I am building a predictive model using logistic regression to predict if an applicant should be given a credit product based on their telecommunication data of the previous eight months postpaid subscribers only. But the behaviors I am trying to capture can vary with time for the entire population. Example: data usage, spending power, etc will vary with time. So the question is how to build a model that performs consistently across time. Else, a suggestion on how to refresh/update the model with time should help as well.

  • $\begingroup$ Interesting question. It would help to give a little more detail about the predictors involved. Some may be extensive, some intensive with respect to time (Intensive and extensive properties). If I understand right, the issue with, say, data usage is that an aggregate measure over a short interval is subject to fluctuations that convey little or no information about someone's suitability for the credit product, whereas an aggregate measure over a long interval includes information that may no longer be relevant. $\endgroup$ – Scortchi - Reinstate Monica Apr 28 '16 at 9:58

Every time you conduct the analysis, you could use the latest obtained data points, for instance, data collected in the last 180 days. I suppose the applicants' data is stored and updated on database maintained by you or a third party. So a script would help to do the routine work, i.e., data manipulation, etc.

  • 2
    $\begingroup$ Alternatively you can fit time as a flexible nonlinear effect and include all the data. $\endgroup$ – Frank Harrell Apr 28 '16 at 11:37
  • $\begingroup$ @FrankHarrell: Thanks Frank. Your suggestion sounds like a good one. Could you give more pointers on how to do this? I tried searching on the internet but didn't get much. A link/pdf or any resource/example would help $\endgroup$ – Clock Slave May 2 '16 at 9:35
  • $\begingroup$ Regression splines are described in my course notes at biostat.mc.vanderbilt.edu/rms $\endgroup$ – Frank Harrell May 2 '16 at 12:44

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.