Professional Documents
Culture Documents
Dennis Glennon
The opinions expressed in this paper are those of the authors and do not necessarily reflect those of
the Office of the Comptroller of the Currency. All errors are the responsibilities of the authors.
Filename
Agenda
2
2
Filename
Models Risk: What is it?
4
4
Filename
Managing Model Risk
5
5
Filename
Model Risk
6
6
Filename
Model Risk: Sound Modeling Practices
7
7
Filename
Models as Decision Tools
8
8
Filename
Model Selection: Which model is better?
Model 1 Model 2
1 3 5 7 9 1 5 4 4 11
y
y
1 1
[0.1] [0.3] [0.5] [0.7] [0.9] [0.08] [0.45] [0.44] [0.67] [0.92]
[bad rate] [bad rate]
11 6 5 2 1
9 7 5 3 1
0
0 10 20 30 40 50 60 70 80 90 100 0
0
10
20 30 40 50 60 70 80 90 100
Score (quintiles) Score (quintiles)
[#B / (#G + #B)] 9
9
Filename
Models as Decision Tools
4
3
3 Odds: 33:1
Bad %: 3.0%
2 2
Odds: 12.2:1
Bad %: 7.6%
1 1
Score: 253
0 0
320 300 280 260 240 225 309 289 269 249 229 209
score
10
Filename
Illustrative Example
Risk-Rating Model
7
ln(20/1) = 3.0
3
bad rate = 5%
2
ln(4/1) = 1.4
1
bad rate = 20%
0
644 653 665 675 684 693 706 715 725 739 753
Score Bands
11
Filename
Models as Decision Tools
13
13
Filename
Model Validation
The classification objective is the weaker of two
conditions.
There are well-developed methods outlined in the
literature and accepted by the industry that are used to
assess the validity of models developed under that
objective.
In practice, we see:
Development
KS / Gini used as the primary model selection tool
These evaluated on the development, hold out, and out-
of-time samples
Validation
KS / KS
Stability test (e.g., PSI, characteristic analysis, etc.)
Backtesting analysis 14
14
Filename
Model Validation
Almost all scoring models generate KS values that reject the null
that the distribution of good accounts is equal to the distribution
of bads.
15
15
Filename
Model Validation
The test that have been developed, however, tend to
be sensitive to sample size. Given the size of
development and validation samples, very small
changes may be statistically significant.
16
16
Filename
Model Validation
Predictive models are developed under a model
accuracy objective.
As a result, a goodness-of-fit test is required for model
selection.
Common performance measures used to evaluate
predictive models:
Interval Test
Chi-Square Test
18
18
Filename
Vasicek Test: An Example
Upper Bound
Segment Accounts Estimated PD Actual PD Vasicek 95% CI
19
Filename
Model Validation: Vasicek Test
If is too high the bands are too wide: too
many models would pass the test
21
21
Filename
Illustrative Example
Default Rates
40.00
30.00
default rate
20.00
10.00
0.00 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
actual 34.11 22.75 16.18 13.63 11.44 9.84 9.07 7.60 7.35 6.83 6.37 5.72 5.41 4.49 4.41 3.57 3.44 2.93 2.27 1.14
predicted 34.56 22.55 16.59 13.27 11.48 9.86 8.65 7.90 7.16 6.54 5.97 5.50 5.04 4.64 4.14 3.76 3.38 2.96 2.46 1.43
score range
22
22
Filename
Illustrative Example
HL stat 27.0433
p-value 0.0782
23
23
Filename
Illustrative Example
p-values
1.00
0.80
0.60
p
0.40
0.20
0.00 11
13
15
17
19
1
24
24
Filename
Interval Tests with Large Samples
Conclusion:
Statistical difference: significant
Solutions?
Reduce the number observations using a
Interval test
Focus on capital
25
25
Filename
Interval Tests with Large Samples
(5)
(4)
(3)
(2)
(1)
-1% 0 +1%
26
26
Filename
Interval Test
Restate the null as an interval defined over an
economically acceptable range
If the CI1- around the point estimate is within the in
interval, conclude no economically significant
difference
May want to reformulate the interval test in terms of
an acceptable economic bias in the calculation of
regulatory capital
27
27
Filename
Conclusion
28
28
Filename