A method for mortgage and closed end loan portfolio management
in the form of an analytic tool designed to improve analysis of
past and future performance of loan portfolios. In accordance with
one aspect thereof, the invention aggregates loan units into loan
vintages, wherein the loans in each vintage originate within a predetermined
time interval of one another. The invention compares different vintages
to one another in a manner such that the ages of the loans in the
different vintages are comparable to one another. An early warning
component of the system predicts delinquency rates expected for
a portfolio of loans during a forward looking time window.
1. A computer implemented process for predicting the performance
of a loan portfolio, wherein each loan portfolio comprises a plurality
of loan units, each of the loan units having a borrower, the computer
implemented process comprising: separating the loan units of the
loan portfolio into a plurality of loan groups such that a date
of origination of each of the loan units included in a loan group
are all within a first time interval; assigning a group date of
origination to each of the loan groups; selecting an analysis time
interval; selecting a plurality of analysis points in time within
the analysis time interval; determining historical bad performance
by identifying the loan units in each loan group that have experienced
bad performance at each analysis point in time, a bad performance
is determined if payments are in arrears at the particular analysis
point in time; obtaining a credit score for each of the borrowers
of the loan units; and determining a projected bad rate for each
loan unit by combining the historical bad performance for each loan
unit and the credit score of the borrower of the at least one loan
2. The process of claim 1, in which the step of combining the historical
bad performance and the credit scores is carried out by using a
logistic regression formula.
3. The process of claim 2, in which the logistic regression formula
is Log(P/(1-P))=A+(B1*AGE)+B2*C+B3*D1+B4*D2+B5*SCORE+B6*NO SCORE)
wherein P is a projected bad rate for a particular loan unit, AGE
is the age of the loan group to which the particular loan unit belongs,
SCORE is the credit score for the borrower of the particular loan
unit, C0, D1, D2 and NO SCORE are dummy variables, and A, B1, B2,
B3, B4, B5 and B6 are estimated coefficients.
4. The process of claim 3, wherein the value C0 is assigned a value
of one if the particular loan unit is current at the beginning of
the first time interval and zero otherwise, the variable D1 equals
one if the loan unit is one month past due at the beginning of first
time interval and is zero otherwise, D2 is assigned a value of one
if the loan unit is two months past due at the beginning of the
first time interval and zero otherwise, and NO SCORE equals one
if the loan unit has no credit score available at the beginning
of the first time interval and zero otherwise.
5. The process of claim 3, further comprising: applying the logistic
regression formula to each loan unit in a particular loan group;
and aggregating the results of the logistic regression formula into
a projected bad rate for the particular loan group.
6. The process of claim 5, further comprising using the projected
bad rate for the particular loan group to decide whether to purchase
the particular loan group.
7. The process of claim 5, further comprising using the projected
bad rate for the particular loan group to decide whether to include
the particular loan group in the portfolio.
8. The process of claim 5, further comprising using the projected
bad rate for the particular loan group to decide servicing rights
with respect to the particular loan group.
9. The process of claim 5, further comprising: repeating the applying
and aggregating steps for each of the plurality of loan groups in
the portfolio and aggregating the projected bad rates for the plurality
of loan groups into a projected bad rate for the portfolio.
10. The process of claim 5, further comprising using the projected
bad rate for the portfolio to decide whether to purchase the portfolio.
11. The process of claim 5, further comprising using the projected
bad rate for the portfolio to decide servicing rights with respect
to the portfolio.
12. The process of claim 3, further comprising: defining different
loan groups; applying the logistic regression formula to each loan
unit in the different loan groups; and aggregating the results of
the logistic regression formula into predicted bad rates for the
different loan groups.
13. The process of claim 12, wherein the step of defining different
loan groups further comprises separating the loan units of the loan
portfolio into a plurality of different loan groups such that the
loan units included in a particular different loan group share a
common loan characteristic.
14. The process of claim 13, wherein the common loan characteristic
is the location of the origination of the loan units.
15. The process of claim 13, wherein the common loan characteristic
is the rate type of the loan units.
16. The process of claim 13, wherein the common loan characteristic
is the loan to value ratio of the loan units.
17. The process of claim 1, further comprising the step of graphically
depicting the projected bad rates in the form of a first curve.
18. The process of claim 17, further comprising the step of producing
a bar chart showing current mean bad rates and forecasted mean bad
rates and superimposing the first curve over the bar chart.
19. The process of claim 18, further comprising the step of creating
the first curve by creating a quarterly bad rate curve, smoothing
the quarterly bad rate curve by averaging the values thereof with
one another and further smoothing the quarterly bad rate curve by
taking a risk ratio thereof.
20. The process of claim 19, further comprising the step of creating
markers on the first curve including markers which show the changes
in the first curve at a positive to a negative slope transition
thereof and markers which show jump points of a predetermined size.
21. The process of claim 1, wherein the first time interval is
a calendar quarter.
22. The process of claim 1, wherein the analysis time interval
is a two year interval.
23. The process of claim 1, wherein the step of obtaining the credit
score further comprises obtaining the credit score from credit bureau
CROSS REFERENCE TO RELATED APPLICATIONS
This application is based on U.S. patent application Ser. No. 08/893,389,
filed on Jul. 11, 1997, entitled METHOD FOR MORTGAGE AND CLOSED
END LOAN PORTFOLIO MANAGEMENT, the entire disclosure of which is
hereby incorporated by reference.
BACKGROUND OF THE INVENTION
The present invention relates to banking and, more particularly,
to a loan performance analytic tool designed to improve analysis
of past and future performance of loan portfolios.
Financial institutions such as banks own large portfolios of mortgage
and other closed end loan instruments. Further, there is a constant
influx of applications for new loans and mortgages and, moreover,
existing loans are treated by banks as commodities or products which
they trade among themselves. Banks underwrite loans and/or purchase
loan portfolios of other banks or sell portions of their own loan
portfolios. In doing so, banks customarily continually assess and
reassess the quality of various loan portfolios, which quality depends
on the interest rates earned on those loans, the customer payment
history on the loans and other criteria.
As regards newly originated loans, the process begins with loan
applicants submitting applications to financial institutions which
then triggers an investigation by either the bank and/or related
service organizations which check the credit history of the applicant
before the loan is approved. Typically, the decision to grant or
not grant a loan implicates various credit screens that examine
such factors as the loan to value ratio (LTV) of the particular
application or the debt to income ratio (D/I) of the applicant and
other historical facts, which shed light on the commercial worthiness
of the given loan transaction. Once a loan is granted, it becomes
part of an aforementioned vast portfolio of loans which a given
financial institution owns and/or services. The "quality"
of the particular loan heavily depends on the interest fees earned
by the financial institution on each loan and on the performance
of the loan which is dependent on the timeliness of the payments
by the loan applicant and/or on loan prepayment.
Loan portfolios represent to banks two separate and distinct lines
of business or sources of income. One business line or source of
income flows from the ownership of the loans and the earning of
interest fees thereon. The second line of business involves the
servicing of the loan, for example, the keeping of records, collection
of periodic payments, enforcement in the form of loan foreclosures,
etc. Banks can earn fees on servicing of loans which they either
own outright or which they service on behalf of other financial
institutions. This is because it is traditional in the banking industry
to attribute to each loan a basic cost of servicing which is included
in the interest fees charged to the customer. If a bank is able
to carry out or perform these servicing tasks at a cost structure
which is below the originally attributed servicing cost, the bank
is able to realize a profit from its loan servicing business.
It is not uncommon for large financial institutions to immediately
turn around and sell to other investors portions of the loans that
they have booked, to spread the credit risks and in order to diversify
the types of loan instruments that they are holding. The same is
true of the services end of the business with respect to which decisions
are constantly made as to whether retain or sell the servicing components
of various groups of loans.
The loans that are retained for servicing are assigned to a subsidiary
of the financial institution which is a purely service organization
that has developed the methodology and procedures for servicing
loans. A portion of the loan portfolio can be sold to third party
loan servicing bureaus. It is common for banks which sell loans
to retain ownership of the servicing rights to earn the fee income
thereon. In addition, many financial institutions may decide to
purchase servicing rights from other financial institutions.
In any case, bank managers are responsible for managing loans totalling
billions of dollars both as pure loan instruments and as products
that require servicing. The decisions whether to retain different
groups of loans or whether to sell them off to other investors and,
on the other hand, whether to purchase loan portfolios from other
institution for ownership or servicing purposes are bottom line
decisions that have the potential to affect the financial institution's
profits and/or losses to the tune of tens or even hundreds of millions
of dollars. Hence, loan portfolios are constantly examined by bank
managers very carefully since different vintages of loans can perform
quite differently from one another.
For example, a portfolio of loans representing mortgages granted
in a particular locality during a particular time frame might be
deemed to represent high quality loan instruments, as for example
in the situation where the history of these groups of loans has
shown that the rate of default for that group of loans has been
extremely low and the interest rate on those loans is high compared
to present interest rates. Conversely, another portfolio of loans
granted in another region of the country which may have suffered
economic decline may result at some future date in large rates of
default. Assuming further that these loans were issued at a low
interest rate, it is not difficult to understand that the particular
"product"--the portfolio of loans--would be deemed to
possess low value and be a good candidate for being sold. Alternatively,
a shrewd bank manager might see future value in a presently poorly
performing loan portfolio and seek to buy at its current low price
structure for its potential improvement. In the same vein, the "servicing"
of such loans may be more difficult and expensive due to higher
default instances. A bank might wish to sell off the ownership component
of such a loan portfolio, or the servicing rights thereof, or both.
Sometimes, however, a financial institution which has a "servicing"
subsidiary that is being underutilized may be willing to accept
loan portfolios of servicing rights considered unattractive by other
In the prior art, bank managers entrusted with making the aforementioned
decisions have often resorted to and relied on manual research and
their intuition in their attempts to predict, manage and select
loan portfolios for ownership and servicing purposes. The prior
art approach has failed to provide a straightforward and easy to
comprehend and administer system for assessing the past performance
and future likely course of loan instruments.
SUMMARY OF THE INVENTION
Accordingly, it is an object of the present invention to provide
a system and method which improves the understanding of the past
performance of loan portfolios.
It is another object of the invention to provide a system which
enhances the ability of financial institution managers to choose
which mortgage and other debt instrument applications to underwrite.
Yet another object of the present invention is to provide a system
and method which enhances the ability of financial institution personnel
to make decisions whether to retain or dispose of different groups
It is also an object of the present invention to provide a system
and a method which is able to dynamically and automatically evolve
loan underwriting criteria.
It is yet another object of the present invention to provide a
dynamic underwriting model which is capable of being implemented
in a general purpose computer.
It is also a further object of the present invention to provide
a system and methodology which enables automatic processing of loan
applications through a system that feedbacks information from a
dynamic processor and which allows loan acceptance decisions to
be made automatically and rapidly.
The foregoing and other objects of the invention are realized by
a system and process which is tailored to analyze and select loan
portfolios for either continued or future investment by a financial
institution. Each loan portfolio comprises a plurality of loan units
and the system operates by separating the loan portfolios into a
plurality of loan vintages, in a manner such that the loans included
in each loan vintage have origination dates that are on average
of the same age. The system of the invention produces an analysis
of the past performance of loan portfolios, as well as an indication
of the future performance thereof in two different formats.
As to past performance, the invention develops the loan vintages
in a manner such that vintages of different years can be compared
to one another meaningfully because the loan units in each of the
different vintages are actually of the same comparative ages. For
example, when 1993 and 1994 loan vintages are compared, the loans
units that are being compared are of the same age to provide more
meaningful comparisons. This is referred to in the ensuing description
as the Crus Classes analysis system. In one embodiment of the Crus
Classes system, output results are graphically depicted by means
of a curve which represents the difference between the delinquency
rates of loans in the two yearly vintages. To improve the reliability
of the results, an area of uncertainty is superimposed over the
difference to allow users to focus their analysis on those locations
on the difference plot which lies outside the area of uncertainty.
This increases the reliability of the analysis and the ability to
trust its results. The area of uncertainty can be calculated as
a +1 and -1 standard deviation, but the actual size thereof is a
matter of personal choice.
The early warning system (EWS) constituent of the invention is
one of the systems and processes which predicts the percentage of
the loans in a given loan vintage which are likely to enter a "bad"
state within a predefined forward looking time window, for example,
the next two years. The prediction is calculated by using a logistic
regression formula which has been developed in part on the basis
of the analytic results obtained from the Crus Classes analysis
component of the invention.
Finally, the so-called matrix link component of the present invention
is generally similar to the aforementioned early warning system
in that it is a prediction tool. It differs from the early warning
system in the respect that it is capable of forecasting the percentage
of loans that are likely to be bad at a date certain within the
aforementioned forward looking time window. In all cases, the results
of the analysis can be graphically depicted by comparing vintages
to one another, using various curves, bar charts and the like in
a manner described herein. For imposing the integrity of the results
it is desirable that the number of loan units in the analysis be
large, preferably in the hundreds of thousands of loan units and
preferably at least 50,000 loan units.
As a general note and definition applicable to throughout the present
specification and claims, the term "loan portfolios" means,
includes and/or refers to booked loans, applications for loans for
which underwriting decisions have to be made and the aforementioned
loan servicing rights.
Other features and advantages of the present invention will become
apparent from the following description of the invention which refers
to the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is an overall block diagram of a dynamic underwriting system
and method in accordance with the present invention.
FIG. 1A is an explanatory chart which shows an example of the delinquency
rates of various loans by vintage year for a given loan portfolio.
FIG. 1B is a tree chart and exemplar of an actual family of different
types of loan groups and shows the rate of delinquency associated
FIG. 2 illustrates a prior art method for assessing the profitability
and performance of a portfolio on a yearly vintage basis.
FIG. 3 illustrates a novel method of assessing the past performance
of loan portfolios based on yearly vintages in accordance with the
FIG. 4 is a graphic that illustrates a method of the present invention
involving assessing the relative value of different vintage loan
FIG. 4A is a table which illustrates calculations performed to
obtain data for the graphic of FIG. 4.
FIG. 5 illustrates a further graphical method of the present invention
for showing both the past performance and the future predicted performance
of loan portfolios.
FIG. 6 is an explanatory graph provided for explaining how a portion
of the graph of FIG. 5 is obtained.
FIG. 7 illustrates the format of a roll rate delinquency table
for one year which is used in the matrix link analysis module of
the present invention.
FIG. 8 is a table which shows the methodology and equations used
in forecasting bad rate probabilities in the matrix link component
of the present invention.
FIG. 9 is a plot of the final results obtained with the matrix
link component of the present invention.
FIG. 10 is a hardware/software block diagram of key components
of the present invention.
DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION
By way of background and introduction, the general environment
for the method and system of the present invention can be better
appreciated by initial reference to FIG. 1. As illustrated therein,
home buyers and refinanciers 12 typically submit applications for
loans to one or more financial institutions 14. These institutions
include loan granting departments that decide whether or not to
book given loans by applying various credit screens, i.e. criteria.
One screen may focus on the applicable LTV (loan to value) of a
transaction, the D/I (debt to income) ratio of the involved transaction
and/or on the credit history of the particular applicant.
Based on the aforementioned and other criteria, a decision is made
to accept or reject a particular loan application. Each loan that
has been accepted is added as another loan unit to a large portfolio
of similar families of loans, e.g. conforming loans, jumbo loans,
government loans, etc. A loan has typically a loan start date and
a date by which the loan is expected to be fully paid up, as is
typical of home mortgage loans. A loan that is issued for a fixed
amount and period of time is known in the trade as a closed loan.
These closed loans 16 are artificially split and treated as two
business securities or entities--namely as a "loan" entity
and as a "servicing" right, as indicated at 32.
Each loan unit or instrument represents to the financial institution
an opportunity to earn a profit on the differential between its
cost of money and the amount of interest earned from the borrower.
Another profit component is realizable from the servicing element
of each loan entity. That is, a finite budget for labor and equipment
use must be allocated when the loan is issued to service each loan
over its life time. The banking trade has traditionally derived
substantial revenues from the servicing of loan portfolios, to the
extent that they were able to service loans at a cost below the
originally calculated service allocation. Consequently, banks and
other financial institutions sometimes trade loan "servicing"
contracts. These contracts are routinely purchased and sold in large
units since they represent income opportunities. For example, a
bank which lacks a servicing department might contract with another
bank to service its loans at a set, per loan pricing arrangement.
The bank that purchases the contract does so with the expectation
of earning a profit on the project. If it develops later that a
particular loan portfolio experiences a large rate of defaults,
the extra servicing needed to collect funds on the loans might render
the particular servicing contract unprofitable. In such a situation,
the service organization might attempt to resell the service contract
to another service organization which might be interested in it,
for example, at an increased service rate.
With further reference to FIG. 1, block 18 represents the department
of the financial institution which makes the decision whether to
retain or sell a particular loan portfolio. Typically these loans
are sold in very large blocks, each containing thousands of individual
loan units. Those loan units originating at block 12 which are retained
by the given financial institution are represented by block 20.
On the other hand, as indicated by the block 22, a portion of the
book of loans is sometimes sold off to investors and is securitized.
Therefore, it will be appreciated that selling and purchasing loan
portfolios requires careful examination of various loan product
lines to assess their viability, profitability and related factors.
As already noted, another source of profit flows from the servicing
portion of the loans. Block 24 identifies the step which decides
whether to retain or sell the servicing component of a loan portfolio.
Those loans for which servicing is retained are serviced at the
bank which originated the loans as indicated at 26. The servicing
of the balance of the loans procured at block 14 is contracted out
to third parties for services as indicated at block 28. In addition,
the servicing end 26 of the banking business is also able to purchase
the servicing rights as indicated at 32.
As described, the banking industry distinguishes between ownership
of loans and the servicing thereof. Loans that are owned by a given
financial institution can be serviced by that institution's own
servicing subsidiary or the servicing part can be contracted to
third party servicing bureaus. Indeed, not all financial institution
have loan servicing departments. Conversely, a bank with a servicing
organization can purchase the "servicing" component associated
with loans owned by other banks and render the servicing thereon.
In any case, it is self-evident that the profits from earning interest
on loan portfolios and from the loan servicing line of business
is heavily influenced by the performance of various loan groups
vis-a-vis the default rate of these loans over the life of the loans,
foreclosures, collection efforts, loan prepayment and the like.
Loan portfolios which experience low default rates are easy to service
and are highly profitable to financial institutions.
Traditionally, the decision to purchase, retain, sell or create
loan portfolios demands critical analysis of the past performance
of the loan portfolios under consideration. Moreover, such decisions
invariably implicate assumptions and predictions as to how such
loan portfolios will perform in the future. Not surprisingly, the
decisions to book loans at block 14 typically depended on and required
analysis and consideration by highly skilled and experienced persons
having very keen and sharp analytical powers to determine the potential
profitability of loan portfolios being considered.
The present invention departs from the prior art by providing a
dynamic underwriting method and system 30 comprising several key
components including an early warning system 32, a Crus Classes
analysis section 34 and a matrix link 36, all to be described further
on. Essentially, the information obtained from the subsystems 32,
34 and 36 is designed to be applied, via feedback line 38, to the
decision box 14 in a manner which systemizes and provides a standardized
approach to forming the decisions whether to book loans. The invention
substantially increases the reliability, consistency and speed of
the loan acceptance decision process. Further, the dynamic underwriting
system 30 of the present invention can also be applied via feedback
line 40 to the decisional box 32 which addresses the decisions at
block 32 whether to purchase loan servicing rights of loans owned
by other financial institutions. Finally, the feedback line 41 provides
feedback for forming the decisions identified in blocks 18 and 24.
The invention shall now be described with respect to the subcomponents
of the invention, including the aforementioned subsystems 32, 34
and 36, beginning first with the Crus Classes component 34.
Experience has shown that the past performance of a group of loans
is often a key indicator of its future behavior. Therefore, the
first step in the analysis process focuses on providing an improved
analytical tool for the examination of the loans' past performance.
This is the function provided by the Crus Classes system 34. The
Crus Classes subsystem 34 essentially represents a fresh approach
to the analysis of the prior performance of already booked groups
of loans. The early warning system 32 is a forward looking system
which comprises a method and process that is able to predict what
portion of the overall number of loans in a particular loan group
will experience 90+ day tardiness in payments by the borrower(s)
thereof at anytime within a predefined time period, for example,
the next two years. By way of example, if the system predicts that
thirty loans out of a thousand in a given loan portfolio will experience
90+ day delay in payment at any time within the next two years,
the EWS (Early Warning System) 32 will return the value 0.03 to
indicate that it expects 3% of the loans in the given group to go
"bad" at least once during the predefined time period.
Finally, the matrix link component 36 provides a more sophisticated
analytical model, in that it not only assigns a probability to how
many loans will enter a 90-day default but, moreover, calculates
the expected number of loans in default on a specific future date.
Turning first to the Crus Classes system 34, a common technique
used in credit risk management in the mortgage industry is to group
loans by common intervals of origination, e.g. annually, to compare
their performances. For example, the mortgage industry might typically
wish to analyze the performance of 1994 vintage loans. Vintage in
this context means all loans that have been originated in 1994.
