0
$\begingroup$

I am trying to build a non-parametric model to predict pure premium of an insurance policy. Pure premium is simply the expected claim amount of an insurance policy.The problem is, most insurance policies do not result in claims and so the data set has a lot of zeroes in the response variable: over 90% of the observations made no claim. Existing methods to this prediction specifies a tweedie distribution for the response and uses GLM. But I do want to use non parametric methods to do this.

Is there anything that I can do to a response variable like this one since all the machine learning algorithms I have tried so far are giving me horrible results.

$\endgroup$
2
  • $\begingroup$ How correlated your predictors are to the response? $\endgroup$
    – SmallChess
    Feb 10, 2017 at 2:37
  • $\begingroup$ Why don't you want to use the Tweedie distribution? $\endgroup$
    – Peter Flom
    Feb 10, 2017 at 13:49

1 Answer 1

1
$\begingroup$

And why do you want to use non-parametric methods?

I'm not an economist/econometrician, but they do have a bunch of statistical models that handle insurance claims where there are a bunch of 0s. A logical one might be a hurdle model, which employ a binary probability (logit or probit) model for the 0 cases and a truncated count model (i.e. Poisson or negative binomial) for the positive cases. Presumably you are in R, so here's an example package (I use Stata, so I am not familiar with which R packages are best for this model):

https://artax.karlin.mff.cuni.cz/r-help/library/pscl/html/hurdle.html

Healthcare claims tend to have a lot of 0s, but usually nowhere near 90% 0s. People tend to default to GLM or GEE with a gamma distribution and a log link, which I understand is a special case of the tweedie family of distributions. It's not a perfect model but it's a generally accepted one. But, it may also not apply to your case.

$\endgroup$
3
  • $\begingroup$ Thank you. The reason why I want to use non-parametric methods is just to investigate how they will do on data sets like this one. $\endgroup$
    – Scorgio345
    Feb 11, 2017 at 2:55
  • $\begingroup$ Fair enough. But, without assuming a parametric distribution, how are you going to predict premiums? $\endgroup$
    – Weiwen Ng
    Feb 12, 2017 at 3:34
  • $\begingroup$ use some sort of learning algoritm to do it $\endgroup$
    – Scorgio345
    Feb 13, 2017 at 4:27

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.