|
PureBytes Links
Trading Reference Links
|
Title: Message
hi
ferreira,
i see
that you are basing your trading rationale on statistical analysis methods.
white noise wouldnt be familiar to anybody just doing technocal analysis. i
think scientifically these are beyond doing simply technical analysis. and i
definitely respect that being a phd student.
i am
not questioning whether this kind of a scientific approach does work. but at
least hope that it works, cause i will be investing my next 2-3 years in data
mining and plan to use it in my thesis as well.
the
main point that other side, or the side that is critisizing you is that academia
would not work in the 'field', the markets in our case, which has been a
long issue. besides, if you ask the economic departments they dont take
statistics as a science, thats why most of the stats departments have
tobe connected to somewhereelse, if cant stand alone..
as it
is always said in this group there is no best. there are different people with
different colors, there are bars with different colors, and different approaches
to same problem. you make it more scientific and earn more?? it is a very hard
hypothesis.
personally i enjoy trying different things and listen to
everybody. and final decision is mine.
i
think it is very good to watch the discussion.
thanks
for all participants.
Hi
Andrew,
Let us just backtrack a bit. I noted, when coding the T3
and IE/2,
that the IE/2 appeared to be similar to the Theta model, which I
know
to be a good model. So I did not bother to test it, assuming it
also
to be a good model. The Theta model implementation provided is
based
on what can be done in Metastock in a very short time span, and
was
given on request, and has maybe too many shortcuts in it. If
anybody
has ever tested the IE/2, I think we can safely use that as a
proxy
for the Theta's performance and vice versa.
Now, since I
appear to be the defender of the Theta model. We (yes,
I happen to be
part of a team) use the Theta model extensively to
prepare short term
forecasts of monthly data, such as M3, CPI,
wholesale trade and so
on. We use it as part of an array of models
and we never use the
results of just one model, but the Theta model
shines in this capacity as a
good performer and often has a fairly
large dynamic weight allocated to
it. Here performance is measured
in forecasting accuracy, which
usually is a poor indication of
whether it will work in a trading
environment.
But we also use this model, for end-of-day data, in a
trading
environment as part of yet again a suite of models. This is
quite
fashionable and dicated by theory as well - using a suite of
models,
and I am in a way recommending this to the group and also
recommending
the inclusion of the Theta in such a suite.
Now, let us
not run away from the real point, testing the Theta model
as a singular
trading model. I note your observation, as well as
that of some other
members of this group, and can well believe it -
that the Theta did not
perform well when you tested it.
This is true of prediction models in
general, so allow me to expand a
bit. A good prediction model is
supposed to predict where the market
will be in future, say tomorrow.
Now, if it is a good projection
model, then it will be unbaised, so that
the market will be above it
about 50% and below it about 50% of the
time. The residual or error
for a good model will be random. So
if we use a prediction model as
is, we are trading white noise, and should
not get good results. So
we have to apply our minds a bit. I am
thinking aloud, why is the
Theta not performing as I would expect, so
please bear with me.
In our trading model, we do use the Theta model's
prediction as well
as its slope. So we extrapolate the model and note
the slope of this
extrapolation and we use both in the model. We have
noted that when
the Theta long term line (theta = 0) turns, it often
indicates a
turnaround in trend. This could be a better way to build
a trading
model, using the slope of the long term component. The
slope of the
extrapolation is in fact half the slope of the long term
component,
since the extrapolated short term is constant and the Theta
is
( lt + st ) / 2
so
d( lt +
st ) / 2 = dlt / 2
since
dst = 0
Another
note, we often take the log of the series before we calculate
the slope,
but this should not make a big difference in many cases.
Anyhow, try
the following test
linregslope(log(CLOSE),periods)
and optimise on periods. When
this line goes above zero, buy, and
when below, sell. Please let us
know the results.
Note that the parameter should be on the long
side. It should
ideally be above 30 for a number of statistical
reasons that I'd
rather avoid for now. I think a good starting point
would be 50 days
and test up to at least 250.
Regards
MG
Ferreira
TsaTsa EOD Programmer and trading model builder
http://tsatsaeod.ferra4models.com
http://www.ferra4models.com
PS :
I *really* appreciate your opening sentence.
--- In
equismetastock@xxxxxxxxxxxxxxx, "Andrew
Tomlinson"
<andrew_tomlinson@xxxx> wrote:
>
> Let's keep
this within the bounds of polite debate. MG, I've tried
a
couple
> of backtesting runs with this on the S&P and on baskets of
stocks,
over 5,10
> and 15 year periods, and show losses
consistently. Perhaps you could
give us
> an example of the operation
of the system in practice and the securities
> that it can be used on,
so we can verify? It doesn't have to be your
most
> tuned,
proprietary version, but enough to demonstrate that there is some
>
verifiable substance here.
>
> Andrew
| Yahoo! Groups Sponsor |
ADVERTISEMENT
![click here]() | |
![]() |
Yahoo! Groups Links
|