I think his point was that Microsoft's reputation for accurate statistical
calculation is quite poor.
Logigally one might assume that functionality provided by the processor
(trig functions, etc) was not re-invented by Microsoft or other software
vendors. As such, they should be quite good. Limited testing suggests that
they provide machine accuracy (or something very near to it). However
functions that MS had to write from scratch may have algorithm issues. As
the OP correctly notes, there are issues with the accuracy and working range
of most probability and statistics functions as well as Engineering
functions.
Excel 2003 improved the accuracy of most statistics functions (but
introduced a new bug into LINEST, where coefficients of exactly zero are not
to be trusted). Excel 2003 improves the accuracy and working range of
NORMDIST, NORMSDIST, NORMINV, and NORMSINV. 2003 improves the numerical
inversion of other continuous distributions, but that is of limited value
since the accuracy of their distribution functions has not been improved.
2003 claims to imrpove the working range of discrete probability
distributions, but does so at the expense of accuracy, for example there are
POISSON probabilities that 2003 calculates that are wrong without warning by
as much as 0.5 (huge since 0<= probability <=1). Accuracy and working range
issues with engineering functions have yet to be addressed at all by MS.
Excel 2007 fixes the bug in LINEST that 2003 introduced, but otherwise has
made no improvements in accuracy that I was able to find.
Jerry
