)
0.00144***
(0.0000955)
0.00144***
(0.000529)
CLTV
0.0274***
(0.00200)
0.0267***
(0.00200)
0.0390***
(0.00201)
0.0378***
(0.00201)
0.0378***
(0.00884)
No Cash Out Refi
-5.499***
(0.0617)
-5.536***
(0.0619)
-5.285***
(0.0622)
-5.231***
(0.0623)
-5.231***
(0.508)
Cash Out Refi
-6.266***
(0.0495)
-6.286***
(0.0496)
-6.465***
(0.0499)
-6.478***
(0.0501)
-6.478***
(0.643)
Low or No Doc.
2.824***
(0.0422)
2.821***
(0.0423)
2.510***
(0.0424)
2.577***
(0.0441)
2.577***
(0.193)
ARM
2.153***
(0.0437)
2.173***
(0.0440)
1.835***
(0.0446)
1.596***
(0.0460)
1.596***
(0.182)
Option ARM
7.229***
(0.0915)
7.126***
(0.0951)
6.824***
(0.0954)
6.592***
(0.0987)
6.592***
(0.468)
Originated in 2006H1
0.0126
(0.0526)
-0.0472
(0.0525)
0.0803
(0.0532)
0.0803
(0.104)
Originated in 2006H2
0.183***
(0.0570)
0.114**
(0.0570)
0.151***
(0.0580)
0.151
(0.0967)
Originated in 2007H1
0.456***
(0.0664)
0.352***
(0.0664)
0.295***
(0.0680)
0.295
(0.192)
Originated in 2007H2
-0.773***
(0.169)
-0.787***
(0.168)
-0.458***
(0.170)
-0.458
(0.286)
No
No
No
1563223
No
No
No
1563223
Yes
No
No
1563223
Yes
Yes
No
1563223
Yes
Yes
Yes
1563223
6.426
0.0300
6.426
0.0301
6.426
0.0346
6.426
0.0359
6.426
0.0359
FICO
Balance
State Fixed Effects
Underwriter Fixed Effects
SEs Clustered by State
Number of Loans
Percent Misrepresented
R2
20
Table IA.VII
Non-Owner-Occupancy Misrepresentation and Loan Default
The table presents OLS estimates from regressions in which the dependent variable takes a value of one if the
mortgage ever defaults (goes 90 days past due on payments) in the first two years since origination, and zero
otherwise. Panel A shows the results with Misreported Non-Owner-Occupant as a control variable. Panel B shows
the results with both Misreported Non-Owner-Occupant and Reported Non-Owner-Occupant as control variables,
where the excluded category is loans truthfully reported as for owner-occupied properties. “Other Controls” include
origination variables reported to investors used in Table IA.VI such as FICO and LTV ratios. Squared and cubed
terms for FICO and LTV ratios are included to account for potential nonlinear effects. The estimates are in
percentage terms; standard errors are in parentheses. * p < 0.10, ** p < 0.05, and *** p < 0.01.
Misreported Non-Owner
Occupant
Panel A: Mortgage Default and Misreported Non-Owner-Occupant
All Years
2005
2006H1
2006H2
9.35***
4.04***
10.13***
15.58***
(0.11)
(0.12)
(0.25)
(0.3)
2007H1
14.00***
(0.39)
2007H2
14.38***
(0.94)
Other Controls
Yes
Yes
Yes
Yes
Yes
Yes
Half-Year Origination
Cohorts
Number of Loans
Percent Default
Percent Misrepresented
R2
Yes
No
No
No
No
No
1,563,223
15.28
6.426
0.157
743,827
7.150
5.904
0.0727
327,128
17.02
6.903
0.108
282,904
26.51
7.167
0.149
187,352
27.36
6.535
0.184
22,012
17.11
6.528
0.214
Panel B: Mortgage Default and Misreported versus Reported Non-Owner Occupants
(1)
(2)
(3)
Misreported Non-Owner
10.72***
9.992***
9.48***
Occupant
(0.11)
(0.11)
(0.11)
(4)
9.48***
(0.71)
4.02***
(0.075)
3.57***
(0.073)
3.45***
(0.074)
3.45***
(0.62)
Other Controls
Yes
Yes
Yes
Yes
Half-Year Origination Cohorts
No
Yes
Yes
Yes
State Fixed Effects
No
No
Yes
Yes
No
1,827,497
14.89
6.424
0.103
No
1,827,497
14.89
6.424
0.153
No
1,827,497
14.89
6.424
0.161
Yes
1,827,497
14.89
6.424
0.161
Reported Non-Owner Occupant
SEs Clustered by State
Number of Loans
Percent Default
Percent Misrepresented
R2
21
Table IA.VIII
Non-Owner-Occupancy Misrepresentation and Mortgage Interest Rates
The table presents OLS estimates from regressions in which the dependent variable is the mortgage interest rate at
origination, with Misreported Non-Owner-Occupant and Reported Non-Owner-Occupant as control variables.
