|
PureBytes Links
Trading Reference Links
|
> VBatla@xxxxxxx wrote:
>
> > I've heard a lot about over optimization in system design. I also heard that
> > inputs should be constant and work with almost all markets...
>
> At 10:15 PM -0400 5/15/99, Robyn Greene wrote:
>
> >
> >Perhaps I'm out of the mainstream - but I think that the person who can
> >design a
> >system that will trade bonds and internet stocks wonderfully with the same
> >inputs
> >is a genius. Since I've never found that system yet - I think anyone who
> >thinks
> >he has such a system is a fool. Robyn
>
>
> To build a good trading system, you need to find something that happened in
> the recent past that correlates well with what the price is going to do in
> the future. This is fundamental. Without it, it is just gambling.
>
Bob, you are right on! Having spent 3 years discovering what aspects
of my trading system predict which aspects of the market, I can confirm
that you NEED some event (very preferably a MAJOR event) to implement
these discoveries. The first week I turned the system on, I was running
only Close of Market models, as I was using a 300-day history and the
computer took an hour to run one portfolio (the system uses an n^2
algorithm). One Thursday night, I noticed an imbalance (== unusual
pattern) in a set of the model's outputs. The next day the DJIA dropped
300+ points. A possible predictive correlation was noted (!). Six
weeks later, the same imbalance and drop even occurred. I became
a believer.
This was repeated time and time again. As I began to see that the
prediction delays were hours and not days, I ran some tests and cut
the history back to 70 days and the model run time back to about 10 minutes,
so that I could run models every 15 minutes instead of once a day. Let's
say we have 15 "indicators" per model run. That's 60 numbers per hour
to keep track of when trying to see which number(s) predict which market
moves. I'm a slow learner so it took 2.5 years to "see" the most
reliable (a hypothesis as yet untested) indicator --- a 45 minute
lead on the DJIA. All together, without use of proper dynamic systems
models, I have hammered out (my great-great-grandfather was a blacksmith)
enough crude indicators to provide 45, 60, and 90 minute predictions
which are fuzzy-linear (larger indicator values ==> larger market move)
and a fuzzy-picture of the next-day's intraday ticks. Along the way,
I put aside other indicators which seemed good a spotting merger/acquisition
plays and which came pretty close to predicting the closing or 3 PM change
in the DJIA.
> That said, can you find some universal characteristic that works on all
> markets? There obviously are some. You can apply a moving average to the
> price and buy when it turns up and sell when it turns down. That is pretty
> universal but will probably not be particularly profitable. Most of the
> universal things are pretty apparent to lots of people so probably do not
> work very well.
There is a dynamic and a physics to the market. Because our young
scientific civilization has spent the past few centuries worrying about
the (apparently) linear physical systems (planets, rockets, reciprocating
engines) of the real world, the science is just catching up with the
situations where agglomerations of little things (quantum physics,
market trades) affect our lives. This is to say that there ARE, indeed,
invariants (== universal characteristics) of markets-in-general. We just
don't yet know how to get ahold of them. The mathematics we inherited
and were trained in simply isn't up to the task. New math concepts
and new perspectives on how markets work are needed.
>
> Then there are those who try to classify the mode of a market - trending
> vs. congestion, etc. I have never had much success with this approach since
> markets do not seem to switch cleanly between the different modes making
> the modes are very hard to identify. And techniques that work in trending
> markets get killed in congestion, etc.
It's in the transition between modes that chaos can set in ...
>
> So I would suggest that the most interesting and most profitable
> characteristics on which to base a trading system are specific to an
> individual market, or possible a combination of two markets. As an analogy,
> a voice recognition system that works on all languages is almost certainly
> likely to be poorer on English than is one designed to work only on
> English.
>
However, Bob, running (say three) models of different markets side-by-side
can provide additional perspective on any particular one of them that
you want to focus on. I traded only OEX, but ran NDX and DOW because
the three markets share components and their dynamics are correlated.
> Bob Fulks
>
>
Cheers,
Robin Lake
rbl@xxxxxxxxxxx
|