mur

					 Borrower Misrepresentation and Loan Performance




                                      Mark J. Garmaise
                                      UCLA Anderson




                                             Abstract



Misrepresentations by borrowers are associated with grave loan outcomes. Residential
mortgage borrowers who reported personal assets just above round number thresholds
were more than twice as likely to experience subsequent delinquency. Significantly more
borrowers report assets just above these thresholds than just below. Both these findings
are true only for applicants with unverified assets, which is consistent with reported asset
targeting by misrepresenting borrowers. Misrepresentation has the greatest impact on
delinquency for properties that later become potentially underwater. Misrepresentation
is most common for mortgages that result in a high debt load and in neighborhoods in
which refinancings are relatively frequent.



JEL Codes: G21, G32, D23

Correspondence to: Mark Garmaise, UCLA Anderson, 110 Westwood Plaza, Los Angeles, CA, 90095.
E-mail: mark.garmaise@anderson.ucla.edu. I thank the U.S. financial institution for providing the data.
I am grateful for comments from Zhaohui Chen, Tomek Piskorski, Amit Seru and an anonymous referee
and from seminar participants at the SFS Cavalcade, the Summer Real Estate Symposium, the WFA,
the Peruvian SBS and UCLA.
      Risk assessment and control are core functions of financial institutions. In any set-
ting with asymmetric information, one central concern is misrepresentation risk, the risk
that other parties will provide false or misleading information. Allegations of misrep-
resentation in financial transactions have become quite common, but there is relatively
little evidence clearly identifying it and quantifying its extent. How serious is the impact
of misrepresentation? What can be done to detect or deter it? Providing insights into
these questions can help the support the long-run stability and performance of financial
institutions.

      In this paper I offer an empirical description of personal asset misrepresentation by
a bank’s mortgage borrowers and show that it had grave consequences for loan outcomes.
Using a regression discontinuity approach, I show that borrowers who claimed a personal
asset level just above a round number threshold (e.g., $203,000) were more than twice
as likely to experience subsequent delinquency as those who claimed personal assets just
below the threshold (e.g., $198,000). I also find that many more borrowers reported
assets above, rather than below, these thresholds. Both these results are consistent
with the idea that some borrowers misreported their personal assets to be above round
number thresholds and that these borrowers were much more likely subsequently to
become delinquent. The impact of misrepresentation on delinquency is much greater for
loans that later became potentially underwater (i.e., loans for which the home equity
may have become negative). The severity of misrepresentation risk varies with the loan
terms and the location of the underlying property, suggesting that financial institutions
may be able to take steps to shield themselves from its effects.

      Information collection and risk mitigation by banks have been the themes of a
broad stream of research. Some of this work has focused on a bank’s sourcing of soft
information and its incorporation into loan provisions (Stein 2002, Petersen and Ra-
jan 2002 and Berger et al. 2005). Recent intra-bank studies have analyzed the role
of hierarchy (Liberti and Mian 2009) and the impact of moral hazard on within-firm
communication (Hertzberg, Liberti and Mian 2010). In this paper, I present mortgage


                                             1
data from a bank along with detailed information provided by the borrower. These data
allow me to investigate the main determinants of misrepresentation risk. Reducing this
risk will help improve bank screening and thereby raise lending standards (Ruckes 2004,
                                                                            o
Dell’Ariccia and Marquez 2006, Norden and Weber 2010 and Maddaloni and Peydr´
2011), a policy goal that has received much greater attention lately in light of the events
of the last several years.

      The data are from a U.S. bank that originates and retains residential mortgages.
Relating the borrower’s personal assets (excluding the house to be financed) to loan
outcomes, I observe a large discontinuous increase in delinquency risk specifically at
round number thresholds. (Very few borrowers reported asset levels at round numbers
precisely; the delinquency results are not driven by the use of a rounding heuristic
by a class of applicants.) This jump in delinquency is observed only for borrowers
whose asset claims are not documented; borrowers with verified assets experience no
discontinuity at thresholds. This suggests that only borrowers with undocumented assets
engaged in misrepresentation. These findings are robust to the inclusion of both month-
of-origination and zip code fixed effects.

      The distribution of reported assets offers additional support for the hypothesis of
misrepresentation. The density of unverified assets exhibits a sharp discontinuity at
round number thresholds, with significantly more borrowers reporting assets just above
the thresholds than just below. The density of verified assets displays no discontinuity at
thresholds. This is precisely the pattern that would be generated by systematic claims
of above-threshold unverified assets by misrepresenting borrowers, and it indicates that
misreporting plays an important role in this setting, as does in many other financial
environments (Fich and Shivdansani 2007, Povel, Singh and Winton 2007, Grahan, Li
and Qiu 2008 and Bollen and Pool 2009 and 2012). By contrast, borrowers were asked
to report typical monthly income (rather than the current actual number, as for assets),
and many responded by providing a rounded estimate. Misreporting borrowers may
have targeted their income as well, but they are hidden in the mass of presumably


                                            2
honest borrowers with round number income reports. The widespread use of a rounding
heuristic (Goldreich 2004) in the income data thus makes it less suitable for detecting
misrepresentation.

     Borrowers had two incentives to report high asset levels. First, assets were re-
quired to exceed a certain multiple of the monthly payment for loan acceptance (these
were not, however, round number thresholds). Second, assets were used in determining
loan terms and applicants with higher reported assets received loans with lower interest
rates. These reasons provide a general motive for misrepresentation. Why did misrep-
resenting borrowers choose to state assets just above a round number threshold? There
is substantial behavioral evidence that numbers above these thresholds are perceived to
be significantly larger than numbers just below them (this is related to the phenomenon
that retail prices generally end in 99 cents: Kalyanam and Shively 1998, Anderson and
Simester 2003 and Thomas and Morwitz 2005). Once a borrower has elected to misrepre-
sent assets, the cost to choosing an above threshold number relative to a below threshold
number is likely trivial, and the potential benefits from reporting a considerably higher
perceived asset level could have been material.

     Loan terms and borrower risk characteristics, however, do not significantly differ
between above- and below-threshold mortgages. This has two implications. First, the
bank was clearly unaware of this specific asset reporting practice, and did not even
partially adjust upwards the price of loans to reflect the misrepresentation risk. Sec-
ond, misreporting borrowers may have over-estimated the positive effects of stating
above-threshold assets. Nonetheless, given the likely minor cost of misreporting above-
threshold rather than below-threshold assets, and the insignificant effect on loan terms,
it is unlikely that they were harmed by this strategy.

     The extent of the misrepresentation problem varies significantly with certain cross-
sectional aspects of the data. While low subsequent housing returns and high borrower
indebtedness (accounting for initial borrowing and potential negative amortization) are
both associated with generally poor loan performance, I do not find that either alone

                                           3
generates worse misrepresentation-driven outcomes. The combination of the two, how-
ever, in the form of returns so low and potential indebtedness so high that the borrower
may have negative equity, is a strong determinant of the impact of misrepresentation.
That is, misrepresentation appears much more likely to lead to delinquency if the bor-
rower is potentially underwater, for it is just in that case that it is appealing to exercise
the strategic default option.

     Some loan contracts are associated with worse misrepresentation risks. New loans
which result in a high total loan-to-value ratio on a property (including any other per-
sisting debt) exhibit greater evidence of misrepresentation. This is consistent with the
argument that borrowers seeking a high overall debt load were desperate for financing
and were especially willing to engage in misreporting.

     There is significant heterogeneity in the impact of misrepresentation across different
neighborhoods. In particular, there is clear evidence that misrepresentation was more
frequent and serious in zip codes with high refinancing activity relative to purchase
loans. These areas are in essence hot spots of misrepresentation. Most refinancings
(over 80%) involve a net cash payment to the borrower. A relative concentration of
these cash out refinancings may indicate that there is a rush to withdraw home equity
from a certain neighborhood, and in their haste to extract value some borrowers may
be willing to misrepresent their personal assets. Within a given zip code, by contrast,
refinancings are no more likely to involve misrepresentation than purchase loans. At
the neighborhood level, however, a high frequency of refinancings apparently sends a
negative signal about misrepresentation risk within the zip code.

     The bank also suffered from more frequent misrepresentation during days with
high origination volume. This may reflect either a strain on the evaluation capacities of
loan officers or an increased possibility of misrepresentation when sourcing loans from
new markets during times of expansion.

     Taken together, these cross-sectional findings indicate that the extent of misrep-


                                             4
resentation risk varies with the specific mortgage terms of the loan, the neighborhood
characteristics of the underlying property and the bank’s daily volume of originations.
While these associations may potentially be affected by unobserved variables, they also
suggest that contract design, borrower selection and origination volume management
can have an impact on a financial institution’s exposure to misrepresentation.

     Mortgages brokers and loan officers do not appear to be complicit in promoting
this form of misrepresentation. Deals in which brokers receive large compensation for
completing the transaction are no more likely to exhibit misrepresentation. Also, there is
no evidence of disproportionate misrepresentation in originations from brokers suspended
or terminated for cause. It seems unlikely that loan officers, bank insiders who were
aware that asset targeting above round numbers yields no tangible benefits, would have
encouraged borrowers to adopt this practice.

