Dynamic Econometric Loss Model

A Default Study of US Subprime Markets

C. H. Ted Hong, PhD

Beyondbond, Inc.

Abstract

The meltdown of the US subprime mortgage market in 2007 triggered a series of global
credit events. Major financial institutions have written down approximately $120 billion
of their assets to date and yet there does not seem to be an end to this credit crunch. With
traditional mortgage research methods for estimating subprime losses clearly not working,
it now requires revised modeling techniques and a fresh perspective of other macro-
economic variables to help explain the crisis. During the subprime market rise/fall era,
the Housing Price Index (HPI) and its annual appreciation (HPA) had been the main
blessing/curse attributed by researchers. Unlike the traditional models, our Dynamic
Econometric Loss (DEL) model applies not only the static loan and borrower
characteristic variables such as loan terms, Combined-Loan-To-Value ratio (CLTV), Fair
Isaac credit score (FICO), as well as dynamic macro-economic variables such as HPA to
projects defaults and prepayments, but also the spectrum of delinquencies as an error
correction term to add an additional 15% accuracy to the model projections. In additional
to our delinquency attribute finding, we find that cumulative HPA and the change of
HPA contribute various dimensions that greatly influence defaults. Another interesting
finding is a significant long-term correlation between HPI and disposable income level
(DPI). Since DPI is more stable and easier for future projections, it suggests that HPI will
eventually adjust to coincide with DPI growth rate trend and that HPI could potentially
experience as much as a 14% decrease by the end of 2009.
Contents

Introduction......................................................................................................................... 4
Figure 1. What happened in 2007 .......................................................................... 5
Model Framework............................................................................................................... 6
Figure 2. Number of Securitized Alt-A and Subprime Mortgage Origination ...... 6
Default Modeling ................................................................................................................ 8
Default Modeling Factor Components: .......................................................................... 8
Seasoning ........................................................................................................................ 8
Figure 3. Seasoning: CDRs by Date and Vintages of ARM 2/28 .......................... 8
Why is the 2005 seasoning pattern faster than prior vintages?................................... 9
Figure 4. Seasoning: CDRs by Age and Vintages of ARM 2/28 ........................... 9
Payment Shock – Interest Only (IO)............................................................................... 9
Figure 5. IO Payment Shock: CDRs by Date of ARM 2/28 .................................. 9
Combined Loan-to-Value (CLTV): .............................................................................. 10
Figure 6. Stratified seasoned CDR over CLTV ranges ........................................ 10
Figure 7. CDRs by Date and CLTVs of ARM 2/28............................................. 11
FICO ............................................................................................................................. 11
Figure 8. Stratified CDR by CLTV and FICO of ARM 2/28 .............................. 12
Figure 9. FICO: CDRs by Date and FICOs of ARM 2/28 ................................... 12
Debt-To-Income Ratio (DTI) andLoan Documentation (DOC)................................... 12
Figure 10. Stratified CDR by DTI ranges .......................................................... 13
Figure 11. DOC: Actual vs. Fitted for 2004....................................................... 13
Figure 12. Unemployment Rate 1979-2007 ....................................................... 14
Figure 13. Unemployment vs. T-Bill 3mo ......................................................... 14
Loan Size ...................................................................................................................... 14
Figure 14. Loan Size Stratification..................................................................... 14
Figure 15. Size: Actual vs. Fitted CDR for 2004 ............................................... 14
Lien ............................................................................................................................... 15
Figure 16. Lien Stratification ............................................................................. 15
Figure 17. Lien: Actual vs. Fitted CDR for 2004............................................... 15
Occupancy..................................................................................................................... 15
Figure 18. Occupancy Stratification.................................................................. 16
Figure 19. Occupancy: Actual vs. Fitted CDR for 2004 .................................... 16
Purpose.......................................................................................................................... 16
Figure 20. Purpose Stratification........................................................................ 16
Figure 21. Purpose: Actual vs. Fitted CDR for 2004 ......................................... 16
Dynamic Factors: Macro-Economic Variables............................................................. 17
Housing Price Appreciation (HPA) .............................................................................. 18
Figure 22. 30-day Delinquency of ARM2/28 by Vintages ................................ 18
Figure 23. HPA versus CDR of ARM2/28 2005 vintage................................... 18
Multi-dimension HPI Impacts................................................................................... 19
Figure 24. HPA versus HPA2D, Actual and Extreme Simulation..................... 19
HPI and DPI.............................................................................................................. 20
Figure 25. HPI & HPA QoQ 1975-2007............................................................ 20
Figure 26. HPI vs. DPI ....................................................................................... 20
Geographical location and Local HPI....................................................................... 21
2
 Figure 27. Geographic Components of HPA by CBSA..................................... 21
Prepayment Modeling....................................................................................................... 22
Prepayment Modeling Factor Components: ................................................................. 22
Housing Turn Over and Seasoning ............................................................................... 22
Figure 28. U.S. Housing turnover 1977-2007 .................................................... 22
Seasoning ...................................................................................................................... 23
Figure 29. CPR over various vintages of Discount Fixed Rate, coupon=6% .... 23
Figure 30. CPR over various vintages of ARM2/28 .......................................... 23
Teaser Effect ................................................................................................................. 23
Interest Only (IO) Effect............................................................................................... 24
Figure 31. CPR over various vintages of ARM2/28 .......................................... 24
Refinance ...................................................................................................................... 24
Figure 32. Refinance: Stratification by Coupon,................................................ 24
Burnout Effect............................................................................................................... 25
CLTV Wealth Effect..................................................................................................... 25
Figure 33. Wealth: CPR over various CLTV of ARM2/28 ............................... 25
Figure 34. Fitted CPR over CLTV 81-90 of ARM2/28, 2004 vintage .............. 25
FICO Credit Effect:....................................................................................................... 25
Figure 35. Credit: CPR by FICO of ARM2/28 .................................................. 26
Figure 36. Credit: Fitted CPR............................................................................. 26
Prepayment Penalty ...................................................................................................... 26
Figure 37. CPR over various vintages of ARM2/28 .......................................... 26
Interaction between Prepayment and Default ............................................................... 27
Figure 38. CDR and CPR of ARM2/28, 2004 vintage...................................... 27
Figure 39. Loss projection of ARM2/28, 2004 vintage ..................................... 27
Delinquency Study............................................................................................................ 28
Delinquency, the leading indicator ............................................................................... 28
Figure 40. Default and Delinquency over time for 2003 vintage....................... 28
Figure 41. Cross Correlations of Default and Delinquency for 2000 vintages .. 29
Analysis among delinquency spectrum ........................................................................ 29
Figure 42. Correlations between various Delinquencies.................................... 29
A Delinquency Error Correction Default Model .......................................................... 29
Figure 43. Delinquency Error Model: Actual vs. Fitting ................................... 30
Figure 44. Comparison of Model Explanation Power for 2000-2007 Vintages 31
Conclusion ........................................................................................................................ 32
Why Innovate? .............................................................................................................. 32
Innovation ..................................................................................................................... 33
Advantages.................................................................................................................... 33
Findings......................................................................................................................... 34
Future Improvements .................................................................................................... 34
Business Cycle – Low Frequency of Credit Spread ................................................. 34
Figure 45. Historical TED Spread and Histogram ............................................. 35
Dynamic Loss Severity ............................................................................................. 35
Appendix I Default and Prepayment Definition ............................................................... 36
Appendix II General Model Framework........................................................................... 37
Appendix II – Default Specification ................................................................................. 38
Appendix III – Prepayment Specification......................................................................... 42
REFERENCES ................................................................................................................. 44
3
INTRODUCTION
Subprime mortgages are made to borrowers with impaired or limited credit histories. The
market grew rapidly when loan originators adopted a credit scoring technique such as
FICO to underwrite their mortgages. A subprime loan is characterized by a FICO of
between 640 to 680 or less, versus the maximum of 850. In the first half of the decade,
the real estate market boom and well-received securitization market for deals including
subprime mortgages pushed the origination volume to a series of new highs. In addition,
fierce competition among originators created various new mortgage products and a
relentless easing of loan underwriting standards. Borrowers were attracted by the new
products such as “NO-DOC, ARM 2/28, IO” that provided a low initial teaser rate and
flexible interest-only payments in the first two years, without documenting their income
history.
As the mortgage rates began to increase in the summer of 2005 and housing activity
revealed some signs of a slowdown in 2006, the subprime market started to see some
cracks as the delinquencies began to rise sharply. The distress in the securitization market
backed by subprime mortgages and the resulting credit crisis, had a ripple effect initiating
a series of additional credit crunches. All this has pushed the U.S. economy to the edge
of recession and is jeopardizing the global financial markets.
The rise and fall of the subprime mortgage market and its ripple effects raises a
fundamental question. How can something as simple as subprime mortgages, which
accounts for only 6-7% of all US mortgage loans, be so detrimental to the broader
economy as well as the global financial system?
Before formulating an answer to such a big question, we need to understand what the
fundamental risks of subprime mortgages are. Traditional valuation methods for
subprime mortgages are obviously insufficient to measure the associated risks that
triggered the current market turmoil. What is the missing link between traditional default
models and reality? Since a mortgage's value is highly dependent on its future cash flows,
the projection of a borrower's embedded options becomes essential to simulate its cash
flows. Studying consumer behavior to help project prepayments and defaults (call/put
options) of a mortgage is obviously the first link to understanding the current market
conditions.
This paper focuses on modeling the borrower’s behavior and resultant prepayment or
default decision. A Dynamic Econometric Loss (DEL) model is built to study subprime
borrower behavior, and project prepayment and default probabilities based on historical
data from Loan Performance’s subprime database (over 17MM loans) and prevailing
market conditions from 2000 to 2007.
The paper is organized in the following manner. We started by constructing a general
model framework in a robust functional form that is able to not only easily capture the
impact of individual model determinants, but also be flexible enough to be changed to
reflect any newly found macro-economic variables. We then modeled default behavior
through an individual factor fitting process. Prepayment modeling follows a similar
process with consideration of the dynamic decision given prior prepayment and default
history. The delinquency study builds the causality between default and delinquencies
and the relationship within the spectrum of different delinquencies. We then utilized the

4
delinquencies as a leading indicator and error correction term to enhance the
predictability of the forecasted defaults by 15%. Our findings and forthcoming research
are then drawn in conclusion section.
Figure 1. What happened in 2007
01/18 A real-estate consortium unveils a 07/17 News Corp. reaches a
07/20 The Dow industrials 10/25 Merrill posts a $2.24 billion
$21.6 billion offer for Equity Office tentative agreement to acquire Dow
cross the 14000 m ilestone loss as a larger-than-expected
properties Trust. Jones at its original offer price of
for the first time. $8.4 billion write-down on
$60 a share.
mortgage-related securities
02/10 Fortress Investment Group leaves the firm with its first
LLC’s shares surge 68% in their 07/25 Countrywide
quarterly deficit since 2001.
debut to finish at $31. Financial Corp. says profit
slips 33%, dragged down
by losses on certain types
04/26 ABN Amro Holding NV 06/22 Blackstone Group LP’s of prime mortgage loans.
14,500 IPO is priced at $31 a share, 11/27 Citigroup, seeking to
receives a $98.58 billion
raising as much as $4.6 restore investor confidence amid
takeover approach from a group
billion. massive losses in the credit
led by Royal Bank of Scotland
markets and a lack of
Group PLC.
06/13 U.S. bond yields 08/11 Central banks perm anent leadership, receives
hit a five-year high as pum p money into a $7.5 billion capital infusion;
inventors continue to the financial system 11/27 HSBC’s SIV bailout will
14,000 03/09 New Century Financial sell Treasurys, with the for a second day to move 2 SIV’s of $45 billion to its
Corp.’s creditors force the yield on the benchm ark ease liquidity balance sheet.
subprim e-mortagage lender 10-year not rising to strains.
to stop m aking loans am id 5.249%.
rising defaults.

13,500
10/27 Countrywide
03/23 Blackstone files for Financial Corp.
an IPO to raise about $4 07/18 Bear posts its first
billion Stearns Cos. quarterly loss in 25
Says two years on about $1
13,000 hedge funds it billion in write-
runs are worth downs.
nearly nothing.
12/14 Citigroup 12/21 Bear
08/17 Blue chips fall more
bails out seven Stearns
than 300 points at 845.78
affiliated posts a loss
08/14 Goldman Sachs after foreign markets
structured- of $854
Group Inc. says three of tum ble on certain that U.S.
12,500 06/23 Bear Stearns credit problems could
investm ent million, the
its hedge funds have vehicles, or SIVs, first in its
Cos. Agrees to lend trigger a global slowdown;
seen the net value of bring $49 billion in 84-year
as m uch as $3.2 08/17 Countrywide taps an
their assets fall about assets onto its history. The
billion to one of its $11.5 billion credit line in a
$4.7 billion this year. balance sheet and firm takes a
own troubled hedge bid to shore up its finances,
funds. further denting its $1.9 billion
but its stock falls 11%.
capital base. write down.
12,000
7

7
/0

/0

/0

/0

/0

/0

/0

/0

/0

/0
/0

/0

/0
3

9
/0

/0

/0

/0

/0

/0

/0

/3

/3

/2

/2
/0

/3
01

02

04

05

06

07

08

08

09

10

11

12
0

What happened in 2008

5
MODEL FRAMEWORK
When a lender issues a mortgage loan to its borrower, the loan is essentially written with
two embedded American options with an expiration co terminus with the life of the loan.
The lender will then receive payments as compensation for underwriting the loan. The
payments will include interest, amortized principal and voluntary/involuntary
prepayments along with any applicable associated penalties. The risk for the lender is
they might not receive the contractual payments and will need to go after the associated
collateral to collect the salvage value of the loan. Additionally, the foreclosure procedure
could be costly and time consuming.
Unscheduled payments come in two forms. A voluntary prepayment is usually referred to
simply as ‘prepayment’ and an involuntary prepayment, which is known as ‘default’
(with lags to potentially recover some portion of interest and principal proceeds).
Prepayment is nothing but a call option on some or all of the loan balance plus any
penalties at a strike price that a borrower has the right to exercise it if the option is in-the-
money. By the same token, default is a put option with the property’s market value as the
strike price to the borrower. Understanding the essence of both options, we need to find
the determining factors that trigger a borrower to prepay/default through filtering the
performance history of the loan. A list of determinant factors regarding consumer
behavior theory for modeling default and prepayment will be discussed in the next two
sections.
In order to construct a meaningful statistical model framework for empirical work, the
availability of data and the data structure are essential. In other words, our model
framework is designed to take full advantage of data structure of the Loan Performance
subprime mortgage historical information and the prevalent market information. The
model empirically fits to the historical default and prepayment information of US
subprime loan performance from 2000 to 2007 (over 17MM loans).
Figure 2. Number of Securitized Alt-A and Subprime Mortgage Origination
Type /
Orig. Year ARM OTHER ARM2/28 ARM3/27 ARM5/25 FIXED Grand Total
2000 11,452 187,232 68,430 4,059 390,671 661,844
2001 11,389 261,316 67,018 10,449 477,718 827,890
2002 33,776 434,732 100,939 25,827 605,233 1,200,507
2003 51,548 697,073 164,228 71,839 958,170 1,942,858
2004 221,818 1,239,522 413,366 213,572 1,172,413 3,260,691
2005 496,697 1,577,003 393,020 301,829 1,619,257 4,387,806
2006 490,975 1,137,345 234,344 349,460 1,754,382 3,966,506
2007 99,946 161,480 36,795 160,549 404,278 863,048
Grand Total 1,417,601 5,695,703 1,478,140 1,137,584 7,382,122 17,111,150

Mathematically, our general framework constructs the default and prepayment rates as
two separate functions of multiple-factors that the factors are categorized into two types –
static and dynamic. 1 The static factors are initially observable when a mortgage is

1
There is no industry standard measure for default rate, thus a different definition on default rate will give a
very different number. As there is no set standard we define our default rate based on the analysis in this
6
originated such as borrower’s characteristics and loan terms. Borrower’s characteristics
include Combined-Loan-To-Value ratio (CLTV), Fair Isaac credit score (FICO), and
Debt-To-Income ratio (DTI). Loan terms include loan maturity, loan seasoning, original
loan size, initial coupon reset period, interest only (IO) period, index margin, credit
spread, lien position, documentation, occupancy, and loan purpose. The impacts to the
performance of a loan from the static factors provide the initial causality impacts yet their
influence could diminish or decay as the information is no longer up to date.
Dynamic factors include several macro-economic variables such as Housing Price
Appreciation (HPA), prevailing mortgage interest rates, consumer confidence, gross
disposable income, employment rate, and unemployment rate. These dynamic factors
supply up to date market information and thus play an important role in dynamically
capturing market impacts. The accuracy of capturing causality impacts due to the static
factors and the predictability of the dynamic factors presented a constant challenge during
the formulation of this model.
For each individual factor, a non-linear function is formulated according to its own
characteristics. For example, a “CLTV” factor for modeling default is formulated as the
function of default rate over CLTV ratio. However, a DOC factor is formulated as the
function of multiplier over discrete variables of “FULL” versus “LIMITED” with
percentages of respective groups.
A general linear function of combined multi-factor functions is then constructed as basic
model framework to fit the empirical data and to project forecasts for prepayments and
defaults.2 In the following sections, we will discuss each factor in detail.

paper, “Loss Severity Measurement and Analysis”, The MarketPulse, LoanPerformance, 2006 Issue 1, 2 –
19. Please refer to Appendix I for definition of default we used throughout this paper.
2
See Appendix II for the details of model specification.
7
DEFAULT MODELING

Default Modeling Factor Components:

Seasoning Occupancy
Combined Loan-To-Value (CLTV) Owner
Credit Score (FICO) Second home
Debt-To-Income ratio (DTI) Investor
Payment Shock (IO) Property Type
Relative Coupon Spread Single-Family
Loan Size Multi-Family
Lien Condo
First Loan Documentation
Second and others Full
Loan purpose Limited
Purchase Housing Price Appreciation (HPA)
Refinance State Level
Cashout CBSA Level

Seasoning
Loan information regarding Figure 3. Seasoning: CDRs by Date and Vintages of ARM 2/28
borrower’s affordability is 20

usually determined at 18
2003ACT

origination. As a loan 16
2003Prj
2004ACT
2004Prj
seasons, its original 14
2005ACT
2005Prj

information decays and its 12

default probability starts to

CDR (%)

surge. A seasoning baseline 10

curve with annualized 8

Constant Default Rate 6

(CDR) against its seasoning 4

age would post a positive 2

slope curve for the first 0

three years. 02/03 08/03 02/04 08/04 02/05 08/05

Date
02/06 08/06 02/07 08/07

Figure 3 shows actual CDR

curves and their fitted result of different vintages of ARM 2/28 mortgage pools. They
roughly follow a similar shape to the Standard Default Assumption (SDA) curve. 3 But
as shown in the figure 4, the ramp up curve can be very different for different vintages.

3
SDA is based on Federal Housing Administration (FHA)’s historical default rate and was developed by
Bond Market Association (BMA), now known as Securities Industry and Financial Markets Association
(SIFMA).
8
Why is the 2005 seasoning pattern faster than prior vintages?
Since the seasoning baseline
Figure 4. Seasoning: CDRs by Age and Vintages of ARM 2/28
curve is not independent of
dynamic factors, a dynamic 20

factor such as HPA could tune 18

2003ACT

vintage seasoning curves up and 16

2003Prj
2004ACT
2004Prj
down. In Figure 4, the 2005 14
2005ACT
2005Prj

seasoning pattern is significantly 12

steeper than its prior vintages.

CDR (%)
10
Looser underwriting standards,
deteriorating credit 8

fundamentals can be an 6

important reason. The negative 4

HPA obviously starts to 2

adversely impact all vintages 0

after year 2005. 0 6 12 18

Age (month)
24 30 36

Source: Beyondbond Inc, LoanPerformance

Payment Shock – Interest Only (IO)

The boom of subprime market Figure 5. IO Payment Shock: CDRs by Date of ARM 2/28
in recent years has introduced 25

new features to the traditional

mortgage market. An ARM 20
ACT_NIO
Prj_NIO

2/28 loan with 2-year Interest ACT_IO

Prj_IO

Only (IO) feature has a low

fixed initial mortgage rate and
15
CDR (%)

also pays no principal for the

first two years prior to the 10

coupon reset4.
When the IO period ends, the
5

borrower typically faces a

much higher payment based 0
02/04 08/04 02/05 08/05 02/06 08/06 02/07 08/07

on its amortized principal plus

Date

the fully indexed interest. This

sudden rise in payments could produce a ‘Payment Shock’ and test the affordability to
borrowers. Without the ability to refinance, borrowers who are either under a negative
equity situation, or not able to afford the new rising payment will have a higher
propensity to default. Consequently, we see a rapid surge of default rates after the IO
period.

4
The reset is periodical and the interest rate is set as Index + Margin.
9
The ending of the IO period triggers payment shock and will manifest itself with a spike
to delinquency.5 Delinquent loans eventually work themselves into the defaulted
category within a few months after the IO period ends. Figure 5 shows the difference
patterns and the default lagging between IO and Non-IO of ARM 2/28 pools.

Combined Loan-to-Value (CLTV):

LTV ratio measures the ratio of mortgage indebtedness to the property’s value. When
multiple loans have liens added to the indebtedness of the property, the resulting ratio of
CLTV becomes more meaningful measure of the borrower’s true equity position.
However, the property value might not be available if a “market” property transaction
does not exist. A refinanced mortgage will refer to an ‘appraisal value’ as its property
value. Note that ‘appraisal value’ could be manipulated during ferocious competition
amongst lenders in a housing boom market and would undermine the accuracy of CLTV.
As we know, default is essentially a put option embedded to the mortgage for borrower.
In a risk neutral world, a borrower should exercise the put if the option is in-the-money.
In other words, a rational borrower should default the mortgage if CLTV is greater than
one or the borrower has negative equity.
Figure 6. Stratified seasoned CDR over CLTV ranges
CDR vs. CLTV of ARM 2/28 Non-IO with age>24 CDR vs. CLTV of ARM 2/28 Non-IO with age >24
and FICO between 641 and 680
16
14
14
12
12
10
10
8
CDR%
CDR%

8
6
6
4
4
2
2

0
0
11- 60 61- 70 71- 80 81- 90 91- 95 96-125
11- 60 61- 70 71- 80 81- 90 91- 95 96-125
CLTV Range CLTV RANGE

Source: Beyondbond Inc, LoanPerformance

At higher CLTVs it becomes easier to reach a negative equity level as loan seasons and
its default probability increases. Figure 6 provides the actual stratification result of CDR
over various CLTV ranges. Obviously, CDR and CLTV are positive correlated. In
addition, lower CDR values are observed for top subprime tier FICO ranged from 640 to
680. It shows that the FICO tier granularity is another important factor in modeling.

5
The delinquency rate is measured by OTS (Office of Thrift Supervision) or MBA (Mortgage Bankers
Association) convention. The difference between these two measures is how they count missed payments,
MBA delinquency rate count the missed payment at the end of the missing payment month while OTS
delinquency rate count the missed payment at the beginning of the following month after missing payment.
This difference will pose a 1-30 days delay of record. OTS delinquency rate is the prevailing delinquency
measure in subprime market.
10
However, since CLTV is Figure 7. CDRs by Date and CLTVs of ARM 2/28
obtained at the loan’s 20

origination date, it does not 18

dynamically reflect housing

80-90 Act
16 80-90 Prj
70-80ACT
market momentum. We 14
70-80Prj
60-70ACT

introduce a dynamic CLTV 60-70Prj

12
that includes housing price

CDR (%)
appreciation from loan 10

origination to attempt to 8

estimate more precisely the 6

actual CLTV. This dynamic 4

CLTV allows us to better

2
capture the relationship
between CLTV and default. 0
02/04 08/04 02/05 08/05 02/06 08/06 02/07 08/07

Figure 7 clearly illustrates Date

that different CLTV groups Source: Beyondbond Inc, LoanPerformance

show a different layer of
risk level.

FICO
FICO score is an indicator of borrower’s credit history. Borrowers with high FICO scores
maintain a good track record of paying their debts on time with a sufficiently long credit
history.6
In recent years, people believe FICO is no longer an accurate indicator due to the boom
of hybrid ARM loans and fraudulent reporting to the credit bureaus. Since refinancing is
much easier to obtain than before, issuers are giving out tender offer to borrowers in
order to survive the severe competition amongst them.
CLTV and FICO score are two common indicators that the industry uses to predict
default behavior. 7 We examine the combined CLTV and FICO effects on CDR as shown
in the Figure 8. The figure presentation a same 3-D surface of stratified CDR rates over
CLTV and FICO ranges from two different angles for seasoned ARM 2/28 pools. The
relationship between CLTV and CDR is positively correlated across various FICO ranges.
On the other hand, the relationship between FICO and CDR is somewhat negatively
correlated across various CLTV ranges. However, the case is not as significant. FICO
factor impact is obviously not as important as we originally expected.

6
According to Fair Isaac Corporation’s (The Corporation issued FICO score measurement model)
disclosure to consumer, 35% of this score is made up of punctuality of payment in the past (only includes
payments later than 30 days past due), 30% is made up of the amount of debt, expressed as the ratio of
current revolving debt (credit card balances, etc.) to total available revolving credit (credit limits) and 15%
is made up of length of credit history. Severe delinquency (30 plus) and credit history length make up 50%
of the FICO score. This score reflects people’s willingness to repay. It’s essentially the probability
distribution for people’s default activity on other debts such as credit card and/or utility bills and etc.
Statistically speaking, people with higher FICO score will have lower probability to default.
7
Debt-to-Income ratio is also an important borrower characteristic, but in recent years, more Limited-Doc
or/and No-Doc loans are issued. For these loans, many of them do not have DTI ratio report, so we
consider DTI separately for different DOC type.
11
Figure 8. Stratified CDR by CLTV and FICO of ARM 2/28
CDR vs. FICO and CLTV of Seasoned ARM 2/28 CDR vs. FICO and CLTV of Seasoned ARM 2/28
16
16
14
14
12
12
10
10
CDR%

CDR%
8 8

6 6

4 4

2 2

<= -57
0 >=

54 -60
0

30 0
9 96
81 1 - 9

1
57 64 0

1
<=301 >=96

1
60
541- 71 -9 5 91- 95
571-

1-
64
- 0 81- 90

0
570 601- 61 80 71- 80

1-
68
600 641- - 61- 70

68
640 681- <= 70

1-
>=
<=60

0
>=701 60

70
680 CLTV RANGE CLTV RANGE FICO RANGE

70
FICO RANGE 700

0
1
<=60 61- 70 71- 80 81- 90 91- 95 >=96 <=301 541-570 571-600 601-640 641-680 681-700 >=701

Source: LoanPerformance

In our analysis, CLTV=75 and FICO=640 serves as a base curve, and then we adjust the
CDR according to movements of other default factor.
Figure 9 gives an example of Figure 9. FICO: CDRs by Date and FICOs of ARM 2/28
fitting results based on 18
680-700ACT

ARM2/28 2004 vintage pools. 16

680-700Prj
600-640ACT

The difference between 600- 14

600-640Prj

640 and 680-700 FICO ranges

only make approximately a 1%
12

difference in CDR for a 10

CDR (%)

seasoned pool. 8

0
02/04 08/04 02/05 08/05 02/06 08/06 02/07 08/07
Date

Source: Beyondbond

Debt-To-Income Ratio (DTI) and Loan Documentation (DOC)

The DTI in this paper is defined as the back-end DTI which means the debt portion for
calculating the DTI ratio includes not only PITI (Principal + Interest + Tax + Insurance)
but also other monthly debts such as credit card payments, auto loan payments and other
personal obligations. 8 The DTI ratio shows the affordability of a loan to a borrower and
provides us with a clearer picture of a borrower with an exceptionally high DTI. For
different regions of the country, the DTI ratio could imply a different financial condition

8
There are two major measures of DTI in the industry, Front-End DTI ratio = PITI/Gross Monthly Income;
Back-End-DTI ratio = PITI + Monthly debt/Gross Monthly Income. PITI=Principle + Interest + Tax +
Insurance.
12
 of the borrower because of different living standards and expenses between those of rural
areas and large cities.
DTI is captured and reported as part of the loan documentation process. Loan
documentation, also referred to as DOC, consists of three major groups: ‘FULL DOC’,
‘LOW DOC’, and ‘NO DOC’. Lenders usually require a borrower to provide sufficient
‘FULL’ documentation to prove their income and assets when taking out loans. People
who are self-employed and/or wealthy and/or have lumpy income stream are considered
as borrowers with ‘LIMITED’ (LOW or NO) documentation. In recent years, the fierce
competition pushed lenders to relax their underwriting standards and originated more
LIMITED DOC loans with questionable incomes. This uncertainty regarding income
poses uncertainties in determining the real DTI.
The stratification report shows two very different patterns of default between FULL and
LIMIT documentation categories when analyzing the DTI effect. For FULL-DOC loans,
default probability versus DTI is very much positively correlated, CDR increases as the
DTI increases. Since FULL-DOC loans are loans that have documented income and
assets, it shows the default DTI relationship most clearly as Figure 10. LIMITED-DOC
has weaker relationship compared to FULL-DOC. Figure 11 shows the two different time
series pattern of CDR curves and their fitted values between FULL and LIMITED DOCs.

Figure 10. Stratified CDR by DTI ranges Figure 11. DOC: Actual vs. Fitted for 2004
Stratified CDR of seasoned pools between 2000- Actual versus Fitted CDR curve over time by
2007 by documentation types, FULL and documentation types, FULL and LIMITED for
LIMITED 2004 vintages
12

Lim_ACT

10.5 Lim_Prj
FULL DOC Limited DOC 10
Full_ACT
10
Full_Prj

9.5 8

9
CDR (%)

6
8.5
CDR%

8
4

7.5

7 2

6.5
0
03/04 09/04 03/05 09/05 03/06 09/06
6
Date
<=10 11-15 16-20 21-25 26-30 31-35 36-40 >=41

DTI

Source: Beyondbond Inc, LoanPerformance Source: Beyondbond Inc, LoanPerformance

Since income is one of the main elements in determining the DTI ratio, the macro-
economic variable, unemployment rate, becomes an important determinant that affects an
individual’s income level. We find an interesting result when we plot the unemployment
rate against 3-month U.S. Treasury Bills. They are very negative correlated for the last 7
years. Whether it was a coincidence or not, it suggested that the monetary policy has been
mainly driven by the unemployment numbers.

13
Figure 12. Unemployment Rate 1979-2007 Figure 13. Unemployment vs. T-Bill 3mo

UNEMPLOYMENT vs. TBILL3M

18.00 6.5
16.00
6.0
14.00

UNEMPLOYMENT
12.00 5.5

10.00 99 00 01 02 03 04 05 06 07
5.0
8.00
4.5
6.00
4.0
4.00
3MTBILL
2.00 Unemployment 3.5
0 1 2 3 4 5 6 7
0.00
79 83 87 91 95 99 03 07 TB ILL3M

Source: US Bureau of Census Source: US Bureau of Census

Loan Size
Figure 14. Loan Size Stratification Figure 15. Size: Actual vs. Fitted CDR for 2004
Seasoned CDR by different Loan Size ranges

18
12 600-800ACT
600-800Prj
16 350-417ACT
350-417Prj
10 150-200ACT
14 150-200Prj

12
8

10
CDR%

CDR (%)

6
8

4 6

4
2

0
0
<100 100-150 150-200 200-250 250-300 300-350 350-400 >400 03/04 09/04 03/05 09/05 03/06 09/06 03/07 09/07
Date
Original BalanceRange

Source: LoanPerformance Source: Beyondbond Inc, LoanPerformance

Is bigger better? The conventional argument is that larger loan size implies a better
financial condition and lower likelihood of default. According to the stratification results
based on original loan size in Figure 14, CDR forms a smile curve across original loan
balance. Loans with sizes larger than $350,000 tend to be a bit riskier although the
increment is marginal. Loans with a size less than $100,000 also seem riskier. Larger
loans do not seem to indicate that they are better credits. The original loan size usually is
harder to interpret as it can be affected by the other factors such as lien, property type,
and geographical areas. For example, a $300,000 loan in a rural area may indicate a
borrower with growing financial strength; while the same amount in a prosperous large
city may indicate a borrower with weak purchasing power. Without putting size into the
14
 context of property type and geographic location, the factor could be misleading. This
may explain why we do not see a clear shape forming in Figure 14. Figure 15 shows the
three different time series pattern of CDR curves and their fitted values based on their
loan size ranges. Since the size is mixed for all the property types, the pattern and fitted
results for each category is distorted and the fit is not as good as other factors.

