Neuromongers, don't pass by :) need advice

 

Greetings :) .

The situation is as follows: about half a year ago, I got seriously involved in neuronics. And gradually, step by step, not without help of some forum users, I reached such a picture:

Here is my masthead, if anyone is interested:

GBPUSD pair, forward glued from 10.2001 till now.

Kagba the positive trend is there :) it's good :) I'm happy to no end. But expectation and drawdown are not pleasing at all.

If I trade this way, I get less than 20% per annum with the same maximal drawdown.

Hint: I suspect I can reproduce a similar picture on any major.

The only problem is that it took me 2.5 hours and more than 1000 network retraining to get this chart.


Is there any way to improve these statistics?

I would be glad to hear any suggestions from people who use neural networks in trading.

 

A picture from the chart is certainly good, but it's not enough for any sensible advice and thoughts, and just to talk about the topic, but not about anything else. With your permission a couple of leading questions:

1) What is the frame? Standard learning period length and OOS?

2) The depth of the input data window? And in general about the input data what's not "sorry"...

3) Type of NS?

4) Fitness function?

WZY... And don't promise to swear by all "majorities" yet) For my NS it's GBPUSD and EURUSD that turn out to be the most "major", the rest are definitely worse...

Z.U.2. And it would be nice to have a separate result of the training period and the following OOS, not of the whole period of course, but one piece of the glue, the average.

 
Figar0:

1) What frame? Standard learning period length and OOS?

M15. Window of 25 months, OOS of a month.

2) The depth of the input data window? And in general about the input data what is not "sorry"...

Converted ironed quotes. For the chart above the depth is 60 bars, predicted tail 15. The tail is not traded the full length. This is what was the default setting.

3) Type of NS?

Echo network :) Never mind though. Pretty sure I could get similar results with say FANN, only with more work.

4) Fitness function?

Fitness function of what?

Z.U. And don't promise to swear by all "major" instruments yet) For my NS it's GBPUSD and EURUSD that turn out to be the most "major", the others are definitely worse...

I'll check it out. It's just long.

 

In short:

Learning on History - > Retraining - Profit on History - > On Real - Total Flush = Stages of any neural automaton.

Aren't you tired of making the same mistakes?

Lucky people, so much free time.

 
TheXpert:

Fitness function of what?


The target learning function of the neural network. What is the network implemented in? What is it trained in?

Z.I. Ran to read about "Echo" network) Start link not to share, Yandex and google lead to the wrong place obviously...

Risk:

In brief:

Learning on history - > Re-learning - profit on history - > On real - complete drain = Stages of any neural automaton.

Aren't you tired of making the same mistakes?

Lucky people, so much free time.


The man showed you OOS like that, why should there be a loss on the "real"?)

 
Risk:

In brief:

In short, shoo!
Figar0:

The target learning function of the neural network. What is the network implemented in? What is it trained in?

The target function is a semblance of a tail. The network is implemented in a dll, the implementation is mine. Training and basic service functions are provided by the implementation.

 
Figar0:

The target learning function of the neural network. What is the network implemented in? What is it trained in?

Z.I. I ran to read about the "Echo" network)


The man showed you OOS like, why should there be a drain on the "real"?)


The system is fitted to the story and the better the profit curve, the more the system is retrained. Once the system hits life, that's it ... kills it.

Neurons are only good for finding solutions in basic science.

/ deleted by moderator/
 
TheXpert:

Target function -- similarity of tail. The network is implemented in a dll, the implementation is mine. Training and basic service functions are provided by the implementation.




I think I understand, i.e. the network is trained directly in MT?

Another question is why 1000 retraining, which is clearly more than the number of months of the test? Training month OOS, training month OOS.... . Some training results are culled by the bidding results and retraining takes place?

 
Figar0:


I think I get it, i.e. the network is trained directly in MT?

Another question is why 1000 retraining, which is clearly more than the number of months of the test? Training month OOS, training month OOS.... . Some training results are culled by the bidding results and retraining takes place?


Situation: Event A - > leads to event B, and this has been the case for say the whole year 2009

It's 2010 - Event A - > leads to an event not B.

/deleted by moderator/.

 
Risk:


The system is fitted to the story, and the better the profit curve, the more the system is over-trained. Once the system hits life, that's it ... it's over.

Neurons are only good for finding solutions in basic science.

The ZE Expert just doesn't understand the nature of things, it's like a monkey with a gun.


You don't seem to have worked out what the cause of overtraining is, and the cause of the drain... that's all, and you can solve many problems with nets.

You don't seem to have stepped beyond NSH4 or 5...

 
Risk:


The system is fitted to the story, and the better the profit curve, the more the system is over-trained. Once the system hits life, that's it ... kiddo.

Neurons are only good for finding solutions in basic science.

The ZE Expert just doesn't understand the nature of things, it's like a monkey with a gun.


man, you've been told it's a forward test.