Machine learning in trading: theory, models, practice and algo-trading - page 937
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
Yes, you can have many classes. But I rarely worked with models with more than 2 classes, they always have difficulties and problems, especially if one class exceeds the others. It is possible to try, but I don't hope for a good result, to be honest.
Now it's a bit complicated - one tree is a filter that knows where you can't go long, another tree knows where you can go long, while missing 2 similar trees for shorts. Maybe combining all this information into one tree will overcome disadvantages of using three classes, and the solitaire will come together :)
Yes, you can have many classes. But I rarely worked with models with more than 2 classes, they always have difficulties and problems, especially if one class exceeds the others. It is possible to try, but I don't hope for a good result, to be honest.
Now it's a bit complicated - one tree is a filter that knows where you can't go long, another tree knows where you can go long, while missing 2 similar trees for shorts. Maybe combining all this information into one tree will overcome disadvantages of using three classes, and the puzzle will work out :)
It will be interesting to see the result!
I don't know, maybe I should make new targets with stupid take profits, then there might be more entries. I think there is no certainty about the reversal, but there is certainty about finishing the trend movement, and one can expect flat, i.e. we get noise targets there. Where there is a file with a small number of entries, there should be clean targets, but there will also be different types of entries - the exit from the flat, the entry after the correction to the continuation of the movement, the entry to the reversal (for example, during the formation of the pin bar on the hourly TF). Besides, there is such a nuance as "an open entry window" when the entry is allowed during a decent number of bars in a row - there may also appear noises from accidents. So I think, maybe we should split the targets to make the picture converge?
Have you experimented with kfold alglib? does it improve the results? Saw in one of your old posts that it didn't seem to mix the sample. Does it even have to?
any conclusions as to what overtrains less - forest or mlp? on the same data. I have suspicions that for regression tasks the forest works crookedly and returns errors crookedly (very small), for classification norms.
2. I saw you were interested in OpCl, was there any thought of rewriting NS for it? For example, I gave up GA altogether as some nightmare nonsense, everything is now trained in 1 thread on 1 core. You could speed it up on Cl (although it's as fast as that). Or, if you train on Spark, it's already parallelized there and there's no point.
More or less understood what you put on the git and how you use it. Great interesting work, respect! :)
Hi)
I scanned the code on the washing machine and turned it on to teach it.
Then I sewed the code on the sewing machine three times to train the neural network in depth.
And here is the result - the code has changed beyond recognition, and now I wonder how to use it in trade...
Is the super bot ready yet?
I've been waiting for two years.
are there any conclusions as to what is less overtrained, the forest or the MLP?
In one of the packages there is a demo-example of predictive NS (MLP). Relatively uncomplicated NS, by the way. So it's unreal to retrain it. No matter how you retrain it, everything works.
So, retraining depends, first of all, on the task set for the NS. Well, and on the adequacy of the complexity of the MLP to the task.
I mean specifically about alglib and about tests on the same data of different models
The NS is the same as the NS. It makes no difference, alglib, or what. If NS retrains, then it is inadequate to the task, and it is necessary to change either the task or the structure of NS. And don't bother). As one ensign used to say - f... think here, you have to shake it).
I can't say anything about the forests, I don't work with them.
What if we train the tree on a non-target, non-predictor? Will we be able to identify non-significant predictors, on the basis of good training? That is, if the test model recognizes predictors well, then the other predictors already contain enough information and that predictor is redundant and can be gotten rid of. But what if it is not about the predictor, but about the redundancy of the unique combinations themselves?
For example, I decided to choose the hour number as the target.
The result on the test sample
The result on the test sample
You can see that 10 o'clock is well predicted, can we make the assumption that there are different conditions at that time than the rest of the sample? I think so. I will try to group the remaining hours and see if I can better distinguish the peculiarities of price behavior in these time ranges.
What if we train the tree on a non-target, non-predictor? Will we be able to identify non-significant predictors, on the basis of good training? That is, if the test model recognizes predictors well, then the other predictors already contain enough information and that predictor is redundant and can be gotten rid of. But what if it is not about the predictor, but about the redundancy of the unique combinations themselves?
For example, I decided to choose the hour number as the target.
The result on the test sample
The result on the test sample
You can see that 10 o'clock is well predicted, can we make the assumption that there are different conditions at that time than the rest of the sample? I think so. I'll try to group the rest of the hours to see if I can better distinguish the peculiarities of price behavior in those time ranges.
I don't think a day would be enough.) It's actually a specialty). They have been studying it for years, even all their active lives.
NS, not reminiscent, but widely used in DSP. The question of the application of NS in DSP is the correct formulation of the problem for NS. Say, one of the applications of NS is filtering, including adaptive filtering.
By the way, NS will work well when you give it a quite specific task, not bring this, I don't know what.
I keep getting into it.
I've been studying this fic for 5 years and couldn't even imagine that it is possible to apply on a forum.
In short, NS is an ordinary adaptive non-recursive filter.
There are many varieties.
With a teacher, without a teacher, etc.
Predictors are digital filtering coefficients. It's hilarious, because it's elementary, and they're kind of trampling science... All this is as old as tree socks. (It's from the 1950s.)
You tell me - have you done blind or I/O with training?
//Blind is when you don't know what is on the output and there is nothing to compare it with... They say that in this case the main purpose is to reduce the noise.
//If we are using a sighted one, the forecast is calculated using error vector, which is then filtered out. But there is a delay in processing, at least for one tick.