     Misreporting has also been documented in residential real estate mortgages for
transaction prices and appraisals (Agarawal, Ben-David and Yao, 2012, Ben-David 2011
and Carrillo 2010) and personal income (Jiang, Nelson and Vytlacil 2009). In general,
there are not many broad prescriptions for managing misrepresentation risk other than
the retention of excess capital.

     The findings in this paper indicate that behavioral biases such as round number
targeting may play an especially important role when some agents are engaged in mis-
representation. This is particularly true of interactions between consumers and financial
institutions. A closer study of the form and frequency of these biases may therefore be
quite helpful in mitigating misrepresentation risk, a key aspect of the broader problem
of asymmetric information in finance.

     The rest of the paper is organized as follows. Section 1 details the residential
mortgage data that I use to analyze misrepresentation risk. In Section 2 I outline my
econometric approach, and I describe the empirical findings in Section 3. Finally, Section
4 concludes.


                                            5
1     Data

The data in this paper describe 8,287 residential single-family mortgage loans originated
by a U.S. financial institution in the period January 2004- October 2008. Loans made
to insiders are excluded, as are loans for which the personal asset information of the
borrower is not provided. These loans were retained by the bank and not securitized.
As described in Table 1, the data include pricing information and details on borrower
and property attributes. This bank offers floating rate mortgages, and the mean spread
between the loan interest rate and the underlying index is 3.36 percentage points (various
indices are used, including the prime rate, the Treasury bill rate and LIBOR). Many of
the loans allow borrowers to make payments less than the current interest rate, thereby
causing negative amortization. The mean loan-to-value (LTV) ratio is 72%, the mean
cumulative loan-to-value (CumLTV) ratio (i.e., including any other mortgage persisting
after the new financing) is 73%, the mean monthly borrower income is $16,086 and the
mean borrower FICO credit score is 719.4. This relatively high mean FICO score and
income reflect the fact that the bank made almost no subprime loans (e.g., only 0.3%
of borrowers had FICO credit scores below 620). That these loans were made to high-
quality borrowers and not securitized suggests that this bank was not directly affected
by key drivers of default emphasized in other research (Mian and Sufi 2009 and Keys,
Mukherjee, Seru and Vig 2010). Data is also provided on the purpose of the loan (home
purchase, cash out refinance or rate/term refinance).

      In common with broader market trends, the bank experienced significant delin-
quencies in its residential lending. Specifically, 20% of the loans in the data are delin-
quent (30 or more days past due).


1.1    Origination Process

Essentially all the residential loans made by the bank are presented to them by mortgage
brokers. A loan officer employee of the bank works with the broker and prepares a

                                            6
mortgage file. There are 155 loan officers observed in the data. The base interest rate
charged is determined by a fixed set of loan characteristics (LTV, FICO score, etc.), but
the bank may also adjust the pricing to reflect other perceived risks. The mean of this
exception pricing is 15.2 basis points.

     As part of the application, borrowers state their level of personal assets (excluding
the property to be purchased) and income. Mortgages differ in their level of documen-
tation: a borrower chooses how much documentation to supply and receives a rate that
depends on this choice. Borrowers may provide documentation verifying both personal
assets and income, verifying assets but not income, or neither. Low-documentation
mortgages were designed for borrowers whose assets or income were difficult to substan-
tiate (e.g. owners of small private businesses); the house serving as the loan collateral
was, in any case, regarded to be the main security for the mortgage. Some borrowers
may misrepresent their personal information. While misrepresentation may be costly in
general (and these costs may vary with the level of documentation required), misrepre-
senting borrowers can likely make claims for their asset/income data that are plausible
in some range.

     The bank made use of asset and income information in its approval process in
a manner that is not completely transparent to borrowers. Internal bank protocols
required that assets exceed a multiple of the monthly principal and interest payments
plus insurance and property taxes (PITI) due from the borrower. Borrower income was
also considered in the mortgage evaluation process.



2    Empirical Specification

The focus of my empirical analysis is on the impact of borrower-reported assets on the
eventual delinquency of mortgages. I explore the hypothesis that borrowers who mis-
represent their personal asset holdings are more likely to state asset levels above, rather
than below, round number thresholds. In other words, a borrower who is misrepresent-

                                            7
ing his assets is more likely claim an asset level of $102,000 rather than $97,000. If this
is the case, and if misrepresenting borrowers are more likely to eventually become delin-
quent, then we should expect to see a positive discontinuity in delinquency probabilities
at round number asset thresholds. Under the null hypothesis of no misrepresentation,
or no threshold-gaming, there should be no discontinuity in delinquency probabilities at
round number thresholds.

       I take asset multiples of $100,000 to be the round number thresholds and define
round(x, y) to be the value of x rounded to the nearest positive multiple of y. I define
normalized assets A as



                          A = assets − round(assets, 100000).                                         (1)

       The indicator variable IA denotes mortgages with reported assets above the thresh-
old:

                                            {
                                                 1 if A ≥ 0
                                  IA =                                                                (2)
                                                 0 otherwise

       To analyze the hypothesis of a discontinuity in delinquency probabilities at the
thresholds, I estimate the following formal model:



                                ∑
                                6                    ∑
                                                     6
 Delinquenti,t = α + βIAi,t +         ωj A j
                                       A
                                           i,t   +         ξj IAi,t Aj + γ ∗ controlsi,t + λt + ϵi,t , (3)
                                                            A
                                                                     i,t
                                j=1                  j=1


where Delinquenti,t is an indicator for whether loan i provided in month t subsequently
became delinquent, Ai,t is the personal asset level claimed by the borrower, IAi,t is an
indicator for whether this asset level is above a round number threshold, controlsi,t is a
vector of loan and property controls, λt is a month fixed effect and ϵi,t is an error term.
The controls may include fixed effects for the loan officer, depending on the specification.


                                                     8
      The coefficient of central interest is β, which measures discontinuities in the delin-
quency probability at round number personal asset thresholds. Under the null hypothesis
of no asset misrepresentation (or no systematic misrepresentation around round number
thresholds), we should expect to find β = 0. Under the alternative hypothesis that some
borrowers misrepresent their asset levels to be just above round number thresholds and
that these borrowers are more likely to become delinquent, we should find that β > 0:
there should be a discrete jump in delinquency probabilities just above the thresholds.

      I estimate (3) using OLS, despite the binary nature of the Delinquent variable,
due to the large number of fixed effects along several dimensions and the resulting in-
cidental parameters problem in non-linear maximum likelihood estimation (Abrevaya,
1997). OLS coefficients are estimated consistently even with multiple fixed effects. This
approach is similar to the one used in the models of Card, Dobkin, and Maestas (2004)
and Matsudaira (2008). The specification allows the delinquency probability to be con-
tinuous in personal assets, with the shape of the probability function permitted to be
different on either side of round number threshold.

      For some tests I study cross-sectional variation in the impact of above-threshold
assets on delinquency probabilities by estimating (3) in subsamples that differ, for ex-
ample, by characteristics of the loan or origination process. Significant differences in the
estimates of β in distinct subsamples provide evidence that the severity of the misrep-
resentation risk may have varied with loan contract features or evaluation procedures.


2.1    Bank Policy

As described in Section 1.1, the bank requested personal asset levels from borrowers
for use in determining loan eligibility. The bank did not, however, compare the stated
asset levels to round number thresholds. Instead, the assets were required to exceed
a multiple of the monthly PITI due from the borrower. This bank requirement was
therefore unrelated to the round number thresholds I study.


                                            9
3     Results

3.1    Misrepresentation of Assets

I begin by testing the hypothesis that borrowers that stated asset levels just above
round number thresholds experienced higher subsequent delinquency rates. This test
is motivated by the idea that borrowers who misrepresented their personal assets were
likely to state assets above, rather than below, round numbers, and that misrepresenting
borrowers were more likely to become delinquent. I estimate the discontinuity model
exploring the link between delinquency and reported assets described in equation (3). In
the first test, I regress an indicator variable for delinquency on a dummy for whether the
borrower reported assets above a round number threshold, a sixth degree polynomial in
reported assets and the interaction between the above-threshold dummy and the sixth
degree polynomial. The estimation is via OLS, with robust t-statistics clustered by the
month of mortgage origination.

      As described in the first column of Table 2, I find a positive and significant (t-
statistic=3.47) coefficient on the above-threshold dummy. The coefficient estimate is
0.25, which indicates that borrowers who reported assets just above round number
thresholds experienced delinquency rates 25 percentage points above those who reported
assets just below the thresholds. The coefficient is large in absolute terms, and is also
large compared to the mean delinquency rate of 0.20 observed in the data. The re-
gression results are graphically depicted in Figure 1. The curved lines represent the
fitted polynomials and the connected points describe the average delinquency rates for
each of the buckets of $4,000 in normalized assets. As the figure makes clear, there is
a large jump in the delinquency risk at the round number thresholds (asset multiples
of $100,000) where the normalized assets are equal to zero. This finding of dramati-
cally higher delinquency rates for above-threshold asset reporters is consistent with the
hypothesis that misrepresenting borrowers are both more likely to report assets above
round numbers and are more likely to become delinquent. The higher delinquency rates

                                           10
of misrepresenting borrowers are probably not simply attributable to the fact that they
have lower assets than they report; misrepresentation is likely a signal of wider dishon-
esty and perhaps a greater willingness to engage in strategic default.

