Neural network - page 6

 
Folks, please help me understand a particular neural network! About five years ago my friends and I created a neural network in statistics, then we used it to trade quite successfully. I have some sources, but unfortunately I have never worked with STATISTICA program itself, and for the last three years I haven't tried it. Help sort out where to load what the network output. Ready to throw the archive with the sources on my email. Thanks in advance.
 
StatBars >> :

If we know the sinusoid and can therefore predict it using networks, then create a more complicated formula, the analytical notation will be known to you, so we can predict it too. The market is the same formula, only more complicated and it is not known to WE...

It's not that the market has a very complicated formula, and it's not that we don't know it. Neural networks can approximate functions of any complexity. The point is that the market is changing!

Hence the impossibility in principle of learning with a teacher.

 
marketeer >>:Imho, you've been too quick to discount grids, without actually trying even 10% of what you should have (just judging by the time you started working directly on your project). By virtue of your maths background you have options with what to try and replace the grid ;-). You are welcome to try. But it seems to me that by criticizing the grid you are focusing your attention on the wrong points.

No one has criticised the nets and no one is going to do so. It was about something else.

 
joo писал(а) >>

It's not that the market has a very complicated formula, and it's not that we don't know it. Neural networks can approximate functions of any complexity. The point is that the market is changing!

Hence the impossibility in principle to learn with a teacher.

The market is not changing, it's just your excuse for not being able to cope...

Especially if the market changes track these changes and take them into account.

 
StatBars >> :

The market is not changing, that's just your excuse for not being able to cope...

Especially if the market is changing, track these changes and take them into account.

Where did you read that I am not up to the task? I just don't use training with a teacher. But that doesn't mean I don't use NN! I don't need to keep track of market changes, NN does that.

LeoV, please explain, I am running out of patience.

 
joo писал(а) >>

Where did you read about me not being able to cope? I just don't use teaching with a teacher. But that doesn't mean I don't use NN! I don't need to keep track of market changes, NN does that.

LeoV, please explain, I'm running out of patience.

I'm already chattering...

The impression is that you do something with NS only because you want to, not because there is any real reason, confirmed by some experiments or anything else.

 
StatBars писал(а) >>

For example, an Expert Advisor is optimized on 2 months of history with only 3 parameters, 80% of the positive results are profitable on all the histories.

The same thing with networks...

Is this a stable result six months without overoptimization (overtraining)? You don't need 80% of positive trades to make profit. You can do with much smaller number.....

StatBars wrote >>

In general you're talking about stability of results, not error. If the network is stable in detecting something or making a prediction, and that prediction is good enough for profit, you will make profit on both forward and real account.

If the error is satisfactorily small, then it will lead. What does it mean satisfactorily? For each problem this condition is set separately, I only know the empirical way.

There are no studies, scientific papers or any other confirmations that maximum profit is guaranteed when you obtain the minimum error. These are all assertions that are made up by yourself and taken as an axiom. But this assertion may be wrong.....

StatBars wrote(a) >> The market is the same formula, only even more complicated and it is not known to WE...

Unfortunately, it's not a formula. The market is something more complex than a formula - I'd say it's some "living mechanism" that is constantly changing somehow over time. And if it were a formula, it would have been found long ago......

joo wrote(a) >> Hence the impossibility in principle of learning with a teacher.

I wouldn't say impossible, but simply impossible to understand or calculate what would act as an ideal teacher for a network. The very notion of an ideal teacher for a neural network is not clear. It is absolutely not certain that it will be a function which describes the market perfectly. Perhaps the opposite is true that the network is worse trained using an ideal teacher as it is impossible to perfectly learn the market and a neural network cannot do it (though I don't know, it has its own nuances). By the way, in neural network programs for trading the teacher is also optimized when training the NS because it is impossible to understand what kind of teacher the NS needs to bring profit. It is better to do without it at all - one variable will be less, especially it is not clear good or bad....))))

StatBars wrote(a) >> The market isn't changing, that's just your excuse, the excuse that you can't handle the task...

Well that statement is hard to answer. If the market doesn't change - then write a market formula and be done with it.....))))

StatBars wrote(a) >> Especially if the market changes track those changes and take them into account.

The market is not changing according to any known patterns. Finding and accounting for them all is not possible.....

joo wrote(a) >> LeoV, please explain, I'm running out of patience already.

Well actually I'm getting tired too. If somebody has thoughts and wish to bring all to "single denominator", single formula - well, let them write, teach, experiment - we all would be glad to listen.....))))

 
LeoV писал(а) >> because it is impossible to learn the market perfectly, and a neuronet cannot do it (although I don't know - it has its own nuances).

Here it must be said - I was a little wrong or misunderstood. A NS with a few neurons can perfectly learn almost any curve. But the trick is that further on in the real world, all that it perfectly learns will not repeat perfectly either. Therefore the better or more perfectly a neuronet learns history - the worse it will work in the future. It is a variant of fitting. And neuronet, due to its non-linearity, very easily fits the story. There is a big disadvantage of a neural network in this......

 
LeoV писал(а) >>

Is six months without over-optimisation (over-training) a stable result? And you don't need 80% of positive trades to make a profit. You can get by with a much smaller number.....

/*I*/: 9 years, not half a year! Not 80% of positive trades, but 80% of the optimization results!

//----------------------------------------------------------------

There are no studies, scientific papers or any kind of proofs saying that maximum profits are guaranteed when you get the minimum error. These are all assertions made up by yourself and taken as an axiom. But this assertion could be wrong.....

/*I*/: No, so do them(research...). I say there is a very definite relationship between generalization error and profit, this relationship needs to be established for each task to at least know what to aim for in training the network!!!

//----------------------------------------------------------------

Unfortunately, it's not a formula. The market is something more complex than a formula - I'd say it's some kind of "living mechanism" that is constantly changing somehow over time. And if it were a formula, it would have been found long ago......

/*I*/: It would have been found a long time ago, it would have been made a long time ago... Strange then why the computer wasn't invented 1500 years ago, why anything is still being invented, perfected... It doesn't matter in principle. even you as a living intelligent human being ("living mechanism") can be predicted, with some margin of error, and to think that all discoveries or the like have already been made I think is silly.

We have different notions of what the market changes or does not change... I assume that for you, if the Expert Advisor, the grid... ... I don't think the market has changed, but for me it means the model, Expert Advisor and grid has not been built correctly.

In short, I just wanted to help not to scatter on unnecessary tasks and invented properties of the market... But probably, fuck it...

 
LeoV >> :

To the question about error: Why do you need mediator in the form of direction prediction?

It may be just the situation when the net direction is guessed right but it is local, but for example the global one was wrong.

So here is minus in profit. If you want to profit-train the net, you cannot do it, it is either profit or not.

And so the network as if guessed and as if there is no profit (I hate the turn of phrase "as if", he gives with one hand the other takes away).