Computer Generated Trading Strategies

The benefit of computer generated strategies is not that you will find any one strategy of enormous potential, but more likely that you will find several uncorrelated mediocre strategies that can be combined, leading to lower draw-downs and higher profits.
 
Quote from walterjennings:

Hello, I am currently doing research into generating profitable strategies using an ML system.
More...

What you are trying to do also goes by another name: Curve Fitting.

You take a Set {S} of strategies that can be expressed using a language {L} and you try to validate the following conjesture:

There is Si such that {Si} belongs in {S} AND Si --> {P} when Si is expressed using {L}

where {P} is a set of performance parameters.

For arbitrary {S} this is a NP complete problem. It is also like you are trying to find an algorithm that can determine whether another algorithm is designed the best possible way.

Bill
 
I've always tried to to stay away from using the term 'curve fitting' since it always brings to mind the idea of uselessly 'over fitting' a predictor to a sample set. Curve fitting is essential, imagine trying to create a predictor to differentiate between apples and oranges, say the data we obtain from each element in our sample set is taste, color, texture. then curve fitting to our sample set will most likely give an excellent predictor.

But after some thought. I agree that what we are trying to do is curve fitting. In fact if you are trying to create a predictor on sample data without having any extra fundamental knowledge of the system (what we are trying to do here), the best you can hope for is creating a good transformation function which changes the sample data into a space which is is easier to categorize what you are trying to predict, then you curve fit to that new space and hope your transformation function allows for an 'somewhat' accurate predictor.

Now that we agree that this type of approach is curve fitting (along side any other purely technical trading strategy), NP complete and has an infinite vapnik-chervonenkis dimension, does anyone have any comments on the implementation and possible improvements? :p

I was wondering if anyone has ever tried using SVMs to help find good strategies?
 
It bears emphasis that a system with four degrees of input that works with great reward over a long period of time is less 'curve fitted' than a system that uses twenty degrees of input with the same results over the same period.

Don't mistake 'curve fitted' with 'valid strategy'.

Occam's razor is an excellent tool. We should always prefer the hypothesis that not only fits the data, but is simplest. Remember, with sufficient degrees of freedom, any curve can be fitted perfectly.

Using computers to develop mechanical systems removes what is probably the worst hindrance: inductive bias. If a computer does not evolve a solution, your personal choice of potential inputs actually determines the behavior of your future algorithm.

So what I would recommend is to do exactly what you plan on doing, allowing all levels of inputs ... but by identifying fitness as some sort of measure using both return AND levels of fitness (something like return/# degrees might work)...

Best,
C
 
for future reference we should try to separate the terms 'curve fitting / fitted' from 'over fitted'. curve fitting can be bad, if the 'curve' you are fitting to does not represent a reasonable predictor. or it can be good, as in the apple vs orange example, where it is very rare to find oranges which are any other color, thus making it an excellent predictor(+50%). we use the term 'over fit' when it comes to the negative side of curve fitting.

and remember the 'curve' we are fitting to is not necessarily straight market data, but some transformation of market data through maths. ie maybe some stat arb indicator which results in a high % prediction.
 
Quote from walterjennings:

its nice to see this type of strategy being used in industry. it gives me a tiny shred of hope.
More...

Michael Harris of Tradingpatterns.com in his book "Profitability and Systematic Trading" coined the term Synthesis of Trading Systems for the process of mechanical descovery of trading strategies. The book is out of print but I think a new edition will be published by Wiley & Sons soon. There is an article in his website about this:

http://www.tradingpatterns.com/synthesis.pdf

In his book he talks of a more general class of trading systems found by essentially combinatorial searches. It is a very interesting concept.

Ron
 
Quote from man:

my understanding about this is that on the one hand
we have anomalies that have been existing for ten years.
problem with these is that there is a constant search of
about 1.000 (my personal guess) teams around the
globe trying to find them and exploit them and thus
eliminate them. it becomes tougher and tougher.

the second thing is that anomalies arise and vanish again
within years, months, weeks ... and so forth. to find
those, exploit them and get out once they fade is a
game where you need a lot of "computer generated
trading strategies" ...
More...

The shortest path can be the longest one. The easiest way can be the hardest one.

The longest path can be the shortest one. The hardest way can be the easiest one. :D
 
Quote from ronblack:

Michael Harris of Tradingpatterns.com in his book "Profitability and Systematic Trading" coined the term Synthesis of Trading Systems for the process of mechanical descovery of trading strategies. The book is out of print but I think a new edition will be published by Wiley & Sons soon. There is an article in his website about this:

http://www.tradingpatterns.com/synthesis.pdf

In his book he talks of a more general class of trading systems found by essentially combinatorial searches. It is a very interesting concept.

Ron
More...

I'm looking to buy a legal used copy of APS Automatic Pattern Search v4.8 or higher. Please contact billchen90 --at-- yahoo.com

Bill
 
You must log in or register to reply here.
Back
Top