Re: [Quantlib-dev] gaussianorthogonalpolynomial.cpp : double to bool conversion and possible problem on x64 targets

[Klaus S. <kl...@sp...>]

Hi

On Thursday 12 August 2010 10:33:16 Kakhkhor Abdijalilov wrote:
> Btw, can anyone explain why LS path pricer does calibration and
> pricing separately? Current implementation simply discards all paths
> used for calibration. It is not only wasteful...

If you are using the same paths for both calibration and valuation the 
resulting estimator includes a "foresight bias". A standard approach to get 
rid of the foresight bias is to use simple discard all paths used for 
calibration. Please find more details and other algorithms to remove the 
foresight bias here

http://www.christian-fries.de/finmath/foresightbias/

> EquityOption.cpp example
> uses only 4096 paths for calibration and tries to price with
> tolerance=0.02

Values are taken (more or less) from Glasserman, Monte-Carlo-Methods in 
Financial Engineering. Within the Monte-Carlo error (0.02)  the Longstaff 
Schwartz price (4.481675) is consistent with the other american pricer 
(especially with the finite different pricer, 4.486118). Increasing the 
number of calibration paths towards e.g. 65535 does not change the price 
significantly (4.464887). IMO 4096 calibration paths are enough for this 
example.

best regards
 Klaus