[R-SIG-Finance] [R-sig-finance] garchFit failing to give a solution

michaelb boldinm at wharton.upenn.edu
Tue Feb 19 17:42:51 CET 2008


I saw the messages on garch vs garchFit - minimum sample size  and my
experinces are the same. For < 500 obs garchFit() often fails and < 1000
garch() fails.

In fact, using garchSim to create test data with obs < 2000 I found garchFit
failed more often then I would expect.  I conjecture 2 reasons for this (1)
the alpha1=.10 and beta1=.80 starting values are not good choices for many
cases. (2) the lower bound restriction of 10E-6 on each alpha and beta
should be replaced with  0 < sum(alpha,beta) < 1.

I might have missed how to change these in the garchFit documentation.  Or
maybe there are too embedded in the procedures to change, and if so I
suggest adding more flexible upper and lower bounds settings in the next
version of garchFit, and also better diagnostics on non-converged or early
breaking in the optimization step.  

I see the Rmetrics developers contribute here , so IMHO :
For multiple parmater restrictions such 0 < sum(alpha,beta) < 1 a llh
penalty is usually easy to add and works too, and checking h > 0 for all t
might also be necesssary.  Also an initial step that selects the best
starting values from a set of 4 or 5 would add robustness.   Finally, is
there a way to check the garchFit version I am running?  adding the version
number would be useful to add to the output description, as well as the
method used to compute the parameter standard errors and whether a
convergence criteria was met.
-- 
View this message in context: http://www.nabble.com/garchFit-failing-to-give-a-solution-tp15549288p15549288.html
Sent from the Rmetrics mailing list archive at Nabble.com.



More information about the R-SIG-Finance mailing list