Anyone Trading Coffee KC ?

[AdrianHagh81]

I feel that we cant fully avoid overfitting because we select the best model out of a lot models. So even in pure randomness, we can find a valid model at 5% out of 20 models.
I also look for models corresponding on what relationship I saw on the market. Some of this relationship can be random but we dont know yet. So I have always trading results below out of sample backtests even I use few parameters, simple models and paramets which make sens from a fundamental perspective.

So, all my strategies are based on very simple models because I feel the signal is lost in so much noise.
Did you experience any great success using SVM on real trading ?

My CL system is running live from an SVM resuits.

What I did for the KC system, was take the trades from Collective 2 arabica system

Throw Massive silicon power on it, select the best models that resembles the strategy output.
then use those as my initial population, then run SVM + GP on it again with the 6 years period back test

and then select the best models with the simplest parameters.

 

[AdrianHagh81]

This Is an example of an overfit system with over 15 parameters.
It's for the CL, with 16000+ trades in 6 years.

 

[Rachmaninov]

Do you get any slippage between the live SR you get on your CL model and the SR you had on your backtest ?
How many models did you test on your KC models ? I don't if 10 models with one parameters is as much overfitting as 1 model with 10 parameters. But it looks similar because you have 10 free parameters.

I ll try to do some work on it and publish it.

 

[AdrianHagh81]

the cl system im using limit orders. not the above posted ec.
on the kc system the system evaluated about 2 million price to price models in total.