Subject: re : backtesting for different percentiles
vlady , i enclosed the file with 2 backtesting plots ( you saw them before ) .
the following table shows what was the percentage of the days when pnl fell
below var 95 , var 90 , var 85 .
these results are based on the real ng forward prices from 1 / 1 / 99 to 6 / 7 / 00
for 2 different portfolios :
- portfolio 1 contained the positions equal to ng - price - prc portfolio
positions on 6 / 6 / 00 ,
- portfolio 2 consists of the positions equal to storage - prc positions on
5 / 25 / 00 .
portfolio 1
var 95 var 90 var 85
implied vols 2 . 93 4 . 11 5 . 57
historical vols with decay = 1 7 . 62 12 . 02 15 . 54
historical vols withdecay = 0 . 97 6 . 75 12 . 02 15 . 54
historical vols withdecay = 0 . 94 6 . 45 12 . 02 15 . 54
portfolio 2
var 95 var 90 var 85
implied vols 4 . 1 6 . 74 9 . 97
historical vols with decay = 1 7 . 04 11 . 14 15 . 84
historical vols withdecay = 0 . 97 6 . 74 10 . 56 16 . 13
historical vols withdecay = 0 . 94 7 . 04 11 . 14 15 . 84
this shows that when we have more observations ( columns corresponding to
var 90 and var 85 )
compared to the column corresponding to var 95 the frequency of curve shift
being lower than var
becomes closer to the theoretical value ( 5 % , 10 % and 15 % ) . the numbers in the
column " var 85 " are
very close to 15 % . this is the argument in favor of using historical vols .
and also the results do not depend on the decay factor in this experiment .
also notice : the numbers in column " var 95 " are higher than 5 % and this is an
indication of fat tails .
let me know if you have any questions .
tanya .