Lien
As we all know that the 2nd mortgage/lien has lower priority to the collateral asset than 1st
lien mortgage/lien in the event of a default. Thus, the 2nd lien is riskier than the1st lien.
The 2nd lien borrowers usually maintain higher credit score, usually with a FICO greater
than 640. We sometimes see a very mixed effect if the layer risk is not put into
consideration. In Figure 16, 2nd lien loans are significantly riskier than 1st lien loans
when measured against comparable FICO ranges for both liens. Figure 16 shows the two
different time series pattern of CDR curves and their fitted values based on their liens.

Figure 16. Lien Stratification Figure 17. Lien: Actual vs. Fitted CDR for 2004
Seasoned CDR of 1st versus 2nd Liens

14 20

1Lien_ACT 1Lien_Prj
18
12 2Lien_ACT 2Lien_Prj

16

10
14

8 12
CDR%

CDR (%)

10
6
8

4
6

4
2

2
0
1st 2nd 0
LIEN 02/04 08/04 02/05 08/05 02/06 08/06 02/07 08/07
Date

Source: LoanPerformance Source: Beyondbond Inc, LoanPerformance

Occupancy
Occupancy consists of three groups: ‘OWNER’, ‘INVESTOR’, and ‘SECOND HOME’.
The ‘OWNER’ group views the property as their primary home, rather than as an
alternative form of housing or an investment. This group will face emotional and
financial distress if the property is in foreclosure or REO. Thus, this group has a lower
propensity to default compared with others if all other factors remain the same. On the
other hand, ‘INVESTOR’ and ‘SECOND HOME’ groups would be more risk neutral and
are more willing to exercise their options rationally. In other words, they should have
higher default risk.
Figure 18 reports an occupancy stratification regarding the default risk profile. The result
evidently supports the risk neutral idea with respect to the ‘INVESTOR’ group and
‘INVESTOR’ does show the highest default risk among all three groups. The ‘OWNER’
group however, is not the lowest default risk group. Instead, the ‘SECOND HOME’
group is the lowest one. The observation is interesting, but not so intuitive. It indicates
15
 that when a borrower faces the financial stress, a ‘SECOND HOME’ will be sold first
even at a loss to support his/her primary home. Thus the default risk of ‘SECOND
HOME’ is actually reduced by incorporating a borrower’s primary home situation and
cannot be simply triggered by the risk neutral idea. Figure 19 shows the two different
time series pattern of CDR curves and their fitted values between ‘OWNER’ and
‘INVESTOR’.
Figure 18. Occupancy Stratification Figure 19. Occupancy: Actual vs. Fitted CDR for 2004
Season CDR by Occupancy types for ARM 2/28 and
FIXED 2000-07 vintages
20

18
InvestorAct
12
16 InvestorPrj
ARM2/28 FIXED

10 OwnerAct
14

OwnerPrj
12
8
CDR (%)
CDR%

10
6
8

4
6

2 4

2
0
Owner Second Home Investor
0
02/04 08/04 02/05 08/05 02/06 08/06 02/07 08/07
Occupany
Date

Source: LoanPerformance Source: Beyondbond Inc, LoanPerformance

Purpose
Figure 20. Purpose Stratification Figure 21. Purpose: Actual vs. Fitted CDR for 2004
Season CDR by Purpose types for ARM 2/28 and
FIXED 2000-07 vintages
20
12
ARM2/28 FIXED
18

10
Refi_ACT
16 Refi_Prj
Purchase_ACT
8 Purchase_Prj
14
Cashout_ACT
CDR%

Cashout_Prj
12
6
CDR (%)

10

4
8

2
6

0 4

Purchase Cash Out Refi-Other

2
Purpose

0
02/04 08/04 02/05 08/05 02/06 08/06 02/07 08/07
Date

Source: LoanPerformance Source: Beyondbond Inc, LoanPerformance

Loan Purpose classifies three key reasons for borrowing a loan as ‘PURCHASE’,
‘CASHOUT’, and ‘REFI’.9 ‘PURCHASE’ means the borrower is a first time home
buyer. ‘CASHOUT’ refers to a refinance loan with extra cash inflow to the borrower due

9
For simplicity sake, we put refinance, 2nd mortgage, and other miscellaneous types as ‘REFI’.
16
to the difference between new increased loan amount and the existing loan balance.
‘REFI’ uses the loan for refinancing the outstanding balance without any additional funds
draw from the equity in the property.
‘CASHOUT’ and ‘REFI’ usually reflects an intention to rollover the IO period or benefit
from a lower mortgage rate. They can only be afforded by borrowers in good financial
condition. ‘REFI’ is a group of borrowers with a higher FICO, LTV as compared to the
other two categories. So we expect the ‘REFI’ Loans to have a lower default rate than
‘PURCHASE’ loans. The argument seems correct for the fixed rate mortgages. ‘REFI’
borrowers have much lower default probability than ‘PURCHASE’.
Beginning in 2007, the credit crunch hit the market and most of the lenders tightened
their credit standards. Hybrid ARM types of loans, such as ARM 2/28, facing new resets,
borrowers who no longer qualified for refinancing were in danger of defaulting. If these
people can no longer afford the payment after IO and/or reset, they will eventually enter
default. ARM 2/28 loans show a significant increase in defaults for ‘REFI’ purpose as
compared with FIXED rate loans. Figure 21 shows the three different time series pattern
of CDR curves and their fitted values among various purpose types.

Dynamic Factors: Macro-Economic Variables

As we have mentioned in Model Framework, macro-economic variables such as HPI,
interest rate term structure, unemployment rate, and etc. that supply up to date market
information can dynamically capture market impacts.
In theory, an economy generally maintains its long-term equilibrium as “Norm” in the
long run and a handful of macro-economic variables are usually used to describe the
situation of the economy. While the economy is in its “Norm” growing stage, these
macro-economic variables usually move or grow very steadily and the risk/return profile
for an investment instrument can be different depending on its unique investment
characteristics. Because of that, a diversified investment portfolio can be simply
constructed based on relationship of the correlation matrix. Thus the macro-economic
variables usually are ignored during the “Norm” period. However, when an economy is
under stress and approaches a “bust” stage, many seemingly uncorrelated investments
sync together. The same the macro variables become the main driving forces that
crucially and negatively impact the investment results. The current credit crunch is
creating mark to market distress for investments across not only various market sectors
but also credit ratings, clearly describing our view regarding to these macro-economic
variables.
Since the severe impacts due to these variables mostly occur in economic downturn, cross
correlation could provide a preliminary result in understanding the causality and the
magnitude of their relationship. The dynamic interaction between these variables and
consumer behavior would then provide a better sense of prediction and therefore either
prevent the next downturn or efficiently spot an investment opportunity based on the next
market recovery.

17
Housing Price Appreciation (HPA)
The Housing Price Index Figure 22. 30-day Delinquency of ARM2/28 by Vintages
(HPI) has been the most 20

quoted macro-economic 18
variable that causes the 16
high delinquency and 14
default rates since the

Delq Rate (%)

12
beginning of subprime
crisis.10 Thus Housing 10

Price Appreciation 8

(HPA) which measures 6

2000 2001
the housing appreciation 4 2005 2006
rate compared with a 2 2007
year earlier has become
0
the most important 1 5 10 18 26 34 42 50 58 66 74 82 90
indicator within the U.S. Age
housing market. By
Source: OFHEO, LoanPerformance
comparing the 30-day
delinquency across vintages, the delinquency rates unidirectionally increase after the
2005 vintage.
When we look at our Figure 23. HPA versus CDR of ARM2/28 2005 vintage
seasoning pattern across
14 16
2000-2005 vintages, the
2005 seasoning pattern 12
14

starts to surge after 18- 12

month of age or the 3rd 10

10
quarter of 2006. 8 8
CDR%

Coincidentally, HPA CDR of ARM228 in 2005 HPA

6
started to decline in the 6
nd
2 quarter of 2006. 4
4

Although a similar HPA 2

rd
pattern appeared at the 3 2
0
quarter of 2003, the main 0 -2
difference was that the Feb-05 Jun-05 Oct-05 Feb-06 Jun-06 Oct-06 Feb-07 Jun-07 Oct-07

former on was the up- Distribution Date

trend of HPA, but the Source: OFHEO, LoanPerformance

latter was on a down-
trend. Defaults in the 2003 were obviously lower than in 2006 with comparable loan
features and seasoning/age. In order to capture this subtle trend difference, we studied
HPI and it’s various dimensions in addition to HPA level which proved illustrative.

10
The Housing Price Index HPI used in this paper is published by Office of Federal Housing Enterprise
Oversight (OFHEO) as a measure of the movement of single-family house prices. According to OFHEO,
The HPI is “a weighted, repeat-sales index, meaning that it measures average price changes in repeat sales
or refinancing on the same properties”. See website of OFHEO www.ofheo.gov for details.
18
Multi-dimension HPI Impacts
To systematically identify Figure 24. HPA versus HPA2D, Actual and Extreme Simulation
the impacts of HPA, we 15
D Lags4 YoY
5

measure HPA in three

D Lags4 YoY2D
U p w a rd T re n d 4
10
D o w n w a rd T re n d

aspects regarding each loan: 3

•
A s s u m H P A d r o p 3 0 % in 2 -

Cumulative HPI, an y e a r a n d th e n f la t 2

accumulative HPA since

0
1
M S M S M S M S M S M S M S M S M S M S M S
ar ep ar ep ar ep ar ep ar ep ar ep ar ep ar ep ar ep ar ep ar ep
-0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -0 -1 -1
0 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7 8 8 9 9 0 0

origination, is calculated -5
0

based on HPI levels to -1 0

-1

capture equity gain for S tr e s s e d S c e n a rio

-2

borrowers. -1 5
-3

• HPA, the change rate of -2 0 -4

HPI, captures the pulse Source: OFHEO, LoanPerformance

of housing market.
• HPA2D, the change of HPA, is used to capture the trend/expectations of housing
market.
The HPA factors form multi-dimension impacts to reflect loan’s up-to-date capital
structure, current housing market conditions, and future housing market prospects. We
embedded the ‘Cumulative HPI’ into CLTV to build a dynamic CLTV to reflect the
dynamic equity value to the property. In a risk neutral analysis, an option model can be
easily applied to project the default probability. HPA is already a leading market
indicator in explaining defaults. HPA2D basically serves as the second derivative of HPI;
it allows us to capture the general expectation on home price movements and market
sentiment.
The negative impact due to HPA2D in the 3rd quarter of 2006 is apparently different
from the 3rd quarter of 2003 even the HPA numbers are at the similar level.11 HPA2D
undoubtedly offers another dimension that reflects consumers expectations about the
general housing market. When HPA2D is negative, the probability of borrowers holding
negative equity increases.
The remaining challenge lays in the deterioration of the housing market which is
producing unseen record-low HPI levels. While the HPA continues decreasing, HPA2D
plunges even faster. Our multi-dimensional HPA empirical fitting merely relies on a very
limited range of in-sample HPA data. To extrapolate HPA and HPA2D requires
numerous possible market simulations to induce a better intuitive sense of numbers. The
shaded area in Figure 24 draws a sample of simulated extreme downturn housing market
that assumes a 30% drop of HPI level based on 4th quarter of 2007 level and then a
leveling-off. Based on the simulation results, the HPA2D starts to pick up at least one-
quarter earlier than HPA and one-year earlier than HPI level. While a two-year HPI
downturn is assumed, the consumer’s positive housing market expectation reflected in
HPA2D effectively reduces their incentive to walk away from their negative equity loans.

