Forex-ai.com

[Batchboy]

RE: "Holy Mackerel!":

After doing latest optimization run for a fixed .1 lot (1 mini) and got my usual ho-hum $5760 gain in 5 months and 16% dd, I then wanted to try to get that auto-lots turned-on with same optimization results, I had to put a Thou in tester and set to 10% to make it fly (Gary's manual says to be modest and only do in range of a few percent to be safe), and whooooaaaaaaa! Mama! profit=$118K, but the drawdown % is 39%, which didn't matter that much because half the time it carefully crawled out of the .1, .2, .3, .5 lot sizes til it started to zoom to final lot size of 14, albeit the larger lot sizes get the equity curve to swerve some up 'n down in the remaining half of the time, buuuuut, by then it's gravy-train anyway by then!!! Soooo, now that means two separate real life plays to try! The first one is to see if a $500 account can really get to range of $5K like back-tests show, optimizing often as the weeks go by, and the second scenario is if I ever had a Thou to throw away sometime down the pike I could try this Holy-Moly plan, LOL Guess I could set-up a second std-demo live running with Auto on.
jb

[jsp]

There's a problem, though. I did some true walk-forward testing with this AI. For example, optimize from January-June 2007 then test your optimized parameters in July 2007. That way, there's no cheating--it's just like you had been optimizing and trading in real-time.

The results weren't pretty. I didn't systematically test for every month, but the 8 or so periods I tested were all losers. I suggest others do the same kind of testing. If you just optimize on recent history, you'll always get great results.

[Batchboy]

Originally Posted by jsp

There's a problem, though. I did some true walk-forward testing with this AI. For example, optimize from January-June 2007 then test your optimized parameters in July 2007. That way, there's no cheating--it's just like you had been optimizing and trading in real-time.

The results weren't pretty. I didn't systematically test for every month, but the 8 or so periods I tested were all losers. I suggest others do the same kind of testing. If you just optimize on recent history, you'll always get great results.

jsp,
I passed your post to Gary to see what he says.
That could be leaving it too on auto, e.g., I just had on my demo finally a loser out of 4, first 3 were winners. So, I try to stay on top of it, I did optimization again because of this loser showing up (just did it yesterday), and run back-test from best selection.......same loser there, so "it's all part of the plan" loser occurence.
I'll have to see how this "plan" goes, what is your next step/plan? Just trash it because month by month optimize back-tests didn't work? Obviously a month is too long, could be this all doesn't work, but won't know if the thing works good somehow too uinless "try plan X" that could possibly work.
jb

[jsp]

I'm not using the EA any more. My tests were enough for me. Originally, this EA was supposed to be based on monthly optimization. I've done enough testing/optimization to know that weekly vs. monthly shouldn't make that much difference if the core concept is robust.

I also think the whole gold or oil correlation is questionable. The EURUSD was highly correlated to gold for awhile, but sometimes this completely breaks down. At this point , oil and gold are almost inversely correlated. When I ran this EA live, it seemed to jump into trades based purely on gold's movement.

Having said all this, it's an interesting EA and I think Gary is an honest seller. I just it's success is based more on the type of market than anything else. It can catch huge gains in a trending market...sometimes. It can also get whipsawed for enough losses to offest these gains and then some.

If I were to trade it live, I'd do a lot more walk-forward testing first. I'd use weekly data (no offense, but changing it mid-week after a loss or two is curve-fitting IMO). Walk-forward testing through the past is a time-consuming process, but it's the best gauge of future performance. Whenever you optimize on fresh data, the NN will always "find" a few parameter sets that worked great...in hindsight. That's the nature of optimization. The test is how well this works in real-time or in an out-of-sample period.

[Batchboy]

hi jsp,

Originally Posted by jsp

changing it mid-week after a loss or two is curve-fitting IMO

Not really, curve fitting has to do with how the training is going, for all we know the optimization process may be too curve-fitting, hopefully not. No, this doing it as often as possible is just keeping up with any possible changes added into the picture, don't matter weekly, daily, but it's not curve-fitting.

Originally Posted by jsp

the NN will always "find" a few parameter sets that worked great...in hindsight. That's the nature of optimization.

I believe most of us here know that.

Originally Posted by jsp

The test is how well this works in real-time or in an out-of-sample period

Yep again!
So I'm just doin' as you were, checkin' it out "seein' if it's naughty or nice"! You did the out-of-sample method, that takes a bit of time, and I'm trying real-time, which is even much longer patience required, daily and monthly watchin' 'n doin' So your check was strike-one, be nice if this second "up to bat" I'm doin' isn't also a strike-out, but one never knows for sure, eh?

So jsp, you not a dummy re: auto trading 'n EAs, any EA you use and is your going-forward$$$ EA you're happy with and understand and have a grasp on?

Best,
jb

[jsp]

Batchboy,

Thanks. I hope it goes well and that you keep us posted. "Live" walk-forward testing is certainly valid, too, as long as you test through different market conditons.