      Figure 1 extends from -$100,000 to +$50,000 because borrowers with less than
$50,000 in assets have normalized assets of below -$50,000. All other borrowers are
within $50,000 of a threshold. Excluding the borrowers with assets below $50,000 has
little effect, as they are quite distant from the normalized asset threshold of zero and
have little impact on the discontinuity estimate. In the specification without these low
asset borrowers, the estimated coefficient on the above-threshold dummy is 0.256 with
a t-statistic of 3.20.

      I next consider a second implication of the misrepresentation hypothesis, namely
that it should be more severe for loans for which misrepresentation was more feasible.
Specifically, misrepresentation should be expected to be a more serious problem for loans
in which the borrower stated his assets without supplying verification. I divide the loans
into two samples, one for which asset documentation was not supplied (consisting of 3,276
loans) and one for which it was (composed of 4,994 loans). I then regress delinquency
on the above-threshold dummy, the sixth degree polynomial in assets, the interaction
of the above-threshold and the polynomial and the full set of controls. The controls
include the interest rate charged, the loan to value ratio of the mortgage, the credit
score of the borrower and fixed effects for the month of origination. These regressions
are estimated via OLS with robust standard errors clustered by month of origination.
As detailed in the second column of Table 2, in the sample of mortgages submitted with
unverified assets, the coefficient on the above-threshold indicator is 0.405 and significant
(t-statistic=3.42). This result is described visually in Figure 2. The general shape of
the delinquency curve resembles that in Figure 1, but the magnitude of the jump at zero
is much bigger in the unverified assets sample: the asset misrepresentation problem is
indeed more severe when supporting documents were not required.

      Housing returns were clearly an important determinant of delinquency during the

                                           11
sample period. For each loan, I use zip code level pricing indices to calculate the min-
imum housing return experienced by the associated property, as measured from the
origination date until the sample close. (I do not take into account subsequent loan
payoff or default dates, as these are clearly endogenous.) As shown in the third column
of Table 2, including the housing return variable in the unverified assets delinquency
specification has little impact on the estimated coefficient on the above-threshold indi-
cator. Column four of Table 2 displays the results for the specification that includes zip
code fixed effects, which have little impact on the main finding.

     In the sample of mortgages with verified assets, however, the results displayed in
the fifth column of Table 2 show that the coefficient on the above-threshold variable is
both very small (-0.023) and statistically insignificant (t-statistic=-0.29). Figure 3 shows
clearly that there is no meaningful jump at zero normalized assets in this sample. Taken
together, these findings provide strong evidence of misrepresentation of assets: there is
a very large jump in delinquency at round number asset thresholds, but this jump only
exists for the set of mortgages for which asset verification was not provided.

     As the figures make clear, the central finding of a large significant jump in delin-
quency probability at the round number threshold for the unverified asset mortgages
and no significant jump in the verified asset sample is robust to a variety of specifi-
cations. In the basic sixth-degree polynomial specification described above the jump
coefficients are 0.41 and -0.02 in the unverified and verified asset samples, respectively.
In the fifth-degree polynomial specification, the estimates are 0.48 and -0.03, and in the
seventh-degree polynomial specification the estimates are 0.48 and 0.04. There is some
curvature in the underlying data, so higher order polynomials offer a more stable fit,
though the particular choice of polynomial length does not have a material effect. The
results are not driven by data points that are far from the threshold. Reducing the
sample to just those data points with normalized assets within $10,000 of the threshold
yields a jump estimate of 0.57 in the unverified assets sample and an estimate of -0.04
in the verified assets sample. All the above estimates are statistically significant in the


                                            12
unverified sample and insignificant in the verified sample.

        Was misrepresentation confined to asset levels or was there also systematic target-
ing of reported income? To consider this question, I estimate the equivalent of equation
(3), substituting the reported monthly income for assets, and comparing it to thresholds
that are multiples of $1,000. For the sample of 6,017 loans with unverified income re-
ported by the borrower, I regress the delinquency indicator on the income threshold, a
sixth degree polynomial in income, the interaction between the above-threshold income
dummy and the polynomial and the full set of controls. As documented in the sixth
column of Table 2, there is no significant discontinuity (t-statistic=-0.19) in delinquency
rates around round number income thresholds. In unreported results, I conduct the
analogous test for annual reported income and thresholds of $10,000 and again find an
insignificant effect.

        To explore the contrasting findings on assets and income, I turn to an examination
of the distributions of these two variables.


3.2      Distributions of Reported Assets and Income

The results in the previous section document a strong increase in delinquency at round
number asset thresholds, but I find that the data on reported income do not display
this pattern. In this section, I analyze the distributions of reported assets and income
to provide further evidence on the effects of misrepresentation.


3.2.1     Round Number Heuristic

There is a tendency in financial markets for participants to make use of round number
heuristics. Goldreich (2004), for example, shows that dealers in U.S. Treasury auctions
frequently submit bids with a final digit of zero. He attributes this to boundedly rational
investors making use of a heuristic. Do what extent do borrowers make use of this
heuristic in reporting assets and income? One might also ask if the discontinuity in

                                               13
delinquency at round number asset thresholds described in Table 2 is driven by the use
of this heuristic by a subset of borrowers.

     The evidence is quite different for reported assets as opposed to reported income.
Applicants were requested to provide their actual current asset levels, not rounded es-
timates. As a result, out of 8,287 data points, only 23 loans (fewer than 0.3%) have
normalized assets of zero. That is, very few of the borrowers submitted round number
values for their reported assets. Excluding these 23 loans yields discontinuity effects that
vary little from the base case: the coefficient estimate is 0.46 in the unverified asset sam-
ple (significant at the 1% level) and 0.01 in the verified asset sample (insignificant). The
results on reported assets are not determined by data points precisely at the threshold.

     For reported income, by contrast, applicants were requested to supply their typical
or stable monthly income. The bank, like other mortgage providers, wished to exclude
unusual income in a given month from the underwriting decision, so it did not request
the borrower’s monthly income precisely at the time of application. Many borrowers
apparently regarded this instruction as an invitation to supply a rounded estimate of
their monthly income. Consequently, 979 (11.7%) of the borrowers in the full sample
have normalized reported income of zero. For the sample of unverified assets, 15.5% of
the borrowers reported income as a round number. When asked to provide their typical
monthly income, borrowers made far greater use of the round number heuristic.

     The behavioral evidence on retail price perceptions shows that it is the first (left-
most) digit of a number that matters most (Anderson and Simester 2003 and Thomas
and Morwitz 2005). It may well be the case that misrepresenting borrowers targeted their
reported income to have a first digit that made the number look larger, as they did for
reported assets. Given the request for stable monthly income, however, misrepresenting
borrowers, and many other honest borrowers, may have elected to simply provide a round
number. Given the large mass of honest borrowers with normalized income of zero, there
is not a significant effect on delinquency at this threshold. A round number threshold
approach for identifying misrepresentation requires that borrowers make precise data

                                              14
reports and that not many borrowers utilize the round number heuristic. In other
words, the income data, given the evidently common use of the round number heuristic,
are not suitable for detecting the relatively small number of misrepresenting borrowers
who may have targeted their reported income as well.


3.2.2     Prevalence of Above and Below Threshold Reported Assets

If the striking increase in delinquency risk just above round number asset thresholds is
driven by borrower misrepresentation, as I have argued, then the distribution of unver-
ified reported assets should exhibit a discontinuity at the threshold. Specifically, there
should be relatively few borrowers with reported assets just below the threshold and rel-
atively many borrowers with assets just above the threshold. Misrepresenting borrowers
should cluster around normalized unverified asset levels of just above zero.

        To examine this hypothesis I employ McCrary’s (2008) test for a discontinuity in
a density function. McCrary’s methodology is well-suited for this setting, as it allows
for distinct kernel density estimates on both sides of the threshold, while correcting
for boundary bias.1 McCrary also proposes a test to evaluate the log difference in the
density heights on either side of the threshold point. If this log difference is significant,
it indicates that the density is discontinuous at the threshold.

        The estimated kernel density of reported unverified assets is described in Figure 4.
The thick line represented the density estimate and the surrounding thin lines depict the
95% confidence interval. The circles describe scaled frequencies (i.e., they are analogous
to histograms). The bin size of 1,342.8 and bandwidth of 14,638.3 are selected using
McCrary’s automatic algorithm.

        As the figure makes clear, there is a sharp discontinuity at normalized assets of zero,
with significantly more reported assets above the threshold than below. The estimated
   1
    Other methods for assessing discontinuities are considered in Burgstahler and Dichev (1997), Gar-
rod, Pirkovic and Valentincic (2006) and Bollen and Pool (2009).



                                                 15
log difference in kernel heights is 0.81 (t-statistic= 3.49). This discontinuity is not driven
only by borrowers with normalized assets of zero (i.e., reported assets equal to a round
number); excluding those borrowers yields an estimated log difference in kernel heights
of 0.49 (t-statistic=2.11).

      By contrast, the estimated kernel density of reported verified assets is described
in Figure 5. Verified assets should be difficult to manipulate, and in Table 2 I find
no significant discontinuity in delinquency at normalized verified assets of zero. This
suggests that there should not be a discontinuity in the density of reported assets for
the verified sample, and Figure 5 is consistent with this hypothesis. The estimated log
difference in kernel heights is -0.09 (t-statistic= -0.65). In the sample excluding reported
assets equal to a round number the log difference in heights is -0.20 (t-statistic=-1.32).

