Machine learning in trading: theory, models, practice and algo-trading - page 584
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
You didn't even know how to determine the importance of predictors in RF, by giving some nonsense about annealing and so on without any explanation (what does this have to do with anything?).
Who said where are the benches exactly for forex applications? why Ada and not GBM? your answers are too much fuzzy abstractions. in reality the gain will not be more than 5% with more overtraining.
At the level at which this discussion is taking place.
Let me clarify the level of ALGLIB - the level of the collective farm, villages near Novgorod. You have repeatedly written that this level suits you. It is quite possible that for your purposes it is enough, but why be offended?
You are not so sure aboutthe level of your income, but you are very happy with it.
You're wrong.
If about the selection of predictors, I've tried almost all of R and this annealing is the most effective.
Why Ada and not GBM? your answers are too many fuzzy abstractions. in reality the gain will not be more than 5% with more overtraining.
Because I've tried them and not just them. I still have the protocols.
Yes, the best is ada? Yes, 5%, maximum 7% in relation to the forest. And I don't know anything better than that.
And what is "great overtraining"? What are you talking about? About overtraining, I can't recall a single post from you in which you show that your models are not overtrained!
I can only repeat that overtraining does not depend at all on the model, it depends on:
At the level at which the discussion is taking place
I clarify the level of ALGLIB - the level of the collective farm, the village near Novgorod. You have repeatedly written that this level suits you. It is quite possible that for your purposes it is enough, but why be offended?
I am not so sure, but I am sure that you will be satisfied with it.
You're wrong.
If about the selection of predictors, I've tried almost all of R and this annealing is the most effective.
Why Ada and not GBM? your answers are too many fuzzy abstractions. in reality the gain will not be more than 5% with more overtraining.
Because I've tried them and not just them. I still have the protocols.
Yes, the best is ada? Yes, 5%, maximum 7% in relation to the forest. And I don't know anything better than that.
And what is "great overtraining"? What are you talking about? About overtraining, I can't recall a single post from you in which you show that your models are not overtrained!
I can only repeat that overtraining does not depend at all on the model, it depends on:
What is the difference between binning and boosting? In binning there is less fitting initially and more element of randomness, while in boosting one is fitted on the leftovers of the second, then on the third and so on. And you end up with a total overfit. I.e. RF can already be quite "coarse" from the very beginning, but I'll have to double-check it, I haven't had time yet.
All my models are retrained :) since I haven't found any permanent patterns for them yet
Alglib has almost everything - convolution, PCA, clustering, neural network ensemble, forrest... so, according to the classics, everything is there, what else you need - I don't understand :) more modern things, of course
And the author writes that he doesn't treat neural networks etc. with great reverence, but considers them as normal tools of cassification/regression and doesn't distinguish them from other methods. I like this realistic approach.
Regarding annealing and stuff like that, I don't get it either - is it some universal way for all models or what? Each model should have its own way of evaluation, through which this particular thing can be trained in the best way
Well, what is the difference between banging and boosting? In banging there is less fitting initially and more element of randomness, and in boosting one is fitted on the rest of the second, then on the third, and so on. And you end up with a total overfit. I.e. RF can already be quite "coarse" from the very beginning, but I'll have to double-check it, I haven't had time yet.
All my models are retrained :) since I haven't found any permanent patterns for them yet
Alglib has almost everything - convolution, PCA, clustering, neural network ensemble, forrest... so, according to the classics, everything is there, what else do you need - I don't understand :) more modern things, of course
And the author writes that he doesn't treat neural networks and so on with great reverence, but considers them as normal tools of cassification/regression and doesn't distinguish them from other methods. I like this realistic approach.
Regarding annealing and so on, I also don't understand - is it some universal way for all models or what? Each model should have its own way of evaluation, through which this particular thing can be trained in the best way
I've tried several times to explain some elementary things to you, from my point of view. Unsuccessfully.
I can only advise: spend a couple of months on caret and you will have a different way of thinking, a qualitatively different outlook.