The classification of loans into yearly vintages by the prior art
has often resulted in significant distortions of analytical conclusions.
Unlike wines for which classification into yearly vintages makes
sense, lumping all loans originating in the same year into a same
"vintage" distorts results because there are several exogenous
factors which affect how these loans perform and these factors intrinsically
vary over time in a manner which can produce significant quarterly,
and even monthly loan performance fluctuations.
The present inventors have opted to use the term "Crus Classes"
for its similarity to the wine industry. But Crus Classes, as used
herein, differs from and departs from the prior art approach of
grouping loans by annual origination dates. The invention overcomes
some of the statistical inaccuracies associated with the prior art's
attempts to lump loans into yearly vintages solely on the basis
of the origination of a loan in a given year.
Traditional vintage techniques in the mortgage industry allow bankers
to gauge the quality of mortgages as they are "aging".
However, the inventors have added certain statistical procedures,
such as hypothesis testing, used in the process control manufacturing
environment, that allow the method of the invention to test for
the statistical significance of the differences in performance among
the "vintages". The result and benefits of the Crus Classes
method to be described below is that it provides several advantages
over the typical, prior art vintage analysis. For example, it incorporates
a measure of dispersion. Further, it sets an analysis interval time
shorter than a year to increase accuracy. This produces several
advantages over traditional vintage analysis: (1) it automatically
adjusts the comparison to account for different numbers of loans
and for different size loans; (2) the Crus Classes method also allows
management to set the confidence intervals; and (3) it automatically
adjusts the year-to-year comparisons for loans with different credit
However, prior to describing the specific features of the Crus
Classes method of the present invention it is worthwhile to introduce
the following background information. Mortgage companies are vitally
concerned with the performance of the loans they service and own.
Active management of any loan portfolio requires the information
needed to properly categorize the performance of the underlying
loans. On an extremely broad and wide-sweeping comparison, delinquency
rates are generated and compared to various classes of mortgage
loans and summarized on a national level.
It is common in the industry for different financial institutions
to share data about the total number of loans serviced by them and
the appropriate number of loans that are in some form of delinquency.
Delinquency categories or "buckets" range from the least
serious, e.g. one payment past due, to the most serious category--namely,
in foreclosure. The following is an example prepared by the MORTGAGE
BANKERS ASSOCIATION of a delinquency chart:
TABLE-US-00001 Total Loans 1 Payment 2 Payments 3 Payments Loans
in Total Loans Serviced Past Due Past Due Past Due Foreclosure Delinquent
or or or or or or Outstanding 30 Days 60 Days 90 Days F/C Total
of of of of of of 22,426,005 733,330 163,710 141,284 230,988 1,269,312
% Delinquent 3.27% 0.73% 0.63% 1.03% 5.66%
Presenting delinquency performance in this manner is helpful in
quantifying total delinquency, but it reveals nothing about the
endogenous factors contributing to the default performance of the
underlying loans. These endogenous factors which affect the performance
of loan portfolios include, but are not limited to, some form of:
Performance history or age of the loan,
remaining time to maturity,
borrower's credit worthiness,
geographic locations, and
underlying collateral type.
However, the effect of "age" on the performance of loans
is a main factor that mortgage originators use to discern whether
a group of loans was (or is) "good" or "bad".
The vintage of the loan refers to the time when the given loan
or family or group or set of loans has originated or has been placed
on the books of the lending institution. In the mortgage industry,
loans are categorized by year of origination, where all of the loans
originated between Jan. 1, 1996 and Dec. 31, 1996 are referred to
as '96 vintage loans. The yearly vintages and their corresponding
delinquency rates are compared to each other to estimate relative
performance and value. FIG. 1A is an example of one such vintage
chart listing multiple vintages. This is a snapshot taken at the
last quarter of 1996. Therefore, the oldest 1995 vintage is twenty
one months old. In the figure, the 1995 curve shows a delinquency
rate on the order of about 3% for the 1995, twenty one month old
vintage. In contrast, the 1994 vintage curve shows a delinquency
rate approximately one half the size of the 1995 vintage for the
same twenty one month vintage. The mortgage default rates of FIG.
1A significantly affect future loan performance. Indeed, a single
percentage rise in the delinquency rate represents many millions
of dollars in losses to the typical financial institution which
carries a very large portfolio of loans.
Further by way of background, mortgage loan portfolios are quite
heterogeneous, with many subtle and changing variations in the basic
product characteristics and behavior. An important consideration
in the methodology of the present invention is the evolvement of
a model that preserves the heterogeneous nature of the mortgages.
Therefore, the inventors have grouped the mortgages by various (endogenous)
characteristics and made inferences about the relationship between
each of these characteristics and the resulting level of default.
A large number of characteristics results in an extremely large
number of combinations of groups to consider.
FIG. 1B illustrates the significance of maintaining proper distinction
lines between various loan instruments based on origination. Thus,
FIG. 1B shows for a given financial institution a total loan portfolio
value of, for example, several billions of dollars, with respect
to which the overall rate of delinquency is 0.86% (at tier 11 of
the loan tree of FIG. 1B). However, for proper analysis the invention
divides that loan portfolio into loan types including "conforming",
"jumbo" and "government" (originated) loans,
as indicated at tier 13. Note that the overall rate 0.86% of default
is the weighed average of the delinquency rate which varies from
1.25% for conforming loans, 0.55% for jumbo loans, and 9.98% for
government originated loans. Still further, each of the broad categories
of conforming, jumbo and government loans are further divided (at
tier 15) into ARM (adjusted rate mortgages) and fixed loans. Note
the significant divergence in the rates of default. The same is
true for the next subdivision (grouping) which hierarchically separates
the third tier loan groups into low LTV loans and high LTV loans.
For example, a government, ARM and low LTV loan at tier 17 has a
rate of default of 0%, whereas a government, fixed and high LTV
loan indicates (for the sample shown in FIG. 1B) a delinquency rate
The invention applies the Crus Classes method on each node of the
loan tree shown in FIG. 1B and then runs a hypothesis test to see
if the performance of each year vintage is better, worse or statistically
the same (at a confidence level of one standard deviation). It is
estimated that there are 308 different combinations and that it
takes approximately 100 megabytes of computer storage memory to
analyze and graph the results for the model shown in FIG. 1B.
Analysis of past performance of loan portfolios requires making
a decision as to what constitutes a delinquent or "bad"
loan, as for example for the purposes of creating a chart such as
in FIG. 1A. In an embodiment of the invention which has been reduced
to practice a first selection was to choose the definition of a
"bad" loan. It was chosen to represent a loan on which
interest and principal payments were at least 90 days delinquent.
That is, loans which are non-accruing or non-performing for a period
greater than 90 days are deemed "bad".
Further, since the industry is accustomed to and prefers to refer
to the "vintage" of a group of loans, for example 1993
vintage, 1994 vintage, etc., the Crus Classes method 34 also produces
and presents its results in terms of loan vintages. But, it groups
and selects vintages differently than the prior art.
