Black Box/Composite model
Additional notes have been added on July 17,2004
Author |
Sergey Tarassov |
Data Length |
Depends on involving planets |
Forecastig Intervlal |
Depends on involving planets |
Technique |
Black Box |
Target Function |
I used Detrended Oscillator (1,50,50) |
Market Trader Version |
Gold,Platinum |
I think it is a good idea to combine the Composite Expert models with Black Box possibilities. The strongest part of the Composite Expert is that it allows to see visually how the planets move the stock market.The strongest part of Black Box (and Neural Net as well) is the possibillity to work with models based on Astrological Language tools.
Now I will demonstrate one example how we can combine these two together. Or, in other words, I will show how to work with Composite-Black Box models.
I downloaded the DJI 1992-2003 year as an example. Let us investigate all planetary pairs that work well for DJI; for that purpose, we use the Composite Expert.
Sun cycle is very strong; it explains at least 38% of DJI movement:

Now I put this composite into Composite Box pressing "+" button:

Then, let us do all the same for Mercury composite (which has correlation 9%).
Mars composite gives 8%:

By the way, look at Mars position Composite. According to it, the DJI reaches maximum when Mars ingresses in Libra. I checked this fact on the undependent interval.
Then let us analyze angles. I have found these angle might be important:
Mars-Mercury, Mars-Jupiter, Mars-Saturn, Mars-Uranus, and Mars-Neptune
I did not use the Sun-Uranus, Sun-Neptune, Sun-Pluto cycles because these cycles correlate to the annual Sun cycle.
It is quite cliear from "Periods" information window (Platinum version):

In total, I chose as important 10 terms and have put them into Composite Box:

Thus, I have created the composite model. Now let us save this composite model pressing "Save"button (I saved this model in MyDJI.cms file).
Now we'll start working with Black Box.
Run Black Box window.
As a price event, let us use the "Relative Price Oscillator (1,50,50)":

To create astrological model choose "Add Model from List" item:

A) Choose FAM model (FAM model - because the idea of FAM is to find sensitive points relevant to some particular conditions).
B) Set "Num of Copies" to 3
C) Set Step to 8 degrees
D) press "Comp" button:

Now, let us find our composite file (it is MyDJI.cms file).
You might be interested in seeing the involved planetary pairs and the orbs. I recommend to use "V" button for that (the step for FAM has the same meaning as the orb):

Now this model is ready for optimization. The result of optimization looks like this:

Note 1: Actually, the "Future leaks" are present in this approach. When we work with Composite Expert, we calculate the correlation on testing interval. Based on these results, we choose the best planetary pairs and after that we use these planetary pairs to create Black Box model. But I think this is not so important in this case.
Note 2: It makes sense to play with different values of Num of Copies and Step parameters. The smaller step allows the Black Box model to see more details, but it gives much more noise. Actually, you can define the step values for each of the involved planets:
Choose Step=Manually and press "Step" button:

In this case, the step will be set according to the fastest planet involved in a composite pair. With this setting, the profile for our model looks like this:

Also I recommend to use Phase zodiac, especially for inner planets.
3) My experience shows that the big amount of terms not always provides better results. Sometimes, 3-4 terms are quite enough.
The weakest point of the proposed technique is an existence of "Future leaks". What does it mean? I'll try to explain it. To estimate the importance of any planetary pair, we calculate the correlation on testing interval. Then we use the same interval to evaluate the Black Box model performance. And it is where the problem arises. (Remember that we are discussing Black Box models based on composites. There is no such a problem for regular Black Box models where we create and test models on different intervals.) Now I will demonstrate how we might avoid this.
First of all, you should define the new type of interval - the verifying interval. You can define it this way:

Now the whole price history file is divided on three intervals: A-optimizing, B-Verifying, and C-Testing intervals:
To avoid the "Future leaks" in Composite models, we calculate the correlation coefficient on verifying ("B") interval:

As described above, we add the most strong cycles to "Composite Box" window:

Pressing "->Cl" button, we put all these planetary pairs to clipboard. Now we should run the "Black Box" window, create "FAM" style model taking the appropriate planetary pairs from clipboard ("from clipboard" button):

After that, you can begin optimization of the Black Box model. But to estimate the performance of this model, you should calculate the correlation on testing ("C") interval:

In other words, we estimate the composite performance and Black Box performance using two independent intervals (B and C).
Additional information about verifying interval you can find here http://www.alphee.com/Research/Composite_BT/composite_bt.htm
We select the most active planetary pairs using the correlation coefficient on B interval (it is our usual routine):

Actually, the program estimates the fitness between two curves: normalized price and the curve calculated from the composite:
We should understand that we have the good coincidence on A interval (because we use these price points to calculate the composite). Also we have a good coincidence on B interval (because we calculate the correlation coefficient using the price points from B interval).
The price points on C interval are not considered here. Under this approach, the program simply does not know that price points on C interval exist.
Thus, when we estimate the performance of the Black Box model, we should use the price points from C interval only:

This interval provides independent information (regarding to A and B intervals) .
Possible mistakes that must be avoided:
a) To estimate the performance of the planetary pairs using the correlation on B+C interval.
b) To estimate the projection line produced by composite pair visually, while looking at C interval. The human eye has an ability to be more attracted to some parts of the picture than to the others (like to good fit between two curves), and psychologically you might forget that this is C interval (future data that your model is supposed to predict not to "see"). If you prefer a visual estimation, try to cover the C interval by paper.
Please bear these things in mind.
June 14. 2004