Discussion of high-frequency trading on MT5 - page 73
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
Alex_Bondar:
A pattern can be recognised from incomplete data, just as a person can be recognised from clothes alone, face alone, cologne, etc., and the missing data can be extrapolated. That is the point of patterns, for market predicates. But we can safely assert a number of funny things about them, the first is that these patterns are very simple, complex patterns do not work, and the simple ones are like momentum, acceleration, past price levels consolidation, etc. in simple combinations, that is, standard TA, this is a subtype of pattern recognition. The market does not "remember" ornate patterns. It turns out that the main point is the primary preprocessing of time series, compressing the information into an optimal distribution of "states" from which the patterns are built and it's not a stupid normalized price and not MACs... This is the second point, preprocessing of BP before its analysis by a neuron is 99% of all work, in fact the neuron is no longer needed if the data is processed correctly, the patterns become obvious to the human brain, and they are simply programmed, and if the normalized price is fed to the neuron, or to a wopper, it is useless, those who understand how neurons work do not expect miracles from them. You can play this game forever, the probability of real success is vanishingly small.
You average a number of such patterns for a given FI, timeframe, time of day season, etc., and get a reference one, with which you can compare with, for example, n-dimensional distance or a scalar product. But it's a dead-end street of eternal experiments. Simple normalization is definitely not the way to go.
Could you clarify for a beginner how to determine the difference between complex patterns and simple ones? I am somewhat familiar with the theory and practice of pattern and speech recognition, I would like to clarify if I understand correctly what we are talking about here, with respect to market patterns.
To begin with, what quantitative criteria do we analyze in the context of information complexity? It can be the length of a vector of price series quantized by time, in the pure form, or filtered in many different ways. It makes sense that the filtering would minimise the information weight. And then compared, as you said, with some average vector, which can also be obtained in many different ways. I am interested in the approximate information weight of such threshold vector(s).
For example, in speech recognition, structural complexity may be quite high, so we want to correlate them and draw parallels. Speech recognition, in general, is already at an acceptable level and why not use ready-made tools. In my opinion there are many parallels.
I apologize if this is a bit of an off-topic.
For example in speech recognition, the structural complexity can be quite high, you want to correlate them and draw parallels. Since speech recognition, in general, is already at an acceptable level and why not take advantage of ready-made developments. In my opinion there are a lot of parallels.
Can you give some useful links to get acquainted with the current state of the art in this field? It would be very interesting to read good (qualified) articles, preferably from the developers of such systems.
Could you please clarify for a beginner how to determine the difference between complex patterns and simple patterns? I am a bit familiar with the theory and practice of pattern and speech recognition, I would like to clarify if I understand correctly what I am talking about here in terms of market patterns.
..........It's subjective, I haven't figured out where to draw the line myself. But the fewer elements in the pattern the better.
A pattern can be recognised from incomplete data, just as a person can be recognised from clothes alone, face alone, cologne, etc., and the missing data can be extrapolated. That is the point of patterns, for market predicates. But we can safely assert a number of funny things about them, the first is that these patterns are very simple, complex patterns do not work, and the simple ones are like momentum, acceleration, past price levels consolidation, etc. in simple combinations, that is, standard TA, this is a subtype of pattern recognition. The market does not "remember" ornate patterns. It turns out that the main point is the primary preprocessing of time series, compressing the information into an optimal distribution of "states" from which the patterns are built and it's not a dumb normalized price and not MA... This is the second point, preprocessing of BP before its analysis by a neuron is 99% of all work, in fact the neuron is no longer needed if the data is processed correctly, the patterns become obvious to the human brain, and they are simply programmed, and if the normalized price is fed to the neuron, or to a wopper, it is useless, those who understand how neurons work do not expect miracles from them. You can play this game forever, the probability of real success is vanishingly small.
yes Alex, I asked for any pattern - I guess it is also a tick pattern - just a pattern (if the neural network detects it, the eye should recognize it anyway)
About preprocessing, about neural network unnecessity and eye recognition, you both might be right, but it is desirable to speed up your frequency, and process at least several hundreds of thousands of patterns per second, then your tandem could well replace the Expert Advisor. And to reproduce the same engine as I have, it's enough to find a couple more enthusiasts - to replace the second EA, and as an arbitrage script the so called EA leader, invite respected newdigital, just he wanted the original algorithm - I'll tell him about it ...))
And on the merits, IMHO I think not reasonable any, calculated preprocessing patterns because it can introduce a lag time series or e.g. distortion of the original information and in general violates the methodological unity within the neuro data processing.
Suppose you do 99% of calculations, in front of a neural network, and what do you charge after that, because 99% of responsibility will automatically fall on those calculations or/and the calculator?
If we are talking about improving the efficiency by introducing additional measurements of patterns, then you don't need to go beyond the neuromodel, for example the engine that I use, has the ability to connect not one but many neural networks trained on patterns of different, correlated symbols, different timeframes, or patterns clustered by trading sessions, days, etc... And in EA external settings you can activate or deactivate signals from these neural networks and / or set the rule of their summation.
