Machine learning in trading: theory, models, practice and algo-trading - page 472
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
In other words, not everything in MT5 has been tested yet :)
I do not understand it. ( And why even try what is already in MT?
If there is a concept of TC - we pick up a toolkit. There is in MT and it interacts well with the rest of the system - we use it. It interacts badly - use the external software.
It's not good to send information back and forth just to use the existing MT functionality.
I do not understand it. ( And why even try what is already in MT?
If there is a concept of TC - we pick up a toolkit. There is in MT and it interacts well with the rest of the system - we use it. It interacts badly - use the external software.
It's not good to send information back and forth just to use existing MT functionality.
There's never a ready concept from the beginning, everything is by instinct, the more toolkit - the better.
I start with a concept, work it out on some software (Excel, SciLab, MatLab, R - sometimes all together). There I test it with my tester (I haven't noticed any significant difference with real software). Then I transfer a part of it to the terminal, and the rest to C++/C# with external libraries.
My first systems (since 2008) were created with VBA-Excel and they cooperated with terminal very well. No problems with performance, even with VBA slowness. In general, 50 ms latency, even for 1 TF, is not time.
By the way, remember in your article on sequences you suggested counting several signals in a row, flip there... signal superposition
I have come up with an interesting solution to implement something like this through fuzzy logic and build it into the learning process... I'll post something later :)
Yes yes that's exactly from that article..... Maybe it was not clearly explained there, but no one cancelled the fact of separation.
It is necessary to see clearly how the division into good and bad is going, if the model is steadily losing or pouring it is already an indicator, the main thing is not the correctness, but the stability of the division.....
SanSanych.
Check out the interesting discussion about "overfittinga".
Good luck
SanSanych.
Check out the interesting discussion about "overfittinga".
Good luck
Everything I write here and use in my practice is some kind of trick to somehow reduce the impact of the basic problem of cognition, which here is called overfitting, overtraining.
The problem of model overfitting is NOT solved in principle, it is not solved theoretically, and it has been known for a very long time. A model is always some approximation, some coarsening of reality, of the external world. Therefore, a model ALWAYS reflects an object with some error. This is the strength and weakness of any model. It follows that it is impossible to create a model without a thorough study of the object being modeled.
If we take a quote.
WHAT are we modeling?
If a random forest, we are automating the search for patterns.
If GARCH, we model statistical characteristics of a quote.
There will ALWAYS be an error. And I was suggesting here some sort of heuristic that this error should NOT change on successive chunks of the time series. If the model literally gives a different error in the next chunk, then it is over-trained (over-fitted) and cannot be used in principle.
For random forests, I have found that if you remove the noisy predictors as I understand them, the trained model gives the same error much longer than with the noisy predictors. But I'm not aware of solutions that would create an all-time model, and I don't need to. I'm fine with retraining the model on the output of the latest data. But training has to start with checking the noise predictors - their set changes over time. That is, not only the model itself changes, but also the input data set. I was doing window runs along the time series, so out of my 25 predictors there is a constant sample of 10-15 predictors and the composition of this group changes, i.e. something changes into noise, and then back again. The error is very stable and ranges from 23% to 27% = always less than 30%. If the error is less than 20% on my predictors, I do not use this model and wait.
PS.
From this I see the reason why I am not interested in other models.
Everything I write here and use in my practice is a kind of trick to somehow reduce the impact of the basic problem of cognition, which here is called overfitting, overlearning.
The position is clear.
Good luck.
http://journal.frontiersin.org/article/10.3389/fnsys.2016.00095/full
http://journal.frontiersin.org/article/10.3389/fnsys.2016.00095/full