1
$\begingroup$

I am trying to predict loan defaults with a fairly moderate-sized dataset. I will probably be using logistic regression and random forest.

I have around 35 variables and one of them classifies the type of the client: company or authorized individual.

The problem is that, for authorized individuals, some variables (such as turnover, assets, liabilities, etc) are missing, because an authorized individual should not have this stuff. Only a company can have turnover, assets, etc.

What do I do in this case? I cannot impute the missing values, but I also can't leave them empty. In the dataset there are about 80% companies and 20% authorized individuals. If I can't impute that data, should I just drop the rows in which we find authorized individuals altogether? Is there any other sophisticated method to make machine learning techniques (logistic regression and random forests) somehow ignore the empty values?

$\endgroup$
3
  • $\begingroup$ Is this really 1 model or 2? Should companies and individuals be in different models? I am not just speaking from the data view but from your business problem view? Are these two entities different, have different behavior, might take different business action depending on the result of the model? $\endgroup$
    – Craig
    Apr 20 at 10:22
  • $\begingroup$ Depending on the result of the model, the bank theoretically will choose whether to grant a loan or not. They indeed are different entities, as individuals do not have assets/liabilities/etc. If I do make 2 models, they will most likely have different accuracy, which is not that reliable. $\endgroup$
    – IcarusX
    Apr 20 at 15:26
  • $\begingroup$ But if many variables are missing in one large group, why will that be reliable? The different models may have different performance characteristics (accuracy being one and often a weak metric). Perhaps there is additional data that can be acquired for the individuals. Imputing or binning, makes it seem to a logistic regression that the value is known. Should put an indicator variable also. Having 1 model for loans to businesses and another for loans to people, each have different behaviors and data, would be a good experiment. $\endgroup$
    – Craig
    Apr 21 at 12:58

1 Answer 1

0
$\begingroup$

Do not ignore missing values. In your case, they carry important information. Consider (1) binning numeric variables, including a separate bin for 'missing', or (2) impute the missing values with 0, introducing a dummy variable for when the variable is 'missing'.

Point (1) results in a loss of information, but is most common and easiest to interpret. Point (2) reduces information loss, but leads to bias. I would consider (1).

$\endgroup$
3
  • $\begingroup$ Thank you for your input. Could I then somehow use the binned variables alongside numeric variables in decision trees and/or logistic regression? $\endgroup$
    – IcarusX
    Apr 20 at 6:55
  • $\begingroup$ Yes, use can use binned and continuous variables together (provided they're not the same variable). $\endgroup$
    – ralph
    Apr 20 at 8:19
  • $\begingroup$ Thank you, you've been super helpful. Just one more question. When binning the "missing", should I bin it as "0", then the other binning categories will start from 1? $\endgroup$
    – IcarusX
    Apr 21 at 8:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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