The difference in "vintage" selection can be appreciated
from the matrix tables of FIGS. 2 and 3. The abscissa axis 50 in
FIG. 2 indicates the yearly quarter of origination, for example,
March '93, June '93, etc. The ordinate axis 52 indicates the end
quarter of a group of loans, for example, June '96, March '96, March
'95, etc. The matrix data in FIG. 2 indicate the number of months
that have elapsed from the quarter of origination to the end point.
For example, a loan originating in March '93 is 36 months old in
March '96 as indicated by reference numeral 54. Similarly, a loan
originating in June '93 is 15 months old in June '94 as indicated
by reference numeral 56.
The approach of the prior art has been to select and aggregate
as 1993 and 1994 loan vintages all of the loans between the bracket
lines 58 and 60 for the respective years 1993 and 1994. Carefully
comparing the precise ages of the 1993 and 1994 vintage loans reveals
two aspects which may undermine and distort the comparisons. First,
the traditional approach reflected by FIG. 2 compares loans whose
ages differ on average by twelve months. Indeed, some of the 1993
loans which are thirty six months old (see reference numeral 54)
are two years older than the twelve month old 1994 loans (reference
numeral 55). The above approach skews the results considerably since
the performance of loans is very age sensitive as can be appreciated
from FIG. 1A. It is far more meaningful to compare loans of the
same age which originate in different years. Therefore, it is far
more relevant to be able to compare the performance of different
loan vintages, as of the time when they were at the same ages. For
example, in seeking to answer the question: which loan vintage 1993
or 1994 is better, it is more relevant to know and compare the comparative
performances of the above noted loan vintages when each was, for
example, two years old. The Crus Classes method 34 of the present
invention is able to do so.
With reference to FIG. 3, the present invention selects as the
1993 and 1994 loan vintages, those loans which are bracketed by
the diagonally extending lines 62 and 64. In the selection method
according to the present invention, the ages of the 1993 and 1994
loan vintages that are being compared are identical to one another.
For example, for the year 1993 the ages of the loans vary between
6 months to 24 months and the same is true of the loans in the 1994
To further increase the accuracy of the comparison, each yearly
vintage is divided into four quarterly portfolios, i.e. first, second,
third and fourth quarters. All of the 1993 portfolios of the same
quarterly age are summed and divided by the total number of all
the loans with the same respective age. Referring again to FIG.
3, the present invention separately compares the 1993 and 1994 24
month old loans, then the 21 month old loans and so on. The actual
mathematical calculations/analysis comparing loans originated in
1993 and 1994 and the manner of calculating bad rates is shown below
in Tables I, II and III.
TABLE-US-00002 TABLE I Age (Months) 3 6 9 1993 # of Loans (n) n.sub.3
n.sub.6 n.sub.9 Vintage BAD rate (r) r.sub.3 r.sub.6 r.sub.9 STD
Sqrt(r.sub.3*(1 - r.sub.3)/n.sub.3) Sqrt(r.sub.6*(1 - r.sub.6)/n.sub.6)
Sqrt(r.sub.9*(1 - r.sub.9)/n.sub.9) 1994 # of Loans (N) N.sub.3
N.sub.6 N.sub.9 Vintage BAD rate (R) R.sub.3 R.sub.6 R.sub.9 STD
Sqrt(R.sub.3*(1 - R.sub.3)/N.sub.3) Sqrt(R.sub.6*(1 - R.sub.6)/N.sub.6)
Sqrt(R.sub.9*(1 - R.sub.9)/N.sub.9) 1994 - Difference (R - r) R.sub.3
- r.sub.3 R.sub.6 - r.sub.6 R.sub.9 - r.sub.9 1993 STD of (R - r)
STD.sub.3 - Sqrt(r.sub.3*(1 - STD.sub.6 - Sqrt(r.sub.6*(1 - STD.sub.9
- Sqrt(r.sub.9*(1 - r.sub.3)/n.sub.3 + R.sub.3*(1 - R.sub.3)/N.sub.3)
r.sub.6)/n.sub.6 + R.sub.6*(1 - R.sub.6)/N.sub.6) r.sub.9)/n.sub.9
+ R.sub.9*(1 - R.sub.9)/N.sub.9) Upper Bound +1*STD.sub.3 +1*STD.sub.6
+1*STD.sub.9 Lower Bound -1*STD.sub.3 -1*STD.sub.6 -1*STD.sub.9
TABLE-US-00003 TABLE II Age (Months) 12 15 18 1993 # of Loans (n)
n.sub.12 n.sub.15 n.sub.18 Vintage BAD rate (r) r.sub.12 r.sub.15
r.sub.18 STD Sqrt(r.sub.12*(1 - r.sub.12)/n.sub.12) Sqrt(r.sub.15*(1
- r.sub.15)/n.sub.15) Sqrt(r.sub.18*(1 - r.sub.18)/n.sub.18) 1994
# of Loans (N) N.sub.12 N.sub.15 N.sub.18 Vintage BAD rate (R) R.sub.12
R.sub.15 R.sub.18 STD Sqrt(R.sub.12*(1 - R.sub.12)/N.sub.12) Sqrt(R.sub.15*(1
- R.sub.15)/N.sub.15) Sqrt(R.sub.18*(1 - R.sub.18)/N.sub.18) 1994
- Difference (R - r) R.sub.12 - r.sub.12 R.sub.15 - r.sub.15 R.sub.18
- r.sub.18 1993 STD of (R - r) STD.sub.12 - Sqrt(r.sub.12*(1 - STD.sub.15
- Sqrt(r.sub.15*(1 - STD.sub.18 - Sqrt(r.sub.18*(1 - r.sub.12)/n.sub.12
+ R.sub.12*(1 - R.sub.12)/N.sub.12) r.sub.15)/n.sub.15 + R.sub.15*(1
- R.sub.15)/N.sub.15) r.sub.18)/n.sub.18 + R.sub.18*(1 - R.sub.18)/N.sub.18)
Upper Bound +1*STD.sub.12 +1*STD.sub.15 +1*STD.sub.18 Lower Bound
-1*STD.sub.12 -1*STD.sub.15 -1*STD.sub.18
TABLE-US-00004 TABLE III Age (Months) 21 24 27 1993 # of Loans
(n) n.sub.21 n.sub.24 n.sub.27 Vintage BAD rate (r) r.sub.21 r.sub.24
r.sub.27 STD Sqrt(r.sub.21*(1 - r.sub.21)/n.sub.21) Sqrt(r.sub.24*(1
- r.sub.24)/n.sub.24) Sqrt(r.sub.27*(1 - r.sub.27)/n.sub.27) 1994
# of Loans (N) N.sub.21 N.sub.24 N.sub.27 Vintage BAD rate (R) R.sub.21
R.sub.24 R.sub.27 STD Sqrt(R.sub.21*(1 - R.sub.21)/N.sub.21) Sqrt(R.sub.24*(1
- R.sub.24)/N.sub.24) Sqrt(R.sub.27*(1 - R.sub.27)/N.sub.27) 1994
- Difference (R - r) R.sub.21 - r.sub.21 R.sub.24 - r.sub.24 R.sub.27
- r.sub.27 1993 STD of (R - r) STD.sub.21-Sqrt(r.sub.21*(1- STD.sub.24-Sqrt(r.sub.24*(1-
STD.sub.27 -Sqrt(r.sub.27 *(1- r.sub.21)/n.sub.21 + R.sub.21*(1
- R.sub.21)/N.sub.21) r.sub.24)/n.sub.24 + R.sub.24*(1 - R.sub.24)/N.sub.24)
r.sub.27)/n.sub.27 + R.sub.27*(1 - R.sub.27)/N.sub.27) Upper Bound
+1*STD.sub.21 +1*STD.sub.24 +1*STD.sub.27 Lower Bound -1*STD.sub.21
Table I contains the calculations and comparisons for 1993 and
1994 loans which have vintage ages of 3, 6 and 9 months; Table II
does the same for loans aged 12, 15 and 18 months; and Table III
does so for loans that are 21, 24 and 27 months. The first line
in Tables I III sets forth the total number (n.sub.3, n.sub.6, etc.)
