I am trying to make model with logistic regression. My results are:

At first my results with CUTOFF 0.5 was: good borrowers predicted: 95.02% and 33.05% for bad borrowers.

then I realize that the predicton of bad borrowers are low only (33.05%) so I try to make better results so then I changed the cut off to 0.2 and the results was: good borrowers predicted: 74.97% and 75.91% for bad borrowers.

My question is why is also important to have high score of good borrowers for model? Are these results considered a good?

enter image description here

enter image description here

Yes it is true that is unbalanced is that not good sample then?

Yes overall model with cutoff at 0.5 does better but only predict 33.05% bad borrowers which is not good.

I do this in Eviews (which I regret) because you can not do ROC curve inside this software.

Why is important to predict good borrowers also not only bad borrowers? enter image description here

  • 2
    $\begingroup$ Sorry, but I simply do not understand... What is the question? $\endgroup$
    – Repmat
    Commented Mar 30, 2017 at 12:14
  • $\begingroup$ Why is important to also have high score dep 0 (good borrowers)? If I set cut off to 0.10 I will get the better results for dep 1 (bad borrowers) but will decrease score of good borrowers but my mentor told me that this score should not be below 0.69%. My question is why is this important? $\endgroup$
    – Roga Men
    Commented Mar 31, 2017 at 9:21
  • 1
    $\begingroup$ What are the consequences of calling a bad borrower a good one & vice versa? If there aren't any consequences you might as well not bother to specify a cut-off. If there are, you can use their relative desirability to decide an appropriate cut-off. $\endgroup$ Commented Mar 31, 2017 at 10:01
  • $\begingroup$ If the cutoff is 0.5 then model predict only 33% of bad borrowers. If the cutoff is 0.2 then model predict 75.91% of bad borrowers. I would like to make model that will avoid investing money to bad borrowers. So it is clear that you need to adjust cutoff right? And make the % of bad borrowers higher.... But I am wondering why it is also important to have high score of good borrowers. $\endgroup$
    – Roga Men
    Commented Apr 2, 2017 at 7:52

1 Answer 1


Your dataset is very unbalanced - there are a lot more good borrowers than bad borrowers in your dataset. In fact, 80.41% of samples are good borrowers.

Therefore, your model could trivially just predict that all borrowers are good and achieve 80.41% accuracy. Your model with cutoff at 0.5 does slightly better than that, with 82.89% accuracy (but what is the confidence interval?).

The cutoff you choose depends on what you are trying to get out of your model. You may want to plot a ROC curve to help you make that decision.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.