Quote from Swarm:
@Laissez Faire, many different types of strategies were automatically constructed that when reviewed could be classified, by a human, as breakout, mean reverting, trend following etc. over a full range of market conditions as appeared in that data.
@logic_man, I agree that 7 billion is only a small subset of the overall search space but I did use some very efficient search algorithms like PSO and DGP to ensure that the most interesting looking areas were searched. Obviously there is a limitless way in which constants, logic, arithmetic and probabilistic operators and existing TA functions can be combined to create a strategy but I consider this at least a reasonable attempt to find a solution. Also many solutions boil down to be very similar and the fact that they kept re-occurring suggests that you are finding the same artefacts in the data via different solutions. In fact I had to put code in to ensure that solutions were sufficiently different from each other otherwise I would have ended up with the same solutions many times over.
Because of the sheer number of combinations and the fact that markets are constantly generating yet more random data, you'll never be able to say categorically that TA doesn't work. I just think that I made more effort than most to find useful TA strategies and failed.
Conversely, nobody can ever prove that it works in any statistically significantly way.
Taleb's book 'Fooled by Randomness' probably has much to teach us here.
Do you have the ability to define what the statistically significant boundary is?
If you saw some results, could you judge their significance.
I'm sure that I am a nobody. On the otherhand, occasionally, taking the lesser travelled route can be a branch point in moving forward.
In RDBMS the beginning point only has ten cases that fall into two very unbalanced classes. 24 tables are required to move through a trend to take the full offer of any trend.
Most successful early TA systems only had indicators with six cases and again they fell into two classes.
Later when the PC was invented, the markets only changed in a manner where there was less real time between events.
Events, for sufficiency, fall into only five classes. Also, the way the Orders Of Events mingle is a very rigid invarient system of ranking.
I read what you wrote. I was looking at how you prevented yourself from examining how markets work. I may be that finding this out first, then sets you up for working algorithmically with "in kind" HS's amd their PM's.
Very basically, all markets will always have "granularity". This precludes anything but "finite sets" What turns out to be most rewarding, is understanding that there is NOT a one to one relationship among the variables. If there were, then markets could not "cycle".
this thread is only a few years old and most of the comments have little worth. It is time to measure.
