|
PureBytes Links
Trading Reference Links
|
<BLOCKQUOTE
>
Do you see the potential problem here? The problem is not really
"reoptimization" -- the problem was picking the "best" instead of a more
"robust" value. Re-optimization just made a bad approach worse. The
problem is not re-optimization, because if I re-optimized with the goal of
picking one of the 3 middle values, I would still end up with a "good" but
not necessarily the "best" value for the coming year. In fact, I would not
expect re-optimization to result in chancing the value of choice --
because 15 would still have be near the middle of the good cluster in each
re-optimization.If you take the time to learn how to take a better
approach to optimization, I think you will find 2 things:1.
Re-optimization will no longer make a system worse.2. Re-optimization,
at least frequent re-optimization, will be unnecessary. In fact, frequent
re-optimization will be a waste of time because if the original
optimization was done properly re-optimizations should confirm rather than
change the settings one is using. There is a place for occasional
re-optimization. Do you see why #2 would be true? If you do, you are
will on the way to getting out the "box" you mention.You might
wish to read a less pointed post on this issue by Al V. <A
href="">http://groups.yahoo.com/group/amibroker/message/50303
All the best,b
well said
(and not too pointed for my taste, btw (:-)). I've been trying to think out what
exactly I do when I explore possible parameter settings manually, and it's
something like what you're suggesting, picking "something in a nice
neighborhood" rather than picking "the best".
<FONT color=#0000ff
size=2>
I'm
completely fine with repeated re-optimization not changing the settings used.
what I wanted in the first place was a way to believe that the various
supposedly good settings we hear about weren't just curve fits over some set of
data. I wanted mechanical system evaluation process that when followed over
different time frames would come to profitable conclusions more often than not.
I thought, and still think, that if I couldn't do that, I didn't know how to
make these decisions in the first place, and I'd better sit down and think some
more. maybe the main function of the auto-optimization framework, at least in
its current state, is to debunk the idea of simplistic
optimization.
<FONT color=#0000ff
size=2>
I still do
think that reevaluating system performance and settings periodically is a good
idea, even if settings don't change, because if the market has changed
significantly, you'd want to respond. for example, separate from
auto-optimization, I've seen a lot of previously viable trading rules do
radically worse in '03. something is different. in an ideal world, our ways of
looking at our systems would lead us to different settings or systems in
different times.
<FONT color=#0000ff
size=2>
one
question I have is why good performance over an adjacent range of
parameter values is the equivalent of good performance over a range of somewhat
different data. the real question isn't what'll happen if our settings get
randomly changed somehow, it's what'll happen when tomorrows data isn't the same
as yesterday's. or are parameter ranges just an easier-to-manage proxy for data
variation?
<FONT color=#0000ff
size=2>
to follow
that through in an auto-optimization context, maybe it's worth repeatedly
randomizing the data somewhat, trying each of those sets of data with each
parameter combo, and looking for best average performance over all those data
variations. what do you think?
<FONT color=#0000ff
size=2>
I'm not sure
if we can vary the data without destroying some of the patterns that trading
rules try to pick out. for instance, candlestick analysis watches for buy and
sell patterns that hopefully predict future price action a profitable percentage
of the time. say every day's open, high, low, close and volume were each
randomly 10% different, and that predictive power lessens. what does that tell
us?
<FONT color=#0000ff
size=2>
thanks again
for being a part of this ongoing investigation, and to everyone for putting up
with it (:-). hopefully we'll learn something...
<FONT color=#0000ff
size=2>
<FONT color=#0000ff
size=2>dave
Yahoo! Groups Sponsor
ADVERTISEMENT
Send BUG REPORTS to bugs@xxxxxxxxxxxxx
Send SUGGESTIONS to suggest@xxxxxxxxxxxxx
-----------------------------------------
Post AmiQuote-related messages ONLY to: amiquote@xxxxxxxxxxxxxxx
(Web page: http://groups.yahoo.com/group/amiquote/messages/)
--------------------------------------------
Check group FAQ at: http://groups.yahoo.com/group/amibroker/files/groupfaq.html
Your use of Yahoo! Groups is subject to the Yahoo! Terms of Service.
|