CAPITAL REQUIREMENTS

Preparing for Basel II

Modeling Requirements
Part 4: Stress Testing
by Jeffrey S. Morrison

arlier articles in this series have focused on the development

and validation of PD and LGD models. In contrast, this
fourth and final article will introduce a different modeling

approach that employs aggregated data to predict the impact of economic and portfolio changes on bank default losses.

he term stress testing

describes a range of techniques used to assess the
vulnerability of a portfolio to
major changes in the economic
environment or to exceptional but
plausible events. Stress tests
make risks more transparent by
estimating the potential losses on
a portfolio in abnormal markets.
They complement the internal
models and management systems
used by financial institutions for
capital allocation decisions.1
More simply, stress testing is
a way to produce alternative scenarios using sensitivity analysis.
Banks as well as other businesses
have been using sensitivity analysis for years, even if only in an ad
hoc framework. However, stress
testing as referred to in the new

Basel Capital Accord uses more

quantitative approachesmethods where assumptions can be
empirically evaluated. Stress testing should be able to link dramatic changes in the economic environment to the banks portfolio:
297. A bank must have in place
sound stress testing processes for
use in the assessment of capital
adequacy. Stress testing should
involve identifying possible events
or future changes in economic
conditions that could have unfavorable effects on a banks credit
exposures and assessment of the
banks ability to withstand such
changes.
New Basel Capital Accord (2001)

Unfortunately, Basel has not

yet established specific guidelines

on how to do this, which is not surprising given the range of issues to

be covered, such as data availability, portfolio diversity, and standardization of model inputs and outputs. This article does not intend
to present a survey of all possible
techniques for stress testing. All
such techniques have their
strengths and weaknesses, depending on the banks resources and
portfolio structure. Instead, this
article introduces a methodology
that is practical, easy to implement, understandable, and has a
statistical foundation that has been
around for years. The approach can
be particularly useful for the retail
side of the business because of the
greater number of defaults, and it
could also prove beneficial on the
commercial side of the house.

2003 by RMA. Jeff Morrison is vice president, Credit MetricsPRISM Team, at SunTrust Banks Inc., Atlanta,
Georgia.
2

The RMA Journal September 2003

Preparing for Basel II

Modeling Requirements
Pa r t 4 : S t r e s s Te s t i n g
Before designing any stress
test approach, the first thing we
want to know is whats on the
test. In other words, what questions do you want answered?
Some questions might include the
following:
What would happen to our
risk level if we went into a
deep recession?
Would the recession affect us
immediately, or would there
be a delay?
How do differences in local
economies affect our risk?
What would happen to our
risk level if interest rates
went up significantly?
What impact do new accounts
have on our portfolio risk
level?
If we were to enter a major
recession, could we mitigate
its impact by focusing on specific geographies? Would it be
better to loosen lending policies in some areas while tightening them in others?
What would be the effect of a
significant increase in property values in an area?
What would happen in the
event of a significant shift in
the industry composition of
our portfolio?
Is our small business portfolio
more risk sensitive to changes
in the economy than our middle-market accounts?
How does the economy affect
our large customers? Which
industries are more sensitive?
Questions like these not only
highlight the need to understand
the impact of economic scenarios
on your portfolio, they also point
to the need to quantify strategies
for mitigating potential risk.

Given the availability of data,

some of these questions may be
unanswerable except by judgment and intuition. However, a
great deal of data from loan
accounting systems is often available, making model development
a distinct possibility. Therefore, a
flexible stress-test approach
should have dials not only to turn
down the levels of such external
recessionary factors as employment and income (outside the
control of the bank) but also to
adjust for internal factors (within
the banks control, reflecting
efforts to mitigate risk).
A Different Kind of Regression
As noted in previous articles
in this series, the statistical techniques recommended for PD
(probability of default) and LGD
(loss given default) were regression
models. These models statistically
quantify the correlation of a set of
predictor variables with the
default condition, or the percentage of dollars not recovered. They
use account-specific characteristics at a particular moment in time
to estimate account-specific predictions. In a PD model, for
example, the default definition for
an account is determined over a
one-year time frame. If a default
occurs anytime within a year, then
the dependent variable gets
assigned a value of 1. Otherwise,
it is assigned a value of 0. No
attempt is made to link variations
in risk over time to explanatory
factors that also change over time.
From that perspective, these
types of models are one dimensionalrepresenting the characteristics of the borrower at a single
snapshot in time.