      The statistical significance of the difference in the density discontinuities in the
unverified and verified asset samples can be assessed using a bootstrapping technique.
The discontinuity is significantly greater (at the 1% level) in the unverified sample, both
if round number reported assets are included or excluded.

      Given the evidence in Table 2 and the findings on the distributions of assets and
income, my subsequent analysis will focus on asset misrepresentation in the sample of
loans with unverified asset documentation.


3.3    Asset Discontinuity and Loan and Borrower Characteris-
       tics

The results in Table 2 showing that there is a very large jump in delinquency above
round number asset thresholds control for the rate spread charged by the bank, so they
make clear that the bank did not properly assess the risk of these loans. In fact, the
increase in delinquency risk is so pronounced, that it is clear that if the bank had been
fully aware of it, these loans would not have been made at all. Still, one may ask if the
bank was aware of this risk in any respect. Do the loan terms exhibit any discontinuities


                                             16
at the asset thresholds?

     To evaluate this question, I regress the rate spread on the above-threshold indica-
tor, a sixth degree polynomial in assets and all the previous controls (excluding the rate
spread). The result, described in the first column of Table 3 shows that the coefficient
on the above-threshold dummy is insignificant (t-statistic=-0.09, with standard errors
clustered by month of origination). Not only is the coefficient estimate insignificant, it
is also very small: the asset threshold discontinuity is associated with less than a 1 basis
point increase in the rate spread. That is, the effect is measured quite precisely and it is
clearly very close to zero. The bank is not pricing up the above-threshold loans in any
systematic way, and it is certainly not pricing them up in a manner reflective of their
dramatically higher risk.

     I also regress the exception pricing on the loan, the loan-to-value (LTV) ratio, the
cumulative loan-to-value (CumLTV) ratio (including any other persisting mortgage), the
log of the loan size, the log of the loan maturity and the maximum permitted negative
amortization on the above-threshold indicator, the sixth degree polynomial in assets and
the standard controls (omitting LTV as a control in the LTV, CumLTV and loan size
regressions). The results, displayed in Table 3 columns 2-7 are uniformly insignificant
and small in magnitude. For example, LTVs are 10 basis points smaller (t-statistic=-
0.03) above the threshold and loan maturities are 2.6 percent longer (t-statistic=1.00).
Borrower characteristics also show no apparent discontinuities at asset jumps. The
borrower credit score is 1.4 points lower (t-statistic=-0.10) above the threshold, which
is very small compared to a sample standard deviation of 45.4. There is no apparent
distinction between the observable characteristics of above- and under-threshold asset
applications at the time the loan is extended. The bank was not aware of this specific
practice of asset misrepresentation and did adjust loan features to reflect it.




                                            17
3.4    Benefits of Misreporting Assets

The discontinuity described in Figure 4 in the distribution of reported assets and the re-
sults in Table 2 linking above-threshold assets to much higher delinquency rates together
provide strong evidence of misrepresentation of unverified assets by some borrowers. On
the other hand, the evidence in Table 3 makes clear that reporting above-threshold as-
sets did not lead to significantly better contract terms for borrowers. What incentives
were there for borrowers to engage in this misrepresentation?

      The first point is that borrower assets were required to be above a specific multiple
of the monthly PITI for an application to be accepted. Borrowers with true assets below
these requirements were able to receive loans they would have otherwise been denied by
misrepresenting their assets.

      Second, borrowers were informed that their assets would be used in the underwrit-
ing decision to determine loan terms. Table 4 shows the results from regressing the rate
spread on the borrower’s reported assets, reported income, credit score, LTV, loan matu-
rity, maximum permitted amortization, monthly fixed effects and zip code fixed effects.
As the table describes, an increase in the log of reported assets (column 1) or the level
of reported assets (column 2) are both associated with significantly lower rate spreads.
(Incidentally, income does not have an significant effect on loan terms, which, if this was
understood by borrowers, may help to explain their greater use of the round number
bias in reporting income, as this bias tends to be more pronounced when stakes are lower
(Goldreich 2004).) In other words, borrowers who reported higher assets received more
attractive loan terms. Thus for reasons of both eligibility and pricing, borrowers had a
general incentive to report high asset levels.

      Why did misrepresenting borrowers report assets just above round number thresh-
olds? There is extensive empirical and experimental evidence in the marketing litera-
ture that numbers just below round number thresholds are perceived to be significantly
lower than those at the threshold or above (Kalyanam and Shively 1998, Anderson and


                                            18
Simester 2003 and Thomas and Morwitz 2005). This fact is used to explain the ubiqui-
tous odd pricing phenomenon (i.e., that retail prices generally end in 99 cents). Reported
asset levels above round number thresholds were therefore likely perceived by borrowers
to appear significantly larger than those just below. Once a borrower has elected to
misrepresent his assets, there is likely little cost in choosing to report $101,000 rather
than $98,000. The first number, however, appears significantly larger. (As discussed in
Section 3.2.1, borrowers were asked to report actual asset levels, so reporting a round
number would likely have appeared unusual, and very few borrowers did so.) Given the
general benefits accruing from reporting higher assets, the perceived larger magnitudes of
above-threshold assets and the minor cost associated with misreporting above-threshold
rather than below-threshold assets, misreporting borrowers choose to state asset levels
just above round numbers.

      In other words, misreporting borrowers may have over-estimated the positive effects
of stating above-threshold assets. Nonetheless, given the likely trivial cost of misreport-
ing above-threshold rather than below-threshold assets and the insignificant effect on
loan terms, it is unlikely that they were harmed by this strategy.


3.5    Effects of Misrepresentation by Housing Returns and Max-
       imum Indebtedness

The evidence in Table 2 establishes that there were significant discontinuities in delin-
quency at round number unverified asset thresholds, even in the presence of controls
for subsequent housing returns. Nonetheless, it seems reasonable to hypothesize that
misrepresentation should have a greater effect if housing returns are low, as strategic
default will be more attractive in that case. To examine this issue, I split the sample
into subsamples with minimum subsequent returns that are, respectively, below and
above the sample median (-35.4%). As shown in the first two columns of Table 5, the
coefficient on above-threshold assets is slightly larger in the low housing return sample,
but it is not significantly different from the coefficient in the high housing return sample.

                                            19
     Maximum potential indebtedness, the product of the initial LTV and the maxi-
mum possible negative amortization, may also be expected to exacerbate the impact of
misrepresentation on delinquency. (The full profile of payments for each loan is not avail-
able, so I make use instead of the maximum possible negative amortization.) Columns
three and four of Table 5 display results from splitting the sample based on the max-
imum potential indebtedness (the sample median is 87.4%). I find a larger coefficient
on above-threshold assets in the sample with high potential indebtedness, but it is not
statistically different from the coefficient in the sample with low potential indebtedness.

     Next I consider whether it is neither housing returns nor indebtedness alone that
creates the incentive for strategic default, but rather the combination of the two. I
classify a loan as potentially underwater if the minimum subsequent estimated housing
value falls below the maximum potential indebtedness. Comparing the coefficients on
above-threshold assets in the potentially underwater and not potentially underwater
samples, as shown in columns 5 and 6 of Table 5, makes clear that misrepresentation
had a significant impact only in the potentially underwater sample. The difference
between these coefficients is significant at the 5%-level. Misrepresenting borrowers were
far more likely to become delinquent, but this was only true in cases in which strategic
default was in their interest.

     This finding helps to explain how misrepresentation was able to persist during the
sample period. The large magnitude of the impact of misrepresentation on delinquency
indicates that the bank was not aware of this specific borrower practice, or it would
have put an end to it. In other words, misrepresentation of this severity likely repre-
sents an out-of-equilibrium phenomenon. The impact of misrepresentation, however,
was so substantial only because of the likely unanticipated very negative housing mar-
ket performance. As a result of this particular housing market outcome, the underlying
misrepresentation problem was revealed.




                                           20
3.6    Misrepresentation by Loan Characteristics

In this section, I consider whether asset misrepresentation was more severe for loans
with certain characteristics. In analyzing the question of cross-sectional differences in
misrepresentation, I will examine the relative magnitudes of discontinuities at round
number thresholds in both delinquency probabilities and in the density of reported
assets. Evidence that some deal types had larger discontinuities in both delinquency
and in the density of reported assets could suggest that the prevalence and depth of
misrepresentation varied with the characteristics of the mortgage or borrower or it might
indicate that the bank had varying skill in mitigating misrepresentation risk across
different mortgage types. In the analysis below, I will explore both of these possible
explanations of the cross-sectional findings.

      First, I split the sample into loans with LTV above and below the median of 78%.
I estimate the discontinuity regression of delinquency on assets described in (3), with
the previous full set of controls in each sample separately. As reported in the first
two columns of Table 6, the coefficient on the above-threshold coefficient is larger in
the high-LTV sample, but the difference is not statistically significant. An analysis of
the discontinuity in the density of reported assets yields an estimated log difference in
kernel heights of 1.07 (t-statistic= 3.37) in the high-LTV sample, and an estimated log
difference of 0.41 (t-statistic= 1.34) in the low-LTV sample. These two discontinuity
estimates are not significantly different. That is, there is no statistically compelling
evidence either that the delinquency effect is greater in the high-LTV sample or that
there is more targeting of reported assets just above round number thresholds.

