SAS Modeling Tools

To Assess Credit Risk

July 2011

1
Introduction

• Presenter: Kevin Hougas / [email protected]

• Company: U.S. Bank
• Line of Business: Credit Administration Risk Analytics
• Key Responsibility: Assessing Credit Risk

• Discussion Points
– Overview of Credit Risk
– Proc LOGISTIC
– Proc REG
– Proc ARIMA

2
Credit Risk

• There are 3 main analyses we perform for Credit Risk

1.) Basel II Regulatory Capital
– Extreme loss event using Stochastic process
2.) Economic Capital
– Extreme loss event using Deterministic process
3.) Stress Testing
– Multiple loss events using macroeconomic scenarios

• In a nutshell…
Produce loss estimates that are sensitive to multiple risk drivers

3
Methods Used to Produce Loss Estimates

• Component Method (Proc LOGISTIC, Proc REG)

– Probability of Default (PD)
– Loss Given Default (LGD)
– Exposure at Default (EAD)

For any given borrower,

Loss = PD x LGD x EAD

• Aggregated Method (Proc ARIMA)

– Work directly with observed losses as a percentage of outstanding balance

Loss % = Loss ÷ Balance

4
Model to Produce PD Estimates (Proc LOGISTIC)

• Probability of Default (PD) represents the likelihood that a borrower will

default on their obligation over the next year.
• Proc LOGISTIC estimates the sensitivity of this likelihood to different risk
drivers using logistic regression
– Estimates from logistic regression are constrained between 0% and 100%.

• Data for specifying a logistic regression model can be seen in the table
below.

Identification Variables Modeling Variables

(Dependent Variable) (Independent Variables)
Borrower # Status Beginning of Year Status End of Year Default Indicator Risk Rating S&P 500
1 Active Active 0 2 1,000
2 Active In Default 1 6 700
3 Active Active 0 1 1,200
4 Active Active 0 2 1,300
5 Active Active 0 3 900

5
Sample SAS Code for Proc LOGISTIC

PROC LOGISTIC DATA=Modeling_Data DESCENDING;

CLASS Risk_Rating (PARAM=REF REF='8');

MODEL
default_indicator=

Risk_Rating
SP500;

output out=scoreseg p=phat lower=lcl upper=ucl;

RUN;

6
Sample Output from Proc LOGISTIC

LOGIT = a0 + a1 X 1 + a2 X 2 + ...an X n

1
PD =
1 + exp(−( LOGIT ))

Analysis of Maximum Likelihood Estimates

Standard Wald
Parameter DF Estimate Error Chi-Square Pr > ChiSq
Intercept 1 0.490 0.089 31 <.0001
Risk Rating 1 1 -15.238 77.848 0 0.845
Risk Rating 2 1 -5.126 0.303 287 <.0001
Risk Rating 3 1 -4.462 0.145 947 <.0001
Risk Rating 4 1 -3.933 0.062 4,079 <.0001
Risk Rating 5 1 -3.152 0.039 6,682 <.0001
Risk Rating 55 1 -2.796 0.100 777 <.0001
Risk Rating 6 1 -2.089 0.054 1,485 <.0001
Risk Rating 7 1 -1.306 0.052 631 <.0001
S&P 500 1 -0.002 0.000 545 <.0001

7
Sample Output from Proc LOGISTIC

• Key Output Metric: AUC = 83.5%

Acutal vs. Estimated PD

PD

8
1

8
-0

-0

-0

-0

-0

-0

-0
-0

-0

-0

-0

-0

-0

-0

-0
n

n
ec

ec

ec

ec

ec

ec

ec

ec
Ju

Ju

Ju

Ju

Ju

Ju

Ju
D

D
Actual DF Estimated PD

8
Model to Produce LGD Estimates (Proc REG)

• Loss Given Default (LGD) is the % of remaining balance the bank will lose if
the borrower defaults on their loan
• Proc REG estimates the sensitivity of LGD to different risk drivers using
linear regression
– Estimates from linear regression are not constrained.

