Discussion of article "Random Forests Predict Trends" - page 12
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
To: CyberCortex
Sorry SanSanych that I'm getting involved.
I couldn't pass by such high-mindedness and aggressive unprofessionalism.
Let's consider the author's statements from his description of the application programme one by one
Quote: "Advantages of CyberCortex in comparison with existing analogues:
The algorithm used. For many traders, technologies from the field of artificial intelligence begin and end with the hackneyed topic of classical neural networks. But the bousting algorithm implemented in CyberCortex is an order of magnitude more powerful and modern tool(?), because: "
Objection:
Where does this dismissively grandiose opinion of many traders come from?Generally this field is called Machine Learning, Artificial Intelligence is somewhat different. I am sure that many traders are familiar with many types of neural networks that you have not even heard of (deep networks, convolutional networks and other modern ones). The topic of neural networks is worn out for those who have limited themselves to studying MLP . I hasten to disappoint you, the topic of neural networks has now received a second breath after the development of deep learning methods and for those who know and can do it, it is far from being a closed topic. Besides, we know many other classification algorithms which are implemented in numerous R language packages.
The bousting algorithm in various modifications has been known for quite a long time (1996) and is in no way more powerful than the neural network algorithm. By the way, you should tell us what algorithm your programme uses. At a quick glance on packages R -
"ADA"- adaptive stochastic bousting. one of my favourites, three modifications of the original algorithm are implemented: Gentle-, Logit-, and Real AdaBoost
.
"adabag" - the package uses the multiclass algorithms Adaboost.M1(Freund and Schapire's), AdaBoost-SAMME(Zhu et al., 2009) , and Breiman's Bagging algorithm ;
"boost" - contains a set of boosting methods such as - 'BagBoost', 'LogitBoost', 'AdaBoost' and 'L2Boost' supplemented by pre-selection of features (predictors) using Wilcoxon test statistic ;
"pga" - contains an ensemble of algorithms combining AdaBoost and Random Forest with an evolutionary algorithm.
If you have an original world-unknown algorithm, please describe at least in general termsthe difference from the existing ones and any advantages over them, and a link to the description would be desirable.
As you can see the choice is quite wide.
For those who are not in the subject. "Busting" (amplification) - a method of resampling proposed originally tosolve problems in areas of science, where initially the number of examples (results of experiments) to use them in statistical calculations were few, to extract new ones is impossible or very expensive. The essence of the method is to extract from the initial small set a sample of smaller size formed randomly.Thisprocedure is repeatedmany times and as a result a sufficiently large total sample is obtained, which is then used in statistical calculations.
First of all. Forex market is not a case of "data deficit". On the contrary, it is a case of "data abundance".
Secondly. Longexperience ofresearch and application of bousting and crossvalidation in practice and in universities has shown that these methods are undesirable to use in the process of training, i.e. when selecting a model by internalerror. But it is very useful in the evaluation stage of the trained model, i.e., obtaining the external error. Therefore, for our case of "data abundance", splitting into train/valid/test during training is ideal. There are many other subtleties that should be taken into account in the process of data selection and pre-training and on which the result of modelling depends to a greatextent, but this is a separate big topic.
From my experience of long experiments with ADA I can say that sometimes excellent results, sometimes disgusting = very unstable.
Quote: "Thealgorithm does not require to reduce the dimensionality of the input data and to find out which features are informative and which are not. On the contrary, the more input data available, the higher the probability of correct classification. Uninformative features are automatically discarded."
Question - how are uninformative featuresmagically weeded out? Determination(evaluation) of predictors' importance is included in every decision tree algorithm, but there are quite a lot of criteria used to determine importance and not all of them give unambiguous results. How do you solve it in the programme?
Quote: "Thealgorithm does not require any preliminary data normalisation or any preliminary manipulations at all. All data are automatically reduced to unit variance and zero mathematical expectation."
The only algorithm that really doesn't require any preprocessing isRandomForest. All the others require various preprocessing. You obviously mean to say that your program does the standardisation process (x minus the mean divided by the standard deviation) for the user? I'm not sure this is a good solution, as there are at least a dozen differentpreprocessing methodsand you have to determine which one is best in each individual case.
Quote: " Thealgorithm, unlike neural networks, does not require tuning the training parameters in order to get the optimal classification result."
Not true. Any known bousting algorithm has at least two parameters that need to be tuned to get the best result.
Quote:"The algorithm is virtually immune to the tendency to overtraining."
False statement. Any machine learning algorithm is susceptible to this disease-overlearning. Reducing the risk of getting an overtrained model is achieved by regularisation. Nowadays various regularisation methods are known and successfully used, including in bousting algorithms.
Quote:"The algorithm allows to classify data into any number of classes".
This is not an achievement of this algorithm. Today almost all machine learning algorithms do this.
Quote:"The peculiarity of the implemented algorithm is that if there are patterns in the data - they will be found. On the contrary, if there are no patterns in the data, then the programme's predictions will be no better than a simple coin flip, which is logical"
I may surprise you, but all algorithms have this feature.
Bottom line: the incomprehensible algorithm that you offer for money can be implemented transparently, simply and in various variants in the R language and note that it is absolutely free, i.e. free of charge. This programme will be 10-15 lines long. This is the first. The second, undeniable advantage of the R language is that it is developed and supported by the world scientific community (practically all universities of the world) and now by the giants of the software business. Microsoft has launched a cloud service "Azure" with Microsoft Azure Machine Learning Studio" which allows you in the cloud to produce all stages of creation, training and evaluation of the model and subsequently get predictions by sending data to the service to your model. The service is based entirely on the R language and has a graphical interface that allows you to reduce the process to "drawing" a picture (I'm simplifying of course). If earlier the language was a de facto standard in research circles, now it has practically become such in the applied and industrial field.
One last thing. I am not against the fact that you monetise your labour embodied in the program. That's fine. I am against your attributing non-existent or exaggerated qualities to the product for advertising purposes. It looks unprofessional.
Be careful. As the classic said.
Good luck
Hello!
"Generally, this field is called Machine Learning, Artificial Intelligence is a little different." - Machine learning is a subsection of artificial intelligence. But I'm sure you just forgot about it. It happens...
I have heard of new types of neural networks:) And, if you noticed, the description of my programme says"classical neural networks", not all networks. By classical I mean MLP, its derivatives, Hopfield network, probabilistic network and others.
"Where does this dismissively grandiose opinion of many traders come from?" - This mega hyper-important opinion comes from the fact that the vast majority of traders who try to use weak artificial intelligence technologies use classical neural networks.
"By the way it would be necessary to voice what specific algorithm your programme runs on." - Adaptive bousting of single-level decision trees (stumps). It's written there. Since I implemented it from scratch myself, without seeing any role model in front of me, and there was no normal description on the Internet, I applied some of my heuristics when developing it.
"For those not in the know. "Boosting is a resampling method proposed by..." - I don't know where you copied this from, but I think what this definition describes is not boosting, but bootstrap:) Boosting is boosting simple classifiers. In bousting, you don't get one large sample N from a small sample m. In bousting, for each subsequent simple classifier, examples misclassified by the previous one are preferentially sampled. The number of examples is always equal to the original number of examples. After that, it's especially funny to read your serious:"For those who are out of the loop." xD
"First of all. The forex market is not a case of "data scarcity". On the contrary, it is a case of 'data abundance'." - Well, it depends on which side you look at it from. If you mean forex quotes and a bunch of "pseudo-mathematical" indicators that smell like old times and various derivatives of them, then yes, there is no shortage. But since you have gone into a different direction, accidentally confusing bousting with bootstrap, your maxim will remain unheeded, as I have no idea what kind of answer you are waiting for.
"Longexperience inresearching and applying bousting and crossvalidation in practice and at universities has shown..." - really? Boring...
"From my experience of long experiments with ADA I can say- sometimes great results, sometimes disgusting = very unstable. " " - Your experience is not an axiom for us, maybe you mixed up some buttons there or something, similar to the confusion just above. I don't know.
