1
$\begingroup$

I have a dataset of about 30 features, most of which are profile information of insurance customers. There are a couple features that are time-dependent (guess you could call it time series) insurance premiums. These premiums exhibit a trend of going up.

Due to certain constraints, my training dataset spans 3 years from 2013 to 2015, but missing data in 2016. I've trained a gradient boosting regressor to predict the ROI on these customers at the current time (2017). But as it is right now, there is a pretty big gap from end of 2015 to today. We know that the premiums have been going up during 2016 (actually more rapidly than what we've seen in 2015 and before). So my model is basically trained on old premiums and it led to underestimated ROI on current customers.

Putting that complication of more rapid growth of premium aside, if we assume that the trend before 2015 continues the same way to today, what technic can I use to incorporate this information into the model training?

I know one way to engineer a time-dependent feature is to use EWMA. But I suspect it won't help much with GBM model. Maybe I should try some other models like GLMM?

$\endgroup$
  • $\begingroup$ What exactly is your premium variable? What are the units? $\endgroup$ – Matthew Gunn Mar 23 '17 at 19:07
  • $\begingroup$ @MatthewGunn It's integer in dollars. $\endgroup$ – breezymri Mar 23 '17 at 19:31
  • $\begingroup$ maybe this will help:datascience.stackexchange.com/questions/2368/… I post an answer about how to construct the aggregate feature on the time-dependent features. $\endgroup$ – wolfe Mar 27 '17 at 14:56
  • $\begingroup$ @feng I read your post. It was not clear to me what you wanted to state. Can you elaborate more? $\endgroup$ – breezymri Mar 28 '17 at 0:38
  • $\begingroup$ Considering you data, if you predict the 2017, there is no regressor data in 2017, how to do it? You must generate the regressor data by leading it then moving average. $\endgroup$ – wolfe Mar 28 '17 at 1:00

Your Answer

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

Browse other questions tagged or ask your own question.