Neural networks. Questions from the experts. - page 7

[Andrey Dik]

At input A,B,C,D,E,F,g,h,I, target - M

Hidden layer activation function tanh

Three points from each row of data, you gave three rows, a total of 3 * 3 = 9 input neurons.

I trained on all of the data you gave me. And I was given 6502 training examples.

[Andrey Dik]

joo >>:
Хорошо, чуть позже (часа через 2-3), попробую обоснованно показать, каким образом профит (или что то другое, не важно, что мы хотим получить от сети) зависит от фитнес функции. А гарантию того, что мы получим профит в будующем, конечно, никто дать никогда не сможет. А вот к чему стремится должна сетка, пожалуй, мы должны определять для неё однозначно.

Task:

Let's say we have three grids/TCs. Each is tested on a 10-learning dataset. The table shows the abstract target values (optimisable). We,ll be interested in the grid/TS that produces the smallest values of the target function as often as possible. It does not matter which optimization algorithm (AO) we will use.

Example1

All AOs have the same sum of errors. We see that if we use root-mean-square error, AO will choose TC #3 because it has the smallest value.

If the root-mean-square error is used, AO will choose TC #2, and the same is true if the median is used.

Example2

The situation here is more interesting.

On the one hand TC#1 is not bad, but the error of 200 spoils the picture. TC3 has stable results, although not the best.

Once again we see that if we use the root-mean-square error, AS will choose TC3, as it has the lowest score.

And if we use the root-mean-square error, AO will choose CU #2, but on the median the choice will stop at CU #1.

Conclusions.

If the aim of network training is to obtain a curve most similar in form to the target one, then the root-mean-square error should be used (Tasks of approximation)

If the goal of training the network is to get the smallest/lowest values of the target function as often as possible, you should use root-mean-square error (Classification/Clustering problems).

[[Deleted]]

Looked at it, now I understand why you have the 6th digit. You took data from the future) 2 points that do not exist in the real world)) In fact, your network made a forecast of column M1 knowing columns A2 and A3))) Hence the increase in accuracy. Although notice, gave the data from the future (such a not-so-sweet hint directly say), and the accuracy has increased from 2-005 to 7-006)))) Funny.

[Andrey Dik]

mrstock >>:
Посмотрел, теперь понял почему у Вас 6-ой знак. Вы брали данные из будущего) целых 2 точки, которых в реале не существует) Фактически ваша сеть делала прогноз столбца М1 зная столбы А2 и А3))) Отсюда и рост точности. Хотя заметьте, дали данные из будушего (такая некислая подсказка прямо скажем), а точность выросла с 2-005 до 7-006)))) Забавно.

The order in which you gave me the data is the order in which the training was carried out. You can put the data in reverse order and you should get the same result. This is an approximation problem, and it makes no difference which direction to train in.

[[Deleted]]

Follow-up. I don't recommend using tangents in statistics at all. These guys work wonders. They once predicted the price 25 bars ahead with apupic accuracy. I was looking for the error at first, and then I realised that these guys just made it, but it was really beautiful)))) I use only identity, it most accurately describes the problems I set and does not suffer from history adjustments.

[Andrey Dik]

mrstock >>:
В догонку. Не рекомендую вообще использовать тангенсы в статистике. Эти ребята творят чудеса. Они мне как то раз спрогнозировали цену на 25 баров вперед с апупительной точностью. Я сначала долго искал ошибку, а потом понял, что эти умельцы, тупо подогнали результат, но было чень красиво)))) Я использую только identity он наиболее точно описывают те кзадачи, которые я ставлю и не страдают подгонкой на истории.

Save the code in C++, take a look, there are no miracles there.

PS I don't use Statistica in trading.

[[Deleted]]

Have a look at the attached file.

Are the values from it sent directly to the NS input, or are they normalized?

I understood on the fxexpert.ru forum in the topic "Neural Network Principles of MTS Creation" that eventually they came to a conclusion,

That it is necessary to normalize the values and not take the indicator values or quotes directly, but their changes.

[[Deleted]]

How could it not be? In fact in column A2 we have Tuesday in A3-Wednesday (conventionally) we predict EMA for Monday (a1) and as the EMA went in reverse order, which depends just on subsequent clauses) So there is a difference. Anyway thanks))))

[Andrey Dik]

Qwer791 >>:

Посмотрел вложенный файл.

Значения из него непосредственно подаются на вход НС, или всетаки нормируются?

На форуме fxexpert.ru в теме" Нейросетевые принципы создания МТС" я так понял,в конечном итоге пришли к выводу,

что необходимо обязательно значения нормировать, и брать не непосредственные значения индикаторов или котировок, а их изменение.

This has been discussed before in this thread. The top starter wanted to work exactly the way... he does.

[Леонид]

joo писал(а) >> OK, a little later (in about 2-3 hours), I'll try to reasonably show how the profit (or whatever, whatever we want to get from the network) depends on the fitness function.

joo wrote >> Conclusions.

If the goal of training the network is to obtain a curve most similar in shape to the target, then the root-mean-square error should be used (Approximation Tasks)

If the aim of training a network is to get the smallest/lowest values of the target function as often as possible, you must use root-mean-square error (classification/clustering tasks)

Honestly, didn't get a sense of how the profit depends on the error....))))