Machine learning in trading: theory, models, practice and algo-trading - page 515
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
Prices are not fed into the model without any conversion.
The scaffold for extrapolation takes the nearest known value. Neuronics or ruler in extrapolation will calculate something according to internal formulas. But in fact, all these models in this situation will merge, so there is no difference.Yes, they do without any conversion, but that's not the point.
The difference is huge in the technical application of these or those models. What does it have to do with the drain, again, that's not the point.It is not convenient to do everything in MT. The best option in my opinion is Python for learning and experiments and DLL with a trained model load for MT. Python has such a cool thing for experiments Jupiter notebook. In addition to interactive code execution, you can take notes in it, which is handy for writing down thoughts on a topic. So you have to choose a library that can work in Python and C++. Python is easy to learn.
It is not convenient to do everything in MT. The best option in my opinion is Python for learning and experimenting and DLL with a trained model load for MT. Python has such a cool thing for experiments Jupiter notebook. In addition to interactive code execution, you can take notes in it, which is handy for writing down thoughts on a topic. So you have to choose a library that can work in Python and C++. Python is easy to learn.
This all takes away from the point, you need to do everything in the same place where you trade, without unnecessary hassle... saving time and nerves. The MT already has all the basic elements of statanalysis, the only thing missing is some complicated models that are connected via dlls, as you said.
But python is cool undoubtedly, R twisted - annoying its sluggishness. As I understand it, machine learning pros sit exactly on python, while R is so dabbling and statanalysis and students to teach. But then again it's all leftovers, when you can connect it directly.
R is, from a programmer's point of view, a strange language unlike anything else. Yes, for machine learning, Python is the standard.
R is, from a programmer's point of view, a strange language unlike anything else. Yes, Python is the standard for machine learning.
You people are weird. One "R from a programmer's point of view is a strange and unlike anything else language". What programmers are you talking about? Another- "R is twisted - it's annoying slow" - maybe you were twisting in the wrong place or the wrong thing?
You probably do not know, but all the modules of Python are available in R , and hence in MT. Moreover, all the latest developments giants TensorFlow (Googl), CNTK (Microsoft) and others instantly offered an API in R and hence in MT. I stressed twice - today it is possible to use all the wealth of developments in machine learning in the terminal via R. It takes an understanding of what you want to do, knowledge and skill. You want to want to do, just do it.
The thread has turned into useless verbiage. Unfortunately.
Good luck
You people are strange. One "R from the programmer's point of view is a strange and unlike anything else language". What programmers are you talking about? Another- "R is turned around - it's annoying because it is slow" - maybe you were turning it in the wrong place or the wrong thing?
This thread has turned into a useless verbiage. Sadly.
Good luck
Sluggish R itself, I'm not talking about packages. Runs slower than python and slower than MT5. Or it's slow shells like RStudio, it doesn't even move windows smoothly, which says it's even slow on its own. I'm not even talking about VS 2017 in conjunction with Ropen, it constantly hangs, package incompatibility, etc. I'm afraid to imagine what will happen if you also use a package to work with python. Many packages for R are written by God knows who and may contain a lot of bugs, there are no unified standards.
There are only few good neural network packages, which you've described in your last article, and which work fine without R. All the rest there about preprocessing and stuff for forex is more mind games than effective use of the advantages of the language. Well, yes, the pictures can be evaluated by eye, while in most situations numerical evaluations for all models are sufficient. IMHO :)
Brakes itself R, I'm not talking about packages. Works slower than python and slower than MT5.
Not slower than MT5... I gave a comparison earlier:
ALGLIB is a terrible brake on learning.
I was waiting for 240-50-1 net for ALGLIB, I didn't wait for 2 days and switched it off.
I trained 70-5-1 network in half an hour. I trained nnet from R in less than a minute with the same data.
Not slower than MT5... I gave a comparison earlier:
You gave a comparison with another NS - it's no longer R but a package, which is written on the pluses and of course it's fast. There is no multithreading in free version in alglib, + for large network you need LBFGS optimizer, maybe you used it. Forests are counted in the alglib are pleasantly fast, for example, I like it... and the quality of models is never worse than in MLP. Boosting is certainly inferior, but not too critical, as far as I understood from the articles on the Internet. I.e. in essence there is 1 universal model is the forest of trees, also fast. All the rest, that some other neuron can do something better, it has not been proved in practice yet (in relation to forex).
The multithreading is not in R, it's in neural network packs, connect to mt5 as well and it will be a multithreading
You gave a comparison with another NS - it is no longer R but a package, which is written on the pluses and of course it is fast. In the free version there is no multithreading in alglib, + for large network you need LBFGS optimizer, maybe you used it. The forests in alglib are nicely fast, for example, I like them... and the quality of models is never worse than in MLP
It seems to be the same MLP, so network is the same in structure and quantity of data. That package didn't use multithreading, it read everything with one thread (I checked it with task manager).
I compared the speed of LBFGS (about 40 min.) and LM (27 min.). According to the descriptions LBFGS has to be faster, but in practice it is vice versa in ALGLIB.
Checked and scaffolding - much faster than NS (4 min.), and the result is about the same. And what's interesting linear regression counts even faster, with the same results.
As someone wrote here - it's all about features.
Seems to be the same MLP, so network is the same in structure and amount of data. That package didn't use multithreading, it read everything with one thread (I checked it with task manager).
I compared the speed of LBFGS (about 40 min.) and LM (27 min.). According to the descriptions LBFGS should be faster, but in practice it is vice versa in ALGLIB.
There, as far as I understand, you can set 1-2 epochs, because it almost always converges the first time... maybe this was an omission? although I have not used it for a long time, I may be confused