[R-sig-ins] Maximum pseudo likehood vs tweedie.

[Christophe Dutang]

Le 5 mai 2010 à 22:19, "RATIARISON Eric" <ERATIARISON at monceauassurances.co 
m> a écrit :

> Hello everybody,
>
> I’m interested in rating premium insurance, with the idea to impleme 
> nt under R the theory linked in the paper attached (only hypothesis= 
> distribution belongs to family exponential,
>
> No a priori specification)
>
>
>
> I’ve already have a gauss (software named…) code which do this  
> with a package named CML (constrained maximum likehood).
>
Hello Eric

To do constrained optimisation you can use the pkg Rsolnp, which I  
test since last week. It is very powerful.
>
>
> With R, I’ve searched similar pack/functions like nlminb, maxLik and 
>  optim or other function  in which I can precise the analytic functi 
> on of loglik, gradient and hessian related to family exponential.
>
> But with  no success (Newton raphson unavaible with fixes gr/hessian  
> (own mistake ?) ,
>
Another solution is to use the genetic algorithm and to put the  
constraint as a  penalty. rgenoud implements the GA.

By the way the package fitdistrplus let you to use a custom  
optimisation function, and give you other statistics on a distribution  
fit.

> often some problem of memory (matrix size  : 550 000 rows x 14  
> covariables, and finally not sure that it concerned pseudo likehood.
>
If you have problem of memory, the only good solution is to have more  
RAM on your computer and/or to use a 64 bit OS, see R-help on this  
topic.

>
>
> Could Tweedie  cover this kind of preoccupation?
>
See the cran task view, I think there is a pkg implementing this  
distribution.

Christophe
>
>
> Anyone has try with  R do to this ?
>
>
>
> Thanks for your help .
>
>
>
> E.ratiarison
>
> <robust_inference_in_ratingmodels.pdf>
> _______________________________________________
> R-SIG-insurance mailing list
> R-SIG-insurance at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-insurance