“Other Controls” include origination variables reported to investors such as FICO and LTV ratios. Column (1)
includes only these variables. Column (2) incorporates half-year origination fixed effects. Column (3) also includes
property state fixed effects. Column (4) clusters standard errors by state. The estimates are in percentage terms;
standard errors are in parentheses. * p < 0.10, ** p < 0.05, and *** p < 0.01.
Misreported Non-Owner
Occupant
(1)
0.258***
(0.004)
(2)
0.212***
(0.004)
(3)
0.229***
(0.004)
(4)
0.229***
(0.005)
Reported Non-Owner
Occupant
0.393***
(0.003)
0.365***
(0.003)
0.351***
(0.003)
0.351***
(0.02)
Other Controls
Yes
Yes
Yes
Yes
Half-Year Origination
Cohorts
No
Yes
Yes
Yes
State Fixed Effects
No
No
Yes
Yes
No
1,827,497
6.563
6.424
0.611
No
1,827,497
6.563
6.424
0.674
No
1,827,497
6.563
6.424
0.680
Yes
1,827,497
6.563
6.424
0.680
SEs Clustered by State
Number of Loans
Mean Interest Rate
Percent Misrepresented
R-Squared
22
Table IA.IX
Non-Owner-Occupancy Misrepresentation: Evidence from New Century Loans
Panel A presents summary statistics of key variables for New Century loans with reported owner-occupant status
that have been merged with the BlackBox-Equifax data set. Panel B shows the percentage of these loans for which
the misrepresented non-owner-occupant status is correctly reported in the New Century internal data.
Panel A: Characteristics Based on Owner-Occupancy Status in Merged Data
Reported as
Identified by Us as Having
Owner-Occupant
Misreported Non-Owner-Occupancy
Mean
SD
Mean
SD
Default within 2 Years
0.210
0.408
0.345
0.477
Interest Rate
8.222
1.140
8.236
1.269
FICO
617.7
58.12
656.3
67.32
Number of Loans
3,160
148
Panel B: Non-Owner-Occupancy Misrepresentations in the New Century Internal Database
Percentage
Count
Misreported Non-Owner-Occupant Reported as Non-Owner0%
148
Occupant in New Century Data set
23
Table IA.X
Misrepresentations and Defaults on Credit Cards
The table presents OLS estimates from regressions in which the dependent variable takes a value of one if the
borrower defaults on credit card debt (60 day payment delinquency) in the first two years since mortgage origination,
and zero otherwise. Panel A displays results with Misreported Second x CLTV >= 100, Misreported Second x CLTV
< 100, Reported Second x CLTV >= 100, and Reported Second x CLTV < 100 as controls. The CLTV term
(computed using credit bureau data) takes a value of one if the loan has a CLTV in the appropriate range, and zero
otherwise. Panel B shows the results with Misreported Non-Owner-Occupant and Reported Non-Owner-Occupant
as controls. Column (1) includes FICO and controls related to credit card debt, such as utilization and monthly
payments. Column (2) includes state fixed effects. Column (3) includes full mortgage controls, such as loan balance
and interest rate. Column (4) clusters standard errors by state. Panel B controls for the reported CLTV. The
estimates are in percentage terms; t-statistics are in parentheses. * p < 0.10, ** p < 0.05, and *** p < 0.01.
