99
4.7.1 Cut-off probability of 0.3
The error rates of Models 1, 2 and 3 when the models are trained using differerent sampler
sizes and the cut-off probability is 0.3 are given in Figure 4.22.
Fig. 4.22
Error rates of Models 1, 2 and 3 when the models are trained using differerent
sampler sizes and the cut-off probability is 0.3.
From Figure 4.22, Models 1 and 3 appear to follow a similar pattern.
The error rate of
Model 3 is always below that of Model 1. The error rates of Models 1 and 3 appear to
decrease as the sample size of the “new” data increases. The Bayesian model with non-
informative prior thus appears to perform better than the logistic regression model. The
100
error rate of Model 2 is relatively stable. The error rate of this model is always below the
other two models (except when the sample size is 150, Model 3 has a lower error). This
shows that making use of prior information is very useful when the sample size is small. It
is expected that the error rates of these three models would converge as the sample size
increases.
When a financial institution is expanding into a new economic location, combining expert
information obtained from experience in the home country can be very useful.
4.7.2 Cut-off probability of 0.48
For a comparison, if a cut-off probability was chosen to minimize the
total error rate on the
validation set, the cut-off probability would have been 0.48. Using this cut-off probability
would mean more risk for the financial institution. When 0.48
is used as a cut-off
probability Figure 4.23 is obtained.
101
Fig. 4.23
Error rates of Models 1, 2 and 3 when the models are trained using differerent
sampler sizes and the cut-off probability is 0.48.
Figure 4.23 again shows Models 1 and 3 following a similar pattern. The trend of the error
rate decreasing as the sample size of the “new” data increases is also clear. The error rates
of Models 1 and 3 decrease as the sample size increases but then appear to level-off. The
error rate of Model 2 is again relatively constant. This graph confirms again that the use of
relevant prior information is very useful.
The total error rates when the cut-off probability is 0.48 are lower than when the cut-off
probability is 0.3 (Figures 4.22 and 4.23). Model 2 performs
better when a cut-off
probability of 0.48 is used.
102
Although the models perform better on the total error rate with
a cut-off probability of
0.48, the error rate amongst the accepted applicants is still
higher than when a cut-off
probability of 0.3 is used. This confirms that opting for a higher cut-off probability results
in accepting more applicants and thus results in more profits. It appears that it is better to
take on more risk and opt for a cut-off probability of 0.48.
Do'stlaringiz bilan baham: