Electronic copy available at: http://ssrn.com/abstract=1706844
Determinants and Consequences of Mortgage Default
∗
Yuliya Demyanyk
†
Federal Reserve Bank
of Cleveland
Ralph S.J. Koijen
‡
University of Chicago
Booth School of Business
and NBER
Otto A.C. Van Hemert
§
AQR Capital Management
January 2011
Abstract
We study a unique data set of borrower-level credit information from Tran s Union, one
of the three major credit bureaus, which is linked to a database containing d etailed infor-
mation on the b orrowers’ mortgages. We find that the updated credit score is an important
predictor of mortgage default in addition to the credit score at origination. However, the
6-month change in the credit score also predicts default: A positive change in th e credit score
significantly reduces the probab ility of delinquency or foreclosure. Next, we analyze the con-
sequences of default on a borrower’s credit score. The credit score drops on average 51 points
when a borrower becomes 30-days delinquent on his mortgage, but the effect is much more
muted for transitions to more severe delinquency states and even for foreclosure.
∗
First version: August 2010. We thank Jane Blomquist, Antoni Guitart, Erik Hurst, Gregor Matvos, Joe
Mellman, Amit Seru, Amir Sufi, Stijn Van Nieuwerburgh, Claudia Wood, and Luigi Zingales for useful comments
and suggestions. The views expressed are those of the authors and do not necessarily r e flec t the official po sitions of
the Federal Reserve Bank of Cleveland or the Federal Reserve Sys tem.
†
Yuliya.Demyan[email protected]g.
‡
Ralph.Koijen@chicagoboo th.edu. Koijen is a lso associated with Netspa r (Tilburg University).
§
Electronic copy available at: http://ssrn.com/abstract=1706844
We study the determinants and consequences of mortgage default using a unique data set of
borrower-level credit information from TransUnion, one of the three major credit bureaus. This
database is linked to the LoanPerformance database, which contains detailed information on the
borrowers’ mortgages. Understanding the determinants of default, that is, which borrowers ar e
more likely to be delinquent or to face foreclosure, is important for lenders and policy makers.
1
We
also study the consequences of default and in particular how a borrower’s credit score is affected
by default. For instance, a commonly heard claim is that credit scores deteriorate substantially in
cases of foreclosure. This may motivate borrowers to avoid foreclosure even if t he value of their
house is lower than the value of their mortgage. Hence, to understand the financial incentives
borrowers face in case of default, it is important to obtain estimates of how delinquencies and
foreclosures affect credit scores.
2
Previous studies on the topic of the determinants and consequences of default face data limi-
tations because they mostly rely on loan-level mortgage databases.
3
As a result, the determinants
of default t hat can be a nalyzed are restricted to borrower and loa n characteristics known at orig-
ination only. The consequences of default cannot be measured using these data, as no borrower
identifier is provided, implying that one cannot follow borrowers after a loan ceases to exist.
We make use of two main data sources to overcome both issues. First, we use individual-
level credit data from TransUnion’s Consumer Credit Database. This is a comprehensive and rich
database summarizing the credit situation of households. Second, we use the LoanPerformance
database from CoreLogic, which contains loan-level data on U.S. subprime and Alt-A mortgag e
loans. This is the main database used by institutional investors to analyze the underlying collateral
of non-agency mortgage-backed securities. CoreLogic and TransUnion have joined forces to provide
a highly accurate match of the their two databases. We use the matched database, TransUnion
Consumer Risk Indicators for RMBS, in this paper. We augment the data with a set of macro-
economic variables. Our results illustrate the improved ability of the merged data ba se to predict
future default and the effects on a borrower’s credit score in response to a default. This is for
instance releva nt to value mortgage-backed securities contracts.
We consider four types of default: 30-day delinquency, 60-day delinquency, 90+-day delin-
quency, and foreclosure. We focus in each case on the first-lien mortgage. For the determinants of
1
For instance, to the extent that for e c losure waves have a negative effect on neighboring house prices, see
Campb e ll, Giglio, a nd Pathak (2010), micro-level data might be useful to identify areas that might be more prone
to foreclosure waves.
2
Even if housing equity is negative, borrowers may decide not to default because of borrowing cons traints, see
Campb e ll and Cocco (2010 ). In a ddition, borrowers may decide not to default for non-financial reasons such as
social stigma, see Guiso, Sapienza, and Zingales (2009).
3
See for example, Demyanyk and Van Hemert (2010), Agarwal, Ambrose, Chomsisengphet, and Sanders (2010),
Amromin and Paulson (2009), Gerardi, Lehnert, Sherlund, and Willen (2009), Archer and Smith (2010), Bajari,
Chu, and Park (2008), and J iang, Nelson, and Vytlacil (2010).
1
default we find that the data from TransUnion have predictive power in addition to Lo anPerfor-
mance data only. First, a low credit score, the VantageScore in case of TransUnion, in the previous
month substantially increases the probability of default, as one would expect. The VantageScore
robustly predicts default for all types of default we consider. Second, we show that the current
VantageScore is not a sufficient statistic, as past credit scores are a lso informative. In particular,
holding constant the previous month’s VantageScore, a higher VantageScore six months before
forecasts a higher probability of default. For instance, if the VantageScore is 700 today, down
from 800 six months b efo re, the probability of default is substantially higher compared to a Van-
tageScore of 700 today, up fro m 600 six months ago. We refer to this effect as the “VantageScore
momentum.”
Figure 1: VantageScore distribution one month before and the month of a transition to worse state
The left panel shows the VantageScore distri bution in the month before and in the month of a transition from current to 30-days
delinquency. Similarly, the right panel shows the VantageScore distribution in the month before and in the month of a transition from
90+-days delinquency to foreclosure.
0 5 10 15 20
Percent
500 600 700 800 900 1000
Month of event 1 month before event
Transition to 30−day delinquency
0 10 20 30 40
Percent
500 600 700 800 900
Month of event 1 month before event
Transition to foreclosure
To analyze the consequences of default, we first focus on how the VantageScore is affected by
default. We then link changes in the VantageScore to the cost of debt as measured by mortgage
rates. To illustrate the impact of default on the VantageScore, we plot in Figure 1 the VantageScore
distribution one month before and in the month of a transition to 30-day delinquency (left panel)
and foreclosure (right panel). For the transition to 3 0-day delinquency, we require the mortgage to
be current the month before. For the transition to foreclosure, we require the mortgage to be 90+-
day delinquent the month befor e, which is the case for 76% of the borrowers that enter foreclosure.
Figure 1 shows that a transition from a current payment status to 30-day delinquency shifts the
VantageScore distribution to the left. This implies that VantageScores decline substantially if a
mortgage payment is missed. However, the VantageScore distribution is hardly affected in the
2
case of foreclosure. We estimate that a transition from current to 30 -day delinquent is on average
associated with a drop of the VantageScore of 51 points. For transitions to 6 0-day delinquency,
90+-day delinquency, and foreclosure, the effect is increasingly muted at 25-, 1 4-, and 6-point
drops, respectively. The common wisdom that “foreclosure ruins your credit score” does not seem
to be the case for most foreclosure cases, which are already 90+-days delinquent in the month prior
to foreclosure.
4
Figure 2: Average VantageScore before and after transitioning to delinquency and foreclosure
The figure shows the average VantageScore during the 12 months preceding, and the 12 months following delinquency and foreclosure.
The sample of borrowers is split into four groups according to the borrowers’ VantageScores in the month prior to the credit event.
Group 0 consists of borrowers with credit scores in the range [501,550]; Group 1: (550,700]; Group 2: (700,800]; Group 3: (800,990].
For each group, we repor t the number of borrowers in our sample by N.
550
600
650
700
750
800
850
900
Average VantageScore
−12 −10 −8 −6 −4 −2 0 2 4 6 8 10 12
Months relative to transition
Group 3; N= 885
Group 2; N= 2,582
Group 1; N= 9,539
Group 0; N= 1,377
Transition from Current to 30 days past due
550
600
650
700
750
800
850
900
Average VantageScore
−12 −10 −8 −6 −4 −2 0 2 4 6 8 10 12
Months relative to transition
Group 3; N= 44
Group 2; N= 356
Group 1; N= 4,274
Group 0; N= 2,106
Transition from 30 to 60 days past due
550
600
650
700
750
800
850
900
Average VantageScore
−12 −10 −8 −6 −4 −2 0 2 4 6 8 10 12
Months relative to transition
Group 3; N= 12
Group 2; N= 181
Group 1; N= 2,382
Group 0; N= 1,887
Transition from 60 to 90 days past due
550
600
650
700
750
800
850
900
Average VantageScore
−12 −10 −8 −6 −4 −2 0 2 4 6 8 10 12
Months relative to transition
Group 3; N= 7
Group 2; N= 106
Group 1; N= 1,332
Group 0; N= 1,548
Transition from 90+ days past due to foreclosure
Figure 2 illustrates the muted response of the VantageScore to more severe credit events. The
4
For examples see several news articles: “A foreclosure will drop the borrower’s c redit sc ore by at least 100
points” (The New York Times, 25 October 2009) and “If your ho use goes into fo reclosure, you might take a hit of
150 points or more on your credit score” (Chicago Tribune, 31 January 2010).
3
figure displays the average VantageScore during the 12 mont hs preceding, and the 12 months follow-
ing, the credit event. The top left panel reports the results for a transition to 30-day delinquency,
the top right panel for a transition to 60-day delinquency, the bottom left panel for a transition to
90-day delinquency, and the bottom right panel for a transition to f oreclosure. We r eport the re-
sults for borrowers with different VantageScores, where we differentiate between four credit groups.
As one may expect, Figure 2 also shows that the impact of any credit event is bigger for higher
credit scores. The main insight is that the va st majority of borrowers that transition to foreclosure
are 90+-day delinquent, and foreclosure has virtually no impact on their Va ntageScores. Hence,
due to the cross-sectional distribution of credit scores prior to foreclosure, the average impact o f a
foreclosure is negligible.
We also study the VantageScore one year after default. Figure 3 shows the (cumulative) change
in the VantageScore f ollowing the event in the month of, and 12 months after, a transition to
30-day delinquency (left panel) and foreclosure (right panel). We compute the cumulative change
as a function of the VantageScore the month prior to the event.
5
Figure 3 illustrates that 30-day
delinquency has a strong negative effect on VantageScores both at the moment of delinquency
and one year a fter. Foreclosures, by contrast, have a small negative effect on VantageScores in
the month of the event. In fact, for borrowers who are 90+-delinquent, we find a significant
improvement in their VantageScores one year later, regardless of whether a fo r eclosure takes place
or not.
To better understand why the VantageScores of borrowers improve even after a foreclosure,
we study delinquencies on bank and credit card debt and the number of credit inquiries in the
last six months before the month of the foreclosure and one year later. One view contends that
by not making mortgage payments, households can use this income to make payments on other
forms of credit. Also, by relaxing the budget constraint, these borrowers may not require new
credit. Consistent with this interpretation, we find a substantial decline in the number of bank
and credit card delinquencies and a similar decline in the number of credit inquiries for borrowers
in f oreclosure.
We then analyze the link between changes in VantageScores and mortg age rates to quantify
the impact of delinquency and foreclosure on future mortgage rates. Our estimates suggest that,
on average, a one-point drop in the VantageScore corresponds to approximately a one basis point
increase in the mortgage rate for fixed-rate mortgages. This point estimate implies that the mort-
gage rate on a new mortgage increases by 51bp following 30-day delinquency, by another 25bp
after 60-day delinquency, by 14bp following 90+-delinquency, and by 6bp following foreclosure.
