1
$\begingroup$

I am interested in modeling the probability of default (PD) of a loan product.

Data

  • I have a dataset going back several years. Most of the loans have reached their terminal state (paid off or default) but there is a considerable number that are still active.
  • Each observation represents a loan.
  • The dependent variable represents whether a loan paid off, defaulted, or is active as of the date the dataset was created.
  • There are also variables which I indend to use as explanatory factors.
  • The age of the loan at the time it paid off/defaulted is unavailable

If I model the probability of default using a logistic regression over the entire dataset, how should I treat the currently active loans? Should they be excluded from the training dataset, or modeled as a non-default state?

Improve this question
$\endgroup$
4
  • 1
    $\begingroup$ Depending on the problem at hand it might be possible to just ignore the active ones. However, in reality it is often the case that this yields models of a bad quality. The area you need to learn about is called “survival analysis”. In particular, there is a way how to turn any Boolean prediction model (RF, logistic regression, ...) into a survival like model (if I understood correctly): see benkuhn.net/survival-trees . You probably should try this or some native models around survival analysis... $\endgroup$
    – Fabian Werner
    Commented Feb 16, 2020 at 18:38
  • 1
    $\begingroup$ Thanks Fabian. I'm familiar with survival analysis, however the problem I have is that the loans in this data were created over the span of decades but I don't have the a time-based variable to estimate a hazard rate. The other wrinkle is that the hazard rate of these loans is hypothesized to have evolved over the years. $\endgroup$
    – MrT
    Commented Feb 16, 2020 at 21:26
  • $\begingroup$ So you don't have the time of origin of the loans? You only know that the loan is active on a given date, not how long it has been active for? There is no hope of obtaining that information? $\endgroup$
    – Matthew Drury
    Commented Feb 17, 2020 at 1:46
  • $\begingroup$ I have the date of origin, so I can derive how long a loan has been active for. However for loans that have become inactive, I do not have the date at which they became inactive. $\endgroup$
    – MrT
    Commented Feb 17, 2020 at 14:30

1 Answer 1

Reset to default
2
$\begingroup$

You could refine your question slightly to assess the probability of default by X months after the loan was created. That way, any loans which are still active after X months yield an outcome of 0.

Censored observations only tell you that they have not defaulted yet, and so to properly account for this, you would need to do a Cox regression.

Improve this answer
$\endgroup$
1
  • $\begingroup$ The issue is that I do not have data on when a loan defaulted. I can infer the current age of an active loan because I know when it was created. $\endgroup$
    – MrT
    Commented Feb 17, 2020 at 14:34

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.