Fortunately, a different type

of regression model is capable of
using more than one point of time.
It is called pooled cross-sectional time
series regression2. It works by using
data that has been rolled up from
the account level into broader levels of aggregation. These levels
might be counties, MSAs (metropolitan statistical area), bank
branches, states, industries, etc.,
and are generically referred to as
cross-sections. Mapping accountlevel data into their proper crosssection or MSA before aggregation
has begun is done through zip
code. Data on each aggregated
group is tracked over time and
placed into a regression framework, where correlation estimates
can be made, tested for statistical
importance, and used for prediction. If MSA were the level of
aggregation, then we might try to
explain the MSA default rate over
time as a function of the following
variables from our loan account
systems, also at the MSA level:
Average LTV.
Average age of the loan.
Percent of loans less than one
year old.
Average risk rating or FICO
score.
Average loan size.
Average credit limit.
Average percentage of loans
renewed.
Average level of delinquency.
Percent of loans with cash
advances.
Average number of loans per
obligor.
However, internal data from
our accounting systems is not
enough. In order to capture outside influences on our portfolio,
we must obtain some external
3

Preparing for Basel II

Modeling Requirements
Pa r t 4 : S t r e s s Te s t i n g
data from other sources. Luckily, a
number of economic service
providers offer historical and forecasted data products at various
levels of aggregation. Some credit
bureaus even offer aggregated
consumer data. There is even free
information on the Internet.
Therefore, we could add the following type of external information to our list of resources:
Average home prices.
Credit card delinquency
rate30/60/90.
Credit card utilization.
Average credit limitcredit
cards.
Bankruptcy rate.
Unemployment rate.
Number of households.
Disposable income.
Median household income.
Total employment.
GDP.
Federal funds rate.
Prime rate.
30-year fixed mortgage rate.

Model Estimation
The estimation of pooled
cross-sectional time series models
is a little more involved than for
PD and LGD models, but it still
involves relating a dependent
variable to a set of explanatory
variables over a historical time
period. The difference is that it
accounts for data across two
dimensionsby cross-section and
time series. In most statistical
software packages like SHAZAM,
LIMDEP, and SAS, the data has
to be arranged a certain way to
estimate the regression. Figure 1
shows the data design in spreadsheet form for a fictitious example
in which the cross-section is MSA.
For example, the cross-section
called MSA #1 could represent the
Atlanta MSA. As can be seen, data
for the first cross-section is listed
first and ordered by date, followed
by the next cross-section, again
ordered by date. The total number
of observations in the data would

be the number of cross-sections

times the number of time periods.
Since we are now dealing with
multiple snapshots in time, we can
also create variables that are lagged.
For example, in Figure 1 the variable Prime Rate Lag (1) represents
a one-quarter lag in the prime rate.
In a regression model where the
dependent variable was the MSA
default rate, the inclusion of this
variable would infer that todays
prime rate takes one quarter for its
influence to be felt in our portfolio.
The last two columns of
Figure 1 represent additional indicators called dummy variables. It
is standard statistical practice to
use these types of variables to
enhance model accuracy. Dummy
variables account for differences at
the MSA level that are not already
captured by the other explanatory
variables in the model. These variables are assigned a value of 1 if
the account is a member of that
particular MSA (cross-section), or

Figure 1

Cross
Section
MSA #1
MSA #1
MSA #1
MSA #1
MSA #1
MSA #1
MSA #1
MSA #1
MSA #2
MSA #2
MSA #2
MSA #2
MSA #2
MSA #2
MSA #2
MSA #2