About annealing and so on, it is also not clear - is it some universal way for all models or what? Each model must have a way of evaluation, through which this particular thing can be trained in the best way
Annealing is annealing in Africa, and the goals/objectives are roughly the same. Allows the model to find not local minmaxes, but sort of global ones.
I don't know about ADA, but annealing gives very good results for NS. I don't like the built-in one, because annealing parameters need to be set in advance, that's why I annealed manually, changing the parameters based on the results of previous training.
By the way, more or less complicated NS without annealing generally not really teach anything.
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 :)
Maxim, stop smoking. Taking things out of context, attributing other people's words to others, etc.
+ some of the comments deleted. Also do not attribute authorship of the word rattlesnake Fa (Fomenko)).
Maximka, stop smoking. Taking things out of context, attributing other people's words, etc.
+ some of the comments deleted. Just do not attribute and authorship of the word rattlesnake Fa (Fomenko)).
I'm just saying :) what's on my mind is on my tongue
And then something will turn up... the working process is impersonal.
I was wrong about the rattle.) SanSanych wrote something similar once... nonsense or something like that
From idleness and complete lack of any ideas for further work, I decided to learn something new, for myself of course - may be it is already very old). I started with RF, through RF I came to Python, because it is compatible (as they say) in both directions with my SciLab software. Now I have come to review packages for Python.
In total, there are over 120,000 packages. Of them on Machine Learning - about 70, on neural networks, including deep learning - about 70. Possibly there are much more - I was searching by rubrics and some packages could appear in other sections.
This is not counting the packages distributed directly by other companies. There are a lot of such packages, too. There are also on subjects interesting for us - I've seen them myself, including those on Defense Ministry, National Assembly of Russian Federation and the ADA.
Among other companies there are machine learning, trees, NS, and something related to ADA.
Many packages are made in C/C++, so there is no need to worry about performance - Python is just the interface (scripting language). Just like R, in fact.
All in all, I am having an interesting time).
Why Ada and not GBM? There are too many vague abstractions in your answers. In reality the gain will not be more than 5% with more overtraining.
When classifying, very often "accuracy" - the percentage of correct answers - is used to evaluate the model. In my opinion this is one of the weakest and most inappropriate evaluations of trading models and should be avoided. I've suggested trying a bunch of others here in the thread - kappa, f-score, logloss.
Ada in R (maybe not only in R) uses a slightly different built-in estimation of classification model in training, which is much better compared to "accuracy".
From idleness and complete lack of any ideas for further work, I decided to learn something new, for myself of course - may be it is already very old). I started with RF, through RF I came to Python, because it is compatible (as they say) in both directions with my SciLab software. Now I've come to package reviews.
All in all, there are over 120,000 packages. Of these, on Machine Learning - about 70, on neural networks, including deep learning - about 70. Possibly there are much more - I was searching by rubrics and some packages could appear in other sections.
This is not counting the packages distributed directly by other companies. There are a lot of such packages, too. There are also on subjects interesting for us - I've seen them myself, including those on Defense Ministry, National Assembly of Russian Federation and the ADA.
Among other companies there are machine learning, trees, NS, and something related to ADA.
Many packages are made in C/C++, so there is no need to worry about performance - Python is just the interface (scripting language). So, in fact, does R.
All in all, I am having an interesting time).
Check out more of this stuff https://cloud.google.com/datalab/
Wicked, the AutoML direction is developing there - the service will pick up the model by itself for certain tasks
When categorizing, it is very common to use "accuracy" - the percentage of correct answers. In my opinion, this is one of the weakest and most inappropriate evaluations of trading models, and should be avoided. I've suggested trying a bunch of others here in the thread - kappa, f-score, logloss.
Ada in R (maybe not only in R) uses a slightly different built-in evaluation of the classification model in training, which is much better compared to "accuracy".
It's hard to evaluate trading ones in general this way, there's deal duration and stop loss levels have to be added to everything else, and it also needs to be retrained periodically... so that's a shame :)