11
We have smoothed HPA and HPA2D series to create better trend lines. A linear weighted distributed
lags of last four quarters are adopted for smoothing the series.
19
 This simulation case example clearly shows how the forecasted HPA and HPA2D
numbers could provide a better intuitive market sense to the model.
The relationship between HPI and consumer behavior that forms the HPI impact to
default and prepayment are then modeled. We illustrate the multi-dimensional HPI
impact through an example as below,
1. HPCUM ↓(below 5%) => CLTV↑ => MDR↑, SMM ↓
2. HPA ↓(below 2%) => MDR↑ , SMM ↓
3. HPA2D ↓(below -5%) => MDR↑ , SMM ↓

HPI and DPI

When the economy is experiencing a potentially serious downturn, generating HPI
predictions going out three years is a much better approach than random simulations.
Since HPI has increased so rapidly since 2000, the current fall could be merely an
adjustment to the previously overheated market. The magnitude and ramp up period of
the adjustment nevertheless determines a consumer’s behavior of exercising their
mortgage embedded options. Finding a long-term growth pattern of HPI thus becomes
very vital for predicting and simulating future HPI numbers.
Based on Figure 26, HPI draws a constant relationship with Disposable Personal Income
(DPI) in the long run. Since DPI is a more stable process, a long term pattern HPI
prediction based on the observed relationship between DPI and HPI provides a better
downturn average number. Based on our long-term HPI prediction, HPI could potentially
drop as much as 14% by the end of 2009.12

Figure 25. HPI & HPA QoQ 1975-2007 Figure 26. HPI vs. DPI
.05 HPI vs. DPI (based on 1975q1-2007q3)
.04 500

.03
.02 400
250
.01
300
200 .00
HPI

-.01
150 200

100
100
50
76 78 80 82 84 86 88 90 92 94 96 98 00
0
0 2000 4000 6000 8000 10000 12000
HPI HPI_QOQ
DPI

Source: US Bureau of Census Source: US Bureau of Census

12
A 5% decrease by the end of 2009 in average plus another 9% based on two standard errors of regression
of HPI on DPI result.
20
 Geographical location and Local HPI
In housing market, geographical location (location, location, location or L3) is
undoubtedly the most important price determinant, as it is globally unique.. While we are
pointing out all HPI impacts in general, HPI in the national level does not reflect the
actual local situation and thus distorts the default impact ignoring the granularity of
detailed local housing market information. The consequence of ignoring this kind of
granularity can be very severe when a geographically diversified mortgage pool’s CLTV
has a fat-tailed distribution in its high CLTV end. Since the detailed local HPI can vary
from the national HPI, loans with negative equity have a higher level of relevance than
the use of the national HPI.
Fortunately, we are able to differentiate HPI impacts by drilling down to the HPI
information on a state as well as CBSA level. Figure 27 shows the examples of actual
levels of HPA on December 2007 and our projection of HPA for June 2008 detailed by
CBSA. We started with a national level HPI model to obtain the long term relationship
between HPI and DPI. We then build dynamic correlation matrix between national and
state as well as national and CBSA levels respectively that dynamically estimates
parameters and generate forecasts on the fly. The CBSA level HPI is especially important
for calculating dynamic CLTV. Since the cumulative HPI (HPCUM) is calculated as the
cumulative HPA since origination to capture wealth effect for generating dynamic CLTV.
The more detailed level information apparently helps to predict if a mortgage has crossed
the negative equity zone.
Figure 27. Geographic Components of HPA by CBSA

21
PREPAYMENT MODELING

Prepayment Modeling Factor Components:

Housing Turnover and Age
Refinancing
Teaser Effect
Interest Only (IO) Effect
Burnout Effect
Seasonality
Loan-To-Value Effect
Credit Score FICO Effect
Prepayment Penalty
Housing Price Wealth Effect

Housing Turn Over and Seasoning

Housing Turnover rate is the Figure 28. U.S. Housing turnover 1977-2007
ratio of total existing single- 9%

family house sales over the

existing housing stock.13 With
8%

7%

the exception of cases in early

80s, the housing turnover rate
6%
Housing Turnover Rate

has been rising steadily for the

5%

last fifteen years until 2005. The 4%

result of a rising housing 3%

turnover rate indicates that home 2%

owners are capable of moving 1%

around more than in the past. In 0%

77 79 81 83 85 87 89 91 93 95 97 99 01 03 05 07
19 19 19 19 19 19 19 19 19 19 19 19 20 20 20 20
the housing market boom era, it year

also indicates the height of

Sources: National Association of Realtors and Beyondbond
speculation. When the housing
boom came to an end, the housing turnover rate started to reduce. The movement of
housing turnover after 2005 shows exactly the same directionality. Since the housing
turnover rate used as the base prepayment speed and could generate a significant tail risk
of principal loss given the same default probability, it is especially crucial for a high
default and slow prepayment environment like the current one.

13
We use five year moving average of ‘Total Occupied Housing Inventory’ based on U.S. Census Bureau
times 0.67 to estimate the total Single-family Housing Stock.
22
Seasoning
The initial origination fee and the loan Figure 29. CPR over various vintages of Discount
closing expenses usually takes a few Fixed Rate, coupon=6%
years to be amortized, and this
discourages the new mortgagors from 30
2003
prepaying their mortgages early in the 2004

mortgage term. This ramping-up effect is 2005

20
the seasoning factor. Figure 29 shows the

CPR (%)
age pattern observed for Fixed Rate loans.
The ramping-up period initially lasts for 10

the first few months and then it starts to

level off or decrease due to other
0
prepayment factors. 0 6 12 18 24 30 36 42 48
Age

Hybrid ARMs exhibit similar patterns Sources: Beyondbond, Loan performance

initially during the first 12 months. For
hybrid like ARM 2/28, the prepayment level climbs up from 0% to around 20-50% CPR
within the first 12 months. After that, the acceleration of the prepayment levels starts to
slow down until right before the teaser
Figure 30. CPR over various vintages of
period ends. The difference in prepayment ARM2/28
levels can be readily observed after the 100

12th month when shorter Hybrids begin to 90

2001 2003 2005

show higher prepayment rates. The reason 80

why ARM 2/28 borrowers show higher

70

60

prepayment levels can be due to the faster

CPR %

50

housing turnover of the hybrid group. 40

After the first 12 months, the prepayment 30

20
generally stays around the same level with 10

a wave-like trend peaking around every 0

0 6 12 18 24 30 36 42 48 54 60 66 72
12 months. The seasoning pattern is AGE

illustrated in Figure 30.14

Sources: Beyondbond, Loan performance

Teaser Effect
The teaser effect is the most distinctive feature of Hybrid ARM products. We define the
term as the behavior that tends to persist right around the first reset where borrowers seek
alternatives to refinance their mortgages or simply prepay them to avoid higher interest
rates. In the following section we will describe the empirical statistics gathered to support
the teaser effect.
Approximately one to two months before the end of the teaser period, a sharp rise in
prepayments occurs. The effect is apparently larger for shorter Hybrids like ARM 2/28
since shorter Hybrids are exposed less to other prepayment factors such as refinancing

14
For the data pooling in terms of its vintage year, we usually use the loan distribution data for grouping
information. It helps to maintain the relationship while examining the relationship with macro-economic
variable for time series analysis. It however distorts the age pattern since the loans within same vantage
year could be underwritten in different months. The seasoning graph is specifically grouped by the loan’s
seasoning age to better understand the age pattern.
23
and burnout before the teaser period. The peak level is reached just about two months
after the teaser period ends. Teaser impact usually observed as a sudden jump in
prepayment levels. This spike happens whenever borrowers are able to refinance with
lower cost alternative.

Interest Only (IO) Effect

During the teaser period Figure 31. CPR over various vintages of ARM2/28
before IO period, the
prepayments of ARM 2/28 120
with or without IO track each IO=0 IO=24
other fairly well. Before the 100

end of the teaser, loans with 80

IO exhibit higher prepayment
CPR%

level than the regular ones. IO 60

borrowers are even more 40

sensitive to the payment level
since they are paying only the 20

interest portion before the 0

teaser. Once the teaser ends, 0 6 12 18 24 30 36 42 48 54 60

they will start to pay not only AGE

higher interest but also an Sources: Beyondbond, Loan performance

additional amount of
amortized principal. Their incentive to refinance is definitely higher than regular ARM
2/28 borrowers. Even worse, if they cannot find a refinancing alternative, they could face
affordability issues and increased default risk. We will address this more in the
interaction between prepayment and default section.

Refinance
The prepayment
Figure 32. Refinance: Stratification by Coupon,
incentive is measured as Fixed Rate, 2003 vintage
the difference between
the existing mortgage 60

Coupon 6.99
rate and the prevailing
Coupon 5.78
refinancing rate, which 50
Coupon 8.18
is commonly referred to Coupon 10.68

as the refinance factor. 40

As the refinancing
factor increases, the
CPR (%)

30

financial incentive to
refinance increases and 20

thus changes
prepayment behavior. 10

When the loans are

grouped by their 0

coupon rates during the 02/03 08/03 02/04 08/04 02/05 08/05
Date
02/06 08/06 02/07 08/07

teaser period, the

differences of Sources: Beyondbond, Loan performance

24
 prepayment levels are quite apparent. They behave in similar patterns but loans with
higher coupons tend to season faster due to the financial incentive to refinance while
loans with lower rates tend to be locked-in as the borrowers have secured the lower rates.

Burnout Effect
The heterogeneity of the refinancing population causes mortgagors to respond differently
to the same prepayment incentive and market refinancing rate. This phenomenon can be
filtered out as the burnout. The prepayment level usually goes up steadily with occasional
exceptions across the high financial incentive region. The major reason for this is due to
the burnout phenomena in which borrowers that have already refinanced previously and
have taken advantage of the lower rates and are less likely to refinance again without
additional financial incentives. To capture such a path-dependant attribute, our
prepayment model utilizes the remaining principal factor to capture the burnout effect in
order to reduce the chances of overestimating the overall prepayment levels.

CLTV Wealth Effect

Figure 33. Wealth: CPR over various CLTV of Figure 34. Fitted CPR over CLTV 81-90 of
ARM2/28 ARM2/28, 2004 vintage
100 70

LTV81ACT LTV81Prj
90
60
80

70 50

60
CPR%

40
50
CPR (%)

40
30

30

20 20

10
10
0
0 10 20 30 40 50 60 70
AGE 0
03/04 09/04 03/05 09/05 03/06 09/06 03/07 09/07
11- 60 61-90 > 90 Date

Sources: Beyondbond, Loan performance Sources: Beyondbond, Loan performance

As a property’s price appreciates, the LTV of a loan gradually decreases. Borrowers with
a low LTV may be able to refinance with a lower interest rate. Some borrowers may even
find themselves in an in-the-money situation where they can sell their property to make
an immediate profit. Historically, home prices continue to increase with age, and more
and more loans will fall into this “low LTV” category which has an increasing likelihood
of prepayment. We use a combination of CLTV, HPA and age to model this effect.

FICO Credit Effect:

Subprime market consists of people with limited credit history and/or impaired credit
score. The high FICO score group usually is offered more alternatives to refinance and
thus has the flexibility to choose between different products. For those people who are on
the threshold of subprime and prime market, they could be upgraded to participate in the
prime market during the course of the loan life. Thus the prepayment is an increasing
monotonic function with respect to FICO.
We can see a combined effect of FICO and CLTV on CPR. Those people who have a low
CLTV and a high FICO score can easily refinance and will have highest prepayment rate;
25
 while people who have high CLTV and low FICO score will be on the side of the
pendulum with lowest prepayment rate. Figure 36 gives a sample CPR fitting result based
on ARM 2/28, 2004 vintage pools.
Figure 35. Credit: CPR by FICO of ARM2/28 Figure 36. Credit: Fitted CPR
2000 to 2004 vintage,CLTV 70-90, DTI 35 -45 FICO 641-680 of ARM2/28, 2004 vintage
100 80
640-680ACT
640-680Prj
90
70

80
60
70

60 50
C PR %

50

CPR (%)
40
40

30 30

20
20
10

0 10

0 10 20 30 40 50 60 70 80 90
AGE
0
02/04 08/04 02/05 08/05 02/06 08/06 02/07 08/07
301-600 601-700 > 700 Date

Sources: Beyondbond, Loan performance Sources: Beyondbond, Loan performance

Prepayment Penalty
A prepayment penalty fee in Figure 37. CPR over various vintages of ARM2/28
the loan structure is no doubt 100

a negative incentive and

deters prepayment. No Prepay Penalty
2 yr Prepay Penalty 80

Prepayment is in essence an
embedded call option with
remaining balance as its strike 60
CPR (%)

price for the option. The

penalty simply adds to that 40

strike price as additional cost

when borrowers exercise the 20
option. That ad hoc additional
cost will be reduced to zero
after the penalty period. 0
0 6 12 18 24 30 36 42 48 54 60

Figure 37 shows the Age

prepayment difference when a

penalty clause is in place. Sources: Beyondbond, Loan performance
Before the 2-year penalty term,
prepayment is consistently slower than no penalty loans. Soon as the penalty period ends,
prepayments surge dramatically and surpass the no penalty loans within 3-months and
then consistently maintain a faster prepayment speed.

26
 Interaction between Prepayment and Default
As we stated in the Figure 38. CDR and CPR of ARM2/28, 2004 vintage
beginning of the model 18 80
framework, prepayment CDR
16
is a call option and PrjCDR
70

default is a put option 14 CPR

prjCPR 60
with its loan balance and 12
collateral value as its 50
10

CDR (%)

CPR (%)
strike price respectively. 40
A borrower will 8

continuously find 6
30

incentives to exercise it if 20
4
the option is in-the-
10
money. 2

When we estimate the 0

02/04 08/04 02/05 08/05 02/06 08/06 02/07 08/07
0

prepayment and default Date

for a pool of mortgages, Sources: Beyondbond, Loan performance

the remaining principal
factor encompasses the entire history of the pool’s prepayment and default rates. Since
estimating losses is the main focus for modeling default and prepayment, it is of
particular importance in a slow prepayment environment. Given the same default
probability, the tail risk to the loss curve will still increase substantially. Figure 39
presents a tail risk example. When the prepayment speeds double, the total loss increases
from 21% to 29 % given the same default speeds.
Figure 39. Loss projection of ARM2/28, 2004 vintage
% Loss (%)
40 0.8

35 CDR 0.7 CPR, 29% total loss

CPR
2xCPR, 21% total loss
30 0.6
2xCPR

25 0.5

20 0.4

15 0.3

10 0.2

5 0.1
Age Age
0 0
40 60 80 100 120 140 160 180 200 220 240 40 60 80 100 120 140 160 180 200 220 240

Sources: Beyondbond, Loan performance

Because the history of prepayment and default rates can seriously affect the remaining
principal factor for any given pool of loans, tracking and rolling the principal factor for
loan pool is one of the most important factors for the model projections and future
forecasts. Prepayments are specified prior to defaults and are removed from the
outstanding balance and are therefore not available to default in the future.

27
 DELINQUENCY STUDY
Delinquency, the leading indicator
Is delinquency a good leading indicator for default? When a borrower is late for his
payment for more than thirty days, a 30-day delinquency is reported. If the late payment
exceeds two month, a 60-day delinquency is reported. After 90-day delinquency, a loan
will start its foreclosure process depending on the judicial status of each state and is
considered to be in default. Since a default is a consequence of delinquency, the spectrum
of delinquencies should be leading indicators of future defaults. We should be able to
simply roll delinquency numbers month to month into actual defaults. The question is
whether there is a constant relationship that can be parameterized or not. The time series
plots of defaults and the spectrum of delinquencies for 2003 vintage are shown in Figure
40. The cross correlations indicate an approximately six-month period for a 30-day
delinquency manifest into default as shown in Figure 41.

Figure 40. Default and Delinquency over time for 2003 vintage

2.8 14 3.2 16
MDR MBA30 MDR DLQ30
2.4 12 2.8 14

2.4 12
2.0 10
2.0 10
1.6 8
1.6 8
1.2 6
1.2 6
0.8 4
0.8 4
0.4 2 0.4 2

0.0 0 0.0 0
2003 2004 2005 2006 2007 2003 2004 2005 2006 2007

3.6 9 2.0 2.0

3.2 MDR DLQ60 8 MDR DLQ90

2.8 7 1.6 1.6

2.4 6
1.2 1.2
2.0 5
1.6 4
0.8 0.8
1.2 3

0.8 2 0.4 0.4

0.4 1
0.0 0 0.0 0.0
2003 2004 2005 2006 2007 2003 2004 2005 2006 2007

28
Figure 41. Cross Correlations of Default and Delinquency for 2000 vintages
100%
98%
mba30 dlq30
96%
dlq60 dlq90
94%
92%
90%
Correlation (%)

88%
86%
84%
82%
80%
78%
76%
74%
72%
70%
0 1 2 3 4 5 6 7 8
Lag (month)

Analysis among delinquency spectrum

Figure 42. Correlations between various Delinquencies
mba30 dlq30 dlq60 dlq90
mba30(-1) 0.974228 0.896283 0.849914 0.819303
dlq30(-1) 0.892006 0.99476 0.989324 0.931421
dlq60(-1) 0.842606 0.980814 0.993112 0.915923
dlq90(-1) 0.8199 0.937639 0.934675 0.898144

The results among delinquency spectrum show a very significant cross correlation
between delinquency and it’s lagged earlier tenor.

A Delinquency Error Correction Default Model

Based on the results shown previously, the spectrum of various delinquencies provides a
good indication and can be parameterized for near-term projections. The benefit of
including delinquency to project defaults is that it does not require specific consumer
behavior theory to be applied. By simply looking at delinquency report, we are able to
project the likelihood of defaults. It however, suffers from the long term view that if a
loan fundamentally carries lower credit-worthy characteristics such as low CLTV it has a
propensity to default. We however are impressed with their short-term forecast ability. In
order to fully utilize the information provided by delinquency and the econometric model
based on consumer behavior theory, we have integrated both and created a delinquency
error correction model.
29
 The fundamental idea is that not only can the long-term view and various scenarios based
on changing view of macro-economic variables be adopted, but also the immediate/early
warning signs from delinquency can be observed and utilized.
Figure 43. Delinquency Error Model: Actual vs. Fitting
2001 2002
1.6 1.6
MDR: Actual MDR
1.4 PRJMDR: Projected MDR
MDRF1: Projected MDR with Error Correction
1.2 1.2

1.0

0.8 0.8

0.6

0.4 0.4
0.2

0.0 0.0
2001 2002 2003 2004 2005 2006 2007 2002 2003 2004 2005 2006 2007

MDR PRJMDR MDRF1 MDR PRJMDR MDRF1

2003 2004
2.0 2.0

1.6 1.6

1.2 1.2

0.8 0.8

0.4 0.4

0.0 0.0
2003 2004 2005 2006 2007 2004 2005 2006 2007

MDR PRJMDR MDRF1 MDR PRJMDR MDRF1

2005 2006
1.0 .5

0.8 .4

0.6 .3

0.4 .2

0.2 .1

0.0 .0
05M07 06M01 06M07 07M01 07M07 08M01 2006M07 2007M01 2007M07 2008M01

MDR PRJMDR MDRF1 MDR PRJMDR MDRF1

30
 Figure 44. Comparison of Model Explanation Power for 2000-
In our error correction 2007 Vintages
model, we start by 0.95
projecting default rate using R-Square
the default function with 0.9
fitted parameters. We then
layer on a 6-month lagged 0.85
30-day delinquency as an
additional exogenous
0.8
variable to regress the fitted
errors. The process is then
0.75
repeated with adding 5-
month lagged 60-day and 4-
month 90-day delinquency 0.7
Default Model Error Correction Model
rates as new regressors
respectively. The results are Sources: Beyondbond, Loan performance
very encouraging when
compared to the base model without error correction. The additional R2 pick-up is around
15%.

