Optimizing optimization.

Took me awhile to figure it out, but I finally got the optimization runs working.

Now comes the hard part, making my optimization efforts more efficient, or at least finish faster. Is it possible to use metatrader for distributed computing? I have a 14 node cluster which I use to render my 3-D animations, and it would be nice if I could get them to interface with Metatrader in order to optimize faster.

People told me this when I started and of course I didn't believe them till I learned my lesson the hard way, as such I don't expect you to listen to me.
But just in case, over optimization and curve fitting will get you squat, if your system is not working before fitting, it won't work after, no matter how much computing power you use. To answer the original question, MT optimization is single threaded and slow, you could divide you parameter space up between different computers but you would have to make sure each mt4 had exactly the same data.

Craig,

I'm very familiar with the troubles of over-optimization and curve fitting. Below is my plan to combat or otherwise reduce the chance that overoptimization will hurt me.

I have 10 years of tick data for cable, I take the first two years, overoptimize for those two years, and then test that against the remaining 8 years of data.

I take the next two years, overoptimize for those two years, and test that against the remaining 6 years of data.

I repeat until I run out of data, then look at the changes in various parameters to determine trends, and use those to predict where the next set of parameters will go.

Would this seem like a sound method to combat overoptimization?

I also tried this, it's kind of complex, where do I start...
Let me say first off that I don't know everything about what your doing so I may miss the point, If I do I apologize.

1. Tick data, you have to be careful with tick data as you could just be fitting meaningless noise within the data, if you have systems that depend on tick data optimizations, they are not going to be robust out of sample. Open/close level data at your chosen time frame is often enough resolution. System designers strive to reduce noise and/or degrees of freedom.

2. Out of sample testing, as soon as you start comparing different out of sample results you are inducing data mining bias (which is what you are trying to negate in the first place by using out of sample testing) which invalidates any trends you may see in the out of sample results as being significant.

3. The system itself, the system needs to have a basis in some type of market behavior, if you are just churning though indicator settings and combinations then both your in and out of sample result are going to be meaningless as will any conclusions you may draw from trends you may see in those samples.

The bottom line is you need something with a low degree of freedom which works without optimization, then you can tutu with parameters.

Alright, allow me to explain my system.

This system is a plain coin toss. I have an ATR-based trailing stop and I'm trying to determine optimum multiple and ATR period.

How should I go about optimizing this system for cable?

Craig, Sorry to wonder off topic, but I just realized your Avatar and the quote in your signature. Thomas Pynchon Mason-Dixon.

Sorry to interupt... A very fine book indeed.

Finally someone!
Indeed it is probably my favorite book ever!

What do you mean by 'determine optimum multiple'?

The trailing stop is calculated by taking the current ATR and multiplying that by some number.

The two values I'm trying to optimize are:

1. ATR period
2. the "some number" multiple.

So the system is random entry with MM?
Edit: I have to go to a meeting now, but if the above is true then your just optimizing random noise no matter which way you slice it, you need a better system.

Yes.

I'm conducting a private study as to effective MM assuming no edge.

What if I opened multiple metatrader terminals, and spread the various parameters among them?

"A screaming comes across the sky..." Gravity's Rainbow - One of my Favorite Thomas Pynchon books...

RR: The basic premiss of MM is that you have a system that is profitable. MM can not make a bad system profitable.

Optimization of a system can be done from a MM point of view of course and I am thinking that is your intent - but again the prereq is a profitable system.

By profitable system, do you mean some edge?