Optimization vs. decent modelling quality???

Ok, been at this for a few years now and have come to the point where I am playing/developing an EA, well, testing and fine-tuning one anyway.

I hear alot about why MT4 optimization is a waste of time as it is just line fitting and anyone after enough iterations can create a money maker. Ok, I get that. I also hear people talking about auto opts to keep the EA or system "up to date".
Right now my little project can make money but the modelling quality is anywhere from 40% to 70% and I'm only getting an average of 30% -45% of trades. The reason I'm still showing profit is money management and filtering bad trades. So less than 50% modelling and less than 50% winning average AND it's an MA crossover idea and it is still possible to make money over a 19 month backtest.
I'm confused but I know this -- anything is possible and anything can make money as long as it suits one's style and it minimizes losses and maximizes gains. It need not be any more complicated than that.

My question is what does this mean? I will test it with pennies in a real account but I'm wondering is the fact that it can make some money even with poor modelling a better sign than making a ton of money while highly optimized? What is to be trusted ? I know, neither because history does not neccesarily equate to the future but I am comforted by the fact that it has been through both trending and ranging periods. Is that not the best test -- how it does over varied market conditions?

Your thoughts very much appreaciated.....

My only suggestion would be to take your "best" parameters from your tests and run it foreword on a demo for a month or so. That is the only way to truly test your system.

The impact of modeling quality on your EA performance depends on how your EA works, if for instance you only open trades on a new bar open modeling quality is not going to impact your results in any great way. However, if you are evaluating every tick as a trading opportunity the modeling quality will impact your results, since the MT4 back tester 'interpolates' any data under 1m, optimizing an EA will cause the back tester to fit the unrealistic volatility generated, and the out of training results will suck.

To verify if your onto something, set aside some of your test data for unseen data testing (testing on data which has not being used in the optimization), this can all be done in MT4 using the to/from date settings, if you have something on the unseen test (on the proviso you are not evaluating every tick) you can be reasonably sure that the EA at least stands a good chance going forward. I believe Merlin wrote something quite comprehensive on this a while ago.

Thank you FXMutant and Craig. And I found that post by Merlin.

Cheers,

Don

If you are defient about being "Mechanical"...
Then consider this...
{I am presently looking for a chart with a Large Spike for you to "OPTIMISE".}

Optimizing to the Max will Ensure You are In THAT PARTICULAR {Spike} TRADE. But those Inputs may not be the way to get Out of the quagmire.
Think;
Research: &
You have to apply Bell Curve Technology.

In other words;
TRY

Bemac, you're returning me to my old days...psychological research for my almost PhD (divorce took care of that). Much of one's enegry in trying to prove something was spent tying to disprove the theory. One never tries to prove anything only disprove.

When you say Bell curve is that the statistical bell curve of how do you say normalcy, 64% being the crowd and then standard deviations on either side of things......

Would you rather use 11 or 18 as a parameter?

Now that's useful!!!
How'd you do that?? That's not possible with MT4 is it?
I think that graphically represents what I was thinking about and trying to say about trying to capture the optimal amount of optimization without excessive line fitting. You want the system to be robust enough to capture a decent profit without going off the deep end with 90%+ modelling. So correct me if I'm wrong here, it's not so much the length of data used for testing but the variablity (including both ranging and trends).

Omega allows you to Export Test Results {sorted in various ways} to XL. Then you can do what you want with the Data.

Not sure about the other providers.

And,

You are correct,

it is very usefull.