of loans of the given vintage. The second line in the Tables sets
forth the bad rates (r.sub.3, r.sub.6, etc.) for each vintage age.
The third line calculates the standard deviation (STD) of the bad
rate, using the indicated equations. The first three lines of Table
I supply the relevant information and calculations for the 3, 6
and 9 month old 1993 vintage. The next three lines of Table I supply
the same information for the 1994 vintage. The bottom four lines
of Table I calculate the differences and compare the results for
1993 and 1994 vintage years. The last two lines calculate upper
and lower bounds for the standard of deviation. The two bounds are
plus one and minus one standard deviation, but management can set
this based on their tolerance for default risk. These calculations
constitute the Crus Classes method 30 of the invention whose effect
can be appreciated from reviewing the analysis results plotted in
That is the values calculated in Tables I, II and III above for
the differences between vintages 1994 and 1993 are plotted in the
graph of FIG. 4 relative to a zero percentage base 71. The curve
70 represents the magnitude of the difference in the "bad"
rates of loans of the same age. The value of the curve 70 equals
r.sub.3-R.sub.3; r.sub.6-R.sub.6; etc. shown in Tables I III. One
would be tempted to assume that the 1994 vintage performs better
than the 1993 anytime the value of the curve 70 goes over zero percent
and vise versa. However, such a mode of analyzing the data would
be subject to reaching wrong conclusions due to statistical variations.
To overcome this drawback, it is more significant to ask whether
a vintage that appears to perform better does so in fact or whether
it merely reflects a temporary phenomenon. To answer the question,
the invention uses a hypothesis testing technique which allows the
analyst to set a confidence interval which is adjustable to allow
for different corporate risk tolerance levels. Thus, the confidence
interval can be equated to the amount of risk tolerance management
will accept in originating, purchasing, retaining or servicing loan
portfolios. These confidence intervals can also be used in product
profitability and capital allocations. Management can then rank
the vintages by product, program, age and size. To this end, the
Tables presented above also calculate the standard deviation of
the difference in performance and sets upper bounds and lower bound
of +1 and -1 standard deviations for each quarterly vintage. These
upper and lower limits which appear in the last two lines of Tables
I, II, III are plotted in the form of curves 66 and 68 in FIG. 4.
The area between the curves 66 and 68 is an area of uncertainty.
With this in mind, since the invention superimposes the curve 70
over the area of uncertainty, one can state with greater certainty
which vintage performs better only in the areas outside the area
of uncertainty. Thus, the graph of FIG. 4 shows that the 1994 loan
vintages are "better" than corresponding 1993 loan vintages
for loans that are 6, 21 and 24 months old. On the other hand, the
1993 vintage appears to be better for loans that are 27 and 30 months
old. During other months, the result is too close to conclude with
the chosen degree of certainty which vintage is better. The chart
of FIG. 4 underscores the fallacy of the prior art in referring
to yearly vintages as better or worse. One must be more specific
as to time and other criteria, since relative performance changes
dynamically with time.
While the invention has been described above in relation to the
consideration of vintages in yearly quarterly units, note that in
the loan industry exogenous factors such as changes in economy,
unemployment and inflation are time varying factors that vary greatly
over an annual interval and therefore the system of the invention
permits analysis based on the choice of any interval unit. The important
thing to realize is that in general, a new mortgage loan is more
sensitive to small changes in delinquency performance than an older
mortgage. This is shown by widening of the confidence interval bands
over time. So in essence, the application of the above described
Crus Classes method corrects for this fact.
The invention also takes and adjusts the vintage rating based on
quality comparisons for different volatilities of default. In essence,
using the system lets the user to set policies with respect to volatilities
of default which is another form of risk management. This is new
to the industry.
The confidence level in the assessment of the difference in quality
between groups of loans depends to a certain degree on the sample
size of the loans. For small groups of loans, one will always be
less certain of their performance. The real question is how much
less certain. This is answered with the Crus Classes method. The
Crus Classes method also automatically adjusts the comparison for
different sample sizes of loans in each node or product. This is
evident in the calculations in the previously presented tables which
always take into account the number of loans. An actual calculation
that has been carried out to evolve the vintage comparison graph
of FIG. 4 in accordance with Tables I, II and III is presented in
As described above, the Crus Classes method 34 delivers a comparison
of two loan vintages either in the form of a graph or tabulated
data which permits one to get a sense of which vintages are performing
better. This information can then be used in making manual or automatic,
computer generated yes/no decisions whether to originate, purchase,
or to maintain and sell various vintages of loan products or servicing
rights as needed at the decisional blocks 18, 24 and 32 of FIG.
The basic premise of the Crus Classes analysis is that the future
performance of these loan vintages will match the past pattern.
This may not necessarily be true. To this end, the early warning
system (EWS) 32 of the present invention further enhances the loan
analysis process by incorporating an application of behavioral scoring
that has been specifically designed to be used on closed end loans
with longer maturities such as mortgages. The EWS 32 is able to
statistically predict the probability that a group of loans will
experience credit performance problems during a future preselected
time period, without waiting for that loan to season. In the case
of mortgages, the time to season is typically three to seven years.
The EWS 34 is intended to provide management with automated analytical
tools which allow making decisions well in advance of the aforementioned
three to seven "seasoning" period.
By utilizing the EWS, a mortgage originator can perform portfolio
analysis and ascertain which product type, program, type of underwriting,
property type, type of customer, origination channel, etc. is at
risk, without waiting for the mortgages to actually mature and enter
default. The only constraint is the amount of data attributes that
the mortgage loan originator keeps on any customer over time, which
for the purposes of the present invention may be two years. The
mortgage originator can then dynamically adjust the flow of origination
by altering any credit criteria derived from a particular attribute.
The EWS 34 constitutes the dynamic component of the underwriting
concept of the present invention. With this concept, the decision
maker can estimate improvements in credit quality for each specific
type or amount of change in a criteria, i.e. he or she can calculate
the marginal contribution of any attribute on record.
