Machine learning in trading: theory, models, practice and algo-trading - page 783
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
How do I run Rattle as described in the article https://www.mql5.com/ru/articles/1165?
I did the installation in the R console(install.packages("rattle"))
The R console only has "load workspace" and the article has "File/Workspace;"
Or it's not done on the R console there, but on a different prog?
How do I run Rattle as described in the article https://www.mql5.com/ru/articles/1165?
I did the installation in console R(install.packages("rattle"))
The R console only has "load workspace" and the article has "File/Workspace;"
Or is it not done on the R console there, but on another prog?
And everything is running.
library(rattle) rattle()Similar cuts (head-on), had both me and Doc(as I understand it), as told.
It was better when I tried to stick to this and so forth. -
https://en.wikipedia.org/wiki/Markov_decision_process
https://en.wikipedia.org/wiki/Partially_observable_Markov_decision_process
Spun it a long time ago, not long, just for interest...
there is also some partially observable MPPR, thanks, I will read
I have 3 RSI on the input here and all, I think that can be improved by other predictors a little will
> Is it possible to see in R the whole history, how a forecast of ARIMA came true, to look through the best periods for it?
Yes, for example, I can save bar history from mt5 to csv file, import it to R, and use a sliding window to train on some interval and test it on the next one, and it will shift the training window cyclically.
Thank you!
of those who have gone into "packages", no one has ever come back (it's kind of ambiguous).
until you have a robust idea you'll have nothing... all you've been doing all these years is nothing
because you've been running it with your supposedly good features, but it's not your features that are good
the regularities exist in the real world, not in the world of bags, so look for them, and if you find them, you can even code them yourself
or better yet, forget it.
Really not quite clear from the words above what I mean, so I'll probably start in order, and where expert help is required in R will be pulled up by those who will be interested in my deductions. In any case, the moment will come when it will be necessary to check my statements there and see if we are a team as a whole. So here we go. Regression.
Do not believe me, all my life I worked exclusively with the classification and somehow suddenly became bored. I had to redo everything 3 times during the night, because the error in the calculation of the deviation (htcgtrnelibrarius) entailed more and more, but already now it is clear that the sleepless night was not in vain. And now it has become kind of boring, but it would be interesting to try something new.
In the old version you essentially fed the MA from the cumulative delta, not the deviation. If the results were good, then the predictor is good as it is. Having MA cum. delta and cum. delta itself - NS itself can find analog of deviation, if it considers it profitable.
I think you are upset for nothing and want to switch to something new. Finish your development, the current results are very encouraging.
In the old version, you basically fed the MA from the accumulated delta, and not the deviation. If the results were good, then the predictor is good in this form. Having MA cum. delta and cum. delta itself - NS itself can find analog of deviation, if it considers it profitable.
I think you are upset for nothing and want to switch to something new. Finish your development, the current results are very encouraging.
That's the thing, the work is completed and suddenly became bored. Yesterday, not a single error and at least 3-4 more working days it will remain. The question is what to do during the TC on automatic, when the desire to explore and there is time to do it?
I updated the java and everything went... Yay... So where were we at???? Ah yes.... regression... So, let's define the problem statement and the choice of initial TC settings.
I will try to draw an analogy from the classification methods, but first let's define the conditions.
Regression implies forecasting of a future parameter value. Let's select change of price as a parameter. It is enough to have a forecast for 1 bar in advance, but it is boring. Let us set the problem in the following way:
We should make a prediction of the price change for 10 bars ahead. That is, the result of the model will be a value above or below zero (direction of change), as well as the degree of this change. That is how much it will be...
(I remind you that I only set the tone of the direction of work, if you do not agree with the condition proposed by me and you have it ALTERNATIVE, actively speak up. All discussed and corrected).
As you have already guessed, this is the target function for our model. Let us first define it and fix it. On the basis of what I have said above, we do the following.
Close[i]-Close[i+10] at first let's calculate the change of the current close relative to the close 10 bars ago.
Lead=Close[i-10]-Close[i] now shift our function 10 bars back. This operation is possible ONLY for target functions and cannot be used in real trading. I will show you what we have in the end on the chart.
The green is Chenge and the blue is Lead. The lead indicator looks into the future for 10 bars ahead. It looks like this:
We have a forecast window between the two lines. We can only read the change on the last bar of this window. That is the green line. But we need to know this value on the first bar of this window (just as an example), so we shift the chart back as a blue line, where on the first bar of this window we ask our network to give us the value that is yet to come. Then we can take advantage of it.
Questions Questions on changes in the choice of the target is?????
In support of this approach I will say that a lot of research was done long ago on the target for regression and came to the conclusion that the simpler the function the better. Any increase in the complexity of the target does not improve the model's performance at all, and significantly worsens it. We proceed from the statement "Everything is simple in genius". If the model works properly, this target will be enough. I await feedback and continue...
It is natural that the indicator tip from 1 to 10 bars will be missing in the target (blue line). It is up to the AI to bring it up to 0 bar. This will be the forecast for the future.