Panel A: Credit Card Default and Misreported versus Reported Second-Lien
(1)
(2)
(3)
Misreported Second x CLTV < 100
-3.105***
-2.640***
-1.814***
(0.384)
(0.382)
(0.38)
(4)
-1.814***
(0.591)
Reported Second x CLTV < 100
-2.759***
(0.156)
-2.462***
(0.156)
-1.601***
(0.163)
-1.601***
(0.418)
Misreported Second x CLTV >= 100
-2.462***
(0.263)
-1.155***
(0.262)
-2.123***
(0.271)
-2.123***
(0.663)
Reported Second x CLTV >= 100
-2.023***
(0.124)
-0.696***
(0.125)
-0.854***
(0.152)
-0.854
(0.61)
Other Controls
Yes
Yes
Yes
Yes
State Fixed Effects
No
Yes
Yes
Yes
Mortgage Controls
No
No
Yes
Yes
No
1,023,725
38.36
5.233
0.0769
No
1,023,725
38.36
5.233
0.0859
No
1,023,725
38.36
5.233
0.109
Yes
1,023,725
38.36
5.233
0.109
SEs Clustered by State
Number of Loans
Percent CC 60 DPD
Percent Misrepresented
R2
Panel B: Credit Card Default and Misreported versus Reported Non-Owner-Occupants
(1)
(2)
(3)
Misreported Non-Owner Occupant
7.636***
7.305***
5.959***
(0.156)
(0.155)
(0.155)
(4)
5.959***
(0.319)
0.433***
(0.102)
1.144***
(0.103)
-1.024***
(0.106)
-1.024***
(0.364)
Other Controls
Yes
Yes
Yes
Yes
State Fixed Effects
No
Yes
Yes
Yes
Mortgage Controls
No
No
Yes
Yes
No
1,709,386
37.28
6.580
0.0760
No
1,709,386
37.28
6.580
0.0831
No
1,709,386
37.28
6.580
0.112
Yes
1,709,386
37.28
6.580
0.112
Reported Non-Owner Occupant
SEs Clustered by State
Number of Loans
Percent Default
Percent Misrepresented
R2
24
Panel A: Cumulative Defualt Rate by Occupancy Status
Panel B: Loan Interest Rate by Occupancy Status
Figure IA.3. Owner-occupancy status, cumulative default rates, and interest rates. Panel A of the figure shows the estimated cumulative default rate of a
mortgage in the first eight quarters since origination for subsamples of loans differing based on their owner-occupancy status, holding constant at their overall
sample mean all other observables used in Table IA.V such as FICO. Panel B shows the kernel density plots of the observed difference between actual and
predicted origination interest rates for the same subsamples of the overall data set. The predicted rates used in Panel B are given by a specification similar to that
used in Column (4) of Table IA.VIII, with the Reported Non-Owner-Occupant and Misreported Non-Owner-Occupant variables removed. In Panel B the x-axis
shows the error term from these regressions (the difference between the actual origination interest rate and the predicted value).
25
Panel A: Misreported Non-Owner Occupant
Panel B: Misreported Either
Figure IA.4. Misrepresentations across underwriters. This figure plots the percentage of loans with misreported non-owner occupants (Panel A), and
misreported either (Panel B) by underwriter along with a 95% confidence interval. Coefficients result from adding underwriter fixed effects to similar
specifications as in Table II. These levels are obtained by adding each underwriter fixed effect to the level of misrepresentation for the omitted category (Credit
Suisse) with the other covariates at their means. The dashed straight lines show the mean misrepresentation level in the data.
26
III. Additional Results
Table IA.XI
Misrepresentations and Loan Characteristics
The table presents probit estimates from regressions in which the dependent variable takes a value of one if the loan
has the given misrepresentation, and zero otherwise. Columns (1) and (2) show the results with Misreported SecondLien as the dependent variable, while columns (3) and (4) show the results with Misreported Non-Owner-Occupant
as the dependent variable. The estimates show the marginal effects in percentage terms; standard errors are in
parentheses. * p < 0.10, ** p < 0.05, and *** p < 0.01. These estimates can be compared to the OLS estimates
presented in Table III of the published article.
FICO
Misreported Second Lien
(1)
(2)
0.00523***
0.00531***
(-0.00040)
(-0.00040)
Misreported Non-Owner-Occupant
(3)
(3)
0.00794***
0.00777***
(-0.00034)
(-0.00034)
-0.00783***
(0.00013)
-0.00741***
(0.00013)
0.00209***
(-0.00007)
0.00200***
(-0.00007)
CLTV
-0.118***
(0.00270)
-0.112***
(0.00270)
0.0222***
(-0.00188)
0.0213***
(-0.00188)
No Cash Out Refi
-3.74***
(0.04221)
-3.59***
(0.04251)
-3.85***
(0.035422)
-3.87***
(0.035349)
Cash Out Refi
-9.78***
(0.06194)
-9.57***
(0.061551)
-5.43***
(0.039087)
-5.45***
(0.039135)
Low or No Doc.
-1.05***
(0.05019)
-1.04***
(0.049642)
2.61***
(-0.03972)
2.61***
(-0.03980)
ARM
5.14***
(-0.05975)
5.10***
(-0.059316)
2.13***
(-0.040891)
2.15***
(-0.040999)
Option ARM
-2.76***
(0.08425)
-2.87***
(0.085264)
8.27***
(-0.139085)
8.02***
(-0.142755)
Yes
Yes
Yes
Yes
No
854,959
7.131
Yes
854,959
7.131
No
1,563,223
6.426
Yes
1,563,223
6.426
Balance
Confidence Match Controls
Half-Year Origination Cohorts
Number of Loans
Percent Misrepresented
27
Table IA.XII
Misrepresentations and Loan Default
The table presents probit estimates from regressions in which the dependent variable takes a value of one if the
mortgage ever defaults (goes 90 days past due on payments) in the first two years following origination, and zero
otherwise. Panel A shows results with Misreported Second x CLTV >= 100, Misreported Second x CLTV < 100,
Reported Second x CLTV >= 100, and Reported Second x CLTV < 100 as control variables. The CLTV term
(computed from the credit bureau data) in these interactions takes a value of one if the loan has a CLTV ratio in the
appropriate range, and zero otherwise. Panel B shows the results with Misreported Non-Owner-Occupant and
Reported Non-Owner-Occupant as control variables. “Other Controls” include origination variables used in Table
III such as FICO, interest rates, and LTV ratios. Squared and cubed terms for FICO and LTV are included to
account for potential nonlinear effects. Panel B also controls for the reported CLTV to investors. Column (1)
includes half-year origination fixed effects. Column (2) incorporates property state fixed effects. Column (3) clusters
standard errors by state. The estimates show the marginal effects in percentage terms; standard errors are in
parentheses. * p < 0.10, ** p < 0.05, and *** p < 0.01.
Panel A: Mortgage Default and Misreported versus Reported Second Lien
(1)
(2)
Misreported Second x CLTV < 100
8.20***
8.95***
(0.375)
(0.383)
(3)
8.95***
(0.467)
Reported Second x CLTV < 100
8.81***
(0.159)
9.19***
(0.16)
9.19***
(0.887)
Misreported Second x CLTV >= 100
13.4***
(0.263)
14.3***
(0.269)
14.3***
(1.64)
Reported Second x CLTV >= 100
15.5***
(0.148)
16.2***
(0.15)
16.2***
(2.11)
Yes
Yes
No
No
1,109,250
17.10
5.212
Yes
Yes
Yes
No
1,109,250
17.10
5.212
Yes
Yes
Yes
Yes
1,109,250
17.10
5.212
Other Controls
Half-Year Origination Cohorts
State Fixed Effects
SEs Clustered by State
Number of Loans
Percent Default
Percent Misrepresented
Panel B: Mortgage Default and Misreported versus Reported Non-Owner Occupants
(1)
(2)
(3)
Misreported Non-Owner-Occupant
9.03***
8.34***
8.34***
(0.13)
(0.13)
(0.78)
Reported Non-Owner-Occupant
3.89***
(0.08)
3.72***
(0.08)
3.72***
(0.76)
Other Controls
Half-Year Origination Cohorts
State Fixed Effects
SEs Clustered by State
Number of Loans
Percent Default
Percent Misrepresented
Yes
Yes
No
No
1,827,497
14.89
6.424
Yes
Yes
Yes
No
1,827,497
14.89
6.424
Yes
Yes
Yes
Yes
1,827,497
14.89
6.424
28
Table IA.XIII
Second-Lien Misrepresentation among New Century Loans: Characteristics of New
Century Loans Merged with the BlackBox-Equifax Data set
This table presents summary statistics of key variables for New Century loans reported as having no second lien to
the RMBS trustees that have been merged with the BlackBox-Equifax data set.