31
CONCLUSION
Why Innovate?
Traditionally, practitioners have observed consumer behavior through historical defaults
and prepayments while building an econometric model with several quantifiable factors.
These factors include seasoning patterns, underlying loan characteristics, such as
mortgage coupon, FICO score, loan-to-value, and debt-to-income ratio, and macro-
economic variables, such as prevailing mortgage rate, housing price appreciation. In
order to fit the historical data, non-linear functions are usually constructed with
parameters around the factors to explain default and/or prepayment probabilities. During
the process of historical sample fitting to the econometric model, the traditional modelers
usually miss the following:
1. Traditional models focus on fitting in-sample data with a unique parameter set by
vintage. Although the in-sample data fitting provides a much easier fit of the
parameter set, it assumes that borrower’s behavior varies given same loan
characteristics and loan age. It creates a disconnect among vintages and cannot be
simply applied to new loans.
2. Borrower behaviors underlying, LTV, FICO, and DTI were implicit, but not fully
quantified in dynamic form by traditional models. Since loan information such as
LTV, FICO, and DTI levels are not periodically updated after the loan origination
date, the accuracy of projecting performance of seasoned loans diminishes as time
passes as it is based on original loan information.
3. Out-of-sample projections may produce counter-intuitive results. Since macro-
economic variables, such as HPA, Unemployment, Personal Gross Income future,
etc. can be very important factors for in-sample fitting, they however, do not
provide insight for new scenarios. If a new scenario has not occurred historically,
a stress test for the new scenario should be thoroughly pre-examined.
4. Traditional models focus at the national level rather than drill down to local
housing markets. Since housing prices are highly dependent on its location, a
model with more detailed housing information can make a dramatic difference to
its forecast accuracy.
5. Traditional models treat prepayment and default independently, it ignores the
complexity and interaction between put and call options. For example,
prepayments slow down substantially (burnout) when the principal factor reduced.
6. Traditional models do not dynamically quantify feedback from other leading
indicators such as delinquency rates.
Because of the credit crisis, we now know we must have missed something in the
traditional models. It required us to take a hard look at the models and methodologies
employed today and see what was needed to provide a better interpretation of the data
and current conditions.

32
Innovation
Having addressed the pitfalls that traditional models fail to address, we have built a
Dynamic Econometric Loss (DEL) model framework with the following innovations:
Consistent parameter set for all vintages via the addition of consumer behavior factors.
1. Dynamic consumer behavior factors
a. CLTV ratio (via cumulative HPA since origination) which reflects
housing market wealth effects during housing boom/bust eras.
b. DTI ratio (via unemployment rate forecasts) which addresses housing
affordability.
2. Complete study of HPA index prior to model-fitting
a. HPCUM as the cumulative HPA since origination to capture wealth effect.
b. HPA to capture the pulse of the housing market.
c. HPA2D as the change of HPA to capture the trend of the housing market.
HPA2D successfully captures the timing of the defaults for 2005 to 2006
vintages.
d. In-sample and out-of-sample HPA fit testing to ensure the model’s
robustness.
3. A CBSA detailed level HPA model allows us to not only better understand local
housing markets, but also generate more precise projections.
4. Recursive calculations along seasoning paths while estimating/projecting
prepayments and defaults.
5. An error correction model that systematically builds the linkage between
delinquency and default which enhances our default forecast.

Advantages
The implementation based on our model framework will capture the loss pattern during
the recent period but can also forecast future prepayments, defaults and losses based on
various macro-economic market scenarios. The implementation advantages are:
1. Multiplicative and additive factors for each non-linear function (boot-strapping
Maximum Likelihood Estimation)
2. Comprehensive consumer behavioral economic theory applied in practice
a. Develop a consumer behavior based economic theory.
b. Estimate consumer behavior via an econometric model.
c. Apply the econometric model to prepayment and default.
3. Fully utilize HPA time-series information
a. A built-in time-series fitting model that dynamically estimates parameters
and generates forecasts on the fly. For example,
i. HPCUM ↓(below 5%) => CLTV↑ => MDR↑, SMM ↓
33
 ii. HPA ↓(below 2%) => MDR↑ , SMM ↓
iii. HPA2D ↓(below -5%) => MDR↑ , SMM ↓
4. Multiple built-in time-series fitting models at the national, state, and CBSA level
that dynamically estimate parameters and generate forecasts on the fly.
5. Built-in recursive calculator along seasoning paths for projecting prepayments
and defaults.
6. A set of error correction fitting models that estimate parameters within the
spectrum of delinquencies and defaults that are generated on the fly.

Findings
In order to understand how a loan prepays or defaults, we investigate consumer behavior
via loan characteristics utilizing static factors and relevant macro-economic variables as
dynamic factors. For each factor, we have constructed a non-linear function with respect
to magnitude of the factor. We then built the default/prepayment function as the linear
combination of all factors to justify the impact of each factor accordingly. Since a loan
can either prepay or default over time, we then continue to ensure that the principal
factors are rolled properly for prepayment and default forecasts.
During the fitting, a list of interesting findings were noted:
When the level of HPA is considered the main blessing/curse for the rise and fall of
subprime market, we find that cumulative HPA and the change of HPA contribute
additional dimensions to effect prepayment and defaults.
1. HPI is significantly correlated to DPI over a long-term period. Since DPI is a
more stable time series, it suggests that HPI will eventually adjust to coincide
with DPI growth rate.
2. Default is strongly correlated to the spectrum of delinquency rates. By applying
the fitted parameters between default and delinquency rate to an error correction
model is able to effectively improve default predictability.

Future Improvements
So far, our model provides us a set of better tools to explore consumer behavior and
various impacts due to selected macro-economic variables as dynamic factors and thus
project default and prepayment probabilities in a precise and timely manner.
Nevertheless, modeling the embedded mortgage options for default and prepayment is an
on-going learning process. While we are encouraged by our findings, there is a myriad of
new questions for us to address, with an aim to continuously improve and fine tune the
model in the future. Some example areas of further investigation are briefly described
below.

Business Cycle – Low Frequency of Credit Spread

While studying the dynamic factors in the Default Modeling section, we focused mainly
on the HPI impact on consumer behavior and introduced the DPI as another macro-
economic variable to determine the long-term growth of the economy. In the beginning of
this paper, we were wondering how a relatively small volume of loans could result in a
34
 subprime crisis that proved to be detrimental to the entire U.S. financial markets and
global financial system. We believe that the subprime crisis is merely the tipping point of
unprecedented credit market easing since early this century. During the extreme credit
ease era, yield hungry investors needed to enhance their returns through investment on
either highly leveraged securities or traditionally highly risky assets such as subprime
loans. Through rapid growth of the credit default swap in derivative markets and RMBS,
ABS, and CDOs in the securitization markets, subprime mortgage origination volume
reached record highs beginning after year 2003. The credit ease impacted not just the
subprime market. All credit based lending from credit cards to auto loans, and leverage
buy-out loans were enjoying a borrower friendly lending environment as lenders went on
a lending spree. While the credit default rates reached their historical low in last decade
and resulted in extremely tight spreads among credit products, a longer view of the
history of business cycles started to reveal warning signs of the potential downside risk.
For example, the TED Spread dramatically widened after August 2007 which was a re-
occurrence of the late eighties market environment. Over the past 20 years, traditional
calibration models only focused on shorter time frames have missed the downside “fat-
tail”. The improbable is indeed plausible. Is there a better method to mix the long-term
low frequency data with their short-term high frequency data and then provide a better
valuation model?
Figure 45. Historical TED Spread and Histogram
%
3 2.0

2.5
1.6

2
1.2
1.5

0.8
1

0.4
0.5

0 0.0
85 87 89 91 93 95 97 99 01 03 05 07 0.0 0.4 0.8 1.2 1.6 2.0 2.4
SPREAD_TB3M

Source: Beyondbond, Inc.

Dynamic Loss Severity

Traditionally, prepayment and default modeling is the main focus of the fundamental
research for mortgages while loss severity and timing lags for loss recovery is simply run
as a given. The detailed HPA information provided at the CBSA level and better detailed
information provided by Servicers in recent years has allowed us to create a more robust
dynamic loss severity estimation and we will continue work with Servicers to develop
improved estimates in the future.

35
APPENDIX I DEFAULT AND PREPAYMENT DEFINITION
We consider a loan to be in default if it meets both of the following criteria:

1) The loan is not able to generate any future investor cashflow

2) The loan has been in foreclosure, REO or reporting loss in prior reporting period

The Monthly Default Rate (MDR) is defined as the percentage of defaulted amount as a
sum of all default loan balance compared with the aggregate loan balance of that period.
SMM (Single Month Mortality) is calculated by formula:
Scheduled Balance - Current Balance
SMM =
Scheduled Balance

If we have MDR and SMM, then we can simply derive CDR and CPR from them by
using the formula:
CDR = 1 − (1 − MDR)12
CPR = 1 − (1 − SMM )12

36
APPENDIX II GENERAL MODEL FRAMEWORK
(s) K
y t
= ∑ϕ (X k
(k )
t
|α m , β
(k ) (k )

m
; m ∈ [ 0, M
(k )
]) ⋅
k =0

∏ λ (X |α m , β
(i ) (i) (i ) (i)
i t
; m ∈ [0, M ]) ⋅
m
i=0

∏η ( X |α m , β
( j) ( j) ( j) ( j)
t ,m
; m ∈ [ 0, M ]))
j m
j=0

∑ϕ (X |α m , β
(k )
= k
(k )
t
(k )
; m ∈ [ 0, M
(k )
]) ⋅
m
k =0

∏ λ (X |α m , β
(i ) (i) (i ) (i)
i t
; m ∈ [0, M ]) ⋅
m
i=0
( j)
J
M
∏ ((1 + ∑ X β ) |α m , β
( j) ( j) ( j) ( j) ( j)
t ,m
; m ∈ [0, M ]))
m m m
j =0

Where
(s)
y t
is an observable value at time t for dependent variable type s

ϕ k is a spline interpolation function with pair-wise (α m , β m ) knots

(i ) (i )

(k )
X t
is an observable value of factor k at time t

K is the number of additive spline functions

λ is a spline interpolation function with pair-wise (α m , β ) knots
(k ) (k )
i m

(i )
X t
is an observable value of factor i at time t

I is the number of multiplicative spline functions

( j)
M
η X β
( j) ( j)
is equal to (1 + ∑ t ,m
) and is a linear combination function with
j m m

multiplier β
( j) ( j) ( j)
m
of X t ,m
; where X t ,m
is an observable value of the type m factor

at time t, while β
( j)
is the composition ratio of the distinct factor j of type m
m

J is number of linear functions

37
APPENDIX III – DEFAULT SPECIFICATION
A whole loan mortgage starts at t0 and matures by tn, its MDR by time t can be driven by
two type of variables – static and dynamic.
Collateral characteristics such as mortgage rate, loan size, IO period, teaser period, loan
structure, term to maturity, geographic location, FICO, and CLTV are static factors since
their impacts diminish over time while the loan is getting seasoned.
The macro-economic variables over time such as Housing Price Index, mortgage interest
rate, unemployment rates, Gross Disposable Income, and inflation rates are dynamic.
They are publicly observable and will tune forecast of the default rate based on the
scenario assumption.
We formulate our default function MDR as follows:
Dt = ϕ LTV (vt LTV j ,ht ) + ϕ FICO (c j ) x

λrate (rt | WACt ) x λage (ai | a0 ) x λDTI (d j | DTI j , DOC j ) x

λIO (g t | IOj , ai ) x λsize (s ) x λHPA (HPA) x λH 2 D (H 2 D ) x

η DOC (Docm ) xη LIEN (LIEN m ) xη PURPOSE (PURPOSEm ) x

Where
φ’s are spline functions in MDR % and are additive to form a base value
λ’s are spline functions as multipliers for the MDR adjustments
vt : CLTV by time t where initial CLTV is assumed at time t0

rt : Ratio spread of WACt over original WAC rate

cj : FICO score of loan j
ai : Age of loan j
dt : DTI
gi : Remaining IO period if IO exists and is positive
lj : Size of loan j

