[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[amibroker] Re: FW: Optimize/OverOptimize



PureBytes Links

Trading Reference Links

Hi,

Quite a few articles and books have been written on the topic of 
system optimization.  Most of the work done in this area concludes 
that optimizing individual systems does not provide a superior 
result.  If you take, for example, a MACD oscillator and optimize it 
over a year of data, and then trade that system in the future, it's 
odds of success are meager....

You have to optimize many individual systems and arrive at a 
collection.  You can then use Back Test scores from the systems to 
form a consensus to create Entry/Exit signals.  This is very 
different and much more advanced, than just optimizing an MACD.  You 
can try "fixing" the systems, but in all cases the accuracy of the 
signals will be less than that achieved by optimizing....

No matter which systems you use, they will fail at some time or 
another.  However, you will find that Forward Test results of a 
fixed, non-optimized system are dramatically less than an optimized 
system if you use a collection of optimized systems and form a 
consensus....

Prediction is difficult, especially of the future...  

                                               Niels Bohr

Regards,

Pal

--- In amibroker@xxxxxxxxxxxxxxx, "Dave Merrill" <dmerrill@xxxx> 
wrote:
> let me throw the discussion of optimization in another direction 
for a sec,
> into equity feedback, something I've been looking at a lot lately.
> 
> just to be clear what I mean, the idea is to pick your trades from 
among the
> stocks (or whatever) that have historically done the best at that 
same
> method, or something as similar to it as possible. for example 
(don't put
> money on this!), say you went long or short when MACD(13, 21) 
crossed zero,
> but picked stocks with the most strongly positive returns to date 
by that
> method.
> 
> in a sense, this amounts to automatic optimization, by the most 
seemingly
> relevant means possible: grading performance under the same system. 
another
> way to look at it is that it automatically selects a universe of 
stocks to
> run on, including only the ones that perform best with the strategy 
used.
> 
> given those general mechanics, you'd think this would produce good 
results
> frequently, with a variety of common indicators, over many time 
frames and
> universes of equities. my experience is that this isn't the case. 
on the
> contrary, for example, it's hard to find system that are juicily 
profitable
> from '92 to present on the NASDAQ 100. works better on the whole 
NASDAQ, I
> think because a larger total universe has a better chance of 
containing some
> stellar choices, but it's still hard to do outstandingly well.
> 
> what does it mean if a strategy doesn't perform well when managed 
like this?
> seems like one of two things: either not very many tradable stocks
> (sufficient liquidity and price etc) perform well with the 
strategy, or,
> past performance with the strategy doesn't correlate well with 
future
> performance.
> 
> if not many stocks do well with the strategy, it's just not that 
generally
> applicable. perhaps it has parameters that need to be tailored more 
to each
> individual stock or time frame, more than the algorithm itself 
does. if so,
> the next direction to pursue might be more auto-compensating 
metrics within
> the strategy. OTOH, the more complex and rube goldberg things get, 
the more
> they seem like fragilely over-optimized special cases, unlikely on 
principle
> to be robust beyond the specific conditions under which they were 
tested.
> 
> on the other hand, what if past performance doesn't predict future
> performance? maybe some other metrics would, but how would you 
know? any
> backtests you did to find that out couldn't be relied on to predict 
the
> future themselves.
> 
> is there a third hand? or have I screwed up my equity feedback 
code, and
> everyone else here has tons of highly profitable systems like this 
that
> they're not talking about?
> 
> thoughts?
> 
> dave


------------------------ Yahoo! Groups Sponsor ---------------------~-->
Rent DVDs from home.
Over 14,500 titles. Free Shipping
& No Late Fees. Try Netflix for FREE!
http://us.click.yahoo.com/ArdFIC/hP.FAA/ySSFAA/GHeqlB/TM
---------------------------------------------------------------------~->

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 http://docs.yahoo.com/info/terms/