Default within 2 Years
Interest Rate
FICO
Number of Loans
Reported as Having
No Second Lien
Mean
SD
0.288
0.453
8.409
1.356
594.9
55.65
10,924
29
Identified by us as Having
Misreported Second Lien
Mean
SD
0.403
0.491
7.642
1.082
645.2
41.69
1,279
Table IA.XIV
Second-Lien Misrepresentation and Pool Characteristics
This table presents OLS estimates from regressions in which the dependent variable is the percent of loans in a pool
with misreported second liens. The controls are the pool-level averages of the loan characteristics reported to
investors (e.g., FICO, CLTV, etc). The estimates are in percentage terms; standard errors are in parentheses. * p <
0.10, ** p < 0.05, and*** p < 0.01.
Percent Misreported
Second Lien
(1)
(2)
FICO
-0.0719***
(0.014)
-0.0707***
(0.014)
Orig. Balance
-0.00736***
(0.0018)
-0.00795***
(0.0018)
CLTV
-0.931***
(0.133)
-0.944***
(0.132)
No Cash Out Refi
-0.136***
(0.045)
-0.131***
(0.044)
Cash Out Refi
-0.253***
(0.037)
-0.251***
(0.037)
Percent No or
Low Doc.
0.00954
(0.015)
0.00767
(0.015)
Percent ARM
0.00299
(0.008)
0.00188
(0.008)
Percent Option ARM
0.0492**
(0.019)
0.0409**
(0.019)
Half-Year FE
Number of Loans
Percent Misrepresented
No
333
4.289
Yes
333
4.289
R2
0.220
0.233
30
Table IA.XV
Heterogeneity of Misrepresentation among Underwriters
The table presents the OLS estimates from regressions in which the dependent variable takes a value of one if the
loan has the given violation and zero otherwise. Column (1) shows results with Misreported Second as the
dependent variable, column (2) shows the results with Misreported Non-Owner-Occupant as the dependent variable
(see section II of this appendix for more details), and column (3) shows the results with misrepresentation on either
dimension as the dependent variable (either misreported). Underwriter RMBS Volume/Assets is the volume of loans
in our base sample underwritten in 2005 by the loan’s deal underwriter divided by the underwriter’s assets as of the
end of 2005. Underwriter Years in Subprime is the number of years between the loan’s securitization year and the
first subprime deal underwritten by the deal’s underwriter, excluding deals prior to 1999. Commercial Bank
Underwriter is a dummy that takes a value of one if the underwriter is a commercial bank and zero otherwise. RMI
is an index of the risk-management strength of an institution based on Ellul and Yerramilli (2013). Bonus to Salary
Ratio is the ratio of cash bonus to salary for the underwriter’s five highest-paid executives. These measures are
discussed in section I of this appendix. “Other Controls” include origination variables used in column (3) of Table
III, such as FICO, interest rate, and CLTV ratio. Standard errors are clustered by underwriter; the estimates are in
percentage terms; standard errors are in parentheses. * p < 0.10, ** p < 0.05, and *** p < 0.01.