More specifically, the forward looking feature of the EWS component
of the invention attempts to forecast the likelihood the borrowers
will enter a 90+ days past due delinquency on their mortgages. This
condition--the occurrence of a 90-day past due delinquency--is defined
as a "bad" condition relative to any loan. The EWS calculates
the probabilities of bad conditions occurring by combining loan
information with the credit bureau's current behavioral score for
the given borrower. In other words, the EWS combines the borrower's
current mortgage status (default status and age) with a forecast
that is based on the borrower's performance on other obligations
and uses this information to forecast the bad condition. The EWS
system makes three major assumptions: The future performance pattern
of defaults will be the same as in the past; The future performance
depends upon the current loan characteristics and is dependent on
past performance only through the credit bureau scores. Therefore,
the EWS also carries all the assumption of the credit bureau's score
that was used; and The EWS employs a logistic regression model to
accurately and sufficiently predict default behavior.
The aforementioned "bad" condition is a discrete (yes
or no) event that occurs when and if an individual loan is at least
once three payments past due at any point during a forward looking
preset time period, for example, two years. Bad loans are assigned
the value "1" and good loans the value "0".
How many times the loan enters "bad" is not considered
in the EWS. A good loan is never three payments past due over the
aforementioned two year time frame and therefore is assigned a loan
of a value "0". Preferably, in order to provide reliable
information using the EWS system, the underlying portfolio should
have at least 100,000 loans and the loans should consist of different
distributions of ages, types, locations, etc. In an embodiment of
the invention which has been reduced to practice the number of loans
in the portfolio exceeded one million.
The EWS probability of loans entering the bad condition is developed
or calculated on the basis of looking backward in time through a
development period which may similarly constitute a two-year time
period. The EWS formula considers the age of the loan at the beginning
of development period; the credit bureau score for the loan at the
beginning of the development period; the delinquency status at the
beginning of the development period; and the type of product, for
example, whether a government or conventional or adjustable rate
The following logistic model has been applied to the underlying
portfolio (government and conventional loans being considered separately).
In the formula shown below, P is the probability of a loan becoming
bad at any time in the coming two years: Log(P/(1-P))=A+(B.sub.1*AGE)+B.sub.2*C0+B3*D.sub.1+B4*D.sub.2+B.sub.5*SCO-
In the equation, AGE is defined in categories of quarters from
1 40. Therefore B.sub.1 and AGE are 40 dimension row and column
vectors respectively. SCORE is the mortgage score from a credit
bureau rating company such as the well known Equifax rating bureau,
at the beginning of the two-year time period, i.e. August 1994.
Note, if no such score is available, the invention assigns the lowest
possible value, namely 200. The Equifax scoring scores vary from
200 to 1000. The dummy variables in the above equation are defined
as follows: .times..times..times..times..times..times..times..times..times..times..ti-
es..times..times..times..times..times..times. ##EQU00001## The coefficients
A, B.sub.1, B.sub.2, B.sub.3, B.sub.4, B.sub.5 and B.sub.6 are estimated
by running the model over the underlying portfolio.
All of the forecasting is done at the individual loan level and
then the results are aggregated into the portfolio of interest by
defining or grouping certain loan characteristics (location, rate
type, maturity, LTV, etc.) to make comparisons. Even though the
invention presents the mean probability for a predicted group, the
information contained in the individual loan level data is preserved
because the invention explicitly considers the dispersion around
the mean of the bad rate over time.
To forecast the probability of an individual loan entering 90+
day default during the predetermined time period (i.e. the two year
time frame), all one need do is insert the estimated coefficients
and the characteristics of that individual loan into the logistic
equation for "P" presented earlier.
The logistic model in the form of the aforementioned equation provides
numerical results which are suitable of being graphically presented
as shown in FIG. 5. The graph enables management to readily interpret
and form decisions based on the predictions which it contains. In
accordance with a further embodiment of the invention, the results
are feedback by the computer which then provides yes/no decisions
based on predetermined default risk or profit criteria set by the
FIG. 5 is a snapshot taken in 1996 and depicts loan experience
looking both backwards to the past two years and forwards over a
similar two year time span. The vertical solid bars 72 represent
the current mean (expost) bad rates for a particular group over
the past two years. In other words, these bars show the mean bad
rate percentages of a group of loans that originated in a particular
year. For example, the bar 72 for the loans originated in 1989 shows
a bad rate of about 12.5%. The score for the 1991 loans is just
about 8% whereas the bad rate for 1996 is quite low (under 5%),
reflecting the fact that this vintage of loan has not yet matured
The hatched vertical bars 74 represent the forecasted mean bad
rates (exante) for the same group of loans over the next two years.
The value for the 1996 vintage is somewhere around 7% indicating
an expected delinquency rate of 7% even though the past two-year
performance had an actual bad rate of only about 2.5%. The curve
76 represents the expected bad rate curve that is obtained by modifying
the forecasted bad rates by the risk ratio on nearby vintages, and
this shall be explained more fully later on.
One should not place much emphasis on whether the curve 76 is above
the solid bars 72, since this may merely reflect the normal pattern
of seasoning for mortgages. Nor should one place much emphasis on
the absolute height of each bar, since this may reflect different
expectations among the various groups or types of loans that are
being analyzed. Instead, the graph of FIG. 5 indicates three important
benchmarks for reviewing and forecasting the risk in the given loan
portfolio. First, note the jump which is indicated by the arrow
84. It represents a jump which occurs when the difference between
the hatched bars 74 and the solid bars 72 is greater than one standard
deviation above the historical age weighed performance for that
vintage. The bigger the jump, the more serious the quality problem.
This measure is particularly useful on younger vintages, e.g. the
1996 vintage to which the arrow 84 is pointed.
Second, the size of the portion of the hatched bars 74 which protrudes
above the expected bad curve 76 indicates an unusual level of risk
in the past or the future for that group of loans. Note the arrows
80 and 82 which indicates such conditions. Finally, the arrow 78
indicates a turning point which represents the point at which the
first derivative of the expected BAD rate curve changes sign from
positive to negative: i.e. the first time the bad rate drops as
the loans age increases. The younger the age at which the turning
point occurs, the earlier the portfolio's credit performance will
or has matured.
The manner in which the curve 76 of FIG. 5 is derived may be better
understood by reviewing FIGS. 6 and 7. More specifically, the curve
86 is developed by taking a snapshot at a point in time looking
at loans of different quarterly ages and asking how many in each
age group entered the "bad" state during the preceding
predetermined time period, e.g. two years.
Initially, the EWS 34 develops for each empirical two-year period
of performance the bad rate curve as a function of age in quarters.
See the quarterly bad rate curve 86 in FIG. 6.
Next, using a moving average, the invention smooths each curve
to reduce the randomness of the quarterly performance, thus obtaining
the smooth curve 88 in FIG. 6. In fact, the slopes of the yearly
bad rate curve are calculated as the percentage change in bad rates
from one year to another year, which is known as the risk ratio.
To improve the integrity of the results and protect against possible
statistical aberrations, at least eight two-year time periods are
considered, with the means and standard deviations of the risk ratios
To find the point on line 76 for 1991, the invention uses the point
on line 76 for 1990 which has the expected bad rate for 1990. Multiplying
this bad rate by the corresponding mean risk ratio from FIG. 6 obtains
the expected bad rate for 1991 and its standard deviation. This
expected bad rate is used for the point on line 76 for 1991. However,
if the forecast bad rate (bar 74) is within the expected bad rate,
plus or minus one standard deviation, the invention just substitutes
the expected bad rate by the forecast bad rate so that the line
76 passes through the top of bar 74.
