Machine learning in trading: theory, models, practice and algo-trading - page 582
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 learned that machine learning uses such a thing as feature construction. You can't get far on price alone. The feature in our case is a function of price. The question is which functions to use. Simply going through the indicators with different parameters is not an option. I am interested in materials on this subject. Google usually produces various rubbish, or rather it produces nothing on the subject. I searched on Runet. Maybe someone knows some materials on the subject.
PS. You have to start from the beginning. That's when you learned to construct signs not at random, you can move on to their selection.
Good luck, and read more.
Maxim, do you carefully read the posts of a certainVizard_? This guy does some genius stuff - look at my thread.
You may have a look at this, this, this and maybe this.
Good luck and read more.
I've read it all, the problem is that the significance of features varies greatly from sample to sample, and all these manipulations are good for statistical analysis and not for real trading
optimization, in other words.
Oh, not me... sorry :D
Maxim, do you carefully read the posts of a certainVizard_? This guy does some genius stuff - look at my thread.
Yes, but he is so excellent that I will not drag at this stage :) + he wrote that it's impossible to earn more than 20% per annum... I guess you should always start with such statements, and then go into details :)
Look at this, this, this and maybe this.
Good luck and read more.
I realized that the answer to my question will not find. Where I have been able to read, they write that it is an art. The problem is that there are a lot of indicators, and even more parameters which can be set. Why should I check all possible combinations using the brute force method? I do not know how to do it. I don't know how to do it. You still need to have initial selection rules and then work with selected indicators by attribute selection methods.
I read it all, the problem is that the importance of features varies greatly from sample to sample,
Are you sure the signs even exist? Or, for example, there could be a lot of different ones. Naturally, in this case, you can't pick them up with brute force and all sorts of combinations, and the only way is for the system to determine them itself. In this light, the search for certain predictors and their combinations looks like a crazy business (well, we will always find them successfully in history, of course)).
I see only one way here - the system should identify this set of signs on its own in training, and not for you, but for itself. And our task is only to prepare the data for training. The only task of this preparation is to kind of tell the system some a priori data, thus reducing the area of applicability of the system. Or, in other words, to cut off the intervals where transactions are clearly not feasible.
We should only clearly formulate the problem, but not try to solve it for the system. Imho, everything there is too complicated to even try to solve it with brute force.
I would like to point out that ada gives better results than rf: both more accurate and less prone to overtraining. And you should use ada, not rf.
So it's not just a matter of piling everything up.
GARCH is too complicated. So far I've made my way through ARIMA, and also GARCH and distribution.
Took us into the woods and left us there, and now he says that it was not necessary to go into the woods at all. Well, just a Sussanin.
And I was just about to do them).
Are you sure the signs even exist? Or, for example, there could be a lot of different ones. Naturally, in this case, you can't pick them up with brute force and all sorts of combinations, and the only way is for the system to determine them itself. In this light, the search for certain predictors and their combinations looks like a crazy business (well, we will always find them successfully in history, of course)).
I see only one way here - the system must identify this set of signs on its own in training, and not for you, but for itself. And our task is only to prepare the data for training. The only task of this preparation is to kind of tell the system some a priori data, thus reducing the area of applicability of the system. Or, in other words, to cut off the intervals where transactions are clearly not feasible.
We should only clearly formulate the problem, but not try to solve it for the system. Imho, it's too complicated to even try to solve it with brute force.
I'm never sure about anything :)
OK, let's say that the main attribute is the price itself. Our task (let's assume, classifications) is to find such a combination of buy/sell that it would be stable on history and would give profit, right? and the corresponding trades in the form of some model. At the same time, you, for example, use a certain set of features (about 20). And how do you make the system itself pick out these features?
Basically, this is an ordinary optimization task with no chance of "artificial intelligence". I have a lot of versions of such systems, the last one was finished yesterday. The result is the same - unstable performance outside the sample, and the ability to achieve almost 100% accuracy on training, and any accuracy (to choose), but the decrease in accuracy does not indicate less overtraining. And you don't need to use R and complex abstruse models for this, the result will be exactly the same.
Yes, but he's so excellent that I won't drag him in at this stage :) + he wrote that it is impossible to earn more than 20% per annum... Maybe you should always start with such statements and then go into details :)
I'm never sure of anything :)
Okay, let's assume that the main attribute is the price itself. Our task (let's say, classifications) is to find such a combination of buy/sell, that it would be stable on history and give profit, right? At the same time, you, for example, use a certain set of features (about 20). And how do you make the system select these signs?
Where did you get that from? Where did you get it from?
I don't use signs for the system. With signs I only cut off from the time series (and from training and from the functioning) the areas where there is no need to analyze anything at all.
The NS itself directly chews the time series.
I have already written and even cited a book -NS performs well for highly specialized tasks in combination with conventional methods.