As for preprocessing, neural network unnecessary and eye recognition, you both might be right, only it is desirable to speed up your frequency, and process at least several hundred thousand patterns per second, then your tandem could well replace the EA. And to reproduce the same engine as I have, it's enough to find a couple more enthusiasts - to replace the second EA, and as an arbitrage script the so called EA leader, invite respected newdigital, just he wanted the original algorithm - I'll tell him about it ...))
And on the merits, IMHO I think not reasonable any, calculated preprocessing patterns because it can introduce a lag time series or e.g. distortion of the original information and in general violates the methodological unity within the neuro data processing.
Suppose you do 99% of calculations, in front of a neural network, and what do you charge after that, because 99% of responsibility will automatically fall on those calculations or/and the calculator?
If we are talking about improving the efficiency by introducing additional measurements of patterns, then you don't need to go beyond the neuromodel, for example the engine that I use, has the ability to connect not one but many neural networks trained on patterns of different, correlated symbols, different timeframes, or patterns clustered by trading sessions, days, etc... And in EA external settings you can activate or deactivate signals from these neural networks and / or set the rule of their summation.
invite the respected newdigital, he wanted the original algorithm - I'll tell him about it...))
I didn't refuse to tell about your algorithm either, most people here, except for Negoich, would like you to tell)))) Who knows, maybe someone will leave some useful comments, for example I' m not saying that neurons are not needed at all, only that preprocessing of input vectors should be done more artistically, to clean them, so to say, I thought, based on my own experience of experimenting with neurons.
What if he fails at the end of it all?
Well, if he fails, we'll put him on the spot and send him for advanced training, retrain him and go back to battle... some robots should be dealt with more strictly, because they have already revolted- they rebel. And why, perhaps because they have stuffed them with several strategies, genetic algorithms and a virtual tester... Under these fat monsters and the terminal can slow down and they can sick, all sorts of bad diseases ... (((
By the way, the developers of the engine that I use, promised in the near future to add support for MT5 terminals and the possibility of auto-start in testing mode that will allow a much easier and more flexible to implement the idea of the selection of effective strategies and dynamic optimization described in the above article, namely:
1. When choosing the best strategy, you will be able to vary not with the set of compiled together algorithms, but with the whole list of existing strategies in the terminal which are contained in individual Expert Advisors.
The optimization of strategy-advisor could be done asynchronously in a separate MT4 or MT5 terminal, which will reduce the load on the current trading terminal, and allow to use more options for optimization and more efficient genetic algorithm, as a part of MT standard testers.
3. If your computer has enough resources, you can use optimization of several advisors in parallel running terminals, and if there are not enough resources, you can use cloud services of MT5.
I didn't refuse to tell about your algorithm either, most people here, except for Negoich, would like you to tell)))) Who knows, maybe someone will leave useful comments, for example I'm not saying that neurons are not needed at all, only that pre-processing of input vectors should be done more artistically, to clean them so to speak, as it seemed to me, from my own experience of experimenting with neurons.
Sorry, but I'm not a pro in AI field, that's why I can't describe theoretical background in details, moreover I can't describe algorithms in details - maybe if I'm tortured...((( But it, neural learning on patterns and prediction, at least in that engine, that I have, works and it's enough for me. And the arbitrage algorithm, which I promised to tell you about, when implemented with unrealistic abstraction, is reduced to the selection of the best price and combination of ready-made signals from neural counters and in isolation from them, has no meaning.
Well, if it fails, we put on the mark and sent to improve skills - retrain and go back to battle, with robots must be strictly ... and then, some have a revolt- in revolt. And why, perhaps because they have stuffed them with several strategies, genetic algorithms and a virtual tester... Under these fat monsters and the terminal can slow down and they can sick, all sorts of bad diseases ... (((
By the way, the developers of the engine that I use, promised in the near future to add support for MT5 terminals and the possibility of auto-start in testing mode that will allow a much easier and more flexible to implement the idea of the selection of effective strategies and dynamic optimization described in the above article, namely:
1. When choosing the best strategy, you will be able to vary not with the set of compiled together algorithms, but with the whole list of existing strategies in the terminal which are contained in individual Expert Advisors.
The optimization of strategy-advisor could be done asynchronously in a separate MT4 or MT5 terminal, which will reduce the load on the current trading terminal, and allow to use more options for optimization and more efficient genetic algorithm, as a part of MT standard testers.
3. If your computer has enough resources, you can use optimization of several advisors in parallel running terminals, and if there are not enough resources, you can use cloud services of MT5.
That's all right. And the trading itself? And I remember one case - on a foreign forum someone tried to sell an EA for a lot of money, sold it three times and the buyers backtesting on the tester did not coincide with the trading for the same period - in the tester everything was fine, but in the trading account - on the contrary - there were indicators coded on a closed bar, some on an open bar, also there was price on high/low, and all at once. But since there was no source code and coder wasn't known at all (i.e. - it's hard to believe his word, because nobody knew him) ... the customers packed their bags, and bought plane tickets, and went to the vendor...
I'm just saying that some samples or source versions can be posted here for people to have a substantive conversation. It won't help me (I'm not a coder), but it's nice for people and the discussion will be more fun. Although ... it's just my opinion.