I have never found a system that in a backtesting of 40 FX pairs and at least 500 trades on each, all or most pairs are in profit.
For example I have a system that give me profit on 15 pairs (each backtest have at least 500 trades). On the rest it gives me a loss. I didn't do any "optimization", since I'm not 100% sure of how to protect against curve fitting.
This sound any good?
Anyone that have experts advisors actually working in a real account can tell me if is possible or recommended to perform a backtest in many FX pairs and expect to have most of the pairs in profit (let's say 80% of the pairs analyzed)?
Or this never happen and I should continue testing the system in the "few" pairs that exhibit promising results?
Depends on the time frame.
For example if you trade in weekly time frame a two year period can only give your between 20 to 40 plus signals
Thanks for the comments. So in your opinion is not important or required to have a system that works on most currency pairs (80% of for example 40 fx pairs)?
I would like to see a system that works on many pairs, that would be a vote on confidence for robustness in my view, but I think is very difficult if not impossible.
I'm on the 5 minute timeframe and many FX pairs have larger spread and because of the "small" stop loss in the 5M timeframe, it could be a reason why is not working in many of them.
So is 2 years and more than 500 trades a good sample size in a 5 minute timeframe Sytem?
I would run more tests than that. In my opinion 500 trades are too few trades to evaluate a system. Also, if it runs horribly on any pair, I would try to fix it. It should at least breakeven on the worst situation on any pair.
I usually test my EAs on the majors FX pairs from to 2015-2018, that's 4 years of data. If it runs well, I run on more pairs (around 15 FX pairs, I choose all that I would trade based on my broker's spread numbers).
If all runs well, then I run it from 2007-2018 on selected major pairs. That's thousands of trades.
And just like you, I don't like opmitization. I quitted trying to optimize EAs some time ago.
If a EA can run well with a fixed strategy and fixed parameters on all those tests, I think it will work.