RE: Warning about accuracy of Array functions in TS2000i

["marQ" <mcerar@xxxxxxxxx>]

It is not a question of a margin of error of + or - a few percentage points.
Under the conditions outlined at the beginning of this thread on the array
functions listed, the margin of error is HUGE.  The correlation coefficient
must be a value >= -1.0 and <= +1.0  I remember getting correlation
coefficient values of > 1.0 in TS4.0.  THAT IS A THEORETICAL IMPOSSIBILITY.
It renders the results of the calculation TOTALLY USELESS.

Regards,
marQ

-----Original Message-----
From: M. Simms [mailto:prosys@xxxxxxxxxxxxxxxx]
Sent: Friday, March 17, 2000 2:26 PM
To: Bob Heisler; fritz@xxxxxxxx
Cc: omega-list@xxxxxxxxxx
Subject: RE: Warning about accuracy of Array functions in TS2000i


Thanks Gary....you've proven the point....and this post is in no way
deferential to you....so here is my opinion:

For the incredible uncertainty and risk factors in futures or equity
trading, the single precision seems "good enough" to me.
If I point myself in the right direction and hit the barn door with a
baseball thrown from my arm, I'm a winner.
The double-precision crowd is saying I need a high-powered telescopic lens
and a 300 caliper rifle to hit that barn door right in the middle and I'll
win all that much more.

Bullshit.

The losing trading systems aren't even facing the barn door. How's the "gun"
going to help THEM ?
They'll just miss (and lose) much more accurately....

> -----Original Message-----
> From: Gary Fritz [mailto:fritz@xxxxxxxx]
> Sent: Friday, March 17, 2000 11:48 AM
> To: M. Simms
> Cc: omega-list@xxxxxxxxxx
> Subject: RE: Warning about accuracy of Array functions in TS2000i
>
>
> > I "HEAR" everyone, but can someone "SHOW ME" how significant the
> > difference is ?
>
> This is not exactly rocket science.  TS single precision numbers are
> limited to about 6 1/2 digits of precision, by definition.  If you do
> ANYthing, ANYTHING that requires more accuracy than that, you're
> hosed.
>
> Very simple example:
>
>         value1 = 54321;
>         value2 = value1*value1;
>         print(value2);
>
> Value2 = 2950770944.  Now try it in Excel or in your trusty HP
> calculator.  The right answer is 2950771041.
>
> Here's another one that approximates what functions like StdDev do:
>
>         value1=12345;
>         value2=23451;
>         value3=34512;
>         value4=45123;
>         value5=51234;
>
>         value6 = value1*value1 + value2*value2 + value3*value3
>                   + value4*value4 + value5*value5;
>         value7 = SquareRoot(value6);
>
>         print(value7:5:5);
>
> TS says 80959.50000.  Excel says 80959.4618.
>
> So it's clear the single precision result is wrong, but it's very
> close.  And if "very close" is good enough for you, then don't worry
> about the single precision variables.
>
> But what if you start doing things like looking at the differences
> between two StdDev values?  Then "very close" might get you into a
> lot of trouble.
>
> If you do complex calculations, which many system designers do
> (sometimes without being aware of it because they don't think about
> what functions like StdDev or correlation do), then TS may very well
> be giving you the wrong answers.  Depending on the type of analysis
> you do, this inaccuracy can kill you.
>
> If you're satisfied with those wrong answers, or if ballpark
> estimates are good enough for you, you're fine.  I can't show you the
> difference because "close" is apparently "close enough" for you.
>
> Just be aware that your code may be a lot fussier about the
> definition of "close enough" than you are.
>
> Gary
>