5
We truncate the plots at values of the VantageScore of 800 and 700 for the left and right panels, respectively, as
too few borrowers have a sufficiently high VantageScore the month prior to a transition to a worse state, see Figure
1.
4
Figure 3: VantageScore change in the month of a tr ansition to worse state and 1 2 months after
This figure plots the (cumulative) change of the VantageScore in the month of a transition to a worse state and 12 months after, as a
function of the VantageScore the month prior to the credit event. For the transition to 30-day delinquency (left panel) we require the
mortgage to be current the month before the event. For transition to foreclosure (right panel), we require the mortgage to be 90+-day
delinquent the month before the event.
−150 −100 −50 0 50
Change vantage score
500 600 700 800
Month of event 12 months after event
Transition to 30−day delinquency
−40 −20 0 20 40
Change vantage score
500 550 600 650 700
Month of event 12 months after event
Transition to foreclosure
The cumulative impact from a series of transitions from current to f oreclosure is about 1% in total,
which would increase the annual borrowing costs by $2 ,0 00 for a $200,000 mortgage.
Hybrid mortgages have a fixed-rate period, after which the mortgage rate will be floating and
tied to a benchmark interest rate. The borrower pays the benchmark rate plus a margin during
this lat ter period. Fo r hybrid mortgages, we estimate the increase of the initial mortgage r ate to be
0.6 to 0 .7 basis points for a one-point drop in the VantageScore. For the two-year hybrid contract,
the margin increases by 0.3 basis point, while for t he three-year hybrid contract we estimate this
number to be 0.8 basis point for a one-point drop in the VantageScore.
The first part of our paper, which analyzes determinants of default, complements a large litera-
ture on the topic by introducing variables linked to credit bureau infor matio n such as VantageScore
momentum. Early contributions for instance include Von Furstenb erg (1969) , von Furstenberg
and Green (1 974), and Campbell (1983). R ecent contributions include Deng, Quigley, and Or-
der (2000), Crews and Order (2005), Pennington-Cross and Chomsisengphet (2 007), Ghent and
Kudlyak (2010) , Agarwal, Chomsisengphet, and Liu (2010), and Elul, Chomsisengphet, Glennon,
Hunt, and Souleles (2010).
6
This issue obviously received renewed attention during the great re-
cessions. For papers analyzing the subprime mortgage crisis, see for instance Demyanyk and Van
Hemert (2010), Mayer and Pence (2008), Gerardi, Shapiro, and Willen (2008), Mian and Sufi
(2009), and Keys, Mukherjee, Seru, and Vig (2010). Closest to our paper are Gross and Souleles
6
See also Fay, Hurst, and White (2002) for deter minants of household bankruptcy.
5
(2002) and Elul, Chomsisengphet, Glennon, Hunt, and Souleles (2010) . The former paper uses the
updated credit score to predict credit card delinquency and personal bankruptcy. Gross and Soule-
les (2 002) also find that the most recent credit score is not a sufficient statistic, consistent with
our findings. Elul, Chomsisengphet, Glennon, Hunt, and Souleles (2010) find that the updated
credit score helps to predict 60-day mortgage delinquencies. We show in addition that short-term
trends in the Va ntageScore are highly informative about future delinquency and foreclosure, and
comparable to the most recent VantageScore itself.
The second part of our paper that studies consequences of default for a borrower’s VantageScore,
has fewer directly-related papers, as we use a newly-merged database of credit and mortgage data
for this purpose.
7
We then link these changes to changes in borrowing costs.
This paper continues as follows. In Section 1, we describe the data set we use. We study the
determinants of default in Section 2. In Section 3, we analyze the consequences, in terms of a
borrower’s VantageScore, of delinquency and f oreclosure, and we translate this to mortgage rates
in Section 4. Section 5 concludes. Two appendices describe further robustness checks.
1 Data
In this section, we describe the data sources, the selection criteria we use, and the types of credit
events we study. Table 1 provides an overview of the main definitions.
1.1 Data sources
We use borrower-level credit data from TransUnion’s Consumer Credit Data base. This data set
contains detailed monthly information about the credit situation of mortgage borrowers. The data
cover most borrowers who at some point during the September 2004 to July 2009 sample period
have a securitized subprime or Alt-A mortgage. There are more than 250 attributes. Fo r an
exhaustive list of credit accounts, like mortgage, bank, and department store accounts, we know
the payments status, utilization rates, and requests for new lines of credit. We have monthly
updated information on the VantageScore.
Our second main dat a source is the loan-level LoanPerformance (LP) Securities da tabase pro-
vided by CoreLogic. This data set contains information about loan and borrower characteristics
at origination and monthly loan performance for about 85% of all U.S. subprime and Alt-A se-
curitized mortgage loans. It is the main database used by institutional investors to analyze the
underlying collateral of non-agency mortgage-backed securities. For each loan in the LP data set
7
In a recent paper, Mian, Sufi, and Trebbi (2 010) measure the impact of foreclo sures on house price s, residential
investment, and dura ble consumption.
6
we observe most of the underwriting criteria measured at the time of loan origination: FICO credit
scores, debt-to-income ratios, and loan-to-value ratios. Also, for each mortgage we know the type
(fixed-rate, adjustable-rate, hybrid, balloon, interest-only, et cetera), the structure (prepayment
penalty, timing and types of rate resets, lien, et cetera), the location of the property (zip code
and state), the mortga ge rate at origination and thereafter, and the monthly performance after
securitization. The LP data set does not contain information about other loans, credit accounts,
or the updated credit information after the mortgage has been originated, apart from the payment
status.
CoreLogic and TransUnion recently developed an accurate link between both databases, and we
study this linked database, which is called the TransUnion Consumer Risk Indicators f or RMBS.
The matching alg orithm keys off of overlapping load data between the two databases. Actual
borrower names and addresses are used to minimize f alse positive matches generated by the algo-
rithm. The match rate is exceptionally high in comparison to other matched databases studied
in the literature. The overall match rate of the LP data to credit data is 72%, but this varies
depending on whether a mortgage is active or closed. The match rate on loans that are active is
84%, while it is 68% for mortgages that are closed.
We supplement these data with the ZIP code-level Zillow Home Value Index (ZHVI) to estimate
home values and account fo r housing market trends. Zillow appraises about three out of four homes
in the U.S. several t imes a week and calculates historical values dating back t o 1997. Then, Zillow
aggregates these house-level valuations into indexes at the ZIP code level. The ZHVI does not
require a home to be sold to be included in the calculation. The index is available monthly for
11,799 ZIP codes for t he entire length of our study and loan origination dates.
We use data on monthly county-level unemployment rates fr om the Bureau of Labor Statistics.
For our analysis we create a series of seasonally adjusted unemployment rates using a standard
X-11 seasonal adjustment method. We use the six-month change in the log unemployment rate
in our estimation. Finally, we get the average household income in the ZIP code, based on 2000
Census data , from the U.S. ZIP Code Database.
8
We also use this database to match counties and
ZIP codes.
1.2 Variable definitions
We summarize in Table 1 the definitions of the variables we use in our analysis. For the status
of the mortgag e, we consider the following possibilities: current, 30/60/90+-days delinquent, and
foreclosure.
9
The status is provided by TransUnion. We use the status of the first mortgage, which
8
http://www.zip-codes.com/zip-code-database.asp.
9
Technically, 30/60/90+ days delinquent is defined as being 1/2/3+ months late with mortgage payments.
7
Table 1: Variable definitions
This table present definitions for the main variables used in the statistical analyses. In the table, “TU” stands for TransUnion and “LP”
for LoanPerformance.
Variable Description (sour c e ), a nd range for categorical variables
Mortgage status The status of the borrowers largest mortgage (TU)
C: current
D30, D60, D90+: 1,2,3+ months delinquent
F: in fore closure
VantageScore The VantageScore (TU)
Score G0: score in [501,550]
Score G1: score in (550,700]
Score G2: score in (700,800]
Score G3: score in (800,990]
VantageScore momentum VantageScore current month minus VantageScore 6-months earlier (TU)
Dscore G0: [-489,-100]
Dscore G1: (-100,-30]
Dscore G2: (-30,-30]
Dscore G3: (30,489]
Debt-to-income ratio Debt-to-income ratio, weighted-average of the values reported at origination (LP)
DTI G0: value in [0%,30%]
DTI G1: value in (30%,35%]
DTI G2: value in (35%,40%]
DTI G3: value in (40%,∞)
DTI miss: DTI not provided
Credit utilization Percentage of available credit utilized (TU)
Credit G0: value in [0%,50%]
Credit G1: value in (50%,80%]
Credit G2: value in (80%,100%]
Credit G3: value in (100%,∞)
Equity in the house Computed from CLTV at origination (LP) and ZIP-code level house price data (Zillow)
Equity G0: value in [-Inf%,-2 0%]
Equity G1: value in (-20%,0%]
Equity G2: value in (0%,25%]
Equity G3: value in (25%,∞)
Interest rate Interest rate, updated to re flec t cur rent rate for adjustable- rate mor tgages (LP)
FICO FICO credit score, weighted-average of the values reported at origination (LP)
Year dummy variables We include year dummy variables for all but the first sample year (which is the reference year)
Log income Logarithm of the average household inco me in the ZIP code based on 2000 Census data
Unemployment Six-month change in log county-level unemployment rates (Bureau of Labor Statistics)
8
is defined by the largest mortgage balance.
The other variables in Table 1 are used as potential determinants of a transition from one
mortgage status to another. To curb the impact of outliers in estimating our models and to allow
for non-linear responses, for many of the key variables we use dummy variables that are set to a
value of one if the variable of interest is within a certain range. We typically divide up the possible
range of values into four subsets (groups) and accordingly define the dummy variables G0, G1, G2,
and G3, with G0 for the range with the lowest variable values and G3 for the range with the highest
variable values. In the regression models, we typically omit the G0 dummy, and the coefficients
corresponding to the G1, G2, and G3 dummy variables therefore measure the differential effect
relative to gro up G0.
The VantageScore is provided by TransUnion. We define a variable, which we term “Van-
tageScore momentum” as the six-month change in the VantageScore. This var ia ble can be inter-
preted in a t least two ways. First, lagged VantageScores may be informative to the extent that
deteriorating or improving VantageScores may have predictive power in addition to the current
VantageScore. Second, large swings in VantageScores could identify volatile borrowers. To inter-
pret VantageScore momentum, it is important to realize that VantagesScores are r elative measures,
meaning that the cross-sectional distribution of VantageScores hardly fluctuates over time. Hence,
VantageScore momentum measures how a borrower’s position in the cross-sectional distribution of
VantageScores changes during a period of six months.
The debt-to-income (DTI) ratio is fro m LP and is reported only at the t ime a mortgage is
originated. The DTI ratio is missing for about one-third of the borrowers; hence, we define a
missing DTI dummy.
Elul, Chomsisengphet, Glennon, Hunt, and Souleles (2010) argue that liquidity constraints
and negative equity are important drivers of 60-day mortgage delinquency. We proxy for liquidity
constraints by a high level of credit card utilization, which is reported in the da ta from TransUnion.
Housing equity for borrower i at time t is constructed using the combined loan-to-value ratio at
origination (provided by LP) and the change in ZIP-code-level house prices:
Equity
i,t
= 100% −
Loan
i,0
Value
i,0
×
ZipValue
i,0
ZipValue
i,t
.
This definition abstracts from amortization between the moment of origination (time 0) and the
moment of evaluation (time t). In addition, we proxy for the change in the value of an individual
property (Value
i,0
) by the change in house prices in the ZIP code in which the property is located
in (ZipValue
i,0
/ZipValue
i,t
). An equity value of zero means the loan and the updated value of the
house are equal. A housing equity value of −100% means that the loan is twice the updated value
9
of the house. With this definition, it is possible to have a housing equity value below −100%, but
this is rare in our data set.
