Actuarial Outpost
 
Go Back   Actuarial Outpost > Actuarial Discussion Forum > Property - Casualty / General Insurance
FlashChat Actuarial Discussion Preliminary Exams CAS/SOA Exams Cyberchat Around the World Suggestions


Reply
 
Thread Tools Search this Thread Display Modes
  #21  
Old 04-12-2019, 08:24 AM
Vorian Atreides's Avatar
Vorian Atreides Vorian Atreides is offline
Wiki/Note Contributor
CAS
 
Join Date: Apr 2005
Location: As far as 3 cups of sugar will take you
Studying for ACAS
College: Hard Knocks
Favorite beer: Most German dark lagers
Posts: 65,751
Default

Quote:
Originally Posted by Actuarially Me View Post
Interesting. Isn't that kind of the argument for splitting models by peril or territory?
Gotta remember that there's a trade off between credibility and accuracy when you partition your data.
__________________
I find your lack of faith disturbing

Why should I worry about dying? Itís not going to happen in my lifetime!


Freedom of speech is not a license to discourtesy

#BLACKMATTERLIVES
Reply With Quote
  #22  
Old 04-12-2019, 09:46 AM
Actuarially Me Actuarially Me is offline
Member
CAS
 
Join Date: Jun 2013
Posts: 191
Default

Quote:
Originally Posted by Vorian Atreides View Post
Gotta remember that there's a trade off between credibility and accuracy when you partition your data.
Yeah, I'm wondering if I even have enough credible data for by Peril models.
Reply With Quote
  #23  
Old 04-12-2019, 09:55 AM
Vorian Atreides's Avatar
Vorian Atreides Vorian Atreides is offline
Wiki/Note Contributor
CAS
 
Join Date: Apr 2005
Location: As far as 3 cups of sugar will take you
Studying for ACAS
College: Hard Knocks
Favorite beer: Most German dark lagers
Posts: 65,751
Default

Quote:
Originally Posted by Actuarially Me View Post
Yeah, I'm wondering if I even have enough credible data for by Peril models.
Without know what the perils are, that is tough to tell.

Rough guidelines: look at the Coeff of Var for the data. If it's fairly large (much greater than 1.0), you're going to need a lot of claim data to get something reasonable for your severity model.
__________________
I find your lack of faith disturbing

Why should I worry about dying? Itís not going to happen in my lifetime!


Freedom of speech is not a license to discourtesy

#BLACKMATTERLIVES
Reply With Quote
  #24  
Old 04-12-2019, 10:41 AM
Actuarially Me Actuarially Me is offline
Member
CAS
 
Join Date: Jun 2013
Posts: 191
Default

Quote:
Originally Posted by Vorian Atreides View Post
Without know what the perils are, that is tough to tell.

Rough guidelines: look at the Coeff of Var for the data. If it's fairly large (much greater than 1.0), you're going to need a lot of claim data to get something reasonable for your severity model.
sd(Total Incurred) / mean(Total Incurred) is 6.98 lol

For each peril it's, 10.77, 11.58, and 15.29. The latter makes sense why my initial severity model was complete bonkers.

If I include the prior years, it makes it worse. Plus, the book of business has significantly changed post 2011. What do you do in that case?


Also, how did you go about learning everything? I've read most of the relevant material, but never took the relevant exams. Should I brush up on more of the theoretical aspects of GLMs? Any suggested resources?
Reply With Quote
  #25  
Old 04-12-2019, 10:51 AM
Vorian Atreides's Avatar
Vorian Atreides Vorian Atreides is offline
Wiki/Note Contributor
CAS
 
Join Date: Apr 2005
Location: As far as 3 cups of sugar will take you
Studying for ACAS
College: Hard Knocks
Favorite beer: Most German dark lagers
Posts: 65,751
Default

Quote:
Originally Posted by Actuarially Me View Post
sd(Total Incurred) / mean(Total Incurred) is 6.98 lol

For each peril it's, 10.77, 11.58, and 15.29. The latter makes sense why my initial severity model was complete bonkers.

If I include the prior years, it makes it worse. Plus, the book of business has significantly changed post 2011. What do you do in that case?
Simplest answer: use "year" as a control variable. Or some indicator of "before" and "after" the "significant change" took place.


