Ideas Pls . Input/Training Data

[michael]

Morning all,

I have a model running , not complex .

Well I had 5 weeks of excellent paper trading with it , then all of a sudden it

went haywire , That was the week starting 12th of July .

That was a small drop , nothing too serious.

Anyway I have tried changing the training model length of data ,

Trying to find out what has changed in the out of sample data , they look pretty similar .

So i am trying to find out what would be the signal that the model has broken down ?

Any ideas that you can think of for trying to solve this ?

Any General ideas will be accepted .

Regards

Michael.

[SystemBldr]

I’ve seen this before!

You might look into market cycles, via Dr. Ehlers?

His work shows that markets change abruptly and it is hard to see in the data.

Another approach might be to train across chart pages so you don’t over fit the system to the markets you are working with.  Michael Covel’s book on Trend Following talks all about finding robust rules and systems.  It is counter-intuitive when you have software like NST that can tailor a system exactly to what you want… but that is the danger of over fitting, because just when you think you have it nailed down, it changes out from under you.

 

[michael]

Thank you , I am trying the Ehlers correlation and looking for a break in it.

and looking at taking different time frame correlations and comparing them using the Mcmillan idea of percentile bands and see if it has moved out of that range.