Time Dependent %New

Period Variable Accounts
2001Q1
0.12%
5.00%
2001Q2
0.13%
5.20%
2001Q3
0.14%
5.50%
2001Q4
0.22%
6.00%
2002Q1
0.22%
6.50%
2002Q2
0.28%
6.40%
2002Q3
0.24%
7.00%
2002Q4
0.21%
6.50%
2001Q1
0.34%
12.30%
2001Q2
0.43%
13.60%
2001Q3
0.42%
14.30%
2001Q4
0.51%
12.40%
2002Q1
0.49%
14.50%
2002Q2
0.41%
15.20%
2002Q3
0.32%
16.40%
2002Q4
0.45%
17.90%

Pooled Cross-Section Time Series Data Structure

%
Avg
Avg
Avg
Average % SIC 60 days Unempl Household Bankruptcy
LTV
#67 Deliquent Rate
Income
Rate
Figure 1 - Pooled Cross Section Time Series Date Structure
81.50% 10.10%
1.20%
5.12% $56,405.54
6.30%
83.00% 11.20%
1.33%
5.12% $56,442.37
6.51%
87.40% 12.20% 1.22%
5.13% $56,603.60
6.74%
84.30% 12.20%
1.54%
5.12% $56,897.32
7.08%
88.30% 13.30%
1.44%
6.11% $57,189.21
7.29%
79.20% 12.60%
1.65%
6.43% $57,478.13
7.42%
85.10% 12.80%
2.11%
6.43% $57,684.58
7.69%
85.60% 11.80%
2.01%
7.23% $57,790.08
7.88%
65.40% 15.30%
5.33%
4.32% $35,822.07
7.37%
66.10% 17.40%
4.33%
4.54% $34,066.79
7.86%
65.70% 16.70%
5.32%
5.44% $33,519.87
8.31%
66.50% 18.20%
4.32%
5.01% $34,272.91
8.69%
67.40% 18.30%
4.66%
5.04% $35.031.90
8.96%
65.70% 19.50%
5.67%
5.23% $35,773.35
9.24%
66.10% 20.20%
5.98%
5.21% $36,014.37
9.30%
65.50% 19.50%
6.21%
5.21% $36,201.34
9.09%

The RMA Journal September 2003

Avg
Prime
Rate
8.62%
7.34%
6.57%
5.16%
4.75%
4.75%
4.75%
4.54%
8.62%
7.34%
6.57%
5.16%
4.75%
4.75%
4.75%
4.54%

Prime MSA #1 MSA #2

Rate Dummy Dummy
Lag (1) Variable Variable
1
0
8.62%
1
0
7.34%
1
0
6.57%
1
0
5.16%
1
0
4.75%
1
0
4.75%
1
0
4.75%
1
0
0
1
8.62%
0
1
7.34%
0
1
6.57%
0
1
5.16%
0
1
4.75%
0
1
4.75%
0
1
4.75%
0
1

Preparing for Basel II

Modeling Requirements
Pa r t 4 : S t r e s s Te s t i n g
0 otherwise. Standard practice is to
include in the regression as many
dummy variables as there are
cross-sections, less one.
A Stress Test Walk-Through: A
Simple MSA Example
Typically, two regression models are recommended for each portfolio. One model is to predict the
default rate. This may be defined
simply as the loan dollars defaulted
at the MSA level divided by the
total outstanding dollars in that
area at a particular point in time.
This model, however, will not
account for the condition in which
the obligor draws on credit lines
during difficult economic times.
Therefore, a second model is needed to predict usage. For this model,
usage is defined as the average outstanding balances for the MSA or
cross-section at a particular point in
time. Once these models are estimated, their predicted values can
be multiplied together to obtain
the dollar impact resulting from the
stress test.
As an example, lets perform
a simple stress test reflecting the
impact of a substantial increase in
the unemployment rate.
Step 1: Collect and aggregate
historical data (two years or
more) into MSAs or cross-sections. This includes internal
loan accounting data and such
external data as economic or
credit bureau information. Be
sure to add any dummy variables you created or variables
with time lags.
Step 2: Define the dependent
variable in each regression
(a) average default rate (b)
average outstanding balances.

Step 3: Look at the correlations between each explanatory variable and the dependent
variable. Explanatory variables
with the higher correlations
are good candidates for the
regression. Check the sign of
the correlation. If the correlation is negative, then there is
an inverse relationship
between the default rate and
your explanatory variable.
Does it make sense? Graph
the trends associated with
each variable over time.
Step 4: Pick the variables to
include in the regression.
Make sure the variables you
put in make business sense.
Be wary of using a stepwise
selection feature as you may
have done in your PD model.
The correlations between the
explanatory variables can be
complex and may interfere
with some automatic routines
designed to reduce the number of variables in the regression. Instead, use the t-tests
produced by the software to
give some guidance as to
which variables should remain
in the model. Generally, a tvalue greater than 2 (in
absolute terms) is an indication that the variable has some
importance. Reestimate the
model with different variables
and time lags until you get a
model with valid t-values and
coefficients with signs that
make sense. If you are trying
to predict the default percentage and get a negative LTV
coefficient, does it make sense
that MSAs with higher LTVs
have a lower percentage

default than MSAs with lower

LTVsall other things
remaining equal? If not, go
back to the drawing board. Try
to get a statistic called the
adjusted R-square as high as
possibleall other things
remaining equal. In general,
the higher this value, the better the model.
Step 5: Look at the elasticities
produced by the modelone
for each variable. These are
sensitivity measures produced
by most software packages. A
variable with a high elasticity
(greater than 1 in absolute
value) implies that small
changes to that variable will
result in larger changes to your
dependent variable. These are
the nuts and bolts of your
stress test. They reflect a standardized unit of measure for
the correlation structure in
your model.
Step 6: Produce your stress
tests:
a) Default rate model. Lets
say the model produces an
elasticity of +.42 for the unemployment rate. An elasticity of
+.42 implies that a 1% increase
in the unemployment rate will
lead to a .42% increase in the
default rateall other things
remaining equal. So if we look
at an unemployment rate
shock (stress test) of, say, 25%,
then the default rate would be
expected to increase by 10.5%
(25 .42).
b) Average balance model.
Lets say this model produces
an elasticity of +.22 for the
unemployment rate variable.
5

