Machine learning in trading: theory, models, practice and algo-trading - page 550
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
Well, yes, they are not some ready-made econometric models, but just a set of tools universal for all areas.
Nonparametric econometrics is just about MO and fuzzy logic, but I have not seen any clear-cut things yet, maybe because no general approaches have been worked out. I don't know how to do it and I don't know how to do it :)
I'm afraid to give a false hope, Maxim, but in my opinion - neural networks are not applicable in the way I've read about in my articles. You are working with the price itself, though normalizing it, while it is necessary to work with the probability density of the price. This is where there is an intractable contradiction with Heisenberg's uncertainty principle. But, I will not interfere - it is still interesting.
You do not interfere, write what you want - it's a public. And to your distributions, too, will get, not all at once, as long as there is a number of their "indescribable" ideas :D
I got into the dance only because I had to get the importance of features for LR and RF through python or R, then I got carried away and began to explore further
I have a bot in NS, which I have to periodically feed with "good features" on automatic, taking into account changing market, but otherwise it works fine...
You do not interfere, write what you want - it's a public. And to your distributions, too, will get, not all at once, as long as there is a number of their "indescribable" ideas :D
I got into the dance only because I had to get the importance of features for LR and RF through python or R, then I got carried away and began to explore further
If i've got a robot in NS, i should periodically feed it with "good signs" on automatic, taking into account changing market conditions, but otherwise it works fine...
Maxim, what's going on with your Expert Advisor? Where can you find tests or everything is confidential?
https://github.com/RandomKori/Py36MT5 Here are the sources of Python library for MT5. The only problem is with arrays. Passing an array or receiving it does not work correctly. I debugged the DLL code in Visual Studio. It all works. The question is, it may be a terminal bug. How to work with the library is not written. It doesn't make sense. No one needs it without arrays. Although maybe I screwed up in pythom.mqh file. It will be useful to all.
Real arrays work as they should. Arrays do not work long.
Replaced the long for the int. Now everything works. The library can be used. I will only write comments on usage.
Maxim, what's going on with your Expert Advisor? Where can I see the tests or is everything classified?
I am getting better and better... but slowly, because of the complexity of the subject
Some observations/thoughts on how to correctly build a model for the market (from experience):
For neural network classifiers: Class balancing is mandatory, the number of examples for 2 or more classes must be balanced. If the model is being trained on a trend segment, it is possible to mirror signals and add opposite examples. At the same time signs must not correlate with the target, which is quite understandable.
For the neural network regressors: at least one of signs must correlatestrongly with the target, especially when the model is trained on output prices. If this is not done, the regressor will get lost in a tree and will not be able to correctly predict the price if the sample contains repeated or similar examples with different output prices. To do this, in addition to other features, you can feed the normalized increments with a large lag (more than 50). The larger is the sample, the bigger is the lag needed to exclude repeating variants. It is also desirable to feed several such series with increments shifted relative to each other, then each individual case will be almost unambiguously interpreted.
For random forests as classifiers: the same as for NS.
For random forests as regressors: almost useless, the same as for NS, but it is impossible to give prices for the output (since forests do not extrapolate), and if we give increments for these purposes, the forests will get mixed up in three pines, since there will be many overlapping examples.
Some observations/thoughts on how to correctly build a model for the market (from experience):
For neural network classifiers: Class balancing is mandatory, the number of examples for 2 or more classes must be balanced. If the model is being trained on a trend segment, it is possible to mirror signals and add opposite examples. In this case the signs absolutely must not correlate with the target one, which is quite understandable.
Just the opposite is true. I have evidence all the way down to the actual work. One thing that is clear to me is that you can't build a classification on coffee grounds and Saturn's rings, EVER - it's shamanism. Out-of-sample models give about the same result as in-sample only if you clean the input set of predictors of noise and leave only those that are relevant to the target variable. Furthermore I have an algorithm for cleaning up the noise and calculation results for remaining predictors gives marginal error with which to classify for this set of predictors.
The exact opposite is true. I have evidence all the way up to the real-world work. One thing that is clear to me is that you can't build a classification on coffee grounds and Saturn's rings, NEVER - it's shamanism. Out-of-sample models give about the same result as in-sample only if you clean the input set of predictors of noise and leave only those that are relevant to the target variable. Moreover, I have an algorithm for cleaning up the noise, and the calculation results for the remaining predictors give a marginal error with which to classify for that set of predictors.
For me this question is not obvious at all, if we get a non-linear classifier... what should correlate there with what? ) And if we have categorical features, then the regression will not work at all, but the classification will.
But for regression this is justified