One for the team

I havent made it through every indicator yet, only RSI and macd signal slope
are statistically valid so far. I'll look at Architect's favorite CCI next.
Linear systems do not fit well.

I like the feel of observing indicators but we need something of statistical proof.

This system beats buy and hold by 10:1 since 1985 NDX daily basis.
First cut of a work in progress but you could run with 10:1 odds.

AC... is that a delta fourier transform of the RSI?

I havent made it through every indicator yet, only RSI and macd signal slope
are statistically valid so far. I'll look at Architect's favorite CCI next.
Linear systems do not fit well.

I like the feel of observing indicators but we need something of statistical proof.

This system beats buy and hold by 10:1 since 1985 NDX daily basis.
First cut of a work in progress but you could run with 10:1 odds.

[quote name='TheArchitect' date='Mar 29 2008, 08:35 PM' post='363041']
AC... is that a delta fourier transform of the RSI?


I'm using linear regression, some stepwise and best subset tools to arrive
at the useful terms. The response that I've manually entered at each day
is -2 new sell signal, -1 sell, +2 new buy signal, +1 buy.

Interesting the RSI is the only thing that tests out.
I'm also working now with digitizing RSI levels for example RSI 2 > 95 = -1, RSI 2 < 10 = +1.
This is getting the output up, now at 12.8:1 return vs buy and hold.

To spike the signal strengths, new minor swing lows within and ongoing +1 will get a +2.
And conversely within ongoing sells, minor swing highs will get a -2.
I can play with these values more from there.

The the regression output is just digitized to >0 = +1, < 0 = -1.

I am finding some parameters within the macd that will fit a linear system but are not improving the output or R-sq. I need some non linear event signals to improve it further and some quicker sell condition.

I am finding some parameters within the macd that will fit a linear system
but are not improving the output or R-sq.

I need some non linear event signals to improve it further and some quicker
sell condition.

you aren't data mining are ya?

Well sure it's data mining, or backtesting. But a better way to define it is
multivari analysis, what inputs are statistically related to the response, what inputs are noise
and are discarded because they dont enhance the model. What inputs are collinear and
just corrupt the variance inflation factor?

As a linear function, most of these indicators are useless. We've seen guys
calling stoch oversold and bottoming all the way through the worst parts
of this decline, others calling it a buy signals because macd was rising.

The RSI will
produce p values of .000 in this model. It's the only one that works at all.
The output of the model teaches me how to use the RSI, it likes + RSI 13
and it likes lower RSI 5, interesting? I like it when it confirms my experience
where I learned to like RSI 13 > 45.

To overcome the nonlinear event of overbought RSI 13 I put in a binary RSI 13 sell function.
So the model now contains both
a linear RSI 13 term and a binary RSi 13 term.

The biggest problem, my only concern at this point with the RSI system
is late sell signals from RSI model but it beats buy and hold over 12:1 as it is through
every type of market cycle that has occured since 1985, not too bad for a model.

Really, I abandoned all these collinear indicator concepts years ago and went to the elliot and
channel work.

Most people would do well to never look at an indicator and use something else.

I look forward to testing the CCI and other things in the model but if they dont
improve the model output ?

I am finding some parameters within the macd that will fit a linear system
but are not improving the output or R-sq.

I need some non linear event signals to improve it further and some quicker
sell condition.

you aren't data mining are ya?