The third in the triad of dynamic underwriting tools of the present
inventions, the matrix link system 36, uses the information developed
by the Crus Classes technique 34 and early warning system 32 to
develop a probability based prediction of how many of a given set
of loans will be "bad" at a selected future date.
More specifically, while the Crus Classes method 34 analyzes the
past performance of loan vintages and the EWS system places a probability
on a group of loans entering the bad state within a preset time
frame, i.e. a window, the matrix link system 36 is designed to predict
the default status performance of a group of loans at a preset point
in time within the window of operation of the EWS. For example,
the EWS method is able to say with respect to a group of loans that
3% of those loans will enter a bad state (90+ days in arrears) at
one time or another during a two year window. In comparison, the
matrix link is designed to answer the question how many loans will
be non-accruing, i.e. in the 90+ payment overdue state, at the end
of the first quarter of the two year window or nine months into
the window and so forth, taking into consideration that some loans
may enter the bad state or exit therefrom due to prepayment or on
account of having been matured, or by completing foreclosure.
To perform its functions, the matrix link system 36 predicts how
many borrowers that enter a 90-days delinquency state remain there,
how many loans will return to "good" status and, finally,
in which quarter over the predefined window, e.g. two years, will
these transitions occur. Note that a particular loan can enter a
bad state, return to a good state or remain in a bad state. Loans
may mature, pay off or complete foreclosure during the predefined
window period and thus exit the loan portfolio being analyzed. In
order to provide a quantitative measure of the transitions of loans
between different states, the present inventors have developed a
so-called delinquency transition matrix in a form which, in one
embodiment thereof, appears as in the table in FIG. 7.
In the table of FIG. 7, use is made of a historical file spanning
four years and including vintage years 1993 1996. The table shows
the delinquency performance of a particular group of loans. Further,
the table stratifies the loans by age of origination. Note that
the table lists separately the results for three different types
of loans, namely conforming loans, jumbo loans and government loans.
In each case, it shows the probability of a loan transitioning (a)
from a bad state to a bad state, (b) from a bad state to a good
state; (c) exiting, i.e. maturing and therefore being dropped from
the sample of loans being considered, (d) from good to bad, (e)
from good to good; and (f) from good to exit state.
Using the percentages listed in the delinquency transition matrix
in the Table of FIG. 7, one can then begin to convert the information
obtained through the EWS system 32 into the forecasts of how the
Crus Classes vintages will perform at predefined time periods within
the window. The table in FIG. 8 illustrates the formulae evolved
by the present inventors which are as follows.
First, the current age of the loan at the time of forecast is determined.
In the table of FIG. 7, the age of the loans is indicated in years
(1, 2, 3, 4 . . . 10). However, in an actual implementation of the
invention, the vintages have been stated and the formula is calculated
in terms of quarters to obtain increased accuracy.
For each group of loans of a particular age, the invention uses
a 3-month transition matrix to forecast three months forward, a
6-month transition matrix to forecast six months forward, a 9-month
transition matrix to forecast nine months forward and a 12-month
transition matrix to forecast twelve months forward.
Based on the choice of data in the previous step, the invention
calculates respectively looking forward three, six, nine and twelve
months: 1. how many good loans and bad loans will exist from the
portfolio; 2. how many good loans will turn into bad; and 3. how
many bad loans will remain bad.
From the above data, one obtains the classic "roll-rate"
forecast which provides the first component of the forecast. The
above approach merely projects forward the results that have already
occurred in the past, on the expectation that they will repeat themselves.
However, a greater benefit of the matrix link technique of the present
invention comes from adding the additional information that is contained
in and/or obtained by the early warning system 32.
To this end, the invention: (a) Calculates an empirical ratio obtained
as--the cumulative number of loans which are 90+ at each quarter
(EOP) and divides it by the number of loans that are 90+ at least
once during these four quarters. (b) From the EWS, the invention
obtains or forecasts the "bad" rate for the two-year window
based on the EWS method 32. (c) Using the EWS, the invention forecasts
the bad rate and the empirical ratio above as a new piece of information
to adjust the classic "roll-rate" forecast. This is in
essence what comprises the "matrix link" method 36. FIG.
9 provides the results of the matrix link in graphical form. In
the example shown, the performance of different vintage loans is
predicted one year forward in quarterly installments starting in
For example, the plot 102 shows the two-year performance of 1995
vintage loans which range in age (months) from 0 to 24 months, as
viewed looking backwards in time in 1997. The curve 104 shows the
delinquency rate percentages predicted for the next twelve months.
For example, when the groups of loans attain an age of 27, the delinquency
rate can be read on the ordinate axis. The same is true for this
group of loans when they reach an age of 30 months, 33 months and
36 months. The reason that the curve 104 has a predicted value below
the actual value is that the prediction in the matrix link uses
a moving sum average which weighs down the actual sharp up-turn
in the "bad" rate which has actually occurred toward the
end of the curve 102.
The above remarks are also applicable to the curves 106, 108 which
apply to the 1993 vintage, the curves 110 and 112 which are applicable
to 1994 vintage, and to the curves 114 and 116 which apply to the
Note that the Crus Classes are less static than traditional mortgage
vintage analysis. Therefore, the performance of the last three points
of any vintage can still change somewhat, for better or worse.
The matrix link lines, i.e. the curve 104, 108, 112, and 116 also
show where the inventors expect the last three Crus Classes points
to adjust over the next nine months.
The system of the present invention lends itself easily to being
implemented through use of a general purpose programmable computer
as illustrated in FIG. 10. Thus, the general purpose computer 124
communicates with a local database 122 which receives a wealth of
statistical and specific information about various loans from diverse
sources. For example, the source of the information may be a national
loan database 120 which is maintained by certain industry groups.
The general purpose computer 124 has the usual complement of peripherals
including an operator's console 126, ROM 128, RAM 130 and a hard
The computer 124 operates under control of major software blocks
which perform the dynamic analysis 134 in a manner already described.
The main software components are the software routines 136 which
handle the development and analysis of the Crus Classes associated
with the creation of the loan vintages. The early warning system
block 138 calculates probabilities of loans going bad within a predetermined
forward looking window. Finally, the matrix link software block
140 forecasts the probabilities that a fraction of loans will go
bad within the window at a particular time.
The analytical results developed by these software subroutines
or blocks 136, 138 and 140 are tabulated in the tabulation software
block 150 and outputted through an output software block 152. The
output can be in the form of a signal which drives a printer which
generates a graphical representation of the results in the manner
previously described. Alternatively, the output may supply the results
to the console 126 for visual inspection by the operator. Alternatively,
the operator may program the computer 124 via the console 126 to
provide yes/no answers as to whether an investment should be made
or continued to be made in a particular loan portfolio, again as
Although the present invention has been described in relation to
particular embodiments thereof, many other variations and modifications
and other uses will become apparent to those skilled in the art.
It is preferred, therefore, that the present invention be limited
not by the specific disclosure herein, but only by the appended