Automobile Loan
Credit Model
Matt Bean
Matt Caspermeyer
Venkat Sakruti
Methodology
50% – 50% Data Split into training and validation
Dummy variable creation
Performed discriminant analysis via regression
in SAS
Removed various variables based on following
analysis:
Ensured coefficients of variables were logical
Eliminated highly correlated variables
Eliminated insignificant variables
Final model consisted of 48 dummy values
derived from 13 variables
Scorecard
Variable Name Interval Points Variable Name Interval Points
Intercept (all applicants start with this score) Intercept 668 0 83
$0-1200 -31 1-3 32
$1201-$1600 0 4-6 0
Amount of down payment $1601 + 40 # of trades opened in last 24 months 7+ -62
0-1 0 0 35
2-3 -18 1-5 0
# of derogatory public records 4+ -40 6-15 -35
0-6 -48 # of trades ever rated as bad debt 16+ -88
7-48 0 .01% - 40% 45
Time at residence (in months) 49+ 10 40.01% - 75% 0
0-9 -51 Ratio of balance to credit limit for open auto 75.01% - 90% -69
10-59 0 trades 90.01% + -88
Time at job (in months) 60+ 14 Car 0
0-23 -32 Auto Type Truck 50
24-35 0 0% -39
36-49 39 .01% - 40% 0
50-54 47 40.01% - 50% 42
Customer age (in years) 55+ 99 Ratio of satisfactory trades to total trades 50.01% + 75
0-4 87 0-1 -31
5-6 48 2-5 0
7 21 6-7 43
8 0 # of revolving trades 8+ 93
9 -46 $0 - $1000 0
10-12 -69 Value of trade in vehicle $1001 + 38
Age of car (in years) 13+ -131
K-S Statistic
Measures maximum separation between
the cumulative percentage distributions of
known good/bad applicants based on
model score.
K-S Test – Training Dataset
KS Test for Training Dataset
100
90
80
70
60
Cumulative %
Cumulative % good
50
Cumulative % bad
40
30
20
10
0
E
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
0
97
92
87
82
77
72
67
62
57
52
47
42
37
32
27
20
R
O
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
M
R
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
0
O
95
90
85
80
75
70
65
60
55
50
45
40
35
30
25
01
10
Score Ranges
Maximum separation of .2497 at score range of 651 to 675
K-S Test – Validation Dataset KS Test for Validation Data Set
100
90
80
70
60
Cumulative %
Cumulative % good
50
Cumulative % bad
40
30
20
10
0
E
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
97
92
87
82
77
72
67
62
57
52
47
42
37
32
22
R
O
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
M
R
1
1
1
1
1
1
1
1
1
1
1
1
1
0
0
O
95
90
85
80
75
70
65
60
55
50
45
40
35
30
20
01
10
Score Ranges
•Maximum separation of .2345 at score range of 651 to 675.
•Separation value is 6% lower than training dataset.
K-S Test – Combined Dataset
KS Test for Combined Dataset
100
90
80
70
60
Cumulative %
Cumulative % good
50
Cumulative % bad
40
30
20
10
0
E
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
0
97
92
87
82
77
72
67
62
57
52
47
42
37
32
27
20
R
O
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
TO
M
R
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
0
O
95
90
85
80
75
70
65
60
55
50
45
40
35
30
25
01
10
Score Ranges
Maximum separation of .2362 at score range of 651 to 675
How did the model perform against
known behavior types?
651+ Actual Data
Good Bad Totals
n
io
ct
Good 6147 1908 8055
i
ed
Pr
Bad 3295 2692 5987
Totals 9442 4600 14042
Success Rate 651+ 62.9%
Accepted at 651+ 8055
% Bad at 651+ 23.7%
If Accepted all 14042
% Bad would be 32.8%
Shift acceptance to 626+
Results in 301 additional Good customers with
1.7% more Bad customers than 651+
626+ Actual Data
Good Bad Totals
n
io
ct
Good 6875 2335 9210
i
ed
Pr
Bad 2567 2265 4832
Totals 9442 4600 14042
Success Rate 626+ 65.1%
Accepted at 626+ 9210
% Bad at 626+ 25.4%