Discussion of article "Random Forests Predict Trends" - page 4
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
Model and predictor selection are interrelated. First, one should select a model, and then select predictors based on this model by screening those predictors that have the least "usefulness" in prediction by the same model. Although many articles and textbooks teach otherwise: first we select predictors using some method of calculating the relationship between these predictors and the target series - the output. The most common methods of screening are correlation coefficient between predictors and output and mutual information. Then a model is selected usually unrelated to how the predictors were selected. If you think about it (and econometrics textbooks won't tell you this, you have to think for yourself), the method of selecting predictors by correlation coefficient with output essentially selects those predictors that will have the smallest error in a linear regression model (LRC). The method of selecting predictors by their mutual information with output essentially selects those predictors that will give the lowest error in a Nadaraya-Watson regression based model (abstruse name GRNN).
Totally agree with you about correlation. the effect of predictors on the target variable cannot be established by correlation and is not established by regression. It is done differently. The most popular is the Gini index, but I have managed to use it and use my own considerations and some sequence of actions. while I have managed to pick up some set of predictors for trend prediction, I have not managed to find a set of predictors for predicting price increment.
I would like to draw your attention to my book. The problem of predictors is much more complex than correlation and Gini index. The book clarifies a lot of things.
faa1947:
I would like to draw your attention to my book. The problem of predictors is much more complex than correlation and Gini index. The book clarifies a lot of things.
I would like to draw your attention to my book. The problem of predictors is much more complex than correlation and Gini index. The book clarifies a lot of things.
The book is not there, there is an advert.
There used to be little information about networks here. Folks wanted to study these nets and try them in trading. And now there are a lot of books and articles about networks with step-by-step instructions on how to use them. I look at these books and articles about networks and they not only discourage me from reading, but even cause some disgust to networks. The problem with these writings is that they do not even try to interest the reader: read and use them. And why waste time? Where is the bait? Show in the beginning of the book or article an attractive result of trading using the methods described in the book or article and we will be interested in reading and understanding them. I am looking at a new article about deep nets and I think who will read it except for a couple of specialists who already know about them? I also already know about these networks, and I know that they as well as other networks are not applicable to trading on the market. Even the inventor of these networks, Jeffrey Hinton, recognised this long ago. Listen to his lectures on YouTube.
The book is not there, there is an advert.
There used to be little information about nets here. People wanted to study these networks and try them in trading. And now there are a lot of books and articles about networks with step-by-step instructions on how to use them. I look at these books and articles about networks and they not only discourage me from reading, but even cause some disgust to networks. The problem with these writings is that they do not even try to interest the reader: read and use them. And why waste time? Where is the bait? Show in the beginning of the book or article an attractive result of trading using the methods described in the book or article and we will be interested in reading and understanding them. I am looking at a new article about deep nets and I think who will read it except for a couple of specialists who already know about them? I also already know about these networks, and I know that they as well as other networks are not applicable to trading on the market. Even the inventor of these networks, Jeffrey Hinton, recognised this long ago. Listen to his lectures on YouTube.
The usual task of a trader when developing a TS is to find some set of signals that will predict the future, the so-called pattern. We take ready indicators, buy them, write them ourselves, combine them with each other ...
I argue that there is no such problem. There are algorithms that will find all possible patterns for a given set of predictors. in my article and book, there are about 200 patterns. It is impossible to find anything like that in the traditional way.
Moreover, having mastered R, I have no problem to change one algorithm for finding patterns, for example, neural networks, to deep neural networks, and them to something else. Moreover, there is no need to go into what the algorithm found.
And what is the problem?
And in what you wrote about in your post above: proper selection of predictors. I'll add. Proper preprocessing of predictors. That's a skill. As a result of reading books, you will have that skill, since you have thought about it yourself.
And my book? It is a superficial review of the whole problem, not just specific algorithms for pattern search. I guarantee you one result: you will be fluent in several patterns at a sufficient level for trading, without really delving into these patterns, with all sorts of perseptrons, layers, bugging and bousting - all this will be unnecessary for you. You will focus on predictors.
It's a completely different approach.
And I want to finish by reminding you of the axioms of statistics: "Garbage in - rubbish out". And no model, no algorithm can change that. Therefore, instead of a hyphen, we should put a black box with some name and not worry about it, but deal with trashing.
The usual task of a trader when developing a TS is to find some set of signals that will predict the future, the so-called pattern. We take ready indicators, buy them, write them ourselves, combine them among ourselves ....
I claim that this problem does not exist. There are algorithms that will find all possible patterns for a given set of predictors. in my article and book, there are about 200 patterns. It is impossible to find anything like that in the traditional way.
Moreover, having mastered R, I have no problem to change one algorithm for finding patterns, for example, neural networks, to deep neural networks, and them to something else. Moreover, there is no need to go into what the algorithm found.
And what is the problem?
And in what you wrote about in your posts above: proper selection of predictors. I would add. Proper preprocessing of predictors. That's a skill. As a result of reading books you will have that skill, so you have thought of it yourself.
And my book? It is a superficial review of the whole problem. At the same time I guarantee you one result: you will be fluent in several models at a sufficient level for trading, without really delving into these models, all kinds of perseptrons, layers, bugging and bousting - all this will be unnecessary for you.
A completely different approach.
Do you use these methods yourself in trading? And what are the results? I'm serious, at least hint at the results. For example, I earned enough that I don't need to write books, I bought a villa in Nice or the Bahamas and now I'm on holiday, I'm doing philanthropy, I'm giving away books for free.
If you manage to find a set of predictors, you will realise your list.
PS.
And the book? Allows you to gather a pleasant party, and the price cuts off the grail seekers.
Tell me, is it not possible to cite Out Of Sample at least one?
PS. Sent you an email.
Tell me, is it not possible to cite Out Of Sample at least one?
PS. Sent you an email.
Table 2 in section 5.3 of my article. The rattle() package gives the ALE automatically plus other very useful information, which is shown in the article. In addition, programme code is generated under all this information, which can be used standalone without rattle(). My book is 400 pages long, so everything is chewed up in great detail including the ideology of use, which is not in the original documentation on rattle() and the packages it uses. rattle is a shell, a GUI.
PS.
I replied to your email
Table 2 in section 5.3 of my article. The rattle() package gives the ALE automatically plus other very useful information, which is shown in the article. In addition, all this information is used to generate programme code that can be used standalone without rattle(). My book is 400 pages long, so everything is chewed up in great detail including the ideology of use, which is not in the original documentation on rattle() and the packages it uses. rattle is a shell, a GUI.
PS.
Email replied
I meant Out Of Sample test in MT4 - purely profit from the model is of interest. Let's say that + or - zigzag with an error of 17...20% in practice can turn into another great drainer.
PS. email caught, I hope I can pay soon (need to wait for accamulation of money)
I meant Out Of Sample test in MT4 - purely profit from the model is of interest. Let's say that + or - zigzag with an error of 17...20% in practice can turn into another great drainer.
PS. email caught, hopefully I can pay soon (need to wait for accamulation of money)