Better NN EA development - page 2

barnix
1310
barnix  

Open financial projects:

Open Source Finance Software Links

fxspeedster
434
fxspeedster  

For those of you who understand NN behavior, you can clearly see where and why the sell was taken the way it was taken. Not a lot of elaboration needed. Even with a non-linear behavior.

Files:
hunter113
4
hunter113  

first post

Hi,

I'm writing on this forum first time, but I've been following

quite a lot of things happening here for a while.

I have to tell that what is being formulated here is very

interesting, and I would like to participate as well.

I've good programming experience in several languages that is

what I learnt at university:).

There are quite a lot of valuable information collected already

here on this thread. As I see what needs to be done is

designing the model behind which is I guess the most tricky

part in all this. After it can be programmed easily, I mean

anything can be programmed, that is not an issue.

If the implementation is first done in C++, we can use this

stuff as a dll, or can be implemented in mql if necessary.

What we know about Better system: (at least what I realized:)

-It is trading 2-3 times a day, which is something I really like.

-Using static stop loss of 60, and taking 30-60, sometimes more

pips profit.

-It is using PNN, input some moving avarages, (what time frame,

what length is question)

-It is built up from 3 subsystem (maybe this is the 3 inputs fxspeedster

referring to)

So for me it seems that it is predicting the most probable next 30-60

movement.

I think that the biggest question is that what are the inputs of

the PNN and what is the output it is trained against.

I mean is it some moving avarage value of +t time, or what else

could be?

Here I see that some guys have very good knowledge in NN, could

they start thinking about such a concept, to start implementing

some propototype based on?

Why I beleave in NN? Systems with 100+pips/month can be

created and existing, but these are working for some time, and

after some months or maybe years the market change and the

system stops performing. The advantege of NN is that it can be

always retrained continously adapting to the recent past market

behaviour which is more likely that it will behave like that in the near future.

Hunter

fxspeedster:
project1072, Thanks for your post. First let me say that I have a lot of admiration for your work. You are a very talented programmer. In response to your post I agree 100% that you cannot clone a non-linear EA. The idea here is to mimic its logic and learn from it. We know that this EA is based on PNN and it takes 3 inputs. That in itself is very valuable information. If you analyze trade after trade the author's comments you will start to make sense of the 5 numbers that he displays for each trade. This is also very valuable information. Considering this as a good starting point, one can build on these concepts. We will not have a perfect clone of the EA, but considering its shortcomings on money management, one can end up in the end with even a superior product. One never knows until the development is carried out and tested. As you know from experience, the entries do not play that major of a role in a good EA, but MM does. Take a good EA with excellent entries and poor MM versus a mediocre EA with good MM and you will end up much better resultson the latter. I hope I make some point here.
[Deleted]  

Why MAs?

It has been my experience that using NNs to predict a MA is very difficult. There is simply too much variability (noise) to deliver a profitable model. This is just my experience and I'm obviously wrong if this NN model is going to win the EA contest.

barnix
1310
barnix  

Monte Carlo Method

Next steps:

1.Generating random trades with SL:60 and TP:30-60

(with MSVC6 command line tester or other changed version of this tester)

2.Separate proftable trades and non profitable trades

3.Training (input) data for PNN: Profitable trades with filter values.

Filter Values is:

-MACD val, Kalma filter val, etc.

Files:
nn6.gif  29 kb
hunter113
4
hunter113  

MA crosses instead of random trades?

What if instead of random trades, we would use

all possible MA crossing points like Better's system

and the training input would be those trades

from this set which are profitable?

Would it work that way?

barnix:
Next steps:

1.Generating random trades with SL:60 and TP:30-60

(with MSVC6 command line tester or other changed version of this tester)

2.Separate proftable trades and non profitable trades

3.Training (input) data for PNN: Profitable trades with filter values.

Filter Values is:

-MACD val, Kalma filter val, etc.
barnix
1310
barnix  

Random trade:

Place an order with TP=30 Sl=30 at DateTime= MathRand()

Filter:

MACD, Kalman Filter, RSI etc.

First time we need a simplified command line tester

obtainined from posted msvc6 command line tester!!!!!!

Files:
nn7.gif  34 kb
barnix
1310
barnix  

Monte-Carlo Evaluation of Trading Systems

Files:
flexie
146
flexie  

development tean

nice idea gents. I'd like to toss my hat into the ring and assist where needed. I especially like the idea of less "noise".

I can be reached at pipaholic [at] shaw.ca and skype: mflexie. References are available if needed.

Mark