      As a second split, I divide the sample into loans with CumLTV (total LTV, in-
cluding any other persisting mortgage) above and below the sample median of 79%.
The results detailed in columns three and four of Table 6 show the coefficient on the
above-threshold coefficient is larger in the high-CumLTV sample, and the difference in
coefficients is significant at the 10% level. The estimated log difference in kernel heights


                                           21
is 1.12 (t-statistic= 3.57) in the high-CumLTV sample and 0.32 (t-statistic= 1.05) in
the low-CumLTV sample. These estimates are also significantly different at the 10%
level. Taken together, these findings indicate that loans to borrowers with higher total
LTV after the financing were more likely to exhibit misrepresentation. It is possible
that these borrowers were more desperate for financing, and were thus more willing to
engage in misrepresentation. It is also possible that the bank had less experience in
evaluating these high-LTV loans and was therefore less successful in screening them for
misrepresentation, though it would be reasonable to expect the bank to exercise greater
care in analyzing these risky mortgages.

      The third split divides the sample into loans classified as refinancings and those
originated for the purpose of supporting a home purchase. Columns five and six of Table
6 show that the discontinuity in the delinquency regression is larger in the relatively small
sample of purchases, but the difference in coefficients is not statistically significant. The
estimated log difference in kernel heights is 0.72 (t-statistic= 2.49) in the refinancing
sample and 1.00 (t-statistic= 2.95) in the purchase sample, and this difference is not
significant either. There is no evidence that refinancings are more or less subject to
misrepresentation that loans for home purchases.

      The cross-sectional findings in Table 6 support the argument that asset misrep-
resentation had its greatest impact in mortgages to borrowers who supported a larger
overall loan-to-value burden after the financing. Though I cannot control for omit-
ted variables that may affect the way the bank evaluated loans of different types, this
evidence does suggest that the bank could perhaps have reduced its exposure to misrep-
resentation risk through selection against borrowers with high cumulative loan-to-value
ratios. More broadly, financial institutions may be able to reduce their exposure to
misrepresentation risk by designing contracts that exclude borrowers with specific risk
characteristics.




                                             22
3.7    Misrepresentation Across Neighborhoods

The discussion in the previous section emphasized the role that can be played by ap-
propriate loan contract design in potentially reducing misrepresentation. I now consider
whether there was heterogeneity in misrepresentation across neighborhoods (i.e., zip
codes). I first calculate the mean LTV in each zip code, and sort loans into those
originated in high LTV and low LTV neighborhoods, using the sample median as the
dividing line. The first two columns of Table 7 describe the results from the delin-
quency regressions in each sample. The coefficients on assets above threshold are not
statistically distinguishable. The estimated log difference in kernel heights is 0.57 (t-
statistic=1.81) in the high LTV neighborhood sample and 0.97 (t-statistic=3.14) in the
low LTV neighborhood sample. The difference between these estimates is also not sta-
tistically significant. There is no evidence of greater misrepresentation in either high or
low LTV neighborhoods.

      Similar findings emerge from a split of loans into those originated from high and
low CumLTV neighborhoods. The delinquency regressions displayed in columns three
and four of Table 7 show that the coefficients on assets above threshold are similar in
both samples (and not statistically different), and in unreported results I find that the
estimated log differences in kernel heights are also statistically indistinguishable. Loans
from properties located in areas with high cumulative LTVs were no more likely to
exhibit misrepresentation.

      Columns five and six of Table 7 detail the delinquency regression results from a
split of loans into those originated from neighborhoods with high or low frequencies of
refinancings (relative to purchase loans). As the results make clear, the jump in delin-
quency was much greater in neighborhoods with relatively more refinancings, and the
difference in coefficients is significant at the 5% level. (Clustering at the zip code level,
rather than at the month of origination level, yields a similar result.) The estimated log
difference in kernel heights is 1.75 (t-statistic=3.72) in the high refinancing frequency


                                           23
neighborhood sample and 0.25 (t-statistic=0.95) in the low refinancing frequency neigh-
borhood sample; this difference is significant at the 1% level. Loans from neighborhoods
with relatively more refinancings exhibit a greater delinquency jump at round number
asset thresholds, and the reported assets are also more likely to be targeted above these
thresholds.

      These results suggest that misrepresentation risk was greater in zip codes in which
there were relatively more refinancings than purchase loans. These areas are geographic
hot spots of misrepresentation. The results in Table 6 make clear that, within a given
zip code, the misrepresentation risk was no different for refinancings and purchase loans,
so it is not that refinancings are creating more misrepresentation in and of themselves.
Rather, it appears to be that high refinancing areas have a neighborhood characteristic
that is associated with greater misrepresentation. Most refinancings (82%) are of the
cash out variety, which generate a net payment to the borrower. A high concentration
of cash out refinancings may indicate a press to withdraw home equity from a certain
neighborhood. Perhaps in their rush to extract value from their real estate, borrowers in
these areas are apparently more willing to misrepresent their personal assets. (Sorting zip
codes based only on the frequency of cash out refinancings, and excluding refinancings for
better rates or terms, yields very similar results to those described above.) If homeowners
in these areas wish to withdraw home equity, why don’t they simply sell their houses?
It may be that potential purchasers are better informed informed than the bank of the
likely direction of local home values, and are unwilling to pay high prices. If this is true,
a cash out refinancing may a homeowner’s best way to extract cash from his or her real
estate.

      While the evidence is consistent with this home value extraction hypothesis, it is
also possible that the bank was simply less expert in evaluating loans from high refi-
nancing frequency areas, for reasons driven by unobserved variables. It is clear, however,
that from the bank’s perspective misrepresentation risk does exhibit heterogeneity across
neighborhoods, and its effects appear to be the most severe in areas in which refinancings


                                             24
are relatively common.


3.8    Misrepresentation and Origination Volume

The results in Tables 6 and 7 establish that misrepresentation risk varied with loan
terms and across neighborhoods. Was misrepresentation more severe during certain
time periods? In this section, I consider the impact of overall origination volume.

      Each month I calculate the total number of loan officers who originate at least
one mortgage during the period, and I refer to these as the active loan officers. I then
calculate the total number of originations per day divided by the number of active loan
officers. I split the sample into days with above median or below median originations per
active loan officer. As documented in the first two columns of Table 8, the discontinuity
in delinquency risk at round number asset thresholds is much greater during periods with
a high origination volume per loan officer (the difference is significant at the 5% level).
There is also more asset targeting during these periods. The estimated log difference in
kernel heights is 1.51 (t-statistic=3.80) in the high volume per loan officer sample and
0.30 (t-statistic=1.08) in the low volume per officer sample. The difference in estimates
is significant at the 1% level.

      It is clear that the bank experienced greater misrepresentation during periods in
which its loan officers were very busy. This is consistent with the hypothesis that loan
officers are less able to detect misrepresentation when they are processing many loans.
While the bank did not identify this specific form of misrepresentation during the sample
period, it is possible that loan officers are better able to identify generally suspicious
applications when overall origination per officer is lower.

      It is also possible, however, that borrowers were more likely to submit misrepre-
senting applications during busy periods. Splitting the sample simply into high origi-
nation and low origination days (without controlling for the number of loan officers), I
find, as described in the third and fourth columns of Table 8, that the discontinuity in

                                           25
delinquency at asset thresholds was significantly greater (at the 5% level) during high
origination periods. I also find more asset targeting on high origination days: the esti-
mated log difference in kernel heights is 1.48 (t-statistic=3.74) in the high volume sample
and 0.28 (t-statistic=1.02) in the low volume sample, and the difference is significant at
the 1% level. There is significant overlaps between days with high originations per loan
officer and days with high total originations. It may be that times of high originations
were periods of market expansion during which the bank was especially vulnerable to
misrepresentation. The results discussed in this section do make clear, however, that
the bank suffered much more from misrepresentation during periods of high origination
volume.


3.9      Role of Mortgage Brokers and Loan Officers

3.9.1     Broker Complicity?

What was the role of mortgage brokers in forwarding misrepresenting applications to the
bank? Did they encourage their borrowers to engage in reported asset targeting? There
are two pieces of evidence that brokers did not play a prominent role in the specific form
of misrepresentation I document here. First, the bank received information on the total
broker compensation on each origination. In unreported regressions, I divide the sample
into high and low compensation mortgages, and I find no significant difference in the
impact of above threshold assets on delinquency. There is also no significant difference
in the discontinuities in the density of reported assets in the two samples. That is,
the misrepresentation problem was no more severe for mortgages for which brokers had
explicitly greater incentives to get the transaction consummated. This is not consistent
with brokers steering their clients to misrepresent, since they would presumably be more
likely to do so on deals for which they receive greater compensation.

        Second, a subset of brokers were either suspended or terminated for cause at the
end of their relationship with the bank. These are the brokers who are most likely have


                                            26
engaged in fraud of some sort. In unreported tests, I find that these brokers exhibit
delinquency jumps and reported asset density discontinuities that are not statistically
different from those of other brokers. In other words, the brokers whom one might have
been most likely to suspect of fraudulent activities were no more likely to engage in the
misrepresentation strategy I describe here. There is no compelling evidence of broker
complicity.


3.9.2     Loan Officer Collaboration?

Might it be the case that that loan officers collaborate with misrepresenting borrowers in
an effort to originate more loans, irrespective of quality? I think this is unlikely for two
reasons. First, while it is the case that loan officers were compensated for originations,
it is also true that they were fired if too many of their loans became delinquent. No
loan officer would wish to approve a loan that had a 40.5 percentage point higher than
average likelihood of delinquency, as exhibited by loans just above asset thresholds.
The performance of the misrepresenting borrowers was so poor that it is hard to argue
that loan officers would have willingly originated these loans (though it may perhaps be
argued that they were unaware of just how risky these applicants were). Second, the
specific form of misrepresentation studied here, targeting assets above round-number
thresholds, actually had no impact on the bank’s lending decision, as discussed above.
Loan officers, as bank insiders, would have been aware of this fact, so it would be odd
for them to encourage borrowers to engage in a misrepresentation strategy that would
yield no tangible benefits.

        Borrowers made the specific misrepresentations in their applications, and it does
not appear that they received systematic guidance in doing so from either brokers or
loan officers.




                                            27
4    Conclusion

The experience of the last several years has led financial institutions to think more
carefully about misrepresentation risk. I show that personal asset misrepresentation by
borrowers was a significant risk for a bank making loans in the residential mortgage
market during the 2004-2008 sample period. This risk was present only for loans with
undocumented assets. It had the greatest impact when poor subsequent housing re-
turns and substantial potential borrower indebtedness together created the possibility
of negative equity and strategic default. The severity of misrepresentation was greater
for loans that resulted in higher debt loads, in neighborhoods with relatively frequent
refinancings and on days in which the bank’s total origination volume was high.

     The findings in this paper illustrate the prominent role that a behavioral bias,
in this case the perceived larger size of above threshold assets, can play in explaining
the actions of a misrepresenting agent. Misrepresentation risk is a possible concern
in any setting with asymmetric information. Financial institutions may well benefit
from a closer examination of behavioral biases in devising strategies to detect possible
misrepresentation.

     Successful risk assessment certainly requires the careful measurement and estima-
tion of well-understood hazards. Expanding the range of risks that can be quantified
will also help improve the performance of financial models and the accuracy of forecasts.
Nonetheless, there will always be potential losses arising from the strategic misrepresen-
tations of borrowers and counterparties. Creative methods for detecting these misrepre-
sentations can help preserve the resilience of financial institutions even in the presence
of new forms of risk.




                                           28
References

Abrevaya, Jason. 1997. The Equivalence of Two Estimators of the Fixed Effects Logit
Model. Economics Letters 55: 41–44.

Agarwal, S., I. Ben-David and V. Yao. 2012. Collateral Valuation and Institutional
Pressures: Evidence from the Residential Real-Estate Market, Working Paper, Ohio
State University.

Anderson, E., and D. Simester. 2003. Effects of $9 Price Endings on Retail Sales:
Evidence from Field Experiments. Quantitative Marketing and Economics 1: 93–110.

Ben-David, Itzhak, 2011. Financial Constraints and Inflated Home Prices during the
Real Estate Boom. American Economic Journal: Applied Economics 3(3): 55–87.

Berger, A., N. Miller, M. Petersen, R. Rajan and J. Stein. 2005. Does Function Follow
Organizational Form? Evidence from the lending practices of large and small banks.
Journal of Financial Economics 76: 237–269.

Berger, A., and G. Udell. 2002. Small Business Credit Availability and Relationship
Lending: The Importance of Bank Organizational Structure. Economic Journal 112:
F32–F53.

Bollen, N., and V. Pool. 2009. Do Hedge Fund Managers Misreport Returns? Evidence
from the Pooled Distribution. Journal of Finance 64: 2257–2288.

Bollen, N., and V. Pool. 2012. Suspicious Patterns in Hedge Fund Returns and the Risk
of Fraud. Review of Financial Studies 25: 2673–2702.

Burgstahler, D., and I. Dichev. 1997. Earnings Management to Avoid Earnings De-
creases and Losses. Journal of Accounting and Economics 24: 99–126.

Card, D., C. Dobkin, and N. Maestas. 2004. The Impact of Nearly Universal Insurance
Coverage on Health Care Utilization and Health: Evidence from Medicare. Working
Paper, National Bureau of Economic Research.

Carrillo, P. 2010. Fools or Crooks: Testing for Fraud in the Residential Mortgage Market.
George Washington University Working Paper.


                                           29
Caskey, J. 2009. Information in Equity Markets with Ambiguity-Averse Investors. Re-
view of Financial Studies 22: 3595–3627.

Cremers, K., R. Huang and Z. Sautner. 2011. Internal Capital Markets and Corporate
Politics in a Banking Group. Review of Financial Studies 24: 358–401.

Dell’Ariccia, G., and R. Marquez. 2006. Lending Booms and Lending Standards. Jour-
nal of Finance 61: 2511–2546.

Fich, E., and A. Shivdasani. 2007. Financial Fraud, Director Reputation, and Share-
holder Wealth. Journal of Financial Economics 86: 306–336.

Garrod, N., S. Pirkovic and A. Valentincic. 2006. Testing for Discontinuity or Type of
Distribution. Mathematics and Computers in Simulation 71: 9–15.

Goldreich, D., 2004. Behavioral Biases of Dealers in U.S. Treasury Auctions, Working
Paper, University of Toronto.

Graham, J., S. Li. and J. Qiu. Corporate Misreporting and Bank Loan Contracting.
Journal of Financial Economics 89: 44–61.

Hertzberg, A., J. Liberti, and D. Paravisini. 2010. Information and Incentives Inside
the Firm: Evidence from Loan Officer Rotation. Journal of Finance 65: 795–828.

Holmstrom, B. 1982. Moral Hazard in Teams. Bell Journal of Economics 13: 324–340.

Jiang, W., A. Nelson and E. Vytlacil. 2009. Liars Loan? Effects of Origination Channel
and Information Falsification on Mortgage Delinquency. Columbia University Working
Paper.

Kalyanam, K., and T. Shively. 1998. Estimating Irregular Pricing Effects: A Stochastic
Spline Regression Approach. Journal of Marketing Research 35: 16–29.

Keys, B., T. Mukherjee, A. Seru and V. Vig. 2010. Quarterly Journal of Economics
125: 307–362.

Liberti, J., and A. Mian. 2009. Estimating the Effect of Hierarchies on Information Use.
Review of Financial Studies 22: 4057–4090.



                                          30
                             o
Maddaloni, A., and J. Peydr´. 2011. Bank Risk-taking, Securitization, Supervision,
and Low Interest Rates: Evidence from the Euro-area and the U.S. Lending Standards.
Review of Financial Studies 24: 2121–2165.

Matsudaira, J. D. 2008. Mandatory Summer School and Student Achievement. Journal
of Econometrics 142:829–850.

McCrary, J. 2008. Manipulation of the Running Variable in the Regression Discontinuity
Design: A Density Test. Journal of Econometrics 142:698–714.

Mian, A., and A. Sufi. 2009. The Consequences of Mortgage Credit Expansion: Evidence
from the U.S. Mortgage Default Crisis. Quarterly Journal of Economics 124: 1449–1496.

Norden, L., and M. Weber. 2010. Credit Line Usage, Checking Account Activity, and
Default Risk of Bank Borrowers. Review of Financial Studies 23: 3665–3699.

Petersen, M., and R. Rajan. 2002. Does Distance Still Matter? The Information
Revolution in Small Business Lending. Journal of Finance 57: 2533–2570.

Povel, P., R. Singh and A. Winton. 2007. Booms, Busts, and Fraud. Review of Financial
Studies 20: 1219–1254.

Ruckes, M. 2004. Bank Competition and Credit Standards. Review of Financial Studies
17: 1073–1102.

Scott, J. 2006. Loan Officer Turnover and Credit Availability for Small Firms. Journal
of Small Business Management 44: 544–562.

Stein, J. 2002. Information Production and Capital Allocation: Decentralized versus
Hierarchical Firms. Journal of Finance 57: 1891–1921.

Thomas, M., and V. Morwitz. 2005. Penny Wise and Pound Foolish: The Left-Digit
Effect in Price Cognition. Journal of Consumer Research 54–64.




