2
$\begingroup$

I have a 5 years of Cargo insurance (goods transportation insurance) data.

I need to predict the claim amount based on their policy date and some other variables like mode of transportation, Country Origin, destination, Package mode, Premium paid,Sum Insured, and type of goods.

As I said, my objective is to predict the claim amount based on all the above variables.

For this case we can't directly predict the claim amount, so I built a predictive model (Classification) to predict the claim and non-claim transactions (i.e. the transaction (insurance) which will end up in claim and the transaction which doesn't end in a claim). As of now,the model is 74% accurate in classifying the claim and no-claim transactions.

Now I want to predict the claim amount (in monetary value) ... I don't know in which direction I should move ahead. Can any modelling technique help to predict the exact amount of claim based on all the given parameters?

Thanks

$\endgroup$
  • $\begingroup$ @Glen_b Thx i have edited my content.... $\endgroup$ – leeya Dec 10 '14 at 6:55
3
$\begingroup$

You'll never predict the exact amount of claim. However, you may be able to predict a reasonable amount of the variation in claim size.

A common way to predict claim size would be via a generalized linear model (Gamma claim size is fairly common), or possibly a regression model for log-claim (most typically, lognormal claim size, which would be normal after taking logs) - in either case conditional on there being a nonzero-claim.

Sometimes log-gamma, log-Weibull, Pareto or inverse Gaussian distributions might be used. Many other different distributions are used for claims size in other situations.

If you try to predict claim size including the claims on which 0 payment is made (for whatever reason), you'd need something like a zero-inflated model.

$\endgroup$
  • $\begingroup$ Thanks a lot :) I just read through few papers on zero-inflated model based on ur suggestions,I hope this help my work.Thanks again. $\endgroup$ – leeya Dec 10 '14 at 9:24
  • 1
    $\begingroup$ Consider a Heckman two-stage model, and note that classification accuracy is not a good accuracy measure. $\endgroup$ – Frank Harrell Dec 10 '14 at 11:58
  • $\begingroup$ @Frank Harrell Yeah, its becoz of highly unbalanced data set. Planned to do random sampling,i hope it might help to improve the accuracy. $\endgroup$ – leeya Dec 11 '14 at 2:24
  • $\begingroup$ Frank Harrell wasn't saying "your classification accuracy is poor", he was saying "that's a bad way to measure it". $\endgroup$ – Glen_b Jul 7 '15 at 0:04
1
$\begingroup$

The claims are often modeled as probability of claim and claim amounts. The probability of claims are often modeled either with logit regression or with survival models. The claim amount (given that the claim was filed) is often modeled as a simple regression.

So, if I were you (and I was some time ago), I'd start with a simple regression on relevant variables. I modeled loss given defaults (LGD) on mortgages, and for me the relevant parameters were things like Zip code, house value (indexed), loan outstanding, etc.

Things to be careful about is restrictions such as nonnegativity of the claim and max amount. So, LGD is usually modeled as a percentage of the outstanding balance on a loan, in your case it's the max amount. LGD can go higher than 100%, but not by too much. It should not go below 0, but in the data there are always negative amounts in observations. There are many way when dealing with restrictions. Start with the simplest one: floors and ceilings, i.e. run unrestricted regression, then simply put floor and ceiling to the output.

The key is to start with the simplest models when you're a beginner in this area. You'll progressivly make more sophisticated models as time goes by, and you learn more about the domain.

$\endgroup$
  • $\begingroup$ Thanks ... In my case there are lot of categorical variables than numerical.I have few numerical variable like claim amount,premium paid and sum insured.So,The claim amount is modeled with both classification and regression model.Classification will be used to predict the instance with no claim (Y=0) and claim (Y=1).Next step will be removing the zero claim instances,apply some regression model on the rest.I dont think there is a need for LGD in my case.If you think its still possible please add ur suggestion....Thanks again. $\endgroup$ – leeya Dec 12 '14 at 3:43

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.