

FlashChat  Actuarial Discussion  Preliminary Exams  CAS/SOA Exams  Cyberchat  Around the World  Suggestions 

Thread Tools  Search this Thread  Display Modes 
#11




Quote:
m1 < glm(target~1+predictor1+predictor2 ,data=trainingDataframe ,family=tweedie(var.power=1.6,link.power=0) ,weights=weightVariable) summary.m1 < summary(m1) summary.m1$aic <AICtweedie(m1,2) summary.m1
__________________

#12




Nice, thanks! I'll give it another shot at some point. Glad to see someone else ballparks a 1.6 there as well (which makes intuitive sense anyway).
__________________

#13




Quote:
Thanks! Another quick question. If you're modeling pure premium as the target, do you continue to use the exposure as the weight? I think the math checks out given the tweedie distribution, but intuitively seems like you're double counting exposure. 
#14




Quote:
You could tinker around with weights, though. 
#16




Quote:
I think of it like an aggregated pure premium calculation. To get the pure premium for the whole book, we sum up all of the loss divided by the sum of all of the exposures. Another way to calculate the same figure would be a weighted average of the pure premium for each policy, where the weight is the exposure. If we calculated a straight average of pure premium, it would not be the actual book's pure premium because not enough weight was given to large exposure policies, and too much weight was given to small exposure policies. 
#18




You're welcome. I would recommend trying out some liftbased measures. A double lift curve will hopefully show that your model is better than the current rater. Model accuracy may be very poor from a traditional perspective, but if it differentiates between your best and worst insureds, it should save your company money. The goal here is reducing adverse selection, not predicting actual policy losses.

#20




I would recommend the liftbased measures in Chapter 7 of the GLM monograph for comparing models. Within the Tweedie GLM, I occasionally peak at AIC to make sure that the model is getting better instead of worse when I add a variable. Though, I mainly use lift charts and Gini coefficients for that, too. I don't use cross validation in the automated sense, but sometimes I refit the same variables on different subsets of my training data to check for coefficient and pvalue instability.

Tags 
glm tweedie, weights 
Thread Tools  Search this Thread 
Display Modes  

