Machine learning in trading: theory, models, practice and algo-trading - page 2297
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
https://www.mql5.com/ru/forum/86386/page2108#comment_19209601
Thank you very much, that's what I need!
I am thinking to write over-sampling in MT5 at once. Can anyone here suggest formulas for creating new data elements for oversampling?
Smote as I understand it "new elements are created directly near existing ones":
take the mean, RMS, variance (you can cut out outliers) for each attribute, and then take a random item, add a value within +/- RMS to each attribute of it, and thus multiply as many new ones as you need.
It seems like this should be enough, what do you think?
Aleksey Nikolayev:
Nevertheless, I would recommend tsosnikov to study these tests, and not just run around for years with ideas how to fit a bundle of sine waves to the market)
In fact, I do these tests, looking for differences in noise and price. Apart from vola, I did not find anything serious.
In fact, I do these tests, looking for differences in noise and price. Apart from the vola, I have not found anything serious.
good luck to you
In fact, I do these tests, looking for differences in noise and price. I did not find anything serious except the vola.
I have similar feelings about Fourier series - they are quite suitable for representing daily volatility fluctuations, but are completely inapplicable to the price.
I think a meaningful work with such tests leads to an increased level of mathematical culture, which would be useful for the forum in general (or at least for this thread).
I have a similar feeling about Fourier series - they are quite suitable for representing daily volatility fluctuations, but are completely inapplicable to the price itself.
In my opinion, a meaningful work with such tests leads to an increased level of mathematical culture, which would not be superfluous for the forum as a whole (or at least for this thread).
There are enough claims to Fourier, for example, cycles can be tied to calendar dates, and they do not go well with sine waves, transition to winter time, etc. Counting zigzag knees seems more reasonable. But pf and tsos is a well researched area, many problems have already been solved by someone.
There are enough claims to Fourier, for example, cycles can be tied to calendar dates, and they do not fit well on sine waves, the transition to winter time, etc. Counting zigzag knees looks more reasonable. But pf and tsos is a well researched area, many problems have already been solved by someone else.
It's up to you, I can only advise you to consider the level of significance for the results obtained (as in the tests I mentioned), since SB implementations may well seem like noisy periodic oscillations too.
It is up to you, I can only advise you to consider the level of significance for the results obtained (as in the tests I mentioned), because SB realizations can also quite seem like noisy periodic oscillations.
In fact, I'm not really that keen on csos.
Finding a pattern is half the battle, you still have to use it somehow. The simplest example is circular levels.
Rorschach:
The simplest example are circular levels.
Yeah, it's not even that trivial to imagine them AMO...
but 1.5% quality adds up
Finding a pattern is half the battle, you still have to use it somehow.
Well, it's a rather intimate question and it's unlikely anyone will share their ideas about it (regardless of how much is actually earned on these ideas)
Thank you very much, that's what I need!
I'm thinking to write over-sampling in MT5 at once. Can anyone here suggest formulas for creating new data elements for oversampling?
Smote as I understand it "new elements are created directly next to existing ones":
take the mean, RMS, variance (you can cut out outliers) for each attribute, and then take a random item, add a value within +/- RMS to each attribute of it, and thus multiply as many new ones as you need.
It seems that this should be enough, what do you think?
Easiest option is just to accumulate examples from minor class, you can add a little noise to each. I can't remember SMOTE specifically, I think new examples are created by the closest neighbors. There are a lot of tuning options there.