|
PureBytes Links
Trading Reference Links
|
<SPAN
class=467225700-09102003>[folks, a bit ago steve kernish<SPAN
class=467225700-09102003> posted a presentation by Dave Chamness on
optimization and overoptimization here. I found it interesting, and wrote Dave
with some questions, no doubt pretty basic to Dave and probably to many people
here too. he wrote back, but I forgot to ask him initially if it was ok to post
his replies. tonight I got a msg that it's ok, so here you
go.
<SPAN
class=467225700-09102003>- dave merrill]
<FONT face=Tahoma
size=2>-----Original Message-----From: David
ChamnessSubject: Re: Optimize/OverOptimize
<SPAN
> <FONT
color=#000000>Answers are in the text below. Contrary to Steve's
statement, I have only one degree, BS Mechanical
Engineering.
<SPAN
>
<SPAN
>Dave
Chamness
<SPAN
>
<SPAN
>-----Original
Message-----From: Dave Merrill
<SPAN
><SPAN
>Subject:
Optimize/OverOptimize
<SPAN
>
<SPAN
>Dave, I hope
it's ok to contact you on this. steve kernish<SPAN
class=467225700-09102003> posted a presentation of yours on
optimization that I found very interesting, though I'm afraid I don't
get all of it. this is a topic I'm thinking about pretty much constantly
these days, with quite a bit of accompanying frustration. IMVHO, most of the
world gives way too much weight to optimizations that seem like curve fitting to
me, but I haven't figured out how to move beyond that.
<SPAN
>
<SPAN
>a couple of
questions, if I might:
<SPAN
>
<SPAN
>- can you
explain the scatter plots on slides 3 and 4? what exactly is plotted on x and y?
the punch line, which I'm too ignorant to see, is that the system fails with out
of sample data. the one part I understand, I think, is that the correlation
coefficient, presumably between in and out of sample results, is poor. is
that right? how does the plot itself show this?
<SPAN
>
<SPAN
><FONT
color=#000000>They show the In-Sample gain as % of perfect trading on the x axis
versus the out of sample gain on the y axis. Each data point is a separate
stock with a separate system. In sample gains were 15% of perfect on
average. Out of sample were near zero on average. Perfect trading
wins all close to close changes. There are 2 years in and out of
sample.
<SPAN
>
<SPAN
>- slide 24
mentions "Trend Following on Commodities", as "100 day lookback, trade 34%
before breakout". I don't understand what this means. something about MA or
EMA(100), maybe, but what's the 34% piece? how does it get around the parameter
settings limitations that sink other systems? is this method, or something based
on related principles, tradeable in stocks and/or mutual
funds?
<SPAN
>
<SPAN
><FONT
color=#000000>Breakout buys a new high, sells a new low. Near Breakout
trades sooner. 34% before breakout buys in the top third of the 100 day
high-low range, sells in the bottom third. Specifically, the 34% means 34%
of the high-low range.
<SPAN
>
<SPAN
>- how would I
compute the daily standard deviation of the S&P500, in AmiBroker for
instance, in a way that gives the same .95%/day figure you mention? is
that the average std dev of daily close price change over some specific period
of time? I ask so I can generate comparable figures for other
markets.
<SPAN
>
<SPAN
>Compute the
standard deviation of all the close to close
changes.
<SPAN
>
<SPAN
>- the
parameters I get optimizing today compensate for transient market behaviors that
will eventually end, and eventually it will do very poorly. but if those
behaviors persist, at least somewhat, for a little while, might the system
to do better than average in the short term? if so, is constant re-optimization
worth exploring, or even switching whole trading systems in a mechanical way
based on recent performance?
<SPAN
>
<SPAN
><FONT
color=#000000>I find little tendency for trading systems to work in the
future. Try to identify a simple nonrandomness. Try to find markets
that simple systems work on. Don't pick an impossible market like S&P
500 and try to fit a complex bunch of rules to it.
<SPAN
>
<SPAN
><FONT
color=#000000>Commodities have long term trends. Stocks show short term
2-10 day reversals.
<SPAN
>
<SPAN
>thanks again
for writing and sharing this. makes me wish I lived somewhere near the
meetings you haunt...
<SPAN
>
<SPAN
>dave
<BLOCKQUOTE
>
<SPAN
>Dave is an Agilent,
triple-degreed, engineer. Two weeks ago, he presented this work to our
Denver Trading Group's weekly meeting (actually, this group meets every
Thursday and most Saturday's). Once a month, I moderate a
SIG on mechanical trading (and I haven't seen less than eighty people in
the room since I've been attending).
<SPAN
>
<SPAN
>Although, I don't agree with
certain aspects of his presentation and I somewhat object to his assigning my
name to the "Karnish System" (it has become a bastardized off-shot of my
work), I still believe that there is a lot of merit to aspects of his
work. The "Karnish System" has become the moniker for systems (along the
front range of <FONT face=Arial
size=2><SPAN
>Colorado<FONT
face=Arial size=2><SPAN
>) that stochastically
smoothes a momentum oscillator that initiates buy and sell signals using
symmetrical triggers.
<SPAN
>
<SPAN
>I neither want to endorse, defend
or criticize Dave's work...but, offer this for group members to stimulate
thought.
<SPAN
>
<SPAN
>Take
care,
<SPAN
>
<SPAN
>Steve
Yahoo! Groups Sponsor
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.
|