Quote:
Originally Posted by Actuarially Me View Post
Also, how did you go about learning everything? I've read most of the relevant material, but never took the relevant exams. Should I brush up on more of the theoretical aspects of GLMs? Any suggested resources?
Doing lots and lots of models. Keep in mind that the best way to better understand your data wrt modeling is to create different models (all-peril, by-peril, pp, f x s, etc.) and look at results and see if you can explain (and verify your explanation with the data) the differences.

Reading source material is only a start, but it's not going to get you the understanding you're going to need to be successful.
__________________
I find your lack of faith disturbing

Why should I worry about dying? Itís not going to happen in my lifetime!


Freedom of speech is not a license to discourtesy

#BLACKMATTERLIVES
Reply With Quote
  #26  
Old 04-12-2019, 11:03 AM
Actuarially Me Actuarially Me is offline
Member
CAS
 
Join Date: Jun 2013
Posts: 191
Default

Quote:
Originally Posted by Vorian Atreides View Post

Doing lots and lots of models. Keep in mind that the best way to better understand your data wrt modeling is to create different models (all-peril, by-peril, pp, f x s, etc.) and look at results and see if you can explain (and verify your explanation with the data) the differences.

Reading source material is only a start, but it's not going to get you the understanding you're going to need to be successful.
Thanks!

Another problem with including the older years is that data is incomplete for some variables, so I'd have to impute or drop most of them anyways.

Also, I forgot to filter out the zeros when doing CV. Looks much better now.
Reply With Quote
  #27  
Old 04-12-2019, 04:30 PM
Actuarially Me Actuarially Me is offline
Member
CAS
 
Join Date: Jun 2013
Posts: 191
Default

Thanks for all of the advice for who contributed so far.

I'm finally feeling like I'm heading in the right direction. I wanted to update with some results of my progress for anyone who comes across the post in the future. With advice taken and some other research on proper diagnostics, I tested a poisson, quasipoisson, zeroinflated, and negativebinomial. I also included more data and used Policy Year as a control.

Turns out negative binomial was a better fit for my model.

I think the crunched residuals look a lot more random



I also used have an influence plot that points out 6 points. Looking at the individual policies, they are either outliers in terms of the explanatory variables or ones with a large number of claims. I don't see any reason to exclude them, but maybe someone sees something I don't?

https://imgur.com/0bGZt2j

I also found the DHARMA package that creates scaled quantile residual plots using a simulation. The QQ plot looks much better.

https://imgur.com/sswmXT5


It also has a different test for over/under dispersion using the simulated vs fitted residuals. This doesn't look too bad.

https://imgur.com/Wwy3tq9

A test of zero inflation looks good.

https://imgur.com/Wwy3tq9

So while there's more iterations to do to improve the model, I feel I have a better foundation to start with. For some reason the pictures weren't embedding, so I provided the links. Thanks everyone for your help!
Reply With Quote
  #28  
Old 04-12-2019, 05:43 PM
Tacoactuary's Avatar
Tacoactuary Tacoactuary is offline
Member
CAS
 
Join Date: Nov 2014
Location: Des Moines, IA
College: Vanderbilt, UIUC
Favorite beer: Yazoo Sue
Posts: 1,591
Default

Not well versed in DHARMA but that qq plot seems...too good to be right
__________________
ACAS 7 8 9 FCAS
Reply With Quote
  #29  
Old 04-13-2019, 07:13 PM
magillaG magillaG is online now
Member
 
Join Date: Jun 2007
Posts: 3,001
Default

One nice trick is to simulate losses based on your original model fit. Then fit that simulation. This will show you about what your model diagnostics should look like if your model is correct. (Doing this many times and looking at the distribution of results is sometimes called parametric bootstrapping.)
Reply With Quote
  #30  
Old 04-15-2019, 08:38 AM
Actuarially Me Actuarially Me is offline
Member
CAS
 
Join Date: Jun 2013
Posts: 191
Default

Quote:
Originally Posted by Tacoactuary View Post
Not well versed in DHARMA but that qq plot seems...too good to be right
It does lol. I'm somewhat skeptical too. I should look into the math of what's going on, but was excited to find something relevant.
Reply With Quote
Reply

Tags
glm, peril, pure premium

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off


All times are GMT -4. The time now is 08:43 PM.


Powered by vBulletin®
Copyright ©2000 - 2019, Jelsoft Enterprises Ltd.
*PLEASE NOTE: Posts are not checked for accuracy, and do not
represent the views of the Actuarial Outpost or its sponsors.
Page generated in 0.79566 seconds with 9 queries