Live trades for NN EA idea - page 3

New comment
 
project1972:
That's correct, I did my first test on NN, I used the data of almost 10 indicators to train a NN based on market patterns and Indicators signals.

Instead of learn how to trade, the NN memorized the past data.

The result is amazing performance in the training period, but totally fail in the out of sample data. The best curve-fitting results I saw in my life.

I am still a newbie in that field and more I learn more I discover that I still need to learn.

My big concern is that sooner we will see an avalanche of Sellers selling the best curve-fitting systems you can ever imagine.

That's a pretty dire out-of-sample result!

I bet if you use a lot less inputs, the results will be noticably better - but afterall, that wouldn't be difficult

As for being flooded with curve-fitting Sellers - too late methinks, it has already happened...

 

thoughts on optimisation

Hey guys - thanks for your interesting posts, is certainly food for thought and hopefully will make this thread move in the right direction. You are right that much of neural nets is curve fitting which is why I am anxious to test the results of my system live. I think that curve fitting can be useful but it can also give false hopes of unrealistic returns ! In any case better (olexander) seems to be onto something with his championship EA which took $10k into $115k so far

 
omelette:
That's a hell of a memory you've got!

But yes, he was really good at explaining 'complicated' stuff. What I took from his postings was that yes, NN can be useful when applied properly and probably more important, selectively.

And to do this you've got to understand what you're doing! I spent many a wasted hour playing with NeuroSolutions demo but never really achieving anything...

I'd completely forgotten his name until I saw you mention his user name at the wealthlab forums then it all came flooding back !

Ive played with neural networks for years, for both trading and non related trading problems. My best mate at university did his final year project using them for character recognition, probably 20 years ago when this stuff was cutting edge science.

Ive used ANN in the past quite successfully on a wide range of projects including predicting the characteristics of certain materials based on their chemical composition, and manufacturing parameters, Ive used them in image analysis problems, and credit risk assessment.

The main problem I have with using them in trading, is selecting a suitable and meaningful output layer. For example, in most of the projects Ive used them on, there's a range of inputs, and one or two simple outputs. For example inputs could be the chemical composition of an alloy, and the output could be the tensile strength of the alloy. Or a range of personal details such as salary, age, credit history etc, and an output such as the probability they'll default on a loan.

Although I think they're a useful tool, I cant even begin to see how you can use these things with time series data without the whole thing degenerating into an over optimized curve fit.

Given the fuzzy nature of technical analysis signals, I suppose that they could potentially be useful to identify potential set up conditions, its something Ive looked into superficially, but I'm not sure they provide anything that cant be achieved using other simpler methods, and if you really want a completely over optimized non linear curve fit, GA's are probably even better

 

My simple ideas

Trading with NN's for some 7 years now, I thought perhaps some of my ideas may be of interest here.

I completely agree with the earier post of Project1972 - that curve is just about what one would expect - well, in my experience it is.

However - there is a way to overcome the out of sample data flat line which I find quite successful namely to, have an array of NN's and a catagorizer in front, deciding which NN to use in any specific case.

It is like - well, to use the army example of earlier, if I have a problem between cloudy/sunny days, it would mean that I need two NN's - one for cloudy and one for sunny days. I also need a catagorizer to decide which one to use and then (well, with market price at least) I find some interesting results.

Also - using a single NN the result is like a (very nice) curve fit but as soon as the market pattern change, the pattern is not applicable anymore.

However, when I implement a structure like the above, the output seems to be much more accurate in following the market.

I think we can also see this when Ehlers says that (for a short period of time) the market may follow a certain pattern - then the pattern will change.

Implementing NN's (I think) one may be more able to follow the pattern easier and faster - while the catagorizer actually may try decide which pattern is playing out at this moment.

Just my thoughts.

 

Leeb

cut that 0.01 lots out.

sorry but it makes you sound like an ameteur

Get rid of it

oilfxpro

 

Hi

Hi leeb,

Any backtest with your ea? and is it running 24/5 cause 2day It traded till 10.00 ibfx server time.

Good job .

Metcalfe

 

Hi,

looks very nice for now, will be following it the upcoming time!

 

New version

Hi, thanks for the comments guys. Oilfxpro - I have taken your points on board, I will leave the original version running but have coded a new one which will use larger than 0.01 probably 0.1 but up to 1.0 and above, I have created a new subsystem too, login details:

Login : 1746432

Investor : bl6tuip (read only password)

Remember the other version will still carry on running also

 

Hi Metcalfe, prefer to see how it runs on live account, backtests can be strange ! Cheers

 
zupcon:
I'd completely forgotten his name until I saw you mention his user name at the wealthlab forums then it all came flooding back !

Ive played with neural networks for years, for both trading and non related trading problems. My best mate at university did his final year project using them for character recognition, probably 20 years ago when this stuff was cutting edge science.

Ive used ANN in the past quite successfully on a wide range of projects including predicting the characteristics of certain materials based on their chemical composition, and manufacturing parameters, Ive used them in image analysis problems, and credit risk assessment.

The main problem I have with using them in trading, is selecting a suitable and meaningful output layer. For example, in most of the projects Ive used them on, there's a range of inputs, and one or two simple outputs. For example inputs could be the chemical composition of an alloy, and the output could be the tensile strength of the alloy. Or a range of personal details such as salary, age, credit history etc, and an output such as the probability they'll default on a loan.

Although I think they're a useful tool, I cant even begin to see how you can use these things with time series data without the whole thing degenerating into an over optimized curve fit.

Given the fuzzy nature of technical analysis signals, I suppose that they could potentially be useful to identify potential set up conditions, its something Ive looked into superficially, but I'm not sure they provide anything that cant be achieved using other simpler methods, and if you really want a completely over optimized non linear curve fit, GA's are probably even better

You certainly seem to be pretty well up on NN's. Your character recognition mention caused a flashback to my college attempts at doing something similar (no graphics tablets then, just 'crude' light-pens...) - and the timeframe is about right as well .

I also had a go (way back) at coding NN but after a few futile attempts at coding back-propagation, I decided to leave it to the experts!

12345678910...14
New comment
Reason: