Discussion of article "Neural Networks Cheap and Cheerful - Link NeuroPro with MetaTrader 5"
I would like to take this opportunity to draw attention to my article that describes random forests. The point is that the article uses the Rattle package, which has not only random forests, but also a number of models, including neural networks. And the package provides a possibility to compare different models with each other, which is its undoubted advantage in the light of this article.
I don't know neural networks, so I can't compare the networks in Rattle and in the article. But with the help of Ratte it will be possible to substantiate the choice of a particular model and, if it is a neural network, to switch to a specialised package.
the article is great, thank you .
but transforming formulas through notepad is beyond good and evil ))))
Is anyone else doing neural networks at this level?
With such a variety of advanced tools available.
I'm dumbfounded. Reminds me of the '90s.
Is anyone else doing neural networks at this level?
With such a variety of advanced tools available.
I'm dumbfounded. I'm reminded of the '90s.
Is anyone else doing neural networks at this level?
With such a variety of advanced tools available.
I'm dumbfounded. It brings back memories of the '90s.
Have the neuron and axon as such undergone any changes due to the development of "advanced tools"?
G has remained the same. Or do your nerve cells perceive stimulus signals differently?
I have a question, what if we apply this approach not to one symbol, but to three correlated symbols(EURUSD, USDJPY, EURJPY).
Data for all of them should be uploaded and processed simultaneously... I wonder what the results will be... I will definitely test it.
Meaning? What do you mean by "advanced advanced" means? How about a google cat classifier?
"Developed advanced" can be viewed from two perspectives:
1. development of the NS itself (I can't say anything).
2. development of other machine learning algorithms.
On the second question.
Taking my article. Rattle, which includes 6 qualitatively different algorithms. I take the file attached to the article. I remove variables zz35 and zz75. I fit 4 models: ada (gain model), random forest, support vector machine SVM and neural network from nnet package. Here is the result of the trend prediction error identified from ZZ.
ada = 18.69%
random forest = 16.77%
SVM= 16.92%
neural network = 24.37%
PS.
The caret framework for machine learning algorithms includes more than 140(!) different models.
"developed advanced" can be viewed from two sides:
1. development of NS proper (I can't say anything)
2. development of other machine learning algorithms.
On the second question.
Taking my article. Rattle, which includes 6 qualitatively different algorithms. I take the file attached to the article. I remove variables zz35 and zz75. I fit 4 models: ada (gain model), random forest, support vector machine SVM and neural network from nnet package. Here is the result of the trend prediction error identified from ZZ.
ada = 18.69%
random forest = 16.77%
SVM= 16.92%
neural network = 24.37%
PS.
The caret shell for machine learning algorithms includes over 140(!) different models.
=========================================================================
Point by point answer
1.Neural networks of the second generation, reached their limit of capabilities about ten years ago, little by little have left the scene. The third generation of neural networks, the so-called "deep neural networks", which have appeared and become widespread in many practical applications, show very good results and lack the main disadvantage of "shallow" neural networks. You can dig in this direction.
2. any variant of trees or forests gives better results than any neural network (or their ensemble).
3. Good results are obtained using hybrid ensembles (bagging). This is when different models work simultaneously in one harness.
4. Regarding the article you blogged about, where 140 classification models are compared. I read a review of the developer of the caret package about this article. If it is interesting I will find the link. According to his experience the best results come from bousting and bagging. From my experience the best models are "ada" from the package of the same name and RFnear from the package "CORELearn". The latter is very fast by the way. And absolutely did not show itself SVM, well, except that a very long training.
Everything depends on the choice, preparation of input data and their corresponding output data. Here is the main field for research.
I did comparative results of neural network and RF a few years ago, posted on the forum. RF is unequivocally the first place. Besides, now the RF direction itself has expanded and branched out, there is a lot to choose from. I do not see the need to do it now. It should be said that there are applications in which neural networks show decent results, for example, in regression. But I only deal with classification, and neural networks are not strong in this area.
Maybe my article on this subject will be published at last, and we will discuss it there.
Good luck
SanSanych
Here is a link to the article I was talking about. http://appliedpredictivemodeling.com/blog/2014/11/11/some-thoughts-on-do-we-need-hundreds-of-classifiers-to-solve-real-world-classification-problems
Also in the article there is a link to an earlier article by David Hand on an issue you and I have discussed before - poor results after training on real data. Very interesting thoughts. Maybe you could do a shortened translation?
Going through the archives and found another article on the topic of comparing different machine learning algorithms.
http://www.cs.cornell.edu/~caruana/ctp/ct.papers/caruana.icml06.pdf
Good luck
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
New article Neural Networks Cheap and Cheerful - Link NeuroPro with MetaTrader 5 has been published:
If specific neural network programs for trading seem expensive and complex or, on the contrary, too simple, try NeuroPro. It is free and contains the optimal set of functionalities for amateurs. This article will tell you how to use it in conjunction with MetaTrader 5.
The NeuroPro program was written in one of Russian institutes back in 1998 and is still relevant today. It efficiently runs on Windows XP, Vista and Windows 7. I cannot tell how it works on later versions of Windows as I have not tested it.
Version 0.25 is free and can be found on many websites on the Internet. NeuroPro can create multilayer neural networks with the sigmoid activation function. If you have just started learning neural networks, you do not need more features at this stage. It should be kept in mind that the interface of NeuroPro is in Russian and has not been translated into any other languages.
Neural network can be trained on one data array and then tested on another one. It is an essential feature for traders as it allows to understand quickly if the selected network structure is prone to overfitting and if it can consistently trade outside historical data, i.e. on a real account.
Those who like to dig deeper have an opportunity to see neural network weights as well as which network inputs influence the result of network operation most of all. Beginners do not need that and they may skip this part of the program. This information is useful for experienced traders looking for the Grail because it lets them assume what pattern was identified by the neural network and see where they can continue their search.
Other than that, there are no significant features in NeuroPro except various settings and useful utilities like minimizer of the network structure. These menu sections are not compulsory to use so novices do not have to complicate things and use only the default settings.
From a trader's point of view, NeuroPro has only one disadvantage - absence of integration with MetaTrader 5. Actually, this article is mostly dedicated to loading market and indicator data from MetaTrader 5 to NeuroPro and then turn the received neural network into an Expert in MQL5.
Advancing the topic, I can say that the neural network that we are going to create with NeuroPro will be converted with all the neuron weights into an MQL5 script (unlike the system of include DLL like in any other neural network program). It will ensure fast work and minimum use of computer resources. That is a clear advantage of using NeuroPro. It can be used for creating any trading strategies, even scalping ones with their requirement for the Expert to make decisions nearly immediately.
Trading Strategy
In this article we are not going to consider scalping because the process of creating, training and testing scalping Experts has a lot of peculiarities and goes beyond this article.
For educational purposes we shall create a simple Expert for the H1 timeframe and popular currency pair EURUSD. So, let our Expert analyze last 24 bars i.e. market behavior in the last day, forecast the direction of the price movement in the following hour and then trade based on that information.
Fig. 22. Equity chart after the Expert Advisor has been tested in MetaTester
Author: Andrew