[R-sig-finance] fSeries_221.10065 and garchFit+sqp makes R lock up

Diethelm Wuertz wuertz at itp.phys.ethz.ch
Sun Apr 16 15:46:04 CEST 2006


GARCH-Modelling is not easy, and indeed for your dataset the default
"Sequential Quadratic Programming" solver doesn't converge. I observed
this also for some other time series. There is already an updated 
version on
the server, https://svn.r-project.org/Rmetrics/trunk/fSeries/ which uses 
improved
control parameter settings as default values. With this version there exist
no convergence problems. What can you do? Download the updated
version from the repository, or just use the alternative optimization 
"nlminb"
until the next version of "Rmetrics" becomes published.

regards Diethelm Wuertz


garchFit() # Update - Default Settings
           Estimate  Std. Error  t value   Pr(>|t|)   
mu       -0.016772    0.020792    -0.807     0.4199   
omega     0.008898    0.004055     2.194     0.0282 * 
alpha1    0.047233    0.011134     4.242   2.21e-05 ***
beta1     0.936329    0.014828    63.146    < 2e-16 ***  
Log Likelihood:
 1045.871    normalized:  1.045871
 

garchFit(algorithm = "nlminb") # Current Version
          Estimate  Std. Error    t value   Pr(>|t|)   
mu       -0.016772    0.020793    -0.807     0.4199   
omega     0.008898    0.004055     2.194     0.0282 * 
alpha1    0.047233    0.011134     4.242   2.21e-05 ***
beta1     0.936329    0.014828    63.145    < 2e-16 ***
Log Likelihood:
 1045.871    normalized:  1.045871

 
garchOxFit()
Coefficient(s):
              Value   Std.Error    t.value
Cst(M)   -0.0166990   0.0207920   -0.80315
Cst(V)    0.0089064   0.0040545    2.19670
ARCH(1)   0.0472270   0.0111270    4.24430
GARCH(1)  0.9362900   0.0148290   63.13900




Monty B. wrote:

>Thanks to Sean and Diethelm for pointing out that the link was not working.
>
>The data can be found here:
>
>http://host-a.net/getfile.php?usern=upppload&file=fGARCH_crash.csv
>(click in the yellow box to receive file)
>
>Sorry about the quirky download site. It was the best I could do right now..
>
>
>Many thanks,
>
>Monty
>
>
>On 4/16/06, Monty B. <montezumasrevenge at gmail.com> wrote:
>  
>
>>Dear all,
>>
>>I am fitting garch models to a sliding window of observations of the
>>USD/NOK exchange rate. I've been provided with the Ox/G at RCH package,
>>but I am not entirely happy with it's scriptability, so I thought I
>>would give fSeries a go. The package seems to work well for some
>>series, but for others, it locks up R.
>>
>>This code:
>>
>>library(fSeries)
>>y <- read.table("fGARCH_crash.csv")
>>fg <- garchFit(formula.mean =~ arma(0,0), formula.var =~ garch(1,1),
>>                cond.dist = "dnorm", y, trace=T, title="USD vs NOK")
>>
>>and the file:
>>
>>
>>http://us.f13.yahoofs.com/bc/44422dee_a419/bc/My+Documents/fGARCH_crash.csv?bfcOjQEBfGO1k9on
>>
>>makes R crash giving no output when the default settings are used.
>>Changing the algorithm to "nlminb" seems to provide estimates. BUT, I
>>am a bit skeptical about changing defaults when I do not know what the
>>difference between sqp and nlminb is.
>>
>>Any suggestions? Should I use the non-default optimization? Can anyone
>>refer me to literature on what the difference is? Will the parameter
>>estimates be of worse quality?
>>
>>BTW: I am using R for windows 2.2.1. I have tested both the standard
>>2.2.1 and the patched 2.2.1 versions with this code.
>>
>>Thanks for any input,
>>
>>cheers,
>>
>>Monty
>>
>>    
>>
>
>	[[alternative HTML version deleted]]
>
>_______________________________________________
>R-sig-finance at stat.math.ethz.ch mailing list
>https://stat.ethz.ch/mailman/listinfo/r-sig-finance
>
>  
>



More information about the R-sig-finance mailing list