I understand that there won't necessarily be a great answer to this, but I'd like to hear what people would do in this situation.

Here's the data situation: I have about 30K policy records of which there are about 400 claims associated with them. I have an additional 400 claims that can't be matched to a policy. This is due to really old data for which the policy information wasn't saved in a database. The point of the analysis is to see what the Loss Ratio would be for a potential policy - the loss ratio being loss/premium.

At this point, I've thought about maybe creating a loss distribution using all 800 claims and bootstrap from this distribution to get additional losses. Then I'd perhaps just generate random policies (along with some variables that I'd be testing) and just randomly assigning claims to them, and try doing a GLM from there. Obviously, this introduces quite a lot of bias, and I'm not really sure if this is the 'right' way at all... My statistical knowledge is limited, so please throw me some ideas that I could possibly try.

  • $\begingroup$ Do you have a sense if claims unmatched to a policy are more likely to come from certain types of policies? $\endgroup$ – Matthew Gunn Dec 20 '16 at 16:49
  • $\begingroup$ @MatthewGunn Claims unmatched to a policy is due to really old data which weren't saved in the database. $\endgroup$ – Farellia Dec 20 '16 at 17:04
  • 2
    $\begingroup$ What's the statistical question attached to the data? If you had sufficient data, what would you be doing? $\endgroup$ – Jon Dec 20 '16 at 17:10
  • 1
    $\begingroup$ What makes you feel like 800 isn't sufficient enough to calculate a loss ratio? Are you not using the entire population of claims? Bootstrapping wouldn't give you much more than what you already have. At best, some nice confidence intervals. $\endgroup$ – Jon Dec 21 '16 at 16:45
  • 1
    $\begingroup$ A sample size of 400 (50% of the population) should be sufficient. It's not big data, but you don't need a lot of data to estimate your statistics. As long as your estimates converge well, then you shouldn't have much to worry about. I work in fraud modeling, and because of the size of the data and computational effort needed to process the data, we can only use a very very small percentage of actual data. I don't see a need for concern in your situation. $\endgroup$ – Jon Dec 21 '16 at 23:10

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.