Machine learning in trading: theory, models, practice and algo-trading - page 381
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
MO is always a ready-made meaningful model. Sometimes it is so meaningful, that you don't even understand how it works. Here is an article about gradient bouncing for examplehttps://habrahabr.ru/company/ods/blog/327250/ There is an article, there are descriptions and formulas, but my desire to transfer this into mql I haven't been able to realize, it's too complicated.
It's a little different, not in the sense, but in the narrow specialization.
Arima and Garch work directly with prices without indicators and TA. They have a built-in algorithm for turning a price series into a stationary vector, and there are even some subtleties such as correction of predictions depending on previous errors (MA component). But they are useless for other (not price) data, these models can't classify pictures.
If we pass a time series of prices to the neural network for training, it will not search for autocorrelations, seasonal and trend components of the price, as the neural network is not able to do. It will simply remember what it was given, and for new data during a test or real trading, it will "remember" similar price vectors from the past, and trade as it was before, and this in forex means a loss.
Neuronka needs help in predicting the price - first of all it should find the indicators that, like Arima, will be able to determine autocorrelation, trend, seasonality, and the values of these indicators to send to the neuronka. Then it will have at least a small chance to be comparable to arima and garch.
Another important thing is that arima makes predictions based on time. This model remembers exactly in what order the prices arrived and uses in its forecasting a sort of sliding window, taking several latest prices and making forecasts using them. In contrast to neuronics, which works with the whole training table at once with no idea in what order the prices came.
+100
I thought we're dealing with some useful things here but not with some understandable truths.) Let him make a prediction with time in mind, show me some working examples, since so many people are working on it! here we go :) It's a simple model, like garbage, which is studied in the 2nd year of technical high school... Unfortunately, I was studying humanities... and there's nothing like that there
Let's not generalize your knowledge to the rest of the world.
Judging by the publications, this is the mainstream of trading. And harch is used by high-skilled scientists who studied harch in their childhood and have been doing it all their life - it's not even a simple thing.
Let's not generalize your knowledge to the rest of the world.
Judging by the publications, this is the mainstream of trading. And harch is used by high-skilled scientists, who studied harch in their childhood and have been doing it all their lives.
My knowledge in this area is not much, it's an elementary level, I just orient by the indirect signs always, I know many traders also (far from stupid).
And when the model is more complicated the development in one is not possible of course, i.e. it's not a question of private trading any more
And with a great complication of the model the development in one is not possible of course, i.e. we are not talking about private trading any more
This is a very good point. Sometimes there are some projects, but after figuring it out - there would be 5-6 programmers and a year of work + funding. And there is no project)).
A private person needs something simpler, even less effective. I'm now thinking about combining classical strategies (I mean, on the logic) and MO in one package. Classics is already solving a lot of things on its own, and if it is supplemented and additional functionality is assigned to an MO, then both of them can be simplified as a result. As a matter of fact, in technical applications of MO this is how it is done.
But I do not know yet how to unite it and distribute tasks.
That's a very good point. Sometimes there are some projects, but after estimation - here are 5-6 programmers and a year of work + financing. And there is no project.)))
The private sector needs something simpler, even if it is less efficient. I am now thinking about combining classical strategies (in the sense of logic) and MO in one bottle. Classics already solves a lot of things by itself, and if it is supplemented and additional functionality is assigned to MO, then both can be simplified in the end. Actually, in technical applications of MO this is what is done.
But how to combine it and distribute tasks, I don't know yet.
My idea is as follows: I have read all the articles on this forum :))) I have gathered something I think necessary, all kinds of transformations into more or less stationary form, + my experience and skills, now I shove it all into classifiers and NS and see what will happen, + selection of parameters through genetic algorithm. In short, my first impression is that it is very difficult to achieve stability over a long interval while maintaining profitability, even if you do not train the NS, you will always have to retrain it. It is possible to make money in the short term, but it is not clear how it will look like in the long term....
The focus now is in the direction of adaptive indicators, the values of which can be stuffed into the ns... that would not retrain the ns, and the indicators themselves would be rebuilt depending on volatility... but the task is not trivial, the same Garch would probably help, but I don't know yet.
What machine learning algorithm should be used?
What machine learning algorithm should I use?
Well, yes, for classifikashon the method of reference vectors and Bayesian and Random Forest is the best... quick and easy and without retraining. You can also try convolution.
And Microsoft has its own Random Forest, I forget the name... They say it's cool. The Jungle of Solutions or something like that
Stock prediction:
https://gallery.cortanaintelligence.com/browse?s=stock
Arima:
https://gallery.cortanaintelligence.com/CustomModule/Train-Score-Timeseries-1
https://gallery.cortanaintelligence.com/Experiment/Time-Series-Forecasting-8
Diplerning:
https://gallery.cortanaintelligence.com/Experiment/Neural-Network-Convolution-and-pooling-deep-net-2
Well yes, for classifikashon the method of reference vectors and Bayesian and Random Forest is the best... quick and easy and without retraining. You can also convolutionalize.
And Microsoft has its own Random Forest, I forget the name... They say it's cool. The Jungle of Solutions or something like that
https://gallery.cortanaintelligence.com/browse?s=stock
Arima:
https://gallery.cortanaintelligence.com/CustomModule/Train-Score-Timeseries-1
https://gallery.cortanaintelligence.com/Experiment/Time-Series-Forecasting-8
It's all cool, scientific, beyond the comprehension of normal people, but has nothing to do with the markets, where psychology and puppetry rule.
Gentlemen, stop fooling yourself and inventing highly intellectual toys, well at least if you play in the scientists at least realize that this is just a game, like shooting or racing, and not comprehending the Market, which is beyond mathematics and rigorous formulas.
Why do you need all this? I've been told many times that prediction is not necessary, you can do it without 50/50, the main thing is money management and nerves of steel, no one knows where the price will go in the next moment, except for the insiders, and they do not know 100%, they take risks too, but they have steel balls and deep purses, they know how to take risks, and do not even know about the neural networks and the random forest, or they themselves promulgate such false strategies to distract the "meat.
This is all cool, scientific, beyond the comprehension of normal people, but has nothing to do with the markets, where psychology and puppets rule.
There is a category of traders called "clickers", i.e. the main market meat for brokerage companies. You seem to belong to this category with your primitive categories like money management and nerves of steel.
A great example is Timofey Martynov, who is still trying to curb his psychology and manual dexterity :) The algotraders have no problems with it and the risks are minimal, because they have long gone through all this crap with psychology and all sorts of dolls and other nonsense.