(1)
Misreported
Second
0.546
(-2.600)
(2)
Misreported
Non-Owner Occupant
1.319*
(-0.753)
(3)
Either
Misreported
1.528
(-1.273)
Underwriter Years in
Subprime (Post-1999)
-0.299
(0.543)
-0.0341
(0.155)
-0.200
(0.215)
Commercial Bank
Underwriter
-0.0161
(1.610)
-0.464
(0.545)
-0.110
(1.222)
RMI
-0.434
(3.338)
-0.581
(0.806)
-0.636
(1.817)
Bonus to Salary Ratio
-0.0126
(0.052)
-0.00423
(0.010)
-0.0141
(0.028)
Half-Year Origination
Cohorts
Yes
Yes
Yes
Other Controls
Number of Loans
Percent Misrepresented
R2
Yes
853,954
7.131
0.0672
Yes
1,561,625
6.426
0.0304
Yes
1,700,264
9.192
0.0401
Underwriter RMBS
Volume/Assets
31
Figure IA.5. Distribution of interest rates across second lien status. This figure shows the kernel density plots of
the observed difference between actual and predicted origination interest rates for different subsamples of the overall
data set. The predicted rates are given by a specification similar to that used in Table II, with the Misreported
Second x CLTV >= 100, Misreported Second x CLTV < 100, Reported Second x CLTV >= 100, and Reported
Second x CLTV < 100 variables removed. The x-axis shows the error term from these regressions (the difference
between the actual origination interest rate and the predicted value).
32
A Simple Example of Equilibrium with Asset Misrepresentation
Consider a simple one-period economy with an equal measure of households and bankers.
The utility function of both households and bankers is linear in consumption (endowment good):
u(c)=c. Each household is endowed with two units of the endowment good. The endowment can
be consumed or invested. Each banker has no initial endowment but can invest on behalf of
households. Hence, given the equal population of bankers and households, the per capita initial
endowment equals one.
There are two types of banks of equal measure. “Good” banks have access to an
investment opportunity that requires one unit of investment at time 0 and pays households four
units of the consumption good with probability 0.5 or zero units with probability 0.5 at time 1 at
no additional cost. This is a positive NPV project. “Bad” banks have a technology that pays
households four units with probability 0.1 and zero with probability 0.9 from one unit of
investment. Bad banks are also able to use the technology to generate a hidden benefit of
0.1*investment (e.g., by expropriating some part of the investment). This is a negative NPV
project.
A. Pooling Equilibrium with Asymmetric Information
Consider the case in which the bank’s type is unobserved to households (i.e., bad banks can
pretend to be good banks).
The households know that half of the banks are bad. Without loss of generality we can
assume that households evenly split their endowment across banks so each bank gets one unit. A
competitive pooling equilibrium in this case will be such that:
(i)
households give all their endowment to bankers,
(ii)
bankers offer investment service at zero cost,
(iii) the aggregate payment to households that invest two units (average expected payment
to a household) is: 0.5*4+0.5*0.1*4=2.2, and
(iv)
each of the “bad” bankers consumes 0.1 .
Note that in this equilibrium households cannot tell apart positive and negative NPV projects.
They know that in equilibrium half of the banks that pretend to be good are actually bad. Yet
they decide to invest because the average NPV across projects of good and bad banks is positive,
that is, 2.2>2. Overall, due to banks, the economy moves from one unit of consumption good per
capita to 1.15 units (0.5*2.2+0.5*0.1).
Importantly, despite the fact that households correctly anticipate the extent of average
misrepresentation in equilibrium, eliminating misrepresentation can increase the overall surplus.
To see this, consider the case in which there is no asymmetric information about bank type,
which we discuss next.
33
B. Equilibrium with No Asymmetric Information
When the bank’s type is observed by households (bad banks cannot pretend to be good banks), a
competitive equilibrium in this case will be such that:
(i)
households give all their endowment to good banks,
(ii)
good banks offer investment services at zero cost,
(iii) the aggregate payment to households (average payment to a household is): 2*0.5*4=4,
and
(iv)
each of the “bad” bankers gets no investment and consumes zero.
In this case, due to good banks, the economy moves from one unit of the consumption good
to two units per capita. Comparing this case to the one with asymmetric information, it is clear
that the overall net surplus from the investment is larger (1 versus 0.15 in the previous case).
34