97
Of the 1,662 applicants in the test set, the Bayesian logistic
regression model with non-
informative prior (Model 3) now rejects 282 and accepts 1,380 applicants (Table 4.23). 85
(30.1%) of the rejected applicants are in fact good. 143 (10.4%) of the accepted applicants
are bad. The overall classification error rate is 13.7%.
Comparison of the 3 models
Table 4.24 compares Models 1, 2 and 3.
Table 4.24
Comparison of Models 1, 2 and 3 when the cut-off probability is 0.48.
Model 1 Model 2 Model 3
Accepted
1352
1419
1380
Rejected
310
243
282
Error rate among accepted
10.1%
10.7%
10.4%
Error rate among rejected
34.5%
22.6%
30.1%
Total error rate
14.7%
12.5%
13.7%
The following can be deduced from Table 4.24:
-
Model 2 accepts the most applicants.
-
Model 1 rejects the most applicants.
-
Model 1 has the lowest error rate among the accepted applicants.
-
Model 2 has the lowest error rate among the rejected applicants.
-
Model 2 has the lowest total error rate.
The error rates amongst the accepted applicants are all fairly close for all the models. The
Bayesian logistic regression model with informative prior again has the lowest total error
rate, showing that the use of relevant prior information is beneficial.
98
4.6.3 Comparison of the two cut-off probabilities
The results are now compared when the two different cut-off probabilities are used, 0.3
and 0.48. The following conclusions are reached:
-
In all models, more applicants are accepted when the cut-off probability is 0.48 as opposed
to 0.3.
-
In all models, fewer applicants are rejected when the cut-off probability is 0.48 as opposed
to 0.3.
-
For
all models, the error rate among the accepted applicants
is higher when the cut-off
probability is 0.48 as opposed to 0.3. This shows that the error rate realized by the bank is
lower when a lower cut-off probability is used.
-
For all models, the error rate among the rejected applicants
is lower with the cut-off
probability is 0.48 as opposed to 0.3.
-
For all models, the total error rate is lower when the cut-off probability is 0.48 as opposed
to 0.3.
It appears that 0.48 is a better cut-off probability to use because it exposes the financial
institution to more people who will be good. This means that the financial institution will
be more profitable than one which uses a cut-off probability of 0.3. The difference in the
error rates among the accepted applicants for the two cut-off probabilities is around 2% for
each model. This is not big enough to opt for the conservative
approach of a cut-off
probability of 0.3. The financial institution may want to employ the more risk averse cut-
off probability if it expects the financial markets to become turbulent.
Do'stlaringiz bilan baham: