I've been programming a couple of EAs in the last 6 months which all perform well in backtests but fail in live situation. The result is either a catastrophic loss, either a very large gain, looking too good to be true. In no cases there is a relation between the backtest results and the live results.
So I pursued some efforts to reduce overfitting the curve by choosing as best parameters not those that give the biggest returns, but those that sounds realistic.
This was not sufficient and I'm starting to believe there is a "magic" number in the number on parameters. If there are too many, the EA correspond to a very specific situation which is unlikely to reproduce in the near future.
So, as an example, the robot I'm working on has a few dozens of parameters, due to the number of indicators used. I focus on optimizing only 8 of them, hoping it will make a more "generic" strategy.
What is your choosen number of parameters? Anyone optimizing only 2 or three and having good results? Or is it better to have as many parameters as our CPU can eat?
So I pursued some efforts to reduce overfitting the curve by choosing as best parameters not those that give the biggest returns, but those that sounds realistic.
This was not sufficient and I'm starting to believe there is a "magic" number in the number on parameters. If there are too many, the EA correspond to a very specific situation which is unlikely to reproduce in the near future.
So, as an example, the robot I'm working on has a few dozens of parameters, due to the number of indicators used. I focus on optimizing only 8 of them, hoping it will make a more "generic" strategy.
What is your choosen number of parameters? Anyone optimizing only 2 or three and having good results? Or is it better to have as many parameters as our CPU can eat?