• Data for specifying a linear regression model can be seen in the table below.
Identification Variables Modeling Variables
(Dependent Variable) (Independent Variable)
Date Total Balance at Default Total Amount Recovered Average LGD HPI % Change
Jan-02 10,000 1,000 90% -20%
Feb-02 10,000 2,000 80% -10%
Mar-02 10,000 3,000 70% 0%
Apr-02 10,000 4,000 60% 10%
May-02 10,000 5,000 50% 20%

9
Sample SAS Code for Proc REG

PROC REG DATA=Modeling_Data;

MODEL
LGD =

HPI_Change
;

output out=scoreseg p=phat;

run;

10
Sample Output from Proc REG

LGD = a0 + a1 X 1 + ...an X n

Parameter Estimates
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
Intercept 1 0.9030 0.0026 346 <.0001
HPI_Change 1 -0.4092 0.0228 (18) <.0001

11
 Ja LGD
n-
0
M 2
ay
-0
Se 2
p-
02
Ja
n-
M 03
ay
-0
Se 3
p-
03
Ja
n-
M 04
ay
-0
Se 4
p-
0
Ja 4
n-
0
M 5
Sample Output from Proc REG

Actual LGD
• Key Output Metric: R-Square = 81.3%

ay
-0
Se 5
p-
0
Ja 5
n-
0
M 6
Acutal vs. Estimated LGD

ay
-0
Estimated LGD Se 6
p-
06
Ja
n-
M 07
ay
-0
Se 7
p-
07
Ja
n-
08
12
Loss Estimate Using Component Method

PD - Derived from logistic regression model

LGD - Derived from linear regression model

EAD - Equal to outstanding balance at default

Loss = PD x LGD x EAD

13
Loss Estimates Using Aggregated Method (Proc ARIMA)

• Loss % is the ratio of observed losses to outstanding balance for a given

quarter
• Proc ARIMA estimates the sensitivity of Loss % to different risk drivers using
an Auto Regressive Integrated Moving Average model
– ARIMA requires a time series data source
– Can be used to address serial correlation and specify lag structures

• Data for specifying an ARIMA model can be seen in the table below.
Identification Variables Modeling Variables
(Dependent Variable) (Independent Variable)
Date Balance Loss Amount Loss % Unemployment
Jan-02 10,000 1,000 10% 6%
Apr-02 10,000 2,000 20% 7%
Jul-02 10,000 3,000 30% 8%
Oct-02 10,000 4,000 40% 9%
Jan-03 10,000 5,000 50% 10%

14
 Sample SAS Code for Proc ARIMA

ODS Graphics On;

proc arima data=Modeling_Data;

identify var=Loss_Rate crosscorr=(Unemp);

estimate method=ML plot;

run;

estimate p=2 q=0

input=(

2$Unemp
)

method=ML plot;
forecast lead=12 id=date interval=quarter out=scoreseg printall;

quit;
15
run;
Sample Output from Proc ARIMA

Crosscorrelations to Unemployment
Lag Covariance Correlation -1 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6 7 8 9 1 
0 0.00013 94% | . |******************* |
1 0.00012 88% | . |****************** |
2 0.00010 78% | . |**************** |
3 0.00009 65% | . |************* |
4 0.00007 51% | . |********** |
5 0.00005 36% | . |******* |
6 0.00003 20% | . |**** . |
7 0.00001 7% | . |* . |

Loss _ Rate = a0 + a1 X 1 + ...an X n + Z t

Z t = θ1Z t −1 + ...θ n Z t − n + ε t
Maximum Likelihood Estimation
Approx
Parameter Estimate Standard Error t Value Pr > |t| Lag Variable Shift
MU -0.0004185 0.0056068 -0.07 0.9405 0 Residential_mortgages 0
AR1,1 1.47421 0.15021 9.81 <.0001 1 Residential_mortgages 0
AR1,2 -0.51639 0.14749 -3.5 0.0005 2 Residential_mortgages 0
Unemp 0.12586 0.06794 1.85 0.064 0 Unemp 2
16
Sample Output from Proc ARIMA

• Forecast assumes Unemployment falls to 7.0%

• Key Output Metric: AIC = -414, White Noise = 20%

Acutal vs. Estimated Loss %

Loss %
1

3
-0

-0

-0

-0

-0

-0

-0

-0

-0

-1

-1

-1

-1
ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec
D

D
Actual Loss % Estimated Loss %

17
Review

• Proc LOGISTIC
– Estimates the likelihood of Yes or No events

• Proc REG
– Estimates the relationship between a continuous dependent variable and multiple
independent variables

• Proc ARIMA
– Forecasts time series data while accounting for serial correlation and lag structure

18
Questions?

19