|
PureBytes Links
Trading Reference Links
|
The following does raise some interesting concepts (or should we say,
"conceits")...
>
> >
> > I am totally confused how you could extract any useful training
> > from random data that could perform so well on real data. If
> > it's *really* random, then there is NO useful information in it,
> > by definition!
>
> There are information in it that you can extract.
> But no predictive information.
> The software extracts a complex filtering scheme that is able to detect
trends
> of a minimal amplitude that exist in any random data serie...and that are
> themselves random.
>
This reminds me of that "test" posted here a while ago, wherein we were
suppossed to look at 4 charts of market data and decide which one of the 4
was true market data, the remaining 3 being purely random noise.
The bottom line was that random can look like market data, and vice versa.
My thought at the time was that it didn't matter if market data was created
from random processes or not. What mattered was whether it is possible to
detect and exploit *any* trend with sufficient accuracy and whether a
"system" could be written to take advantage of that.
In this regard, I think that Pierre's slant on this issue is right on
target. One gets to the point in the chase for the Holy Grail where you
suddenly see that pattern filtering, rather than pure prediction, is in fact
what we're chasing here... even when we might "think" we're predicting
something (like a retracement)... current work suggests that in fact, we're
not predicting anything, we're detecting an exploitable pattern.
The more time you spend on this, the better (and faster) your pattern
detection becomes, the more sophisticated the filtering becomes, and the
more effective the money management becomes.
Once you trip over this threshold, it's quite a revelation.
Best regards,
Gene Pope
|