I think limited "live" testing is a real problem for some (not addressing you, but what I've seen in other forums). For example, 2009 has been pretty range -bound so far, and Martingale-grid systems seem to be working again. Some are convinced they can make 10%-20% monthly forever based on the last few months. When I tell them to backtest through the huge moves of 2008, they ignore it. Oh well, I guess they'll learn the next time we see 300-500 pip moves with virtually no retracement!

Take care.

Originally Posted by Batchboy

hi jsp,

Not really, curve fitting has to do with how the training is going, for all we know the optimization process may be too curve-fitting, hopefully not. No, this doing it as often as possible is just keeping up with any possible changes added into the picture, don't matter weekly, daily, but it's not curve-fitting.

I believe most of us here know that.

Yep again!
So I'm just doin' as you were, checkin' it out "seein' if it's naughty or nice"! You did the out-of-sample method, that takes a bit of time, and I'm trying real-time, which is even much longer patience required, daily and monthly watchin' 'n doin' So your check was strike-one, be nice if this second "up to bat" I'm doin' isn't also a strike-out, but one never knows for sure, eh?

So jsp, you not a dummy re: auto trading 'n EAs, any EA you use and is your going-forward$$$ EA you're happy with and understand and have a grasp on?

Best,
jb

[Batchboy]

Originally Posted by rsmereka

I forgot to mention that in when I optimized for lowest drawdown , the profit was only 3K in six months on 10K. I had to changed the optimization method to largest profit which identified between 5-6K. The drawdown in the backtest was low at about 7%.

Since yesterday, the EA has had two major losses for 90 pips each but the EURUSD just spiked upwards and the EA was selling. I think it is/was a news spike.

My demo of EvolutionX is doing much better with the same currency even though the drawdown with this EA is much greater (currently 14.82%).

Rick

hi RS,
How is your forex-ai doing lately?
My $500 mini-demo took an initial many SL hits dive to half its size, but now has had a couple good leaps up to $757. So beginners bad luck at first has had recovery luck now.
My $1000 mini-demo for some reason is only $822, it has missed one or two of the big moves either due to my CNS server house having outage or due to second MT4 talking to same broker fxpro causing an even longer delay talking to Cypress on the other side of the planet! ho-hum. Nothing like auto-trading in reality, and live$ should be even more fun later-on.
jb

[Batchboy]

Forex-ai just availed their v4.0 which allows other periodicities, and I just optimized on 5min +GOLD as alt_symbol and got profit=$1,750 & dd=14% in 26 days, never have I back-tested with such phenomenal results! I'm in comm with author Gary to see how real this really is, as some of you know I prefer if all calculations of trade (incl. TP /SL) be at end-of-bar so that tester replicates live, so not sure this could be so here, but due to being down that at 5min (results cutting down into the quote curve fast enough) it may be "close enough for a G job", LOL (G=government, engineering talk)
jb

[rsmereka]

Originally Posted by Batchboy

hi RS,
How is your forex-ai doing lately?
My $500 mini-demo took an initial many SL hits dive to half its size, but now has had a couple good leaps up to $757. So beginners bad luck at first has had recovery luck now.
My $1000 mini-demo for some reason is only $822, it has missed one or two of the big moves either due to my CNS server house having outage or due to second MT4 talking to same broker fxpro causing an even longer delay talking to Cypress on the other side of the planet! ho-hum. Nothing like auto-trading in reality, and live$ should be even more fun later-on.
jb

Sorry I missed your message jb,

I was just about ready to ditch this EA before the recent bear move by the EURUSD. The EA was finally on the correct side for a change and it currently 389 pips in my favour. I'll see what happens. So far, this EA is not a very good performer using the standard alternate currency .

Rick

[Batchboy]

Comm with Gary of Forex-ai:
--------
hi Gary,
You're right! I used those settings you gave me and my 6-mo 4hr back-data optimized demo forward run is up 20% already! And its only been 3 days since using optimized values! But I shouldn't get excited now I suppose as this no doubt is a moment in equity upswing, nevertheless it's nice to see it doing other than dive like on my tries three times with faster periodicities!
How has longer term running of it faired from lengthier back-data optimizations?


jerry

-----Original Message-----
From: Gary White
Sent: Apr 14, 2009 5:50 AM
To: Jerry Batcheler
Subject: Re: hi Gary,

Hi Jerry,

I enabled the 1-5-15 minutes options because some of our users prefer short term trades. I believe using such short time frames the forex is not easily predictable and it is more of a gamble than trading when using these time frames. This is my personal opinion only.

I will stick to 4H chart for my own trading as I believe this is the best trending one after the daily, weekly charts which are too slow for me.
I am yet to turn on the ver 4 on the live results accounts (I was planning to do it this weekend, but Easter turned out to be busy with family events) and I'll publish the settings I use.

It is likely to be:
4H chart still
90 pips S/L
200 T/P
Trail S/L, not Smart
Smart Exit

Gary