Neuro-forecasting of financial series (based on one article) - page 10
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
I don't care about the articles or their small-minded authors. Especially if the research was done on stationary or stable data.
I base my research on the results of forward-looking tests, not on what some idiot in an article has concocted.
Everyone is entitled to their personal viewpoint without spitting on someone else's.
Everyone has the right to argue that the sun revolves around the earth, but does not have the right to consider dissenters as malacholics or idiots.
This should be considered.
so most are leaning towards nets with 2-3 hidden layers.
All neural network researchers disagree with that statement.
Almost all articles on ns state that the better the network, the more neurons it has, but at the same time, it should not have too many of them.
so most tend towards networks with 2-3 hidden layers.
I don't believe it. Almost all the problems solved by neural networks are solved on a single hidden layer.
Nah, that's right, see above. A network with a single hidden layer will also solve, but sometimes it's easier to add a layer than to fiddle with the size of a single one.
Or vice versa. Start with a generalized mesh, and if successful, simplify the model.
this is something to think about.
Nah, that's right, see above. A network with a single hidden layer will also solve, but it's easier to add a layer than to fiddle with the size of a single one.
OK, I see. I would check it. On artificial and stationary data my tests showed that it is realistic situation when there are not enough neurons in the hidden layer, we can increase the number to get better result on test (validation, as they say in Russian reality) sample. But there is also a real situation when further increasing the number of neurons does not produce a better result. And I never bothered with layers.
In any case, I prefer to reduce the size of the model, following Occam's razor principle.
I don't care about human rights either. Only the results of the forward tests have a say in my opinion, even if they are undemocratic, unconstitutional, oppressive and do not correspond to the publications of morons with degrees and awards from the inventor of dynamite.
The reason is simple: forward test results are closer to the truth. Everything else is closer to misinformation.
Trolling
I do not tolerate rudeness.
It turns out that if you optimize by the minimum drawdown in the deposit currency and then choose this very minimum drawdown from the optimization results, both forwards are successful. If the minimum drawdown is the same for several optimization results, you should choose the one that will have the maximal balance.
I agree, it's possible. But what if we optimize not on the minimum error, but, say, on the maximization of some particular attribute or sets of attributes in the input data. Let's say there is a condition of crossover of MA.
An optimization function should be made that contains the biggest possible number of useful attributes of input data as well as the balance. So we will make the network go in the direction of searching for the maximum number of signs that lead to the balance increase. On the one hand this function should have a balance as a target + n (number of correct MA crossings which lead to profit) which should try to be as big as possible......
Even if we do not obtain the maximal balance for the optimization period, we will obtain the maximal number of signs in the input, which have led to a small, but still increase balance. And then this method should be tested on forwards.... how it will work no....
I.e. optimisation by two parameters. Increase the balance, and increase the number of MA crossings.
Any suggestions on this, or criticism????
I agree, it is possible, but if we optimize not for minimum error, but for maximization of some particular feature or sets of features in the input data. Let's say there is a condition of crossover of MA.
An optimization function should be made that contains the biggest possible number of useful attributes of input data as well as the balance. So we will make the network go in the direction of searching for the maximum number of signs that lead to the balance increase. On the one hand this function should have a balance as a target + n (number of correct MA crossings which lead to profit) which should try to be as big as possible......
Even if we do not obtain the maximal balance for the optimization period, we will obtain the maximal number of signs in the input, which have led to a small, but still increase balance. And then this method should be tested on forwards.... how it will work no....
I.e. optimisation by two parameters. Increase the balance, and increase the number of MA crossings.
Any suggestions on this, or criticism????
Ha!!!!! here we have it. If we can not train the net with this approach (because there is not enough useful information in MA), it will be a sign of uselessness of entering. But it seems to me that there is something here...... I have to check it. By the way, advise here such a thing.
If I have a parameter that is calculated in an indicator. If I want to make optimization of all settings of indicator and variable a reaches 0 on selected plot.
I mean in MT4 how to optimize the parameter, which is calculated and not set????
No... It's not about MA, it's about maximisation of signs, i.e. one part of the optimiser will try to make these intersections as many as possible, and the other part will try to balance as much as possible.....
I.e. we need some kind of optimization function to come up with.....