Yes, you have no bananas. Nothing works on
predetermined issues, all of the time. The single largest trap any
mechanical trader falls into: forcing technical work on specific
markets.
Mechanical systems can be divided into two types
(in general, cut me a little slack here): trend following/breakout or
contra-trend momentum systems. Overall market conditions dictate which of
the "type" of system you should put in play.
We believe one keys is volatility. I publicly
trade a mechanical system and this successful approach has been criticized for
not performing well during certain six-month periods during the past
ten years (but, certainly not any time this year). Examining
the poor-performing periods led us to believe that momentum oscillator
(RSI, StoRSI, etc.) driven systems kick butt in times of low volatility.
Unfortunately, they do not fare as well when rate of change becomes relatively
large. Trend following systems get killed in low volatility and "bring
home the bacon" when volatility is pushing historic highs.
So, first, you have to measure volatility and
understand the pulse of the market....applying the type of approach that suits
the conditions. Then, the fun begins. Traders tend to pick their
universe and then "force fit" mechanical systems on their
favorite issues. Total nonsense. Flawed logic.
Allow your system to select the universe.
Some issues have supply and demand patterns that will never lend themselves to
profitable mechanical extraction. Orderly accumulation and
distribution is the "key" to increased profits. Why trade issues that
lose money? Hopefully, the answer isn't: because it's one of my
favorites. I wonder if your favorite issues love you as much as
you love them?
Once you identify issues that have a history of
"orderly" supply and demand, you can rank your opportunities by a bevy of
measurements: expectancy, % per day return, etc.
Don't fall into the trap of trying to make credible
approaches work on your "favorites". Good luck.
Take care,
Steve
----- Original Message -----
Sent: Tuesday, October 19, 2004 10:17
PM
Subject: [EquisMetaStock Group] Re:
System Tester, Number of Bars
Well I had no idea my question would spur such a
debate. After reading opposing viewpoints, here is what I have converged
to:
I am trying to find the best systems with which to trade my 10
favorite stocks. I have about 50 systems at my disposal. I have
decided to backtest each stock (against all systems) with 5000 bars,
1000 bars (which excludes the dot com bubble years) and 250 bars.
Then, for each security, I will choose a system that performs well
over all three time periods.
So do I have an ice cubes chance in
hell?
--- In equismetastock@xxxxxxxxxxxxxxx, "David"
<junk@xxxx> wrote: > > > While I do respect your
opinion on the matter that more > isn't "necessarily" better given
changes in market conditions from > the past. My view lies more
in the fact that if you can design a > system not only to perform well
in past market conditions, but also > in the dramatic recent changes,
your system is obviously more > robust. And I'm not talking about
a system that performs well in the > past "on average." I
mean consistent gains yearly as much is > possible. I would much
rather have a system that performs just as > well in the past as it is
still doing recently, than having a system > that performs well
recently but not in the past. In that aspect I > believe more is
better. > > But maybe my motives are different. I
look for robust systems that > can be applied to various securities for
diversity and perform > consistently. I'm not looking for max
possibly return. If a > businesss is to be run, you can't expect
to have occasional > profitable results showing up here and there just
when they feel like > it. If your system can only do well in
today's market but not a > decade old market, who's to say that history
won't repeat itself and > the market reverts to old? Not to
say you can't adjust your system > when the time comes, but you cannot
pinpoint that until possibly > years too late. > >
You said that the number of bars used has very little influence on >
curve fitting. In the most ridiculous of examples, if you have only
> one month of data and go test a basket of systems, you will
obviously > come up with a few that bought and sold at the exactly
the right > point. Not necessarily because they are good
systems. So what's > next? You can't have one month of data
represent a whole year of > market movememt, it's not accurate enough
of the whole. What about a > year? That sounds like a
decent amount, but it only represents 10% > of a decade worth of
data. Just as a month is only roughly 10% of a > years worth
of data and thats not accurate enough, then how should > one year be
enough when it's only 10% representative of the market > conditions
over the past decade? Maybe, that then lies more in the > time
frames you plan to choose. If your trade time frame with the >
designed system is short, then superfragalist may be right, more is >
not necessarily better. The short time frame expected to trade might
> be close enough to the previous short tested time, then you might
> make money with the system you designed for it. But I wouldn't
be > willing chance my money on it. So even aside from the
possible curve > fitting issue, I still would find the lack of bars
to be a negative > obstacle given that your system wouldn't have had
time to "prove" > itself in more varying market conditions. As I
said, I respect your > view superfragalist, but the aforementioned
reasons is why I believe > otherwise. But after writing this,
I guess a lot boils down to > personal objectives and trading
style. > > Best Regards, > David > > --- In
equismetastock@xxxxxxxxxxxxxxx, superfragalist <no_reply@xxxx>
> wrote: > > > > Sorry, but I don't agree with this
statement. "I'm sure everyone > > would more than emphatically agree
with me that the more historical > > bars the better to test
on." > > > > While I do agree that using too little data
can be a problem, too > > much data is just a big an issue. Curve
fit is a complex issue and > > the number of bars of data you
use to develop your system has very > > little influence on it.
> > > > I'm not going to go into a long piece on curve fit
because there > are > > many really good systems development
books and internet articles > that > > define, explain and
debate the issue. > > > > Curve fit is easy to test for
using out of sample data in walk > > forward tests. Indicators can
be tested for robustness prior to > walk > > forward testing.
> > > > Curve fit is caused by over optimization, lack of
robustness in the > > indicators, too many variables in the
optimized equation and poor > > selection of variables within the
equation. > > > > Not one of the systems development books
that explore the issue of > > curve fit have a set number of
bars of data that should be tested > to > > reduce curve fit
or to validate equations. > > > > No one says that 500
bars are too few and 2000 bars are too many. > > Everyone has a
different view. However, most authors and systems > > develop people
do agree on what causes curve fit. > > > > Robert Colby in
The Encyclopedia of Technical Market Indicators > often > >
tests using 20 to 40 years worth of data. Does that mean that the >
> best performing systems he has found historically will work well >
> today. Absolutely not. He admits that many of the historcially best
> > performing systems have done poorly in the last few years. Is it
> > because of curve fit? No, it's because his historical data
averages > > out all types of market cycles and the last few
years have been > > anything but average. The point of his book is
not to use what's > been > > great over forty years, but to
look in similar places for current > > versions of the similar
things that will work in these markets. > > > > Sorry I
can't support your opinion. I've gotten a different > > perspective
from studying the issue. > > > > Esignal is slowly
increasing the amount of historical data they > > maintain because
of intraday system's developers requests for the > > data. However,
there has been talk that the historical data will > not > >
be available to users of MS but only to Esignal trading clients. > >
Equis says this is not true, but I've seen some evidence of it. > >
> > Historical one minute data since 1997 on the S&P 500 can be
> purchased > > for about $2500 from Price-data.com. For
people doing intraday > > trading that's reasonably priced. You can
buy individual symbols > for > >
$75.
Yahoo! Groups Sponsor |
ADVERTISEMENT
| |
|
Yahoo! Groups Links
|