Neural networks. Questions from the experts. - page 8
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Честно говоря, не ощутил каким образом профит зависит от ошибки....))))
For example, let's say you are interested in the TS generating as much profit as possible and as often as possible, that is, you are trying to increase the percentage of profitable trades and of course the MO.
A network trained on this principle can be expected to produce profits on OOS as well. You need to apply a root mean square error that accentuates the network on the patterns that contribute to these goals. That is, the network focuses on specific patterns that lead to some effect.
If you use root-mean-square error, you are "averaging" the patterns, not emphasising them.
For example, let's say you are interested in the TS generating as much profit as possible and as often as possible, that is, you are trying to increase the percentage of profitable trades and of course the MO.
A network trained on this principle can be expected to produce profits on OOS as well. You need to apply a root mean square error that accentuates the network on the patterns that contribute to these goals. That is, the network focuses on specific patterns that lead to some effect.
If you use root-mean-square error, you are "averaging" the patterns, not emphasising them.
What is root-mean-square error?
to LeoV
Here's the problem for you, the inverse, on profit maximisation. Which TS will you choose, the values are specified, hmm, in $
to Vinin
The arithmetic mean of the sum of the roots.
Об этом уже говорилось раньше в этой ветке. Топикстартер хотел именно так работать, как... как он работает.
This is the first time I've met someone on this forum who thinks pretty much the same way I do... :)
Single-minded...
Впервые встречаю на форуме человека который мыслит практически также как и я... :)
Единомышлинник...
Thank you. It's nice to meet brothers in intelligence... :)
Profit<-->mistake:
I believe(confirmed experimentally of course).
If the network error is saved on the feedback, the increase/decrease in equity will be saved in cases where the network signal is much better than random.
If the network error is not saved, equity would be random, i.e., equity could rise or fall (slowly/fastly/ jumps) but it would still be random.
The profit<-->error ratio can be determined for each problem.
//-------------------------------------------//
As for the training that LeoV is doing:
Naturally conjecture and mostly reasoning from his posts.
Training is done using genetic algorithm, you can define any fitness function in it...
IN NSH4-5... You train not only profits, but different combinations of system indicators, by maximizing/minimizing we pull the whole system, not only profits, i.e. we get smooth growth of equity.
I forgot to add: "With a genetic algorithm it really doesn't know how to train in order to get stable results on OOS, the sliding control/testing method on an independent sample won't help here anymore, even the opposite will cheat...".
That's why the error may be different but the profit is about the same, the target function is not error minimization.
For me NS 4-5 is a black box, even if I got there a good and stable system with smooth growth of equity, which was tested on QE, I just put it aside for a worse time.
If you train networks not in NSh4-5, but in programs created for more academic purposes, you can at least understand what the mistake is, why the network does not bring profit, and a bunch of different issues, and after finding the answer to which you can confidently talk about neural network trading.
And not so that there is no connection to anything, God only knows if it will work or not, if inputs need to decorrelate or not, if inputs need to remove correlation with output or not, if commit or non-committal is better, etc. ..... Simply put random rambling...
Для меня НШ 4-5 чёрный ящик, даже получив там прибульную стабильную систему с плавным ростом эквити, прошедшую тестирование на ООС, я просто отложил её до худших времён.
Если обучать сетки не в НШ4-5, а в программах созданных для более академических целей, то можно хотя бы разбираться в чём ошибка, почему сеть не приности прибыль, и ещё кучу разных вопросов, найдя ответ на которые, можно будет УВЕРЕННО говорить о торговле с помощью нейронных сетей.
А не так что связей нигде ни с чем нет, на ООС только богу известно будет оно работать или нет, входы надо декоррелировать/не надо, нужно убирать корреляцию входов с выходом или нет, что лучше коммитет или не коммитет и тд..... Проще говоря случайное блуждание...
It is for these reasons that I have, for some time now, completely abandoned 'ready-made' network packages. I prepare what I need myself.
And not so that there are no connections to anything, on the feedback only God knows whether it will work or not, inputs need to be decorrelated or not, inputs need to be correlated with output or not, what is better than commit or non-commit, etc. ..... Simply put random rambling...
Как успехи на этом поприще?
It's too early to talk about successes yet. When you are successful, perhaps I will pay for your round-trip ticket to visit me. :)
For now, I am satisfied with the fact that at least I have full control over the learning process and the learning objectives of the networks (if that's what you mean).
This is a good thing. )))) My pleasure.))) Thank you)))
But basically, what was my question? This topic is very interesting to me. I was asking about the relationship between error and profit on OOS. It is a very interesting topic, due to the fact that talking to many professionals in this business, they do not know the answer to this question. What did you say to me?
Let's say you are interested in TS generating as much profit as possible and as often as possible, that is, trying to increase the percentage of profitable trades and of course the MO.
You can expect the network trained on this principle to produce profits on OOS as well. You need to apply a root mean square error that accentuates the network on the patterns that contribute to these goals. That is, the network focuses on specific patterns that lead to some kind of consequence.
If you use root-mean-square error, you are "averaging" the patterns rather than emphasising them.
и
Regarding profit<--> error:
I believe(confirmed experimentally of course).
If the network error is saved on the feedback, the growth/decline of equity will also be saved, it applies to cases where the network signal is much better than a random one.
If the network error is not saved, equity would be random, i.e., equity could rise or fall (slowly/fastly/ jumps) but it would still be random.
Profit<-->error can be determined for each task.
Well these are not answers, you have to understand))). These are just reflections "on the subject" in general. Ok, we take NS (not trader) or Solution, whatever (for "academic purposes"), make a network (whatever it is) and start training it. Until when do we train it? Until a minimum error? It should be understood, that it will be a 100% overtraining. Not up to the minimum error? Then until what error? What is the profit? And why exactly to this error? Will the profit increase or decrease if we slightly decrease the error? And if you increase the error?
So it's like this.....))))