Resampling Methods for Dependent Data

Wed, Jun 18, 2008 7:19 AM

Wei-han Liu wrote:

Resampling, by definition, only works from the sample available.  If the 
sample has extremes, then the distribution generated by resampling with 
replacement will have heavy tails.  This is a natural result of the rule 
of large numbers...  With a sufficient number of samples, the tails will 
fill out with resampling.

One of the theoretical problems with small initial sample sizes is that 
the tails may not be representative in a resampling context.  If you've 
never seen large loss events, then the distribution resulting from 
resampling may not have heavy tails.  This is why other types of 
analysis are generally employed along with a simple bootstrap such as 
copula or skew-t fitting, or a Monte Carlo simulation with a fat-tailed 
assumption.

Regards,

   - Brian

Resampling Methods for Dependent Data

Thread (3 messages)