|
PureBytes Links
Trading Reference Links
|
The Omega Man wrote:
>
> I've said it before and I'll say it again: This debate over precision is
> ridiculous.
>
> It is ridiculous for two reasons: first, because the input data is not
> precise; and second, because extra precision in calculations is not needed
> for trading.
>
> First, the input data is imprecise. Prices are set by a bunch of screaming
> banshees in a "pit". Do we need to go out to the nth decimal place to
> decide what the price is? Or is a good approximation (which is all that the
> price spotters can give us anyway) good enough? I say that it makes no
> sense to run calculations out to the nth decimal place when the input data
> is +/- 0.1 points.
>
> Second, extra precision is not needed for trading. What's needed for
> trading is not more precision, but less. We need "fuzzy" approaches, not
> precise ones. Carrying calculations out to the nth decimal place in a fuzzy
> environment is ridiculous.
>
> The Omega Man
>
Certainly we don't need a moving average calculated to five decimal
places. Where it might be a problem, is in some statistical calcs like
standard deviation, or the linear regression function? Some of these
formulas take the differences of squared values and then take the sq.
rt. Back in 1982 or so, i remember my stats book giving several
equations for certain calculations, one that showed the "theory" of the
calculation and one or more alternatives that were "easier" to calculate
or less subject to round-off error. I don't know if TS's single
precision would cause trouble or not, since I don't remember exactly
what sp. means. If the errors are small I view it as not a problem (at
least in my very inexact trading). However, if errors are 10, 20 or 50%
then it might matter. For example I wrote an indicator that defines the
trend by plotting the lin. regression slope and std deviation in same.
Maybe no better than a moving average of the same length, but this one
has a neutral zone which is interesting. Anyway, if the slope was 20%
too large and the std dev 20% too small, the net result would be an
error of 40% , and too many uptrends would be indicated. Having said
all of that, I don't know if single precision is really a problem here.
One observation I have made which may be math related. I plot Rsquared,
an indicator which has been posted here. (An interesting alternative to
ADX.) On ED only, it behaves unusually: values sometimes go above 1.00
(which I believe they never should do); other times the Rsquared value
stays low even tho the mkt has some trend. Perhaps the code has an
error. Or, it could be that the true Rsquared doesnt work well on ED
because of Ed's habit of being flat and then having a big jump. Or it
could be TS's math. Any ideas?
Conrad Bowers
|