"How do you solve this in the software? " - every simple classifier focuses on the feature that produces the least classification error. If the feature is not informative, the error is high. That's the magic of it all. Let's challenge this logic with some more scientific nonsense about how universities..... No seriously, write something.
"I'm not sure it's a good solution" - don't you think you're being a bit too categorical? Seems to me the software developer is a bit more discerning about what needs to be added to make it work as intended.
"Not true. Any known bousting algorithm has at least two parameters that need to be tuned to get a better result." - what parameters do you have in R there that need tuning? I'm taking notes...
"Quote:"The algorithm is virtually immune to the tendency to overtraining. "False statement...."" - I like your statements so much :). Well, first of all, it says "practically". In Russian it means "almost". But if there is a stable pattern in the data, for example, as in the experiment with car licence plate recognition, then yes, the algorithm is not retrained. On the training sample the error goes to zero, and on the test sample the error remains constant, about 1%. The questions?
"This is not the achievement of this algorithm. Today, almost all machine learning algorithms do this." - you can tell me it's a disadvantage. It's an advantage compared to a binary classifier. Or don't you think so?
"I may surprise you, but all algorithms have this feature." - Read carefully, "if there are patterns in the data, they will be found." That's my emphasis. The main problem with machine learning algorithms is how many patterns will be found. My programme is fine with this. By your logic, it turns out that having some MLP, people shouldn't develop other algorithms, because a multilayer perseptron is a machine learning algorithm, and you say that"all algorithms have this feature.
"An obscure algorithm you offer for money" - I am not offering an algorithm, but a programme. I have nothing against the R language. But you see what the point is: not all people can and will understand the logic and syntax of the programming language, understand the settings of algorithms, convert some files and connect to the terminal through various crutches to use the result of calculations. If someone can do all that, I congratulate them. Some people just need it to work "out of the box". You'd think I'm running after you and forcing you to buy our software. Use R, I assure you it won't make me upset.
"I don't mind you monetising your hard work in software." - I'm sorry, I forgot to ask your opinion on this. But as long as you're okay with it, I'm fine with it.
"I object to you attributing non-existent or exaggerated qualities to a product for the purposes of advertising. " - and I object to you writing about something you don't know anything about.
Thank you. I'm done. As the classic said.
P.S.: I'm just responding to aggression. No offence. Peace.
P.P.S.: And yes, you shouldn't have started arguing with me:))
Я не понимаю содержание Скрежет обучение режим? В частности, о том, как использовать платформу использовать платформу MT4 MQL предварительной подгото вки?
Guys, hello! Your energy, in the right direction!
Congratulations, you have a strategic error!
I quote the first line of your article: "Initially, the purpose of building a trading system is to predict the behaviour of some market instrument, for example, a currency pair".
I did not even read the article further. It is better to study ALL forecasting methods!
I will tell you a secret: the purpose of building a trading system is to "Create a model of (your) behaviour".
It is not important what the market (leader) will do, it is important what you will do: AND, OR, NOT.
For example: AND - chasing the leader (repeating the leader's actions).
I tried to install rattle on linux (Kubuntu), it didn't work, there are a lot of dependencies for code compilation.
By analysing installation errors I got the following list of packages that need to be installed in the operating system itself before installing rattle in R:
console -
$ sudo apt-get install libxml2-dev unixodbc-dev libssl-dev libgtk2.0-dev
And then you can run the installation in R itself -
> install.packages("rattle", dependencies=TRUE)
And if you still lack some *.h files to install rattle, you can find the required package like this
$ sudo apt-get install apt-file
$ apt-file update
$ apt-file search /someheaderfile.h
Maybe this will help someone else :)
After updating the operating system rattle stopped working, calling rattle() gives an error
Error in method(obj, ...) : Invalid root element: 'requires'
The first way that almost helped was to run rattle with an additional parameter
It turned out that the version of rattle installed from cran is outdated, you need to reinstall rattle by specifying the developers' repository to get a new version.rattle(useGtkBuilder = TRUE)
the programme window opened, but the buttons did not work, the method did not help to the end.
install.packages("rattle", repos="https://rattle.togaware.com", type="source")
And after that everything worked fine.