Actuarial Outpost Modeling loss ratio with GLM, fit frequency as Tweedie?
 Register Blogs Wiki FAQ Calendar Search Today's Posts Mark Forums Read
 FlashChat Actuarial Discussion Preliminary Exams CAS/SOA Exams Cyberchat Around the World Suggestions

 General Actuarial Non-Specific Actuarial Topics - Before posting a thread, please browse over our other sections to see if there is a better fit, such as Careers - Employment, Actuarial Science Universities Forum or any of our other 100+ forums.

#1
07-09-2018, 12:44 PM
 mattcarp Member CAS Join Date: Sep 2016 Studying for Exam 5 College: UC Berkeley Posts: 301
Modeling loss ratio with GLM, fit frequency as Tweedie?

My boss wants to model loss ratio with GLM. I know many sources say that loss cost is best for GLM, but we need loss ratio.

I am thinking of doing a frequency/severity model where frequency = number of claims / premium and severity is as usual. The frequency component (excluding zeros) appears to be gamma distributed, so I'm thinking of modeling the entire frequency component as Tweedie.

What do you all think?
__________________
P FM VEE MFE C S OC1 5 OC2 6 7 8 9
#2
07-09-2018, 03:14 PM
 MoralHazard Member CAS Join Date: Jul 2011 Favorite beer: Sam Adams Rebel Rouser Posts: 94

So the frequency, which you are defining as # claims / premium, seems gamma-distributed. Is this because the claim count (i.e., numerator) is somehow gamma distributed? Or is it something about the distribution of premium (denominator) that is resulting in a gamma-distributed frequency? If the latter, I don't think using the Tweedie in your frequency GLM makes sense here. The distribution is meant to apply to the random element of your outcome -- but the premium is a known quantity. I think you should look at the distribution of claim counts when determining which distribution to use (and probably go with Poisson).
#3
07-09-2018, 05:32 PM
 PrinceNReserve Member CAS SOA Join Date: Aug 2012 Posts: 131

Guessing the reason you are using premium in the numerator is because you are predicting loss ratios?

If it were me, I'd try modeling the loss ratio directly instead. There's plenty of GLMs that produce predictions between 0 and 1. (in theory loss ratio's can be > 1, but I'd hope the premium being charged is at least >= losses in aggregate).
#4
07-11-2018, 02:10 PM
 BassFreq Member CAS Join Date: Jun 2003 Location: Chicago Studying for all eternity Favorite beer: Duff Posts: 1,708 Blog Entries: 2

I would keep it simple and model the loss ratio using premium as weight.
__________________
If at first you don't succeed, you have one data point.
Res ipsa loquitur, sed quid in infernos dicet?