I'm building a auto loan probability of default model where the loan term could be 3 to 7 years and hence default can happen anytime in that interval. But we are a start-up and have only 3 years of historical/performance data. What techniques can i use to effectively build a PD model. Should i use synthetic data? I'm looking for generic answers or methodologies only.

  • $\begingroup$ You say you have 3 years of data. Does that mean you observe each loan from origination to origination + 36 months or does it mean that you observe a basket of loans with various origination dates, terms, and termination dates as they evolve for 36 months? $\endgroup$ – Bill Mar 2 '17 at 20:05
  • $\begingroup$ The latter. maximum i have is 36 months but depending on origination date it could be less than that too $\endgroup$ – IamNotLegend Mar 2 '17 at 21:13
  • $\begingroup$ Do you see any loans, say, starting 12 months after they originated and then see them until 48 months after they originated? Or is 36 months the maximum time from origination to end of data? $\endgroup$ – Bill Mar 2 '17 at 21:33
  • $\begingroup$ The latter. I have loans originating from 2014 Jan to 2016 Dec. So if it originated in Jan 2014 then it will have 36 months of performance, if it originated in 2015 Jan then it would have 24 months of data. $\endgroup$ – IamNotLegend Mar 2 '17 at 22:00
  • $\begingroup$ OK, then it will be necessary to incorporate information from outside your data. That is, if you never observe any loans 38 months old, then you can't estimate how often such loans default. Do you have information like this? Consulting reports. Government statistics. Something which gives default rates (maybe broken out by various factors) for loans older than 36 months? $\endgroup$ – Bill Mar 9 '17 at 17:02

There was a kaggle contest on this. Look at it.


Things to note

  • The test set is randomly ordered, not time-based, so this is essentially not long-time data either.
  • you have full and non-anonymized data, so you can likely engage this in a way that contest participants can't.
  • $\begingroup$ Thanks for sharing this. But did it have the same problem that i have - insufficient performance window. How would i handle that shortcoming in my data? $\endgroup$ – IamNotLegend Mar 2 '17 at 21:53
  • $\begingroup$ The goal is to predict default before the end of loan term, and the advantage of the contest is that it has, for some cases, full-term data for use in training/validation/model-selection. In execution, it will always have incomplete data, because if there were complete data then it isn't prediction anymore. $\endgroup$ – EngrStudent Mar 15 '17 at 19:12

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.