ϕLTV : Original LTV level & HPA t

vt = vt (v0, ht , z j )

H t i : HPI at time ti since origination date t0

z t : Geographic zip code j, e.g. z1 = z(CA ) = 1.3

38
 z 2 = z(OH ) = 1.1
z 3 = z(MI ) = 1.01

z 0 = z(Other ) = 1

the function form of v t

v0 ⋅ H t (i−lag )
vt = . zj
H t ( 0−lag )

h t : the functional form of h t as simple AR(2) model

h t = β0h + β1h h t −1 + βh2 h t − 2 + ε t

Where all the parameters can be independently regressed by h t ’s time series data

z j : the functional form of z j is setup as a dummy variables

z j = βzj ∗ z ( j) if j = “CA” and parameter βzj can be calibrated by default data by

bootstrapping the value

f t : is the actual principal factor and will be either observed for in-sample filtering or
simulated for out-of-sample forecast

FICO: to check if credit scores (original) is a good measure of default

c j : the functional form of c j will be a spline (natural, Linear, tension spline) function with

fixed FICO locators, j’s (suggested only)

[250, 350, 450, 500, 525, 550, 550, 580,
600, 625, 650, 680, 700, 720, 750, 800, 820]
and parameters can be calibrated for Default data base & fine-tuned

AGE: Default probability increases as loan get seasoned but eventually reach to a plateau
given others constant
a t : we will sample linear spline function from 0 to 1 to apply age locators
[0, 1, 5, 10, 15, 20, 30, 45, 60, 120]

39
DTI Effect: Income level will affect default under assumption of DOC if it’s fully
available
β ( UM )
GDPt  UM t 
u t = u0 ⋅  
GDP0  UM 0 
the functional form
λ u (u t ) is a linear spline function of u t

λ DTI (u t , w j ) = (λ u (u t ))
( )
λw w j

where
λ w (w 0 ) = 1 → Full = w 0

λ w (w1 ) = 0.1 → Low = w1

λ w (w 2 ) = 0 → No = w 2

RATE Effect
rt = (WACt − MTG t )

ϕrate (rt ) is a spline function of rt

• WACt is gross coupon which is either observable or can be simulated from index
rates & loan characteristic
• Index rates forecasting will be a spread

y t 's = β0 + β0 y t −1 + β1Swp2Yt + β2Swp5Yt + β3Swp10Yt + β4 LIBOR1M t + ε t

for corresponding index rate LIBOR6M, 1Y-CMT, COFI, 5YY – CMT, …etc.

IO-Payment-Shock Surprised payment increase will increase default

g t = IO0 − a t

λ IO (g t ) = is a Linear Spline function of locators [-30, -20, -10, -5, -2, 0, 2, 0, 2, 5, 10, 20]

40
Crowding Out Measures if the underwriting standard is deteriorated

λ volume is a Spline function

vm t is whole loan issue amount ratio (FICO ≤ 580, 580 < FICO ≤ 700)
*Note: 30-day Delinquency rate for the (12-month) ratio if delinquency report is
available

λ size is a simple step-spline function to if certain loan size after default with locators
[ ≤ 50k, ≤ 100k, ≤ 150k, ≤ 250k, 500k, 800k, 1million]

Occupancy

λocp has 3 kinds of occupancy (Owner, Second Home, Investor,)

Loan Purpose

λ prs has 3 kinds of purpose (Purchase, Refi, Cash Out)

Lien

λ lien has 2 lien positions (First lien, Second lien)

Loan Document

λdoc has 3 kinds of documentation type (Full, Limit, and No Document)

41
APPENDIX IV – PREPAYMENT SPECIFICATION
Single Monthly Mortality (SMM) Rate Function
St = ϕrate (rt ) x
λturnoverrate ( ) x
λteaser (tst ) x
λseasonality ( ) x
λcash − out ( ) x
λage (at ) x

λburnout ( f t ) x
λ yieldcurve ( ) x
λequity ( ) x
λcredit ( )x

λ IO (gt ) x
λcredit (Vt ) x
λissuer ( IY j 's ) x
λ size (l j 's )x

λ penality ( N yes / no )

Housing Turnover Rate

Prepayment based on long-term housing turn-over rate that is composed of existing sales
over single-family owner’s housing stock.

Seasonality

Monthly seasonality is generally believed to affect prepayments. The belief stems from
the mobility of mortgagors, time of housing construction, school year, and weather
considerations. For a specific month of the year and ceteris paribus, prepayment rates are
directly affected by the related month-of-year’s coefficient. Usually, the seasonality
pattern tends to be more active in the spring, rises to peak in the summer, decreases

42
through the fall, and slows down even more in the winter. The pattern may be different
geographically and demographically.

Cash-out

Prepayment is driven by general housing price appreciation.

Rate Factor ϕrate (rt ) (to grab REFI-incentive)

ϕrate : a natural spline function

20 locators [-10, -5, -2, -1, 0, 0.5, 1, 1.5, 2, 2.5, 3, 3.5, 4, 5, 6, 7, 9, 10, 15, 20]

WAC − m t (Fixed)
rt = 
WACD − m t (ARM / Hybrid)

m t : FH 30-yr/10day commitment rate (FHR3010) as prevailing mortgage rate to measure

SATO effect

*Age Factor: PPY has less incentive due to the consideration of initial financing sunk
cost. But the probability increase change time as the 3-yr coast get average out along time.

Age

Mortgages generally display an age pattern.

Burnout Effect

Borrowers don’t behave homogeneously while refinancing opportunities appear.

Some are more sensitive than others. If the borrowers are heterogeneous with respect to
refinancing incentive, the more interest sensitive group will refinance.

43
REFERENCES

Wei Wang, “Loss Severity Measurement and Analysis”, The MarketPulse,

LoanPerformance, 2006 Issue 1, 2 – 19.

Steven Bergantino, Gyan Sinha, “Special Report: Subprime Model Update”, Bear Stearns
& Co. Inc. May 26, 2005.

Rod Dubitsky, Jay Guo, Larry Yang, Rajat Bhu, Sergei Ivanov, “Subprime Prepayment,
Default and Severity Models”, Credit Suisse, Fixed Income Research, May 17, 2006.

Lakhbir S. Hayre, Manish Saraf, “A Loss Severity Model for Residential Mortgages”,
Citigroup Global Markets, Inc., U.S. Fixed Income Strategy & Analysis - Mortgages,
January 22, 2008.

Lakhbir S. Hayre, Manish Saraf, Robert Young, Jaikai (David) Chen, “Modeling of
Mortgage Defaults”, Citigroup Global Markets, Inc., U.S. Fixed Income Strategy &
Analysis – Mortgages, January 22, 2008.

Elizabeth Laderman, “Subprime Mortgage Lending and the Capital Markets”, Federal
Reserve Bank of San Francisco, FRBSF Economic Letter (Number 2001-38), December
28, 2001.

Richard Ramsden, Lori b. Appelbaum, Roy Ramos, Louise Pitt, “The subprime issue: A
global assessment of losses, contagion and strategic implications”, Goldman Sachs Group,
Inc. – Global: Banks, November 20, 2007.

Chris Flanagan, “Subprime Mortgage Prepayment and Credit Modeling”, J.P. Morgan
Chase & Co., April 2, 2008.

Nera Economic Consulting, “At a Glance: The Chilling Effects of the Subprime
Meltdown”, Marsh & McLennan Companies, September 2007.

C.H. Ted Hong, “Modeling Fixed Rate MBS Prepayments”, Beyondbond Inc, October,
2005.

44
C.H. Ted Hong, Mike Chang, “Non-agency Hybrid ARM Prepayment Model”,
Beyondbond Inc, July, 2006, 6 – 17.

Rahul Parulekar, Udairam Bishnoi, Tanuj Gang, “ABS & Mortgage Credit Strategy Cross
Sector Relative Value Snapshot”, Citigroup Global Markets Inc., June 13, 2008.

Stefano Risa, “The New Lehman HEL OAS Model”, Lehman Brothers Inc., December 8,
2004.

Akhil Mago, “Subprime MBS: Grappling with Credit Uncertainties”, Lehman Brothers
Inc., March 22, 2007.

Christopher L. Peterson, “Subprime Mortgage Market Turmoil: Examining the Role of

Securitization – A hearing before the U.S. Senate Committee on Banking, Housing, and
Urban Affairs Subcommittee on Securities, Insurance, and Investment”, University of
Florida, April 17. 2007.

Joseph R. Mason, Joshua Rosner, “Where Did the Risk Go? How Misapplied Bond
Ratings Cause Mortgage Backed Securities and Collateralized Debt Obligation Market
Disruptions”, Mortgage-Backed Security Ratings, May 3, 2007.

45
Disclaimer

This publication contains material that is: (i) for your private information, and we are not
soliciting any action based upon it; (ii) not to be construed as a prospectus or offering
materials of any kind; and (iii) is based upon information that we consider reliable, but
we do not represent that it is accurate or complete, and it should not be relied upon as
such. Opinions, forecasts, prices, yields, and other forward looking statements may be
based on assumptions which may or may not be accurate, and any such opinions,
forecasts or other information are subject to risks and uncertainties and may differ from
actual results. Beyondbond cautions that forward-looking statements are subject to
numerous assumptions, risks and uncertainties, which change over time. Information
provided is current as of the date (s) of issuance and is subject to change without notice.
Actual results may differ materially from those anticipated in forward-looking statements
and future results could differ materially from historical performance. While we
endeavor to update on a reasonable basis the information discussed in this material, there
may be regulatory, compliance, or other reasons to prevent us from doing so. Regarding
the companies or entities mentioned herein, Beyondbond, its affiliates, officers, directors
and employees (including persons involved in the preparation of this material) may, prior
to or concurrent with this publication: (i) have long or short positions in, and/or buy or
sell their securities, or derivatives (including options) thereof; and/or (ii) effect or have
effected transactions contrary to Beyondbond’s views contained herein. The securities
described herein may not have been registered under the Securities Act of 1933, and in
such case, may not be offered or sold within the United States or to US persons unless
they are being sold in compliance with an exemption from the registration requirements
of such Act. The provision of this research by Beyondbond and its affiliates does not
constitute investment advice, and you should not rely on it as such. Neither Beyondbond
nor any of its affiliates makes any representations or warranties with respect to any
securities or investments. You are responsible for exercising your own judgment (either
independently or through your investment advisor) and conducting your own due
diligence with respect to investments and their risks and suitability (including reading any
relevant final prospectus). Beyondbond and its affiliates are not responsible for any losses
that you may incur as a result of your investment decisions, whether direct, indirect,
incidental or consequential. No part of this material may be (1) copied, photographed, or
duplicated in any form, by any means, or (2) redistributed to anyone (including your
foreign affiliates) without Beyondbond’s prior written consent. Derivatives and options
are not suitable investments for all investors. Additional information may be provided
upon request.

46