-----Original Message-----
From: 	Tamarchenko, Tanya  
Sent:	Friday, May 25, 2001 3:20 PM
To:	Kaminski, Vince J
Subject:	FW: "Convolution VAR"

Vince, I contacted Albanese and you can see here the reply I got from him.

The experiment I could suggest to test the "Convolution VAR" for non-normal variables 
would be: to generate prices with non-normal log returns and compare VAR based on simulations
versus "convolution VAR" for chosen non-normal distributions.

The spreadsheet Dr.Albanese attached compares different techniques (Monte-Carlo, Historical, Convolution)
on real data. The results are different. It is hard to come up to any conclusion after such experiment.

Tanya

 -----Original Message-----
From: 	"Claudio Albanese" <albanese@mathpoint.ca>@ENRON [mailto:IMCEANOTES-+22Claudio+20Albanese+22+20+3Calbanese+40mathpoint+2Eca+3E+40ENRON@ENRON.com] 
Sent:	Friday, May 25, 2001 10:45 AM
To:	ttamarc@ect.enron.com
Subject:	Re: (no subject)

Hello Tanya,

our method is meant to be used for non-normal distributions. The assumption
of independence works out very nicely in my experience. The key point is
that the kind of principal component analysis we developed is portfolio
dependent, so we rank directions in risk factor space according to their
impact to portfolio returns. The agreement with empirical VaR or backtesting
performance is excellent. I am attaching a test Excel spreadsheet for our
VaR software library, which includes our fast convolution method together
with a variety of variance-reduced Montecarlo implementation.

If I can be of further help, don't hesitate to contact me.

best regards, Claudio






----- Original Message -----
From: "tanya tamarchenko" <ttamarc@ect.enron.com>
To: <albanese@math.toronto.edu>
Sent: Friday, May 25, 2001 10:02 AM
Subject: (no subject)


> Dear Dr. Albanese,
> my name is Tanya Tamarchenko. I work at Enron's Corp Research department
>
> in Houston. I have a question related to your paper "Fast convolution
> method for VAR
> and VAR gradients". I am wondering if you (or anyone else) have applied
> the "Convolution"
> or "Analytical" VAR method to non-normal distributions.
> We can estimate distributions of the marginals using historical time
> series as you suggest on
> page 3 of the paper, then we would have to assume independence of
> marginals and calculate
> VAR. How different that VAR would be compared to historical VAR based on
> the same time
> series of xi?
>
> Have you heard of any such experiments?
>
> Best regards,
>
> Tanya.
>
>
>

 - mpvartst.zip