Finding a set of indicators to feed into the neural network inputs. Discussion. A tool for evaluating the results. - 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
lea писал(а) >>
The intended purpose is to select a set of variables that are more loosely correlated than the original variables.
I have tried to make a convolutional network somewhat before, for testing I took an M5 time series, a 48-x-48 network, the series was reduced to a symmetric form with respect to 0. The activation function th(). So, I failed to achieve an error on the test sample of more than 0.1 I never could.
I have tried to make a convolutional network somewhat before, for testing I took an M5 time series, a 48-x-48 network, the series was reduced to a symmetric form with respect to 0. The activation function th(). So, I failed to achieve an error of more than 0.1 on the test sample.
It's not really for me to wrap the price series. I already have a set of indicators (i.e. a transformed price series), so the dimensionality of this set needs to be reduced.
I'm not really looking for a price series wrap. There is already a set of indicators (i.e. a transformed price series), so the dimensionality of this set needs to be reduced.
The convolution network and PCA are the same thing, only the terms are different.
Input X, expect it at the output of the network. The number of neurons of the intermediate layer is less than the input/output layer. Value of neuron outputs
The intermediate layer is considered as a mapping of the input set. These data are used in further processing.
lea писал(а) >>
Has anyone dabbled with principal component analysis (aka "principial component analysis" or "pca")?
Thank you. That's a good question.
iliarr писал(а) >>
if the target function is only the number of trades or only drawdown, it will be of little use, because the network will either learn to enter/exit the market often and aimlessly or learn to avoid drawdowns....
you need to optimize the profit, number of trades and drawdowns ... As I remember JGAP allows to have target function with several outputs. My current priorities are: to solve input data and refine recurrence neuronet.
At the moment, as I see it, searching and testing input data with the method I suggested is of little interest to anyone...
Ilya, I think you will be interested in article in Currency Speculator (http://www.spekulant.ru/archive/2004_11_st11.html). The fitness function is discussed there as well (not only by equity alone, but with its dilution by drawdown and number of trades).
If the target function is only the number of trades or drawdown, the result will be useless, because the network will learn to enter/exit the market often and aimlessly, or learn to avoid drawdowns....
I also think it will be of little use, that's why I wrote "for the sake of interest".
iliarr wrote :>>
you have to optimize both profits and drawdowns... as i remember JGAP allows you to have a target function with multiple outputs...
That's what I'm talking about, multi-criteria optimization. Hmm, I thought, "Oh, my God, what a term," but it's been used for a long time... multi-criteria optimization
I think there should not be multiple fitness functions ("...JGAP allows you to have a target function with multiple outputs..."), but one, but with several necessary criteria. I'm just now speculating on how to approach this issue discreetly and grab it by the nostrils... Can anyone advise literature on the subject?
marketeer wrote(a) >>
If a network is trained without a teacher to make hypothetically unlimited profits, it should be kept in mind that the inputs still impose a cap on the size of the profits from above. At the selected period of training we can estimate the amount that cannot be exceeded (by a constant lot, by the selected strategy). So we can calculate the learning ratio of the net at this period as the ratio of the theoretical maximum possible profit to the one the net gives. Then similar estimations are done for the validation period, and the ratios are compared...
I agree with iliarr. You'll get a fit
Hmm, while writing this, Daniil beat me to it, he's talking about the same thing too.
Ilya, I think you may be interested in the article in Currency Speculator (http://www.spekulant.ru/archive/2004_11_st11.html). The fitness function is discussed there too (not only by equity, but with its dilution by drawdown and number of deals).
Interesting article. >>Thank you.
Most interesting to me was:
-the author's definition of the target function "Fit = Profit/(l+Prosadka/ 1 0 + S t o p / 1 0 ) * (CountTrade/10);" I was thinking about a similar one, but here is a ready solution...
- Extracting the MTS with a genetic algorithm. A well-formulated idea is half the solution... There are a lot of advantages to this approach... I'll have to think hard about how to implement it better and easier...
Right now, thanks to:
lea wrote >>
Has anyone dabbled with the principal component method (aka "principial component analysis" or "pca")?
I am very interested in compression of information fed into neural network by eliminating correlation
I tried to make a convolutional network earlier, for testing I took M5 time series, 48-x-48 network, the series was reduced to a symmetric form relative to 0. The activation function th(). So, I failed to achieve an error of more than 0.1 on the test sample.
So I must have used it wrong, or cooked it.
I myself have worked with picture compression. Sometimes the error is zero, sometimes not, depends on the degree of compression (number of principal components) and informativeness of the inputs.
Try with simple examples.
Here are a couple of books on optimisation. Just downloaded, still hot.
........ can't seem to attach. I got it from http://torrents.ru
And you calculated all this in what? MathCad/MathLab?
This is hard for me to believe myself, but the calculation was done in excel. IMHO it's a bit better than Matcad in terms of understanding (visualisation of the calculation process rather than the final state).