The mortgage interest ra te is from LP and is updated monthly for mortgages with a variable
mortgage rate. The FICO score is from LP and available only at the moment of origination of the
mortgage.
1.3 Sample selection
We construct a sample of borrower-level credit data from the merged data of TransUnion and
LP. There are approximately 16.6 million borrowers in the original data set fro m TransUnion.
Approximately 13.8 million of those borrowers have matched subprime o r Alt-A mortgages reported
in the LP data in our sample period, as of December 2009 . These mortgages were originated for
properties located in 34,125 ZIP codes of the U.S. We only select data for those ZIP codes for
which the ZHVI is available. This selection results in approximately 10.6 million borrowers with
8 million loans for properties located in 11,761 ZIP codes. From this data set, we randomly pick
20,000 borrowers.
Our unit of observation is a borrower in a given month. Hence, if several open mortgages
co-exist at a given point in time, we collapse mortgage-level data provided by LP into a single
observation per time period. To this end, we first average the combined loa n-to-value ratio, DTI
ratio, the FICO score at origination, the initial interest rate, and housing equity.
1.4 Summary statistics
We present summary statistics for our data in Table 2. The range of the VantageScore is 501 to 990,
inclusive. Borrowers with higher VantageScores are deemed more creditworthy. The VantageScore
has a mean of 724 and a standard deviation of 123. We also use the FICO score in our analysis.
As with the VantageScore, borrowers with higher FICO scores are deemed more creditworthy. The
FICO credit score is measured only at origination. It has a mean of 658 and a standard deviation
of 71 .
The median (P50) VantageScore change is 0, which is intuitive a s VantageScore is a relative
measure and our sample is representative of the general population. The 5th and 95 t h percentiles
are -109 and +78, indicating that large swings in the VantageScore do o ccur over a six-month
period. The standard deviation of VantageScore momentum is 57, further illustrating that the
VantageScore can be quite volatile. The DTI ratio is reported in percentage points and has an
average value o f 39% with a standard deviation of 9%. The credit card utilization is reported
in percentage points and has an average value of 45% with a standard deviation of 37%. The
95th percentile is above 10 0% at 101%, which can happen if the credit limit is drastically reduced
10
Table 2: Descriptive statistics
This table reports mean, standard deviation, and the 5th, 25th, 50th, 75th, 95th percentile of the distribution of the variables used in
the statistical analysis. The sample period is September 2004 to July 2009.
Mean St. dev. P5 P25 P50 P75 P95
VantageScore 723.61 122.57 530.00 631.00 718.00 809.00 943.00
VantageScore momentum −5.61 57.23 −109.00 −30.00 0.00 24.00 78.00
DTI ratio 39.29 9.43 21.60 33.80 40.80 46.40 52.00
Credit utilization 44.53 37.15 0.00 9.40 38.60 76.50 100.70
Housing equity 16.52 27.61 −31.55 2.56 18.04 33.86 57 .31
Interest rate 7.50 1.75 5.38 6.25 7.15 8.40 10.89
FICO 658.04 70.71 536.00 609.00 659.00 709.00 776.00
Log income 10.76 0.33 10.24 10.5 3 10.75 10.99 11.28
Unemployment 0.08 0.15 −0.12 −0.03 0.06 0.19 0.34
without a commensurate reduction in the amount of credit outstanding. Housing equity is reported
in percentage points and is positive on average, at 17%, but with a standard deviation of 28%.
The interest rate is reported in percentage points and has a mean of 7.50% with a standard
deviation of 1.75%. The interest rate distribution is skewed to the right, as can be seen from the
percentiles, with a rate of 10.89% at the 95th percentile. Log income equals 10.76 on average,
which corresponds to an income level of about $47,000. The average 6-month percenta ge change
in the unemployment rate is positive at 8%.
2 Determinants of mortgage default
We discuss in this section which loan characteristics, credit characteristics, and macro-economic
variables predict mortgage default. We consider several specifications. The baseline specification
(Section 2.1) utilizes the entire sample period. Second, to study the stability of our estimates over
time, we repeat selected analyses, but separately for each year (Section 2.2). Third, we study
the transition to foreclosure from a less severe payment status (Section 2.3). Fourth, we contrast
VantageScore momentum with VantageScore volatility in Section 2.4.
2.1 Baseline-case specification
We estimate probit regressions where the dependent variable is a dummy var ia ble taking the value
one if a borrower transitions to a worse default state. For the baseline-case sp ecifications we present
11
both estimated coefficients, see Table 3, and the estimated marginal effect evaluated at the mean
values of the variables, see Table 4.
To measure the probability of transition from one mort gage status to another, we use for each
probit regression a set of criteria that need to be satisfied for a borrower to be included in the
estimation. The first row displays the selection criteria for the previous month and the second row
displays the criteria for the current month. For instance, to measure a transition from current (C)
to 30-day delinquency (D30), we include only borrowers for whom the lagged status is current and
for whom the status in the subsequent month is either current or 30 days delinquent. The results
for this specification are reported in the first columns of Ta ble 3 and Table 4. Following the same
logic, we have similar inclusion criteria when the dependent variable is a 60-day delinquency status
dummy (D 60), a 90 + -day delinquency status dummy (D90+), and a foreclosure status dummy (F).
For each explanatory variable considered in Table 3 and Table 4, we report the point estimate and,
to assess the statistical significance, the z-score in parentheses. We include year dummy varia bles
in all specifications and cluster standard errors at the borrower level.
10
VantageScore The first block of explanatory variables are the dummy variables for the Van-
tageScore groups. We lag all explanatory variables by one month, except for the year dummy
variables. Recall from Table 1 that we always omit group 0 , which corresponds to the lowest
values for a variable. For example, the coefficient corresponding to “Score G1” measures the dif-
ferential effect of borrowers with a VantageScore between 550 and 700 compared to borrowers with
a VantageScore below 550.
The estimates indicate that borrowers with higher VantageScores have a lower probability of
transitioning to a worse state. The coefficients and marginal effects are monotonically declining
in the VantageScore. The only exceptions to this are the insignificant results for the transition to
90+-day delinquency (D90+).
In Table 4, with marginal effects evaluated at the mean value for the explanatory variables,
we see that the coefficients are smallest for the current to 30-day delinquency transition, which is
merely a reflection of the fact that this transition has the lowest probability among the different
transitions considered. Once a borrower reaches 30-day delinquency, transitioning to more severe
states is more likely.
Based on the estimates reported in Table 4, borrowers with a VantageScore between 550 and 700
are on average 1% less likely to be 30-days past due on their mortgage than borrowers with scores
below 550. Borrowers with scores above 800 are on average 3% less likely than those with scores
below 550 to be 30 days past due. Once a borr ower has already missed one mortgage payment, the
10
We repor t the coefficient for FICO/100 instead of FICO, as the coefficient for FICO would be too small at the
reported precision.
12
Table 3: Determinants of default
Each column reports the estimated coefficients for a multiple probit regression. The dependent variable equals one in case of a transition
to a worse payment status on the first mortgage. The z-score is provided in parenthesis; errors are clustered at the borrower level. Row 4
presents the dependent variable of interest; a status dummy variable, with C = current, D30/D60/D90+ = 1, 2, 3+ months delinquent,
and F = in foreclosure. We focus on transitions to the next worse payment status, and thus require the lagged status to be one notch
better than the dependent variable (inclusion criterion specified in row 1). Also we restrict the status in the eval uation month to be at
most one-notch worse than the month before, and thus omit the rare occurrences of the payment status deteriorating more than one
notch in a month (inclusion criterion specified in row 2). The number of observations, taking into account both inclusion criteria, is
presented in row 3. We include a constant and year dummies (not reported).
Incl. status (lag) C D30 D60 D90+
Incl. status C-D30 C-D60 C-D90+ C-F
Observations 266989 13459 6286 8604
Dependent var. D30 D60 D90+ F
Score G1 (lag) −0.30 (8) −0.12 (3 ) −0.11 (2) −0.09 (2)
Score G2 (lag) −0.75 (19) −0.42 (6) −0.07 (1) −0.21 (2)
Score G3 (lag) −0.99 (22) −0.64 (4) −0.28 (1) −0.36 (1)
DScore G1 (lag) −0.13 (5) −0.21 (6 ) −0.26 (6) −0.24 (5)
DScore G2 (lag) −0.23 (9) −0.32 (8 ) −0.65 (12) −0.51 (11)
DScore G3 (lag) −0.16 (6) −0.52 (8 ) −0.89 (10) −0.71 (9 )
DTI miss (orig.) −0.02 (1) −0.05 (1 ) 0.24 (2) −0.03 (0)
DTI G1 (orig.) −0.02 (1) −0.01 (0 ) 0.17 (1) −0.09 (1)
DTI G2 (orig.) 0.01 (0) −0.13 (2) 0.1 3 (1) 0.05 (0)
DTI G3 (orig.) 0.05 (2) −0.03 (1) 0.2 2 (2) 0.02 (0)
Credit G1 (lag) −0.03 (1) −0.11 (3) −0.04 (1) −0.10 (2)
Credit G2 (lag) 0.09 (4) −0.16 (4) −0.10 (2) −0.04 (1)
Credit G3 (lag) 0.26 (9) −0.17 (4) −0.19 (3) −0.16 (3)
Equity G1 (lag) −0.17 (6) −0.21 (4 ) −0.24 (4) −0.02 (0)
Equity G2 (lag) −0.27 (10) −0.32 (6) −0.38 (6) −0.12 (2)
Equity G3 (lag) −0.32 (10) −0.43 (7) −0.38 (5) −0.27 (4)
Int. rate (la g) 0.02 (5) 0.03 (2) 0.02 (1) 0.0 3 (2)
FICO/100 (orig.) −0.21 (13) 0.09 (3) 0.10 (2) 0.1 1 (3)
Income (lag) 0.08 (3) 0.12 (2) 0.07 (1) 0.00 (0)
Unemp. (lag) 0.22 (3) 0.52 (3) 0.63 (3) −0.15 (1)
13
likelihood of missing another payment increases. Borrowers with a VantageScore between 550 and
700 are 3.6% less likely to miss a second payment on their mortgage than borrowers with scores
below 550. Borrowers with scores above 80 0 are 13.6% less likely to be 60 days past due, again
compared to borrowers with scores below 550.
The VantageScore is designed to capture the creditworthiness of borrowers. It is therefore possi-
ble that the VantageScore is a sufficient statistic to predict default. In this case, other explanatory
variables in a multiple regression would have no additional predictive power. We now turn to the
discussion of other explanatory variables.
VantageScore momentum The second block of explanatory variables contains dummy vari-
ables for the VantageScore momentum groups. Again, group 0, with the lowest VantageScore
momentum, is omitted from the estimation. Hence, the coefficients for groups 1, 2, and 3 measure
the differential effect compared to group 0. The results are strongest f or the transition to 60-day
delinquency, 90+-day delinquency, and foreclosure: Higher VantageScore momentum lowers the
probability of transitioning to a worse state. Put simply, if the VantageScore equals 700, but it de-
clined from 800 six months before, the probability to transition to a worse state is higher compared
to a VantageScore of 700 that increased from 600 six months before. The statistical significance of
the VantageScore momentum variables is among the highest of all the explanatory variables. For
the transition to 90+-day delinquency and foreclosure, its economic and statistical significance far
exceeds the significance of the VantageScore itself.
The results in Table 4 imply that borr owers whose VantageScore improves by more than 30
points during the last six months are 12% less likely to switch from 30 days delinquency to 60 days
delinquency, or 26% less likely to transition from 6 0 days delinquency to 90+ days delinquency,
than borrowers whose VantageScore declined by more than 100 points.
