Neural networks. Questions from the experts. - page 4

New comment
 
joo писал(а) >> If NS allows, try to use root-mean-square error instead of rms. Be sure to report back on your impressions.

The point is that there is no information about the relationship or any patterns between error and profit. Moreover, if this relationship is clear and described in the training section, then in the OOS section there is no information about it at all. Logically it seems that the smaller is the error, the larger is the profit, but in practice this is not so and this is proven by numerous experiments (not only mine, of course). The error or root-mean-square error, it does not matter.

 

Good afternoon!

To LeoV, the point is that there is no forecast as such. In my understanding the prognosis is tomorrow's clause or other values from the FUTURE. There is no future here) The picture above is more of a classification task, if you like. If we know the EMa of row 2 if row 1 and its EMa are known, and they are very closely correlated. That was the task. I'm very skeptical about forecasting with neural networks (in financial markets).

To joo. Ok, I'll try it, but I think the result (by converting the priors back to absolute value) will give the same result in the end. I checked it on other things before and the result was the same)

To integer there was no normalization. As for weights, neurons and number of inputs, it's nonsense (or maybe not) I give only 3 values to input and make one neuron and one hidden layer, the result is 2-005e. ANY other number of inputs and neurons and the same result.

p.s. It would be interesting, if anyone dares to run the above data in their programs and try to get a result. Just wondering what will be with others. That we can compare. Anyone interested? )))))

 
mrstock >>:

to integer Нормализации не было. Что касается весов, нейронов и кол-ва входов, то тут вообще бред (а может и не бред) даю на вход всего 3 значения и делаю один нейрон и одни скрытый слой результат 2-005е. ЛЮБОЕ другое кол--во входов и нейронов и тот же результат.

What exactly are the values (values of what)?

LeoV wrote(a) >>

The point is that there is no information about correlation or any patterns between error and profit. Moreover, if this relationship is clear and described in the training area, then in the OOS area there is no information about it at all. Logically it seems that the smaller is the error and the larger is the profit, but in practice this is not so and this is proven by numerous experiments (not only mine, of course). The error or root-mean-square error is not important.

Which error do you use exactly? If you have a desire, and the topicstarter does not mind, I will say why I think there is a difference.

 
gumgum >>:


Попробуйте при этом поиграть с параметром:

- параметр наклона сигмоидальной функции активации.

Thank you. I play with it. And have been for a long time.

double GetAlfa(double x, double y)
  { 
    return(NormalizeDouble((-1.0)* MathLog((1.0/y)-1) / x, ZDigits));
  }

Where x is the maximum absolute value of 97% of the training sample (for the current input);

y - normalized value (mine is 0.99) corresponding to x

The output is the working alpha for the current input.

 
joo писал(а) >> Which error are you using exactly? If you have a desire, and the topicstarter doesn't mind, I will tell you why I think there is a difference.

These errors are all mathematically related, so c'est la vie... ))))

 
LeoV >>:

Все эти ошибки полюбому математически связаны, поэтому се ля ви...))))

I don't want to prove anything to you. I just wanted to hear the answer, which for some reason you are turning away from. :)

 
joo писал(а) >>

I don't want to prove anything to you. I just wanted to hear the answer, which for some reason you are turning away from. :)

About the mistake? I don't use the mistake at all. I don't even look at it. I look at profitability, equity smoothness, drawdown, number of trades and other nonsense....))))

 
LeoV >>:

По поводу ошибки? Ошибку вообще не использую. Даже не смотрю на неё. Смотрю на доходность, плавность эквити, просадку, колличесво сделок и прочую ерунду....))))

I just have a suspicion, no, a certainty, that you are using RMS error in training the network (NeuroShel does not allow otherwise)

 
joo писал(а) >>

I just have a suspicion, no, a certainty, that you are using RMS error in training the network (NeuroShel does not allow otherwise)

It's not for financial markets. It has no stopping criterion. Stop based on the magnitude of the error? I dont understand how))))

 

Which error criterion (fitness function, if you like) is used determines directly the training results.

PS mrstock, by the way, cannot change this either, Statistica does not allow it.

1234567891011...25
New comment
Reason: