[R-SIG-Finance] pointers on using VaR.gpd with return series?

[Brian G. Peterson]

Krishna Kumar wrote:
> Sylvain BARTHELEMY wrote:
>> I think that there is a problem with the VaR.gpd function, as it works on
>> USD/EUR and not on EUR/USD values
>>
>> library(PerformanceAnalytics)
>> eurusd <- get.hist.quote("EUR/USD", provider="oanda", start = "2006-01-01")
>> usdeur <- get.hist.quote("USD/EUR", provider="oanda", start = "2006-01-01")
>>
>> library(VaR)
>> v1 <- VaR.gpd(as.vector(eurusd))
>> v2 <- VaR.gpd(as.vector(usdeur))
>>
>>
>> output:
>>
>>   
>>> v1 <- VaR.gpd(as.vector(eurusd))
>>>     
>> Error in optim(init, gpd.liklhd, hessian = TRUE, method = "Nelder-Mead") :
>>          non-finite finite-difference value [2]
>> In addition: There were 50 or more warnings (use warnings() to see the 
>> first 50)
>>
>>   
>>
> Ouch the default parameters there are two possible fixes setting the 
> cut-off threshold using p.tr helps.
> 
> (a)  doing the following call on your data comes back with some results..
> 
>  > v1 <- VaR.gpd(as.vector(eurusd),p.tr=0.95)
> 
> (b) The other alternate is to rewrite VaR.gpd and set  hessian=FALSE 
> where it makes the call to maximize the log-likelihood.
> 
> optim(init, gpd.liklhd, *hessian = TRUE*, method = "Nelder-Mead") :
> 
> 
> Neither of these are "the solution"  as this is more an Art than a 
> science. Method (a) relates to the question of  how to pick the 
> threshold. Very few and
>  you have biased fit and too many you are no longer fitting the tail.
> 
> In this context I would point you towards the evir library and the 
> excellent book by Alex Mcneill on this but doing the following should 
> give you some hints..
> 
>  >require(evir)
>  >shape(danish)

Kris and Sylvain,

Thanks for the pointers.  I would have assumed that a VaR function would 
set some reasonable defaults for threshold and p value, but I guess not. 
  Basically, I *know* that a GPD distribution is fitable in a reasonable 
fashion to this data, as I've done it.  I was hoping to cut down on my 
implementation difficulties by using functions already written.

I'll try fitting with evir and QRMlib. (the QRMlib package is the R port 
for functions from McNiel's book, which is well worth owning)

Stay tuned...  as always, we'll share our results.

Regards,

   - Brian