KPI/default in bank risk model, how to deal with extra KPI besides accept and refused?

[lindamtl]

when use proc logistic procedure to set up default model in bank risk analysis, for the current customer, there are 4 kinds of results after application for the loan, accept, refuse , cancel or nouse.

when set up model, to analysis if new customer will default or could approve their application of loan, we will use the current customer loan history.

refused=1,

accept=0

but how to deal with cancel and nouse, i think nouse looked as accepted, but for cancel, how to deal with it? 

[Ksharp]

It is a predicted model. If you have more than two levels of response variable,you could try
Decision Tree:
proc hpsplit
Random Forest:
proc hpforest

or
PROC PLS
or
PROC TRAJ
https://www.andrew.cmu.edu/user/bjones/refpdf/ref2.pdf

[PaigeMiller]

I'm not entirely clear on the meaning of these four words in your situation:

accept, refuse , cancel or nouse

Could you give a brief explanation of each one? And what is KPI?

Then you say

to analysis if new customer will default or could approve their application of loan, we will use the current customer loan history

If you want the logistic regression (or other model) to predict probability of default (is that what you want, even that isn't clear), then  this doesn't make sense to me either

refused=1,

accept=0

Please explain further.

[lindamtl]

I mean, need use current customer's history to predict the new customer. For the current customer, when they apply for loan, there should be only 2 answers from bank, accept or refused.

so could flag,

accept=0,

refused=1

to set up model, and drop the new customer's info into this model to see if accept or refuse the new application. 

But now, there is more value in the column: result of application of loan, accept, refused, cancel, Nouse(look as accept).

in order to set up model, there is only binary value, how to deal with cancel, give it 0 or 1?

[PaigeMiller]

I'm still not following. If you want to create a model which predicts the risk of a new customer defaulting, the model you would build requires a response variable which is whether or not the loan defaulted. Yet you seem to want to use these categories of accept, refuse , cancel or nouse (which I asked you to explain but you have not yet explained) as the response variable.

Or are you talking about these as predictor variables? The whole thing is unclear.

[SASKiwi]

If a customer cancels or refuses a lending application then there will be no loan(s) as a consequence of that application so no loan(s) to model default on. I would be discarding canceled or refused applications from modelling.

If on the other hand a customer has other existing loans or you are rolling up applications across all loans you may want a different treatment. You haven't explained your data enough to draw any conclusions like this though. 

[lindamtl]

when set up model with old customer's loan application data, there is this status, one customer maybe applied many times, for different reason, for mortage, for car or other things, some were accepted, some were refused, some were cancelled by themselves.

Need use model to know if could approve loan to new customer. If similar customer were refused, so this kind of new customer, we will not aprrove.

so these questions i want to know

1, when sent up model, in logistic regression, shoulde put default=1 into model, for refused=1, for noused=1, how to do for cancel, sign 1 or 0?

2, when one customer has different loan application history and with different result, accept/refused, how to define this customer, give **bleep** 1 or 0 to him/her. Or depends on what typ of loan he applied and the result.

3, if want to work out if need approve customer's application, need set up different model based on different loan reason, like mortage, use mortage data to set up a model, then test the new customer; then set up car loan model, then to test customer who apply for car loan?

[SASKiwi]

I assumed you were talking about loan defaults which are defined as a loan being 90 days past due on repayments or 90 days over limit. However it appears you are actually modelling customer applications. What does "default" mean for customer loan applications?

[PaigeMiller]

1, when sent up model, in logistic regression, shoulde put default=1 into model, for refused=1, for noused=1, how to do for cancel, sign 1 or 0?

Are you creating a new model based upon data to predict default? Or are you using an existing model that someone (maybe you, maybe someone else), has created to predict default?

Is this variable that takes on values accept, refuse , cancel or nouse a predictor variable in the model, or is it the response variable? Please (I have asked twice before) define accept, define refuse, define cancel and define nouse.

[lindamtl]

i think should clearly more.

1, Purpose-verify if give approval to new loan application, include current customer or new customer.

2, Need use old application history result to set up a model, to see who is refused, so that use this model to check if need give approval to new application, or refused.

3, For the history record, there are 4 results on application , approval, refused, Noused, Cancelled. Approval and Noused looked as approve, how to deal with cancel, as this percentage is not small, perhaps some are qualified, some not. that is the problem.

4, Need set up customer ID level, on the old history list, one customer may has several records, some for car, some for house, etc, some was approve, some not, so  how to choose one record which is should to be chosen.