Housing equity The effect of housing equity is significant and monotonic for each of the transi-
tions that we consider. The direction of the effect is intuitive: An increase in housing equity lowers
the probability of transitioning to a worse state.
Credit card utilization High credit card utilization (Credit G 3) is associated with a higher
probability of transitioning from current to 30-day delinquency, which is consistent with, for in-
stance, Elul, Chomsisengphet, Glennon, Hunt, and Souleles (2010). We find this result even after
controlling for a host of other variables, most notably t he VantageScore, which takes into account
the recent credit situation. For robustness, we re-estimate the probit regression including only the
credit variables as explanatory variables a nd the year dummy varia bles. We find that the effect of
credit card utilization is much stronger for the probability of transitioning from current to 30-day
14
Table 4: Determinants of default, marginal effects evaluated at the mean
Each column r eports the estimated marginal effects (evaluated at the mean) for a multiple probit regression. The dependent variable
equals one in case of a transition to a worse payment status on the first mortgage. The z-score is provided in parenthesis; errors
are clustered at the borrower level. Row 4 presents the dependent variable of interest; a status dummy variable, with C = current,
D30/D60/D90+ = 1, 2, 3+ months delinquent, and F = in foreclosure. We focus on transitions to the next worse payment status,
and thus require the lagged status to be one notch better than the dependent variable (inclusion criterion specified in row 1). Also we
restrict the status in the evaluation month to be at most one-notch worse than the month before, and thus omit the rare occurrences of
the payment status deteriorating more than one notch in a month (inclusion criterion specified in row 2). The number of observations,
taking into account both inclusion criteria, is pr esented in row 3. We include a constant and year dummies (not reported).
Incl. status (lag) C D30 D60 D90+
Incl. status C-D30 C-D60 C-D90+ C-F
Observations 266989 13459 6284 8604
Dependent var. D30 D60 D90+ F
Score G1 (lag) −0.009 (9) − 0.036 (3) −0.039 (2) −0.020 (2)
Score G2 (lag) −0.021 (1 9) −0.103 (7) −0.027 (1) −0.040 (2)
Score G3 (lag) −0.029 (2 0) −0.136 (6) −0.098 (1) −0.063 (2)
DScore G1 (lag) −0.004 (6) − 0.060 (6) −0.094 (6) −0.049 (6)
DScore G2 (lag) −0.008 (9) − 0.088 (9) −0.222 (13) −0.104 (12)
DScore G3 (lag) −0.005 (6) − 0.122 (11) −0.260 (15) −0.11 1 (13)
DTI miss (orig.) −0.001 (1) − 0.015 (1) 0.090 (2) −0.006 (0)
DTI G1 (orig.) −0.001 (1) − 0.004 (0) 0.063 (1) −0.019 (1)
DTI G2 (orig.) 0.000 (0) −0.037 (2) 0.049 (1) 0.010 (0)
DTI G3 (orig.) 0.002 (2) −0.010 (1) 0.080 (2) 0.005 (0)
Credit G1 (lag) −0.001 (2) −0.031 (3) −0.016 (1) −0.022 (2)
Credit G2 (lag) 0.003 (4) −0.044 (4) −0.037 (2) −0.008 (1)
Credit G3 (lag) 0.012 (7) −0.046 (4) −0.069 (3) −0.032 (3)
Equity G1 (lag) −0.005 (7) − 0.056 (4) −0.085 (4) −0.004 (0)
Equity G2 (lag) −0.009 (1 0) −0.089 (6) −0.137 (6) −0.026 (2)
Equity G3 (lag) −0.010 (1 1) −0.115 (7) −0.134 (5) −0.053 (4)
Int. rate (la g) 0.001 (5) 0.007 (2) 0.007 (1) 0.006 (2)
FICO/100 (orig.) −0.00 7 (13) 0.026 (3) 0.038 (2) 0.025 (3)
Income (lag) 0.003 (3 ) 0.036 (2) 0.024 (1) 0.000 (0)
Unemp. (lag) 0.007 (3) 0.149 (3) 0.227 (3) −0.032 (1)
15
delinquency. We do not report these results for brevity. For transitions to 60-day delinquency and
worse, the effect of credit card utilization is weaker.
Other variables The DTI rat io is a static variable fr om the LP database, which means that it
is not updated after origination. The variable is often missing, hence the inclusion of a missing
DTI ratio group. This variable is in most cases not significant.
A high current interest rate is associated with an increased probability of transitioning to a
worse state, as one might expect, and it is particularly significant for the transition to 30-day
delinquency. The current interest rate captures two effects: (i) it is positively related to the the
size of monthly payments and (ii) it may reflect information about the riskiness of a borrower that
is unobserved to us, but observable to the lender who sets the rate.
The FICO score at originatio n carries a positive coefficient for the transition to 60-day and 90+-
day delinquency and foreclosure. This at first sight may seem odd, as a high FICO score borrower
should be a more creditworthy borrower. Notice however that we run a multiple probit regression
in which we also include the most recent VantageScore. It is therefore yet another manifestation of
credit score momentum, this time measured relative to the moment of mortgage origination. That
is, given a current VantageScore, a borrower starting from a low FICO score (and thus a b orrower
who has made improvements in credit scoring variables since o r ig inatio n) is less likely to transition
to a worse mortgage status.
Log income is the only variable with a counterintuitive sign, as a higher income level is related
to an increased probability of transitioning to a worse state. It is significant only for the transition
to 30- and 60-day delinquency. The univariate effect of income on the probability of default is
negative and thus has the intuitive sign (results not reported).
An increase in the unemployment rate over the past six months increases the probability of
transitioning to 30-, 60-, and 90+-day delinquency. For the transition to foreclosure the effect is
insignificant.
2.2 Transition to 30-day delinquency for different years
For the transition to 30-day delinquency we have enough observations to accurately measure the
effect of the different determinants year by year, see Table 5. The coefficient for VantageScore
(Score G1 -G3) decreases monotonically over time, even though it remains one of the most important
variables. The effect of VantageScore momentum is very stable over time though. Also the effect
of housing equity is stable a s of 2007, while its effect is measured imprecisely in 20 06.
16
Table 5: Determinants of default, transition to 30-day delinquency year-by-year
Each column reports the estimated coefficients for a multiple probit regression. The dependent variable equals one in case of a transition
to a 30-day delinquency payment status on the first mortgage. Each probit regression uses data of only one year (specified in row 3).
The z-score is provided in parenthesis; errors are clustered at the borrower level. Row 5 presents the dependent variable of interest; a
status dummy variable f or 30-day delinquency (D30). We focus on transitions to the next worse payment status, and thus require the
lagged status to be current (C, see inclusion criterion specified in row 1). Also we restrict the status in the evaluation month to be at
most one-notch worse than the month before, and thus omit the rare occurrences of the payment status deteriorating more than one
notch in a month (inclusion criterion specified in row 2). The number of observations, taking into account both inclusion criteria, is
presented in row 4. We include a constant and year dummies (not reported).
Incl. status (lag) C C C C
Incl. status C-D30 C-D30 C-D30 C-D30
Incl. year 2006 2007 2008 2009
Observations 61685 74258 68651 30240
Dependent var. D30 D30 D30 D30
Score G1 (lag) −0.47 (6) −0.36 (6) −0.30 (5) 0.03 (0)
Score G2 (lag) −1.01 (11) −0.85 (13) −0.72 (11) −0.25 (3)
Score G3 (lag) −1.26 (12) −1.11 (15) −0.97 (14) −0.53 (5)
DScore G1 (lag) −0.13 (2) −0.11 (2) −0.16 (4) −0.11 (2)
DScore G2 (lag) −0.27 (5) −0.19 (4) −0.22 (5) −0.26 (4)
DScore G3 (lag) −0.19 (3) −0.12 (3) −0.12 (3) −0.18 (3)
DTI miss (orig.) −0.06 (1) −0.05 (1) 0.01 (0) 0.03 (0)
DTI G1 (orig.) −0.08 (1) −0.05 (1) 0.03 (0) −0.10 (1 )
DTI G2 (orig.) −0.08 (1) −0.03 (1) 0.05 (1) 0.08 (1)
DTI G3 (orig.) 0.05 (1) −0.01 (0) 0.07 (1) 0 .07 (1)
Credit G1 (lag) −0.01 (0) −0.03 (1) 0.00 (0) −0.10 (2)
Credit G2 (lag) 0.10 (2) 0.11 (3) 0.11 (3) 0 .05 (1)
Credit G3 (lag) 0.23 (4) 0.35 (7) 0.24 (5) 0 .20 (3)
Equity G1 (lag) 0.14 (0) −0.16 (2) −0.17 (5) −0.09 (2)
Equity G2 (lag) 0.20 (1) −0.24 (2) −0.29 (8) −0.27 (6)
Equity G3 (lag) 0.16 (0) −0.29 (3) −0.34 (8) −0.38 (6)
Int. rate (la g) 0.03 (3) 0.03 (4) 0.02 (3) 0 .01 (1)
FICO/100 (orig.) −0.25 (8) −0.18 (7) −0.17 (7) −0.20 (6)
Income (lag) 0.17 (3) 0.14 (3) −0.01 (0) −0.04 (1)
Unemp. (lag) 0.19 (1) 0.05 (0) 0.23 (2) 0 .53 (3)
17
2.3 Transition to foreclosure from a less severe payment status
In Section 2.1, we consider a transition to foreclosure from a 90+-delinquency status, which is the
most common status prior to foreclosure (76% of the transitions to foreclosure were preceded by
a 90+-delinquency status). In Appendix A, we also consider the transitions of a current, a 30-day
delinquency, and a 60-day delinquency status to foreclosure, which account for 3%, 4%, and 17%
of all transitions to foreclosures, respectively. These foreclosures are arguably less anticipated.
Interestingly, we find that low housing equity is a strong determinant of foreclosures in these cases.
This suggests that these foreclosures seem more sensitive to changes in negative housing equity
than those borrowers whose status has already deteriorated to 90+-delinquency.
2.4 VantageScore momentum versus VantageScore volatility
We find that VantageScore momentum helps to predict the transition to worse delinquency states
and foreclosure. One may wonder whether VantageScore momentum is effectively measuring Van-
tageScore volatility instead of trends in VantageScores that tend to continue. In Appendix B,
we include VantageScore momentum alongside a measure of VantageScore volatility in the probit
model. We find that for the transition from current to 30-day delinquency, VantageScore volatility
is more important than VantageScore momentum. However, for the transition to worse delinquency
states and foreclosure, VantageScore momentum is more important, and Va ntageScore volatility
has no significant impact.
These results suggest that VantageScore volatility identifies inattentive borrowers who every
now and then miss a payment. For the more important delinquencies and foreclosure, VantageScore
momentum is the more important variable to explain the transition.
3 Consequences of mortgage default
In this section, we study the impact of credit events on a borrower’s VantageScore. We first discuss
the econometric f r amework in Section 3.1 and present the main empirical results in the subsequent
sections.
3.1 Regression specification
We denote the VantageScore o f borrower i at time t as V
it
and let X
it
be a vector of borrower
characteristics. C
t
is a dummy variable corresponding to a certain credit event. C
t
= 1 if the
credit event happens and C
t
= 0 otherwise. We intend to answer questions of the type: “Suppose
a borrower is 30 days delinquent on his or her mortgage. What is the impact o n the VantageScore
18
if the bo r r ower misses another payment and transitions to a 60-days delinquent status?” We focus
on the impact on the VantageScore in the next period, though we also report the impact on the
VantageScore one year later. Formally, the object of interest is:
∆V
it
(k) ≡ E (V
i,t+k
| X
i,t−1
, C
it
= 1) − E (V
i,t+k
| X
i,t−1
, C
it
= 0) .
To this end, we run panel regressions of the form:
V
i,t+k
= β
(k)
0
+ β
(k)0
1
X
i,t−1
+ β
(k)
2
V
i,t−1
+
γ
(k)
0
+ γ
(k)0
1
Y
i,t−1
C
it
+ ε
i,t+k
, (1)
where Y
i,t−1
a set of variables that might affect the impact of the credit event on the borrower’s
future VantageScore. From this regression equation, it follows:
E (V
i,t+k
| X
i,t−1
, C
it
= 0) = β
(k)
0
+ β
(k)0
1
X
i,t−1
+ β
(k)
2
V
i,t−1
,
E (V
i,t+k
| X
i,t−1
, C
it
= 1) = γ
(k)
0
+ γ
(k)0
1
Y
i,t−1
+ β
(k)
0
+ β
(k)0
1
X
i,t−1
+ β
(k)
2
V
i,t−1
,
which we in turn can combine into:
E (V
i,t+k
| X
i,t−1
, C
it
= 1) − E (V
i,t+k
| X
i,t−1
, C
it
= 0) = γ
(k)
0
+ γ
(k)0
1
Y
i,t−1
.
Hence, the estimates of γ
(k)
0
and γ
(k)
1
reveal the difference in VantageScores for households that are
otherwise identical on observable characteristics. We use the VantageScore itself as the dependent
variable.
3.2 Contemporaneous consequences of default
In this section, we discuss the contemporaneous consequences of default in terms of changes in
the VantageScore. In Table 6, we present results for regressions with the VantageScore as the
dependent varia ble. We include the lagged VantageScore as well as dummies related to the lagged
VantageScore, VantageScore momentum, debt-to-income ratio, credit card utilization, and housing
equity. We also include the lagged interest rate, FICO score, income, and unemployment. The
main explanatory variable of interest is a dummy indicating that a borrower transitions to a worse
state. We study the same transitions as in Table 3 and we therefore also use the same criteria for
including observations.
Event dummy The impact on the VantageScore o f a transition from current to 30-day delin-
quency is estimated to be a decline of 51 points (relative to households that stay current and have
otherwise comparable o bserva ble characteristics). For a transition to 60-day delinquency, 90+-day
19
Table 6: Consequences of default
Each column reports the estimated coefficients for a multiple regression, with the VantageScore as dependent variable. The z-score
is pr ovided in parenthesis; errors are clustered at the borrower level. Each regression evaluates the effect on the VantageScore of
transitioning to a worse payment status; the new (worse) status, referred to as “Event” is reported in row 4. We focus on transitions
to the next worse payment status, and thus require the lagged status to be one notch better than the event variable (inclusion criterion
specified in r ow 1). Also we restrict the status in the evaluation month to be at most one-notch worse than the month before, and thus
omit the rare occurrences of the payment status deteriorating more than one notch in a month (inclusion criterion specified in row 2).
The number of observations, taking into account both inclusion criteria, is presented in r ow 3. We include a constant and year dummies
(not reported).
Incl. status (lag) C D30 D60 D90+
Incl. status C-D30 C-D60 C-D90+ C-F
Observations 266989 13459 6284 8604
Event D30 D60 D90+ F
Dependent var. Score Score Score Score
Event −50.90 (79) −25.28 (38) −13.69 (18) −6.27 (8)
Score (lag) 0.96 (647) 0.92 (105) 0.90 (70) 0.9 2 (95)
Score G1 (lag) −2.42 (5) −0.63 (1) 1.56 (1) 1.86 (2)
Score G2 (lag) −3.51 (6) 2.48 (1) 2.25 (1) 5.74 (2)
Score G3 (lag) −2.10 (3) 8.51 (2) 21.82 (3) 2 .66 (0)
DScore G1 (lag) −3.69 (9) −1.02 (1) 0.15 (0) − 1.74 (2)
DScore G2 (lag) −6.26 (15) −2.74 (3) −2.60 (3) −2.27 (3)
DScore G3 (lag) −10.14 (24 ) −6.54 (6) −6.95 (4) −5.19 (5)
DTI miss (orig.) 0.35 (2) 0.47 (0) −1.33 (1) −1.50 (1)
DTI G1 (orig.) −0 .03 (0 ) 1.93 (2) −1.11 (1) −1.45 (1)
DTI G2 (orig.) −0 .31 (1 ) 0.20 (0) −1.77 (1) −1.01 (1)
DTI G3 (orig.) −0 .13 (1 ) 0.26 (0) −1.03 (1) −0.59 (1)
Credit G1 (lag) −2.78 (18) −4.42 (6) −1.92 (2) −2.80 (3)
Credit G2 (lag) −3.50 (19) −5.02 (7) −3.00 (3) −4.04 (6)
Credit G3 (lag) −2.44 (7) −4.39 (5) −3.30 (3) −1.91 (2)
Equity G1 (lag) 1.01 (4) 0.38 (0) 0.97 (1) 0.15 (0)
Equity G2 (lag) 2.43 (9) 3.18 (3) 1.62 (1) 0.89 (1)
Equity G3 (lag) 3.87 (14) 6.03 (6) 5.30 (4 ) 2.31 (2)
Int. rate (la g) −0.28 (7) − 0.46 (3) 0.11 (1) 0.10 (1)
FICO/100 (orig.) 2.16 (19) −0.4 9 (1) 0.10 (0) 0.70 (1)
Income (lag) 1.78 (10) 0.72 (1) 0.08 (0) 0.41 (0)
Unemp. (lag) −3.46 (6) −9.36 (3) −9.66 (3) −5.55 (2)
20
delinquency, and foreclosure, the effect is increasingly muted at 25-, 14-, and 6-point drops, re-
spectively. Hence, by far the biggest hit on the VantageScore occurs at the very first step of the
default process when borrowers transition from current to 30-day delinquency.
Other variables We include the same set of control variables a s in Table 3. The effect of lagged
VantageScore momentum is very significant and mostly monotonic: holding constant the current
VantageScore group, households with positive VantageScore momentum have a tendency to expe-
rience a smaller improvement (greater deterioration) in the VantageScore relative to households
with negative VantageScore momentum. However, for households with positive VantageScore mo-
mentum the subsequent change in their scores also has a smaller standard deviation (not reported).
The latter effect dominates in determining the probability of a large drop in the VantageScore;
hence the presented results are not inconsistent with the previous result that positive VantageScore
momentum is, in fact, less likely to be accompanied by a deterioration in mortgage status (which
would typically be accompanied by a lar ge drop in the VantageScore).
In our control variables, we do not include dummy variables to indicate whether a state is
recourse or non-recourse, or whether foreclosures are judicial or non-judicial. The motivation is
that TransUnion does not use these variables in their credit model. In unreported results, we follow
Ghent and Kudlyak (2010) to classify states as recourse versus non-recourse and judicial versus
non-judicial. We indeed do not find significant differences across states, after controlling for all
other variables.
Event dummy w ithout controls For robustness, we re-estimate the regressions reported in
Table 6, now only including the event dummy variables and not the list of other explanatory
variables. We find quantitatively very similar results, indicating that the measured effect is not
an artefact resulting from a strong correlation with one of the other explanatory variables (results
are not reported). To further illustrate the effect of a credit event without controls, we plot in
Figure 1 the VantageScore distribution 1 month before and in the month o f a tr ansition to 30-day
delinquency (left panel) and foreclosure (right panel). For the transition to 30-day delinquency we
require the mortgage to be current t he month before. For the transition to foreclosure, we require
the mortgage to be 90+-day delinquent the month before. A transition to 30-day delinquency leads
to a dramatic shift in t he VantageScore distribution to the left. For the transition to foreclosure, t he
VantageScore distribution is already heavily tilted toward low values a month before foreclosure,
and the Va ntageScore distribution after fo reclosure hardly differs.
Event dummy variables interacted with VantageScore group dummy variables In Ta-
ble 7 we present the effect on the borrower’s Vant ageScore of a transition to a worse state, for
21
different lagged VantageScore groups. That is, compared to Table 6, we add interaction variables
between the event and the lagged VantageScore group.
The event var ia ble in Table 7 measures the effect for a borrower in the lowest VantageScore
group, group 0, while the interaction variables between the event and groups 1-3 measure the
effect of the event relative to a borrower in group 0. If we first focus on the event variable
without interaction, we find that a tr ansition to 30- day delinquency, 60-day delinquency, 90+-day
delinquency, and foreclosure has an increasingly muted effect on the VantageScore, similar to what
we documented in Table 6. The impact on t he VantageScore of a transition to a worse state is
larger when starting from a higher VantageScore group. Fo r example, transitioning from current
to 30-day delinquency for a borrower in lagged Va ntageScore group 3, leads to a −16−104 = −120
drop in the score on average. For transitions to more severe states we have less statistical power
for the interaction variable between the event a nd the VantageScore group, as most borrower will
be in VantageScore gro up 0 by the time they have reached these more severe states of delinquency.
3.3 Transition to 30-day delinquency for different years
For the transition to 30-day delinquency we have enough observations to accurately measure the
effect on the VantageScore year by year. In Table 8, we include the interaction between t he event
and the lag ged VantageScore group, like we do in Table 7, and estimate the model separately for
each year in our sample. The main takeaway is that the effect of a tra nsition from current t o 30-day
delinquency on the VantageScore is very stable over time. For t he lowest Va ntageScore group 0
the effect ranges from a 15- to an 18-point drop. For t he highest VantageScore group 3, the effect
(additional to the effect for gr oup 0) ranges fr om an 82- to a 116-point drop. Also, the effect of
other explanatory variables included is mostly stable. Hence, we find no evidence to support a
hypothesis that the credit bureau has modified its model to measure the impact of delinquencies
during the great recession.
3.4 Consequences of default after one year
We repeat the regressions presented in Table 6, with the only difference that we use the Van-
tageScore after one year as the dependent variable. More precisely, the dependent variable is
measured in month t + 12, the event is in month t, and the lagged control variables are measured
in month t − 1. The results are presented in Table 9.
Event dummy A year after transitioning from current to 30-day delinquency, the VantageScore
is still 38 points lower, relative to households that did not transition to 30-day delinquency. The
impact is -13 and -3 points for 60-day and 9 0-day delinquency, respectively. Interestingly, for t he
22
Table 7: Consequences of default, event-Va ntageScore group interaction effects
Each column reports the estimated coefficients for a multiple regression, with the VantageScore as dependent variable. The z-score
is pr ovided in parenthesis; errors are clustered at the borrower level. Each regression evaluates the effect on the VantageScore of
transitioning to a worse payment status; the new (worse) status, referred to as “Event” is reported in row 4. We focus on transitions
to the next worse payment status, and thus require the lagged status to be one notch better than the event variable (inclusion criterion
specified in r ow 1). Also we restrict the status in the evaluation month to be at most one-notch worse than the month before, and thus
omit the rare occurrences of the payment status deteriorating more than one notch in a month (inclusion criterion specified in row 2).
The number of observations, taking into account both inclusion criteria, is presented in r ow 3. We include a constant and year dummies
(not reported).
Incl. status (lag) C D30 D60 D90+
Incl. status C-D30 C-D60 C-D90+ C-F
Observations 266989 13459 6284 8604
Event D30 D60 D90+ F
Dependent var. Score Score Score Score
Event −16.03 (18) −11.36 (13) − 10.09 (11) −3.04 (4)
Event×Score G1 (lag) −23.41 (22) −16.8 8 (15) −6.15 (5) −4.89 (3)
Event×Score G2 (lag) −56.12 (35) −32.3 0 (9) −3.74 (1) −1 5.17 (3)
Event×Score G3 (lag) −104.16 (33) −27.37 (2) 29.51 (3) −16.67 (1)
Score (lag) 0.96 (655) 0.93 (105) 0.90 (71) 0.92 (95)
Score G1 (lag) 1.42 (3) 3.63 (4) 3.84 (3) 2.67 (3)
Score G2 (lag) 1.22 (2) 9.06 (4) 3.63 (1) 7.70 (3)
Score G3 (lag) 2.72 (4) 13.53 (3) 14.52 (2) 4.42 (1)
DScore G1 (lag) −3.65 (9) −1.21 (2) 0.16 (0) −1.77 (2)
DScore G2 (lag) −6.19 (15) −3.20 (4) −2.73 (3) −2.33 (3)
DScore G3 (lag) −10.07 (24) −7.4 9 (7) −7.10 (4) −5.43 (5)
DTI miss (orig.) 0.37 (2) 0.27 (0) −1 .66 (1) − 1.33 (1)
DTI G1 (orig.) −0.03 (0) 1.86 (1) −1.29 (1) −1.28 (1)
DTI G2 (orig.) −0.28 (1) 0.03 (0) −2.01 (1) −0.86 (1)
DTI G3 (orig.) −0.10 (1) 0.21 (0) −1.28 (1) −0.47 (0)
Credit G1 (lag) −2.71 (18) −4.74 (6) −1.96 (2) −2.81 (3)
Credit G2 (lag) −3.46 (19) −5.28 (7) −3.17 (3) −4.08 (6)
Credit G3 (lag) −2.77 (8) − 4.55 (6) −3.39 (3) −1.88 (2)
Equity G1 (lag) 0.65 (2) 0.10 (0) 1.05 (1) 0.18 (0)
Equity G2 (lag) 1.91 (8) 2.95 (3) 1.44 (1) 1.01 (1)
Equity G3 (lag) 3.32 (13) 5.47 (5) 5.16 (4) 2.43 (3)
Int. rate (la g) −0.28 (8) −0.45 (3) 0.10 (1) 0.09 (1)
FICO (orig.) 2.15 (19) −0.5 5 (1) 0.12 (0 ) 0.72 (1)
Income (lag) 1.74 (10) 0.83 (1) −0.01 (0) 0.43 (1)
Unemp. (lag) −3.42 (6) −9.19 (3) −10.14 (3) −5.43 (2)
23
Table 8: Consequences of default, transition to 30-day delinquency year-by-year
Each column reports the estimated coefficients for a multiple regression, with the VantageScore as dependent variable. The z-score
is pr ovided in parenthesis; errors are clustered at the borrower level. Each regression evaluates the effect on the VantageScore of
transitioning to a 30-day delinquency (D30), the “Event” reported in row 5, using data from different calendar years (reported in row
3). We focus on transitions to the next worse payment status, and thus require the lagged status to be current, C (inclusion cr iterion
specified in r ow 1). Also we restrict the status in the evaluation month to be at most one-notch worse than the month before, and thus
omit the rare occurrences of the payment status deteriorating more than one notch in a month (inclusion criterion specified in row 2).
The number of observations, taking into account both inclusion criteria, is presented in r ow 4. We include a constant and year dummies
(not reported).
Incl. status (lag) C C C C
Incl. status C-D30 C-D30 C-D30 C-D30
Incl. year 2006 2007 2008 2009
Observations 61685 74258 68651 30240
Event D30 D30 D30 D30
Dependent var. Score Score Score Score
Event −17.71 (7) −14.77 (10) −16.19 (10) −17.71 (8)
Event×Score G1 (lag) −22.76 (8) −25.6 2 (14) −23.80 (13) −17.11 (6)
Event×Score G2 (lag) −50.30 (13) −63.82 (24) −55.89 (20 ) −49.87 (14)
Event×Score G3 (lag) −82.13 (8) −97.0 8 (19) −115.61 (26) −110.39 (17)
Score (lag) 0.95 (309) 0.95 (362) 0.96 (384) 0.97 (291)
Score G1 (lag) 2.87 (2) 3.29 (4) −1.15 (1) −0.14 (0)
Score G2 (lag) 2.64 (2) 3.80 (4) −1.62 (2) −0.89 (1)
Score G3 (lag) 3.94 (3) 5.51 (5) 0.24 (0) 0.1 0 (0)
DScore G1 (lag) −6.58 (7) −3.69 (5) −2.26 (3) −1.36 (1)
DScore G2 (lag) −9.76 (10) −6.23 (9) −4 .23 (6) −3.29 (4)
DScore G3 (lag) −13.73 (14) −10.07 (13) −7.57 (10) −7.21 (7)
DTI miss (orig.) −0.29 (1) 0.75 (2) 0.00 (0) 1.29 (3)
DTI G1 (orig.) −0.74 (2) 0.34 (1) 0.30 (1) 1.04 (2)
DTI G2 (orig.) −1.18 (3) 0.09 (0) −0.21 (0) −0.28 (0)
DTI G3 (orig.) −1.09 (3) 0.00 (0) −0.09 (0) 1.52 (3)
Credit G1 (lag) −2.77 (9) −2.66 (9) − 2.95 (10) −2.7 7 (7)
Credit G2 (lag) −3.29 (9) −3.50 (10) −3.57 (10) −3.64 (7)
Credit G3 (lag) −2.53 (3) −2.87 (4) − 2.95 (5) −2.52 (3)
Equity G1 (lag) 2.49 (1) 0.25 (0) 0.37 (1) 1.07 (3 )
Equity G2 (lag) 3.52 (1) 1.89 (2) 1.56 (4) 1.41 (4 )
Equity G3 (lag) 5.25 (2) 3.41 (4) 2.75 (7) 2.22 (5 )
Int. rate (la g) −0.36 (5) −0.33 (5) −0.27 (4) −0.08 (1)
FICO/100 (orig.) 3.19 (15) 2.49 (12) 1.60 (8) 0.5 1 (2)
Income (lag) 2.10 (6) 1.67 (5) 1.54 (5) 1.40 (3)
Unemp. (lag) −4.14 (3) −1.82 (2) −6.33 (7) − 0.62 (0)
24
Table 9: Consequences of default, effect a fter one year
Each column reports the estimated coefficients for a multiple regression, with the VantageScore 12 months after a prescribed event as
dependent variable. The z-score is provided in parenthesis; errors are clustered at the borrower level. Each regression evaluates the effect
on the VantageScore one-year out of transitioning to a worse payment status; the new (worse) status, referred to as “Event” is rep orted
in row 4. We focus on transitions to the next worse payment status, and thus require the lagged status to be one notch better than
the event variable (inclusion criterion specified in row 1). Also we restrict the status in the evaluation month to be at most one-notch
worse than the month before, and thus omit the rare occurrences of the payment status deteriorating more than one notch in a month
(inclusion criterion specified in row 2). The number of observations, taking into account both inclusion criteria, is presented in row 3.
We include a constant and year dummies (not reported).
Incl. status (lag) C D30 D60 D90+
Incl. status C-D30 C- D6 0 C-D90+ C-F
Observations 207192 9555 4081 4181
Event D30 D60 D90+ F
Dependent var. Score (lead) Score (lead) Score (lead) Scor e (lead)
Event −38.39 (30) −13.23 (8) −3.49 (2) 0.28 (0)
Score (lag) 0.78 (78) 0.70 (25) 0.63 (14) 0.68 (14)
Score G1 (lag) −29.98 (1 4) −10.58 (4) −4.9 5 (1) −2.90 (1)
Score G2 (lag) −31.48 (1 2) −3.08 (0 ) −1.9 3 (0) 2.54 (0)
Score G3 (lag) −24.42 (7) 38.68 (3) 85.9 6 (4) −1.49 (0)
DScore G1 (lag) −8.53 (5) 2.86 (1) −3.08 (1) −0.85 (0)
DScore G2 (lag) −9.14 (5) −0.05 (0) −4.9 0 (2) −6.99 (2)
DScore G3 (lag) −24.09 (1 4) −7.41 (2 ) −11.36 (2) −11.65 (2)
DTI miss (orig.) 2.09 (1) −1.23 (0) −5.49 (1) −6.85 (1)
DTI G1 (orig.) 0.32 (0) −4.60 (1) −7.92 (1) −11.45 (1)
DTI G2 (orig.) −3.15 (1) 1.33 (0) −10.86 (2) −12.29 (1)
DTI G3 (orig.) −4.31 (3) −1.23 (0) −6.16 (1) −5.22 (1)
Credit G1 (lag) −10.28 (9) −8.37 (3) −9.72 (2) −7.65 (2)
Credit G2 (lag) −10.67 (8) −4.70 (2) −6.73 (2) −10.34 (2)
Credit G3 (lag) −3.23 (2) −1.41 (0) −5.83 (2) −6.23 (1)
Equity G1 (lag) 13.3 6 (4) −0.24 (0) 5.08 (1) 0.66 (0)
Equity G2 (lag) 31.7 2 (9) 6.44 (1) 4.85 (1) 0.10 (0)
Equity G3 (lag) 49.8 9 (14) 25 .62 (5) 15 .62 (3) 8.69 (1)
Int. rate (la g) −2.02 (6) −0.93 (1) −0.22 (0 ) 1.00 (1)
FICO/100 (orig.) 16.35 (17) 7.47 (4) 10.06 (4) 11.40 (3)
Income (lag) 7.56 (5) 0.2 5 (0) 0.44 (0) 13.72 (2)
Unemp. (lag) −43.02 (11) −40.39 (4) −46.86 (3) −44.24 (3)
25
transition to foreclosure, the change in in the VantageScore over the subsequent year is similar to
that of a household that did not experience foreclosure, evidenced by the (near) zero coefficient on
the event dummy.
Event dummy without controls We re-estimate the regressions reported in Table 9, including
now only the event dummy variables and not the list of other explanatory variables, a nd find
quantitatively similar results. This indicates that the measured effect is not an artefact resulting
from a strong correlation with one of the other explanatory variables (results are not rep orted). To
further illustrate the effect of a credit event without controls, in Figure 4 we plot the VantageScore
distribution 1 month before and 12 months after a transition to 30-day delinquency (left panel)
and for eclosure (rig ht panel). For the tr ansition to 30-day delinquency we require the mortgage
to be current the month before. For the transition to foreclosure, we require the mortgage to be
90+-days delinquent the month before. We make sure t hat we use the same set of borrowers for
the distribution the month before and one year after, by including o nly borrowers for whom the
VantageScore is given for both periods. A transition to 30-day delinquency leads to a dramatic
shift in the VantageScore distribution to the left, even more dramatic than what we saw in the
month of the transition, Figure 1. This is in the same spirit as the VantageScore momentum effect
we documented: borrowers who take a first step down the road to default often slide further down,
leading to a continuation of declines in the VantageScore. For the transition to foreclosure, the
VantageScore distribution is already heavily tilted toward low values a month before foreclosure,
and the distribution of the VantageScore 12 months after foreclosure actually shows a pronounced
recovery. In particular, the percentage of borr owers in the lowest VantageScore bracket is much
lower one year after the transition to foreclosure.
Recall from our discussion of Table 9 that households that transition from 90+-delinquency
to foreclosure have a 12-month change in their VantageScore comparable to the change experi-
enced by households that did not transition from 90+-day delinquency to foreclosure. The large
improvement in the VantageScore documented in Figure 4 (right panel) for households that tra n-
sition fro m 90+-day delinquency to foreclosure then logically also implies that households that
did not transition from 90+-delinquency to foreclosure experienced a comparable improvement in
their VantageScores. We confirmed that this is indeed the case by constructing a similar figure for
the households that do not transition to foreclosure (not repor t ed). Hence, starting from a 90+
status, 12 months later the Va ntageScore is on average considerably higher, regardless of whether
a foreclosure takes place or not.
We study in more detail what happens to the borrowers that are 90+-day delinquent. As it
turns out, the Vanta geScores of these borrowers are very close to the lower bound of the Van-
tageScore range. This implies that even if a f oreclosure does not occur and the borrower is still
26
delinquent 12 months later, the VantageScore does not deteriorate. However, there are some bor-
rowers that become current again and they experience a significant improvement in VantageScores.
This behavior is similar for borrowers who transition from 90+-delinquency into foreclosure.
Figure 4: VantageScore distribution 1 month before and 12 months after a transition to a worse
state
The figure shows the VantageScore distribution 1 month before and 12 months after a transition to a worse state. For the transition to
30-day delinquency (left panel), we require the mortgage to be current the month before. For transition to foreclosure (right panel), we
require the mortgage to be 90+-day delinquent the month before.
0 5 10 15
Percent
500 600 700 800 900 1000
12 months after event 1 month before event
Transition to 30−day delinquency
0 10 20 30 40
Percent
500 600 700 800 900
12 months after event 1 month before event
Transition to foreclosure
To illustrate in another way how the VantageScore is differentially aff ected by a transition from
current to 30-day delinquency in comparison to a transition from 90+-delinquency to foreclosure, we
plot in Figure 3 the change in the VantageScore in the month of a transition to 30-day delinquency
(left panel) and the cumulative change in the VantageScore 12 months after; we do the same for
a transition into foreclosure (right panel). We plot the changes a s a function of the VantageScore
the month prior to t he event. We truncate the plots above a VantageScore of 800 and 700 for the
left and right panel respectively, as too few borrowers have a sufficiently high VantageScore the
month prior t o a transition to a worse state (see Figure 1).
The patterns are consistent with the results discussed above: the first delinquency has the
largest effect on a borrower’s VantageScore, while foreclosures on average have a small negative
effect on VantageScores. Moreover, going forward, households that are 90+-day delinquent tend
to experience on average an improvement in VantageScores. One interpretation in the case of
foreclosure would be that households are able to make payments on other forms of credit when
they no longer make the mortgage payments. To illustrate the improvement in the credit situation
following a fo r eclosure, we plot in Figure 5 (left panel) the number of months since the latest bank
27
or credit card delinquency, for both the month prior to the foreclosure and 12 months after. In the
month prior to foreclosure more than 50% of households are delinquent on a bank or credit card,
while 12 months after foreclosure this number is less than 30%. More generally the distribution
shows a dramatic shift to the right going from the month before to 12 months after foreclosure,
showing how bank and credit card problems improved following a foreclosure. In Figure 5 (right
panel), we depict the number of inquires for credit over the last 6 months, for both the month prior
to t he foreclosure and 12 months after the foreclosure. Credit inquiries are interesting to look at
not just because they are a measure o f a household’s distress, but also because the credit bureaus
may use them as an input to determine the credit score, where a large number of inquiries has a
negative impact on the credit score. The month prior to foreclosure less than 20% of households
made no inquiry for credit, while 12 months after foreclosure more tha n 35% households made no
inquiry. More generally the distribution shows a dramatic shift to the left going from the month
before to 12 months after for eclosure.
Figure 5: General credit situation 1 month before a nd 12 months after fo r eclosure
We focus on transitions to foreclosure where the mortgage was 90+-day delinquent the month before. The left panel depicts the number
of months s ince the latest bank or credit card delinquency. The right panel shows the number of inquires for credit over the last 6
months. In both Panels, we show both the month prior to the foreclosure and 12 months after the foreclosure.
0 10 20 30 40 50
Percent
0 20 40 60 80
12 months after foreclosure 1 month before foreclosure
Months since bank/credit card delinquency
0 10 20 30 40
Percent
0 10 20 30 40 50
12 months after foreclosure 1 month before foreclosure
Number of inquiries for credit last 6 months
4 The price of credit
In this section, we document the link between the price of credit and b oth the VantageScore
and the loan-to-value r atio.
11
This link is useful for interpreting changes in VantageScores in the
11
Rajan, Seru, a nd Vig (2010 ) also study the link between mortgage rates and both the FICO score and the
loan-to-value ratio.
28
previous section in terms of the cost of obtaining new credit. If households, fo r instance, enter t he
foreclosure process, the change in VantageScores can be translated into the change in mortgage
rates charged on a new mortgage. This argument of course assumes that the borrower is able to
successfully apply for a new mortgage.
We focus on three types of mortgages, namely, fixed-rate mortgages (FRM), 2- year hybrid
mortgages, and 3-year hybrid mortgages.
12
In the case of the last two contracts, the interest r ate
is fixed for two or t hree years, respectively, and floating in all subsequent periods. In the case
of hybrid mortgages, the contract has both an initial interest rate and a margin. The margin is
what the borrower pays over a benchmark r ate during the period in which the rate is floating. We
restrict attention to mortgages that are not refinancings and that are for owner-occupied homes.
For each mortgage type, we first regress mortgage rates on year dummy variables to remove
the aggregate time series variation in mortgage rates. We collect the residuals and regress these in
turn on the loan-to-value ratio, the VantageScore, and the squared VantageScore. It is p otentially
important to allow for non-linearities, as the price of credit may be less sensitive to the VantageScore
for high VantageScores. We measure the VantageScore in the month of the mortg age origination.
We then evaluate the p olynomial at t he average loan-to-value ratio and at different values of the
VantageScore to map out the price of credit.
The results are summarized in Figure 6, which shows that in all cases the initial mortgage
rate declines for higher VantageScores. The figure shows that initial rates on FRM contracts are
somewhat more sensitive to changes in the VantageScore, once we fix the loan-to-value r atio. Both
hybrid contracts result in very similar pricing curves, as might be expected. The R-squared values
corresponding to the regressions of mortgage rates on the loan-to-value ratio, the VantageScore, and
the squared VantageScore is around 30%. In all cases, we find that the pricing curve displays some
convexity, confirming that initial mortgage rates are less sensitive to the VantageScore for high-
quality borrowers. Quantitatively, we find that, on average, a 1-point drop in the VantageScore
results in a about a 1- basis point ( 0.01%) increase in the mortgag e rate, ceteris paribus, for FRM
contracts. This point estimate implies that the mort gage rate on a new mortgage increases by 51bp
following 30-day delinquency, by another 25bp after 60-day delinquency, by 14bp following 90+-
delinquency, and by 6bp f ollowing foreclosure. The cumulative impact from a series of transition
from current to foreclosure abo ut 1% in total, which would increase the annual borrowing costs by
$2,000 for a $200,000 mort gage. Table 7 shows that the impact o f delinquencies and foreclosure
is larger for borrowers with higher VantageScores before the credit event. For instance, 30-day
delinquency already- corresponds to a 120-point decline of the VantageScore if the VantageScore
exceeds 800 before delinquency.
12
Koijen, Van Hemert, and Van Nieuwerburgh (2009) study the determinants of mortgage choice between
adjustable- and fixed-rate mor tgages.
29
Figure 6: Initial mortgage rates for various contract types and VantageScore
The figure shows how initial mor tgage rates relate to a borrower’s VantageScore at origination. We focus on three types of mortgages,
namely, fixed-rate mortgages (FRM), 2-year hybrid mortgages, and 3-year hybrid mortgages. For each mortgage type, we first regress
mortgage rates on year dummy variables to remove the aggregate time series variation in mortgage rates. We collect the residuals and
regress these in turn on the loan-to-value ratio, the VantageScore, and the squared VantageScore. It is potentiall y important to allow for
non-linearities, as the price of credit may be less sensitive to the VantageScore for high VantageScores. We m easure the VantageScore in
the month of origination. We then evaluate the polynomial at the average loan-to-value ratio and at different values of the VantageScore
to map out the price of credit.
550 600 650 700 750 800
−1
−0.5
0
0.5
1
1.5
2
Vantage score
De−meaned mortgage rate (%)
FRM
Hybrid, 2−year fixed
Hybrid, 3−year fixed
30
For the hybrid contracts, we estimate the sensitivity to be around a 0.6-0.7 basis point increase
for a 1 -point drop in the in VantageScore. Finally, we a lso study the impact on the margins in the
case of hybrid contracts. For the 2 -year hybrid contract, the margin increases by 0.3 basis points,
while for the 3-year hybrid contact this number is estimated to be 0.8 basis points for a 1-point
drop in the VantageScore.
31
5 Conclus i on
In this paper we look at determinants and consequences of default on t he first-lien mortgage. We
add to the literature on determinants, by including credit information from Tra nsUnion, one of
the three major credit bureaus. We find that the updated credit score is an important predictor
of mortgage default in addition to the credit score at o rigination. However, the 6-month change
in the credit score also predicts default: A positive change in t he credit score significantly reduces
the probability of delinquency or foreclosure.
We add to the literature on consequences of default, by providing a detailed study on the
impact of default on a borrower’s credit score. The credit score drops on average 51 points when
a borrower becomes 30-days delinquent on his mortgage, but the effect is much more muted for
transitions to more severe delinquency states and even foreclosure.
We also analyze the link between changes in VantageScores and mortgage r ates to quantify
the impact of delinquency and foreclosure on future mortgage rates. Our estimates suggest that,
on average, a one-point drop in the VantageScore corresponds to approximately a one basis point
increase in the mortgage rate for fixed-rate mortgages. This point estimate implies that the mort-
gage rate on a new mortgage increases by 51bp following 30-day delinquency, by another 25bp
after 60-day delinquency, by 14bp following 90+-delinquency, and by 6bp following foreclosure.
The cumulative impact from a series of transitions from current to f oreclosure is about 1% in total,
which would increase the annual borrowing costs by $2 ,0 00 for a $200,000 mortgage.
32
References
Agarwal, S., B. W. Ambrose, S. Chomsisengphet, and A. B. Sanders (2010): “Thy
Neighbor’s Mortgage: Does Living in a Subprime Neighborhood Impact Your Probability of
Default?,” Working Paper.
Agarwal, S., S. Chomsisengphet, and C. Liu (2010): “Consumer Bankruptcy and Default:
The Role of Individual Social Capital Formation Characteristics,” Journal of Economic Psy-
chology, forthcoming.
Amromin, G., and A. L. Paulson (2009): “Comparing Patterns of Default Among Prime and
Subprime Mortgages,” Economic Perspectives, 33(2), 18–37.
Archer, W. R., and B. C. Smith (2 010): “R esidential Mortgage Default: The Roles of House
Price Volatility, Euphoria and the Borrowers Put Option,” Federal Reserve Bank of R ichmond
Working Paper No. 10-02.
Bajari, P., C. S. Chu, and M. Park (2008): “An Empirical Model of Subprime Mortgage
Default fro m 2000 to 20 07,” NBER Working Paper No. w14625.
Campbell, J. Y., and J. F. Cocco (2010): “A Model of Mortgage Default,” Wor king Paper
Harvard University.
Campbell, J. Y., S. Giglio, and P. Pathak (2010): “ Forced Sales and House Prices,” Forth-
coming American Economic Review.
Campbell, T. S. a. J. K. D. (19 83): “The Detrminants of Default on Insured Conventional
Residential Mortgage Loans,” Journal of Finance, 38(5), 1569–1581.
Crews, A. C., and R. A. V. Order (2005): “On the Economics o f Subprime Lending,” The
Journal of Real Estate Finance and Economics, 30(2), 167–96.
Demyanyk, Y., and O. Van Hemert (2010): “Understanding the Subprime Mortgage Crisis,”
Forthcoming Review of Financial Studies.