Preparing for Basel II

Modeling Requirements
Pa r t 4 : S t r e s s Te s t i n g
forecasted assumptions of
the explanatory variables.
Hypothetical Example of a Stress Test
If you are working with an
A
B
C
D
E
F
G
economic data provider
Stress
Stress
MSA #1
MSA #1
MSA #2
that gives you a worst-case
scenario for the economy,
Elasticity Test %
Variable
Current Stressed Difference
you already have some of
Unemployment
Default Rate 0.42%
25%
0.47%
0.52%
0.05%
your assumptions done.
Rate
The remainder of the variAverage
Unemployment
0.22%
25%
$235.00
$247.93
$12.93
ables in your models that
Balance
Rate
may have come from your
# Accounts
10,000
10,000
N/A
loan accounting system
Predicted
must also be forecasted or
Default
$11,045.00 $12,875.98
1,830.98
trended over the forecast
Dollars
horizon.
Figure 3 shows an
% Increase
example of this approach,
In Default
where the historical and
16.58%
$$Default
forecasted data for the
explanatory variables are
An elasticity of +.22% implies
and E, you get a stress test
applied against the regression
that a 1% increase in the
result of $1,830.98. This is
equation. For each model, two
unemployment rate will lead to
16.58% ($1,830.98 / $11,045)
forecasts are run. The first is
a .22 percent increase in averover what you would have seen
the base-case forecast, in
age balancesall other things
if there had been no shock at
which a business-as-usual
remaining equal. So if we
all.
trend of the explanatory varistress test the unemployment
Step 7 (optional): Develop
ables is made and then applied
rate at 25%, then the average
detailed stress test forecasts.
to the regression equation.
balance will increase by 5.5%
The method presented in Step
Next, a stress-test forecast is
(25 .22).
6 is good to produce ballpark
done, in which new values for
c) Calculate percent increase in
stress-test results.
the explanatory variables assodefault dollars. For a particular
Unfortunately, these methods
ciated with the stress test are
MSA, assume there are curdo not easily address the timapplied to the regression equarently 10,000 accounts with an
ing of the stressed event or the
tion. This could include a
average default rate of .47%
performance of stress tests
more complex recessionary
and an average balance of
across a combination of variscenario comprised of a 35%
$235, as shown in column E of
ables at the same time. To
increase in the unemployment
Figure 2. The new stressed
handle the timing and comrate, a 15% decrease in income
default rate is calculated to be
plexity of more advanced
per capita, and even a 12% rise
.52% (.47 1.105) while the
stress tests, the estimated
in the prime rate. Figure 3
new stressed average balance is
model can be placed in an
shows a graphical illustration
found to be $247.93 ($235
Excel spreadsheet along with
of such a default-rate scenario
1.055) as shown in column F.
any historical and forecasted
for a particular MSA. Once
The predicted default dollars
data to provide a more detailed
applied to each MSA or crossare simply the default rate
analysis. This kind of approach
section, the difference
average balance number of
makes for a useful planning
between the base-case forecast
accounts. By subtracting the
tool where multiple scenarios
and the stress-test forecast can
difference between columns F
can be made quickly, based on
be calculated and rolled up to

Figure 2

The RMA Journal September 2003

Preparing for Basel II

Modeling Requirements
Pa r t 4 : S t r e s s Te s t i n g
Figure 3
Stress Test: Major RecessionMSA #1
0.43%
0.38%
0.33%
0.28%
0.23%
0.18%
0.13%
0.08%
2001Q1 2001Q2 2001Q3 2001Q4 2002Q1 2002Q2 2002Q3 2002Q4 2003Q1 2003Q2 2003Q3 2003Q4

Base Forecast

Stress Forecast

determine the overall stress

effect on the entire portfolio.
The what-if capabilities of the
spreadsheet can be extended
to adjust the stress-test forecast for policy measures that
might be taken to mitigate the
impact of the recession.
Assuming that a major recession is on the way, the bank
could begin to simultaneously
stress test loan accounting system variables in the model
that could offset the expected
economic shock to the default
rate. For example, if the
average FICO score is a
variable in the default rate
model, assumptions could be
made to reflect more conservative cut-off score policies. If
the percentage of accounts
60+ days late is also in the

Actual Default Rate

model, assumptions could be

adjusted to reflect a more
aggressive treatment of delinquent accounts. In the average
balance model, assumptions
could be modified to reflect
the lowering of credit limits in
high-risk areas to minimize
exposures.
Summary
The New Basel Capital
Accord requires banks to keep
information on each loan from the
moment it is booked for the purpose of building and validating
risk-rating models. Fortunately,
this data can be aggregated and
supplemented in such a way as to
meet another Basel requirementstress testing. The idea
behind stress testing is to ensure
the bank has the necessary capital

Predicted Historical Default Rate

in reserve to cover unexpected
events. The procedure presented
in this article introduces a type of
regression model that can not only
be used for stress testing purposes, but also can serve as a strategic
planning tool to quantify policy
changes which hopefully would
mitigate recessionary or other
pressures on the portfolio.
Contact Morrison at
[email protected].
References
1 Blaschke, Jones, Majnoni, and Peria, 2001,
Stress Testing of Financial Systems: An Overview
of Issues, Methodologies, and FSAP Experiences,
IMF Working Paper WP/01/88.
2 Pindyck and Rubinfeld, 1981, Econometric
Models and Economic Forecasts, 2nd Edition,
McGraw-Hill.