                                         31
                  Figure 1: Delinquency and Reported Personal Assets- Full Sample
            .6
            .5
Delinquency Probability
      .3    .2
            .1 .4




               −100000                −50000                   0                50000
                                           Normalized Assets




                                                32
  Figure 2: Delinquency and Reported Personal Assets- Unverified Assets
            .6
            .5
Delinquency Probability
     .3     .2
            .1.4




               −100000    −50000                   0              50000
                               Normalized Assets




                                    33
          Figure 3: Delinquency and Reported Personal Assets- Verified Assets
             .6
             .5
 Delinquency Probability
.2       .3  .1  .4




                −100000         −50000                   0                50000
                                     Normalized Assets




                                          34
                   Figure 4: Density of Normalized Assets- Unverified Assets
          .00003
          .00002
Density
          .00001
          0




           −100000                  −50000                   0                50000
                                         Normalized Assets




                                              35
                     Figure 5: Density of Normalized Assets- Verified Assets
          .000015
          .00001
Density
          5.00e−06
          0




            −100000                  −50000                   0               50000
                                          Normalized Assets




                                               36
                                      Table 1: Summary Statistics
Observations are at the loan level. Assets describes the personal assets of the borrower, excluding the property used as collateral.
Assets above threshold is an indicator for whether the normalized assets of the borrower (as defined in (1)) exceed zero. Income is
the monthly income of the borrower. Rate spread is the interest premium paid by the borrower relative to an index. Credit score is
the borrower’s FICO score, the loan amount is given in dollars, LTV is the loan-to-value ratio and CumLTV is the cumulative LTV
including any existing mortgage on the property. Broker compensation consists of a rebate paid to the broker by the bank and may
also include a direct payment from the borrower. In cash out refinances, the borrower withdraws equity from the property, while
rate/term refinances only involve a change in the interest rate or maturity. Assets and income verified are indicators for whether the
borrower provided documentation supporting his asset and income claims, respectively. Delinquency is an indicator for whether a
loan was 30 or more days past due.

                                         .
                                      Mean     Median                 Standard Deviation               1st %     99th %
Assets                              246477.18 53392.00                    1243156.60                 2798.00 3666909.00
Assets Above Threshold                 0.68      1.00                         0.47                     0.00       1.00
Income                              16086.35 10319.00                      34334.46                  2920.00   132100.00
Rate Spread                            3.36      3.50                         0.60                     2.25       4.55
Credit Score                          719.45    716.00                       45.40                    628.00     808.00
Loan Amount                         502195.00 412500.00                    388161.25                122500.00 2000000.00
LTV                                    0.72      0.77                         0.14                     0.27       0.95
CumLTV                                 0.73      0.78                         0.15                     0.27       0.95
Broker Compensation                 10446.55   9215.00                      6604.49                  1130.00    34111.91
Cash out Refinance                      0.65      1.00                         0.48                     0.00       1.00
Rate/Term Refinance                     0.17      0.00                         0.38                     0.00       1.00
Assets Verified                         0.60      1.00                         0.49                     0.00       1.00
Income Verified                         0.28      0.00                         0.45                     0.00       1.00
Delinquent                             0.20      0.00                         0.40                     0.00       1.00




                                                        37
                                 Table 2: Misrepresentation of Assets
Results from the regressions of an indicator for delinquency on borrower and transaction characteristics. The regressors with
reported coefficients are a dummy for whether the normalized assets of the borrower exceed zero (in columns 1-5), the rate
spread on the mortgage (columns 2-6), the credit score of the borrower (columns 2-6), the loan-to-value ratio on the mortgage
(columns 2-6), the minimum housing return in the property zip code in the period subsequent to the financing (columns 3-4)
and a dummy for whether the normalized income of the borrower exceeds zero (column 6). The regressions also include as
controls a sixth degree polynomial in assets (columns 1-5), a sixth degree polynomial in income (column 6), monthly fixed
effects (columns 2-6) and zip code fixed effects (column 4). Reported t-statistics are heteroskedasticity-robust and clustered by
month of origination (columns 1-3,5-6) or zip code (column 4).

                                       Delinq?      Delinq?      Delinq?      Delinq?       Delinq?           Delinq?
Assets Above Threshold                 0.246∗∗      0.405∗∗      0.375∗∗      0.454∗∗       -0.0230
                                        (3.47)       (3.42)       (3.28)       (3.19)       (-0.29)

Rate Spread                                         0.0368∗       0.0309      0.0465∗      0.0455∗∗           0.0462∗∗
                                                     (1.85)       (1.60)       (1.88)       (3.86)             (2.57)

Credit Score                                         -0.740       -0.762       -0.640      -0.814∗∗           -1.284∗∗
                                                     (-1.03)      (-1.11)      (-0.77)      (-3.00)            (-2.67)

LTV                                                 0.241∗∗      0.172∗∗       0.229        0.0717             0.247∗∗
                                                     (3.34)       (2.14)       (1.39)       (1.47)              (4.49)

Housing Return                                                   -0.400∗∗      -0.0475
                                                                  (-9.88)      (-0.16)

Income Above Threshold                                                                                         -0.0154
                                                                                                               (-0.19)
6th-degree polyn. in Assets              Yes          Yes          Yes          Yes          Yes                 No
6th-degree polyn. in Income              No            No           No           No           No                 Yes
Monthly F.E.                             No           Yes          Yes          Yes          Yes                 Yes
Zip Code F.E.                            No            No           No          Yes           No                 No
Sample                                   Full        Unver.       Unver.       Unver.        Ver.              Unver.
                                                     Assets       Assets       Assets       Assets             Income
Observations                             8287         3276         3275         3275         4994               6017
Adjusted R2                              0.008       0.102        0.118        0.125        0.124               0.103
t statistics in parentheses
∗
  p < 0.10, ∗∗ p < 0.05




                                                        38
          Table 3: Asset Discontinuity and Loan and Borrower Characteristics
Results from the regressions of loan and borrower and transaction characteristics on reported asset discontinuities. The depen-
dent variables are the rate spread on the loan (column 1), the exception pricing on the loan (column 2), the loan-to-value ratio
(column 3), the cumulative loan-to-value including any existing mortgage (column 4), the log of the loan amount in dollars
(column 5), the log of the loan maturity in months (column 6), the maximum permitted negative amortization (column 7) and
the borrower credit score (column 8) The regressors with reported coefficients are a dummy for whether the normalized assets of
the borrower exceed zero, the credit score of the borrower (columns 1-7) and the loan-to-value ratio on the mortgage (columns
1,2 and 6-8). The regressions also include as controls a sixth degree polynomial in assets, monthly fixed effects and zip code
fixed effects. Reported t-statistics are heteroskedasticity-robust and clustered by month of origination.
                              Rate Spr.      Exc. Pr.        LTV        CumLTV          Amt.         Mat.      NegAm       Cred. Sc.
Assets Abv. Thresh.           -0.00779        0.0970      -0.000951     -0.00248       0.00983      0.0269     0.0169      -0.00140
                               (-0.09)        (1.37)        (-0.03)      (-0.06)        (0.10)      (1.00)     (0.47)       (-0.10)

Credit Score                    0.0427       -0.770∗∗      -0.971∗∗      -1.153∗∗      -3.371∗∗     0.0833     0.202
                                (0.05)        (-2.13)       (-4.13)       (-4.34)       (-6.95)     (0.39)     (0.45)

LTV                            -0.0506        0.125                                                 0.0209     0.0191      -0.0838∗∗
                               (-0.27)        (0.84)                                                (0.55)     (0.29)       (-4.47)
6th-deg. polyn. Assets           Yes           Yes           Yes            Yes           Yes        Yes         Yes          Yes
Monthly F.E.                     Yes           Yes           Yes            Yes           Yes        Yes         Yes          Yes
Zip Code F.E.                    Yes           Yes           Yes            Yes           Yes        Yes         Yes          Yes
Observations                    3276           3276         3276           3276          3276        3276       3182          3276
Adjusted R2                     0.081         0.072         0.240          0.235         0.602      0.049       0.012        0.082
t statistics in parentheses
∗
  p < 0.10, ∗∗ p < 0.05




                                                         39
         Table 4: Reported Assets and Rate Spread
Results from the regression of the rate spread on the loan on reported assets,
reported income and loan and borrower characteristics. The regressors with
reported coefficients are the log of reported assets (column 1), the log of re-
ported income (column 1), the level of reported assets (column 2), the level of
reported income (column 2), the credit score and the loan-to-value ratio. The
regressions also include as controls the loan maturity and maximum permit-
ted amortization, monthly fixed effects and zip code fixed effects. Reported
t-statistics are heteroskedasticity-robust and clustered by month of origination.

                      Rate Spread                    Rate Spread
Log(Assets)            -0.0233∗∗
                        (-2.35)

Log(Income)              -0.00424
                          (-0.21)

Assets                                                 -0.0848∗∗
                                                        (-3.06)

Income                                                   -1.073
                                                         (-0.69)

Credit Score              -0.164                         -0.258
                          (-0.45)                        (-0.70)

LTV                      -0.00835                       -0.0489
                          (-0.06)                       (-0.38)
Monthly F.E.                Yes                           Yes
Zip Code F.E.               Yes                           Yes
Observations               3173                          3173
Adjusted R2                0.533                         0.534
t statistics in parentheses
∗
  p < 0.10, ∗∗ p < 0.05




                                      40
Table 5: Effects of Misrepresentation by Housing Returns and Maximum Indebtedness
Results from the regressions of an indicator for delinquency on reported asset discontinuities in subsamples varying by subsequent
housing returns and maximum potential borrower indebtedness. The regressors with reported coefficients are a dummy for whether
the normalized assets of the borrower exceed zero, the rate spread on the mortgage, the credit score of the borrower and the loan-
to-value ratio on the mortgage. The regressions also include as controls a sixth degree polynomial in assets, monthly fixed effects
and zip code fixed effects. Columns 1 and 2 split the sample into loans with minimum subsequent housing returns below and above
the sample median (-35.4%). Columns 3 and 4 split the sample into loans with maximum potential indebtedness above and below
the sample median (87.4%). Columns 5 and 6 split the sample into loans that could potentially have become underwater during the
sample period and those that could not. Reported t-statistics are heteroskedasticity-robust and clustered by month of origination.