Deng, Y., J. M. Quigley, and R. V. Order (2000): “ Mortg age Terminations, Heterogeneity
and the Exercise of Mortgage Options,” Econometrica, 68(2), 275–307 .
Elul, R., S. Chomsisengphet, D. Glennon, R. Hunt, and N. Souleles (2010): “What
‘Triggers’ Mortgage Default?,” Forthcoming American Economic Review, Papers and Proceed-
ings.
33
Fay, S., E. Hurst, and M. White (2002): “The Consumer Bankruptcy Decision,” American
Economic Review, 92, 706–718.
Gerardi, K., A. H. Shapiro, and P. S. Willen (2008): “Subprime Outcomes: Risky Mort-
gages, Homeownership Experiences, and Foreclosures,” FRB Boston Working Papers Series,
paper no. 07-15.
Gerardi, K. S., A. Lehnert, S. M. Sherlund, and P. S. Willen (2009): “Making Sense
of the Subprime Crisis,” Federal Reserve Bank of Atlanta Working Paper No. 2009-2.
Ghent, A. C., and M. Kudlyak (2010): “R ecourse a nd R esidential Mortgage Default: Theory
and Evidence from U.S. States,” Federal Reserve Bank of Richmond Working Paper No. 09 -10R.
Gross, D. B., and N. S. Souleles (2002): “An Empirical Analysis of Personal Bankruptcy
and Delinquency,” Review of Financial Studies, 15(1 ) , 319–347.
Guiso, L., P. Sapienza, and L. Zingales (2009): “Moral and Social Constraints to Strategic
Defaults on Mortgages,” Working Paper Chicago Booth.
Jiang, W., A. A. Nelson, and E. Vytlacil (2010): “Liars Loan? Effects of Originatio n
Channel and Information Falsification on Mortgage Delinquency,” Working Paper.
Keys, B. J., T. Mukherjee, A. Seru, and V. Vig (2010): “D id Securitization Lead to
Lax Screening? Evidence from Subprime Loans,” The Quarterly Journal of Economics, 125 (1),
307–362.
Koijen, R. S., O. A. Van Hemert, and S. Van Nieuwerburgh (2009): “Mortgage Timing,”
Journal of Financial Economics, 93, 2 92–324.
Mayer, C., and K. Pence (2008): “Subprime Mortgages: What, Where, and to Whom?,”
Working Paper Federal R eserve Board, Washington, D.C.
Mian, A., and A. Sufi (2009): “The Consequences of Mortgage Credit Expansion: Evidence fr om
the U.S. Mortgage Default Crisis,” The Quarterly Journal of Economics, 12 4(4), 1449–1496.
Mian, A., A. Sufi, and F. Trebbi (2010): “Foreclosures, House Prices, and the Real Economy,”
Working Paper Chicago Booth.
Pennington-Cross, A., and S. Chomsisengphet (2007): “Subprime Refinancing: Equity
Extraction and Mortgage Termination,” Real Estate Economics, 35(2), 233– 263.
Rajan, U., A. Seru, and V. Vig (2010): “The Failure of Mo dels That Predict Failure: Distance,
Incentives and Defaults,” Working Paper Chicago Booth.
34
Von Furstenberg, G. M. (1969): “Default Risk on FHA-Insured Ho me Mortgages as a Function
of the Terms of Financing: A Quantitative Analysis,” Journal of Finance, 24(3), 459–477.
von Furstenberg, G. M., and R. J. Green (1974): “Home Mortgage Delinquencies: A
Cohort Analysis,” Journal of Finance, 29(5), 1545–48.
35
A Transition to foreclos ure from different states
In this appendix, we study the determinants of a tra nsition to fo r eclosure from starting states other
than 90+-day delinquency. Foreclosure is preceded by a current, a 30-day delinquency, a 60-day
delinquency, and a 90+-day delinquency state in 3%, 4%, 17%, and 76% of the cases, respectively.
In the previous analyses we consider only the most common starting state, 90+-day delinquency.
We report the main results in Table 10. The last column (tr ansition from 90+-day delinquency
to foreclosure) corresponds to the last column in Table 3. From our discussion of Table 3, recall
that VantageScore momentum in particular is a strong determinant of foreclosure. Because we
have less foreclosure events preceded by less severe states (current, 30-day delinquency, and 60 -
day delinquency), we have less statistical power. Also we do not have any observatio ns for some
dummy variables, in which case we report an “na” in Table 10. We find less statistical evidence
that foreclosure fro m a less severe state is predicted by VantageScore momentum. Instead, the
housing equity dummies have the most statistical power to predict foreclosures from a less severe
state. The sign is consistent with standard economic intuition: higher housing equity lowers the
probability transitioning to foreclosure. This suggests that these foreclosures seem more sensitive
to changes in negative housing equity than those borrowers whose status has already deteriorated
to 90+ -delinquency.
We do not report results for the consequences of a tra nsition that starts from a less severe
payment status to foreclosure because of a lack of observations and thus statistical power.
36
Table 10: Determinants o f default, transition to foreclosure fr om different states
Each column reports the estimated coefficients for a multiple probit regression. The dependent variable equals one in case of a transition
to a foreclosure payment status on the first mortgage. The z-score is provided in parenthesis; errors are clustered at the borrower level.
Row 4 presents the dependent variable of interest; F = in foreclosure. We vary the lagged payment status, reported in row 1, where C
= current and D30/D60/D90+ = 1, 2, 3+ months delinquent. For this table, the status in the evaluation month is unrestri cted (row 2
shows it can be any of C-F). The number of observations, taking into account both inclusion criteria, is presented in row 3. We include
a constant and year dummies (not reported).
Incl. status (lag) C D30 D60 D90+
Incl. status C-F C-F C-F C-F
Observations 166919 12476 5858 8604
Dependent var. F F F F
Score G1 (lag) −0.69 (5) −0.24 (2) −0.14 (2) −0.09 (2)
Score G2 (lag) −1.30 (7) −0.20 (1) −0.16 (1) −0.21 (2)
Score G3 (lag) na na na −0.36 (1)
DScore G1 (lag) −0.12 (1) −0.33 (3) 0.01 (0) −0.24 (5)
DScore G2 (lag) −0.31 (2) −0.12 (1) −0.14 (2) −0.51 (1 1)
DScore G3 (lag) −0.18 (1) −0.15 (1) −0.35 (2) −0.71 (9)
DTI miss (orig.) 0.12 (1) 0.24 (1) 0.18 (1) −0.03 (0)
DTI G1 (orig.) 0 .25 (1) na 0.33 (2) − 0.09 (1)
DTI G2 (orig.) 0 .29 (1) 0.19 (1) 0.21 (1) 0.05 (0)
DTI G3 (orig.) 0 .23 (1) 0.23 (1) 0.14 (1) 0.02 (0)
Credit G1 (lag) −0.17 (1) −0.23 (2) −0.03 (0) −0.10 (2)
Credit G2 (lag) −0.19 (2) −0.12 (1) −0.08 (1) −0.04 (1)
Credit G3 (lag) −0.30 (2) 0.09 (1) −0.21 (2) −0.16 (3)
Equity G1 (lag) −0.50 (2 ) −0.23 (2 ) −0.11 (1) −0.02 (0)
Equity G2 (lag) −0.51 (4 ) −0.24 (2 ) −0.28 (3) −0.12 (2)
Equity G3 (lag) −0.55 (4 ) −0.33 (2 ) −0.32 (3) −0.27 (4)
Int. rate (la g) 0.04 (2) 0.03 (1) 0.02 (1) 0.03 (2)
FICO/100 (orig.) 0.03 (0) −0.01 (0) 0.03 (0) 0.11 (3)
Income (lag) −0.03 (0) 0.03 (0) −0.03 (0) 0.00 (0)
Unemp. (lag) 0.01 (0) −0.37 (1) 0 .23 (1) −0.15 (1)
37
B Vant age Score Volatility
In this appendix, we study the effect of VantageScore volatility, which we measure as the absolute
value of the VantageScore momentum variable. For this table we choose not to use dummies for
VantageScore and VantageScore momentum, as the measured effect for the highly related absolute
VantageScore momentum variable may depend on the particular choice of cut-offs for the dummies.
In Table 11 one can see that 30-day delinquency is positively predicted by VantageScore volatil-
ity. The VantageScore volatility variable does not significantly predict more severe states of delin-
quency or foreclosure, suggesting that the VantageScore volatility variable helps identify inattentive
borrowers who every now and then are late with their monthly payment but have no real solvency
issues. In contrast, the VantageScore momentum variable has hardly any predictive power for 30-
day delinquency, but it is a powerful determinant borrowers which will transition into progressively
worse delinquency states and foreclosure.
13
13
We report the coefficient for VantageScore/100 rather than the VantageScore, as the coefficient for VantageScore
would be too small at the reported pre c ision.
38
Table 11: Determinants of default, effect of score volatility
Each column reports the estimated coefficients for a multiple probit regression. The dependent variable measures the probability of a
transition to a worse payment status on the first mortgage. The z-score is provided i n parenthesis; errors are clustered at the borrower
level. Row 4 presents the dependent variable of interest; a status dummy variable, with C = curr ent, D30/D60/D90+ = 1, 2, 3+ months
delinquent, and F = in foreclosure. We focus on transitions to the next worse payment status, and thus require the lagged status to
be one notch better than the dependent variable (inclusion criterion specified in row 1). Also we restrict the status in the evaluation
month to be at most one-notch worse than the month before, and thus omit the rare occurrences of the payment status deteriorating
more than one notch i n a month (inclusion criterion specified in row 2). The number of observations, taking into account both inclusion
criteria, is pr esented in row 3. We include a constant and year dummies (not reported).
Incl. status (lag) C D30 D60 D90+
Incl. status C-D30 C-D60 C-D90+ C-F
Observations 266989 13459 6286 8604
Dependent var. D30 D60 D90+ F
Score/100 (lag) −0.39 (31) − 0.17 (5) −0.08 (2) −0.06 (2)
Dscore/100 (lag) 0.03 (2) −0.24 (5 ) −0.46 (6) −0.33 (6)
abs(Dscore/100) (lag) 0.11 (6) 0.00 (0) −0.10 (1) −0.0 6 (1)
DTI miss (orig.) −0.01 (0) −0.05 (1) 0.22 (2) −0.02 (0)
DTI G1 (orig.) −0.02 (1) −0.01 (0 ) 0.14 (1) −0.11 (1)
DTI G2 (orig.) 0.00 (0) −0.14 (2) 0.11 (1) 0.06 (1)
DTI G3 (orig.) 0.05 (2) −0.03 (1) 0.21 (2) 0.03 (0)
Credit G1 (lag) −0.07 (4) −0.11 (3) −0.06 (1) −0.10 (2)
Credit G2 (lag) 0.03 (2 ) −0.18 (4) −0.12 (2) −0.04 (1)
Credit G3 (lag) 0.13 (5 ) −0.20 (4) −0.18 (3) −0.13 (2)
Equity G1 (lag) −0.16 (6) −0.21 (4) −0.23 (4) −0.02 (0)
Equity G2 (lag) −0.26 (9) −0.33 (6) −0.37 (5) −0.13 (2)
Equity G3 (lag) −0.29 (9) −0.44 (7) −0.38 (5) −0.28 (4)
Int. rate (la g) 0.02 (4) 0.03 (2) 0.02 (1) 0.03 (2)
FICO (orig.) −0.15 (9) 0 .07 (2) 0.09 (2) 0.09 (2)
Income (lag) 0.10 (4) 0.12 (2) 0.0 5 (1) 0 .00 (0)
Unemp. (lag) 0.23 (3) 0.53 (3) 0.62 (3) −0.15 (1)
39
