|
PureBytes Links
Trading Reference Links
|
Oops, I should have checked the archives first!
Here is a useful message from six years ago:
<http://www.purebytes.com/archives/omega/2002/msg09083.html>
Perhaps the first worry to be addressed might be,
"is Tradestation using less precision to do these
calculations than my computer's hardware provides,
and is this part of my problem?" I would suggest
implementing the same algorithm in a compiled
language that gives you access to full 64-bit
floats (such as C, C++, et al) and running that
code side-by-side with your Tradestation code, on
the identical same data. Comparing the results
will give an unambiguous yes-or-no answer to this
worry. Why rely on the opinions and vague memories
(which may be inaccurate) of strangers, when you
can easily perform a decisive experiment yourself?
Next, I would consider rewriting the algorithm
as two passes. The current one-pass design is
the traditional implementation (as found in Excel
and many other programs), and it trades speed
for accuracy. In a two pass design you could
calculate mean-X and mean-Y in a first loop,
then in a second loop you can compute deltas-
from-mean and use those to compute the slope,
intercept, goodness of fit, and so forth.
You'll need to find a textbook that discusses the
IDEA of linear regression, rather than just the
one-pass calculation formulae as for example
Excel's Help LINEST barfs out. Something like
Statistics Unplugged by Caldwell, for example.
|