                                       Delinq?        Delinq?         Delinq?         Delinq?         Delinq?          Delinq?
Assets Above Threshold                 0.596∗         0.419∗∗         0.544∗∗          0.332          0.500∗∗           0.102
                                        (1.94)         (2.07)          (2.85)         (1.18)           (3.74)          (0.45)

Rate Spread                            0.115∗∗        -0.0183          0.0442         0.0578∗∗         0.0249          -0.0324
                                        (6.49)        (-0.50)          (0.63)          (2.65)          (1.31)          (-0.88)

Credit Score                           -1.189         -0.221           -1.721          0.0397          -1.182           0.510
                                       (-0.99)        (-0.18)          (-1.05)         (0.05)          (-1.48)          (0.50)

LTV                                   0.481∗∗         0.248           2.672∗           0.186          0.701∗∗          0.131
                                       (2.39)         (1.20)          (1.90)           (0.79)          (2.30)          (1.18)
6th-degree polyn. in Assets             Yes            Yes             Yes              Yes             Yes             Yes
Monthly F.E.                            Yes            Yes             Yes              Yes             Yes             Yes
Zip Code F.E.                           Yes            Yes             Yes              Yes             Yes             Yes
Sample                               Low Hsg.       High Hsg.       High Max.        Low Max.         Poten.         Not Poten.
                                      Return         Return         Indebted.        Indebted.       Underwtr.       Underwtr.
Observations                            1651           1624            1633             1643            2514            667
Adjusted R2                            0.133          0.089           0.110            0.133           0.073           0.099
t statistics in parentheses
∗
  p < 0.10, ∗∗ p < 0.05




                                                        41
                       Table 6: Misrepresentation by Loan Characteristics
Results from the regressions of an indicator for delinquency on reported asset discontinuities in subsamples varying by loan
characteristics. The regressors with reported coefficients are a dummy for whether the normalized assets of the borrower
exceed zero, the rate spread on the mortgage, the credit score of the borrower, the loan-to-value ratio on the mortgage
and a dummy for whether the loan could have potentially become underwater. The regressions also include as controls
a sixth degree polynomial in assets, monthly fixed effects and zip code fixed effects. Columns 1 and 2 split the sample
into loans with loan-to-value ratio above and below the sample median (78%). Columns 3 and 4 split the sample into
loans with cumulative loan-to-value ratio (including any existing mortgage) above and below the sample median (79%).
Columns 5 and 6 split the sample into refinancings and loans originated to support a purchase of a new home. Reported
t-statistics are heteroskedasticity-robust and clustered by month of origination.

                                       Delinq?       Delinq?        Delinq?         Delinq?       Delinq?        Delinq?
Assets Above Threshold                 0.667∗∗        0.323         0.612∗∗          0.238        0.408∗∗         0.706
                                        (2.56)        (1.21)         (2.92)          (1.29)        (2.18)        (0.84)

Rate Spread                             0.0480       -0.00882        0.0718          0.0198        0.0401         0.124
                                        (0.81)        (-0.31)        (1.45)          (0.60)        (1.69)         (0.73)

Credit Score                            -1.610        0.567          -2.029         1.704∗∗        -0.664         5.063
                                        (-1.20)       (0.43)         (-1.43)         (2.12)        (-0.76)        (1.00)

LTV                                     2.413         0.0881         -1.377          0.0787         0.209         2.170
                                        (1.07)        (0.47)         (-1.11)         (0.53)         (1.47)        (0.63)

Pot. Underwater                         0.0975        0.0337        0.0485          0.0399         0.00767       -0.0717
                                        (1.17)        (0.74)        (0.62)          (1.24)          (0.26)       (-0.16)
6th-degree polyn. in Assets              Yes            Yes          Yes             Yes             Yes           Yes
Monthly F.E.                             Yes            Yes          Yes             Yes             Yes           Yes
Zip Code F.E.                            Yes            Yes          Yes             Yes             Yes           Yes
Sample                                   High          Low           High            Low             Refi        Purchase
                                         LTV           LTV         CumLTV          CumLTV
Observations                             1635          1641          1638            1638           2877           399
Adjusted R2                             0.080          0.151        0.086           0.145           0.125         0.159
t statistics in parentheses
∗
  p < 0.10, ∗∗ p < 0.05




                                                          42
                        Table 7: Misrepresentation Across Neighborhoods
Results from the regressions of an indicator for delinquency on reported asset discontinuities in subsamples varying by
neighborhood characteristics. The regressors with reported coefficients are a dummy for whether the normalized assets
of the borrower exceed zero, the rate spread on the mortgage, the credit score of the borrower, the loan-to-value ratio on
the mortgage and a dummy for whether the loan could have potentially become underwater. The regressions also include
as controls a sixth degree polynomial in assets, monthly fixed effects and zip code fixed effects. Columns 1 and 2 split the
sample into loans from zip codes with average loan-to-value ratios above and below the sample median (73%). Columns
3 and 4 split the sample into loans from zip codes with average cumulative loan-to-value ratios (including any existing
mortgage) above and below the sample median (74%). Columns 5 and 6 split the sample into loans from zip codes with
average frequency of refinancings (relative to total originations) above and below the sample median (86%). Reported
t-statistics are heteroskedasticity-robust and clustered by month of origination.

                                       Delinq?       Delinq?       Delinq?        Delinq?        Delinq?       Delinq?
Assets Above Threshold                 0.573∗∗        0.297        0.491∗∗        0.375∗∗        0.744∗∗        0.181
                                        (2.86)        (1.63)        (2.16)         (2.01)         (3.38)       (0.69)

Rate Spread                            0.0837∗∗      0.00703        0.0523         0.0517∗        0.0240        0.0529
                                        (2.31)        (0.26)        (1.43)          (1.95)        (0.79)        (1.57)

Credit Score                            -0.650       -0.386         -0.815         -0.180        -0.945         0.174
                                        (-0.61)      (-0.35)        (-0.81)        (-0.18)       (-0.69)        (0.20)

LTV                                     0.167         0.251         0.109           0.246        0.459∗∗        0.0319
                                        (0.54)        (1.56)        (0.33)          (1.32)        (2.66)        (0.17)

Pot. Underwater                        0.00717       0.00586       0.00552        0.00402         0.0396       -0.0402
                                        (0.14)        (0.13)        (0.11)         (0.09)         (0.87)       (-1.23)
6th-degree polyn. in Assets              Yes           Yes           Yes            Yes            Yes           Yes
Monthly F.E.                             Yes           Yes           Yes            Yes            Yes           Yes
Zip Code F.E.                            Yes           Yes           Yes            Yes            Yes           Yes
Sample                                   High          Low           High           Low            Freq         Infreq
                                         LTV           LTV        CumLTV         CumLTV            Refi           Refi
                                         Nbd           Nbd           Nbd            Nbd            Nbd           Nbd
Observations                             1635          1641          1640           1636           1641          1635
Adjusted R2                             0.125         0.090         0.114          0.116          0.141         0.105
t statistics in parentheses
∗
  p < 0.10, ∗∗ p < 0.05




                                                         43
                          Table 8: Misrepresentation and Origination Volume
Results from the regressions of an indicator for delinquency on reported asset discontinuities in subsamples varying by
daily volume characteristics. The regressors with reported coefficients are a dummy for whether the normalized assets of
the borrower exceed zero, the rate spread on the mortgage, the credit score of the borrower, the loan-to-value ratio on
the mortgage and a dummy for whether the loan could have potentially become underwater. The regressions also include
as controls a sixth degree polynomial in assets, monthly fixed effects and zip code fixed effects. Columns 1 and 2 split
the sample into loans from days with originations per loan officer above and below the sample median. Columns 3 and 4
split the sample into loans from days with total originations above and below the sample median. Reported t-statistics
are heteroskedasticity-robust and clustered by month of origination.

                                         Delinquent?         Delinquent?         Delinquent?         Delinquent?
Assets Above Threshold                     0.661∗∗             0.0232              0.789∗∗             -0.0235
                                           (1.99)              (0.12)              (2.25)              (-0.11)

Rate Spread                                  0.0380              0.0513             0.0527               0.0581
                                             (1.04)              (1.33)             (1.24)               (1.48)

Credit Score                                 1.574               -2.451              0.619               -2.238∗
                                             (1.23)              (-1.64)             (0.60)              (-1.96)

LTV                                          0.409∗              0.0725              0.395               0.204
                                             (1.91)              (0.21)              (1.76)              (0.64)

Pot. Underwater                            0.0519              -0.00556             0.0617              -0.0289
                                           (1.08)               (-0.13)             (1.37)               (-1.11)
6th-degree polyn. in Assets                  Yes                  Yes                Yes                   Yes
Monthly F.E.                                 Yes                  Yes                Yes                   Yes
Zip Code F.E.                                Yes                  Yes                Yes                   Yes
Sample                                  High Vol per         Low Vol per           High Vol             Low Vol
                                         Loan Off.             Loan Off.
Observations                                1620                 1656                1632                 1644
Adjusted R2                                 0.085                0.117               0.094                0.134
t statistics in parentheses
∗               ∗∗
    p < 0.10,        p < 0.05




                                                        44

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:13
posted:9/19/2012
language:English
pages:45