[R-sig-finance] Distribution Fitting

[Krishna Kumar]

mixdist is what you want it works very well and gets you nice gaussian 
mixtures.

The issue is with calibration of these mixtures and I would like to be 
shown otherwise..or just wrong!

But if you have three distribution there could be two different 
combination of distribution parameters that can fit the data equally well.
This becomes a problem of how to divvy up say the variance between two 
of the gaussians..and how good an optimizer you have..

Krishna




Dirk Eddelbuettel wrote:

>On 24 April 2006 at 00:10, Lorenzo Isella wrote:
>|  Dear All,
>| I am experiencing some problems in fitting a trimodal distribution (which
>| should be thought as a sum of three Gaussian distributions) using the nls
>| function for nonlinear fittings.
>| As a test, just consider the very simple code:
>| 
>| rm(list=ls());
>| mydata<-rnorm(10000,0,4);
>| mydens<-density(mydata,kernel="gaussian");
>| y1<-mydens$y;
>| x1<-mydens$x;
>| myfit<-nls(y1~A*exp(-x1^2/sig));
>| 
>| 
>| which I use to get the empirical density (as I would from real experimental
>| data) and test it against a Gaussian ansatz.
>| Well, either R always crashes for a segmentation fault and I have to restart
>| it manually or I get this output:
>| 
>| Error in match.call(definition, call, expand.dots) :
>| '.Primitiv�i�d�������������
...' is not a function
>| 
>| Am I missing the obvious or is there some bug in my R build?
>
>Are you aware of fitdistr() in MASS?
>
>  
>
>>library(MASS)
>>mydata<-rnorm(10000,0,4)
>>fitdistr(mydata, "normal")
>>    
>>
>        mean                  sd            
>  -0.0055755191185632    3.9956609772436904 
> ( 0.0399566097724369) ( 0.0282535897233148)
>  
>
>
>fitdistr() fits a bunch of distribution. I can't recall whether I used this,
>or one of the mixed distribution packages from CRAN when I was doing some
>work on mixtures for fat tails.
>
>Dirk
>
>  
>