How do you achieve a qualitative leap in market analysis? There is an option: - page 5

 
The idea and implementation of neural networks is simple. The idea itself is not obvious. The wheel is simple, but not obvious. Otherwise it would have been invented by all civilisations. Everything brilliant is simple only because the concept of genius can only be given by the majority: ordinary people. The veracity of a statement can only be proved or disproved. Neural networks are great at pattern recognition and many other tasks. I read somewhere that based on evaluation of huge amount of economic indexes, prices of major futures and currency pair quotes, some large company managed to increase its profit by 20% per year using the appropriate neuronet. That was about 15 years ago. There are many publicly available theories that have been and will continue to be applied to market valuation.
 
Reshetov, it is a thankless task to criticise systems. Your system at 'AI' is a good example of a simple neural network implementation. It is a pity that there are only a few dozens of deals in the example you provided. You can not say much about the profitability of the system by such a number. I think that the algorithm must stand the test of hundreds of deals at different quotes. Then we can already speak about its stability for the future, but only very accurately and only presumably. I have already submitted a report of one system in this thread, which shows the stable growth of profit during hundreds of deals in the tester. I cannot say anything about its stability yet.
 
getch:
Reshetov, it is a thankless task to criticise systems. Your system at the 'AI' address is a good example of a simple neural network implementation. It is a pity that there are only a few dozens of deals in the example you provided. You can not say much about the profitability of the system by such a number. I think that the algorithm must stand the test of hundreds of deals at different quotes. Then we can already speak about its stability for the future, but only very accurately and only presumably. I have already submitted a report of one system in this thread that shows the stable growth of profit in hundreds of trades in the tester. So far, there is nothing to talk about its compatibility.
I already answered this question in the thread "Using Artificial Intelligence in MTS". In brief, there are much more calls to the neuron than deals. That is because when a position is set by trend, it is not closed, but the stop loss is pulled up. If the neuronka says it is time to reverse, the next position will appear. Therefore, the number of times the reversal signal was received or the stop loss triggered is equal to the number of trades. No more and no less. MTS has no takeprofits and profit taking should be done only by neural network signals. The grid prefers not to move from side to side in long trends.
 
My system is also based solely on flips, but the flip signal comes more than 10 times more often. What prevents you from training the neural network for more frequent flips? Or does the learning process become much more complicated?
 
getch:
My system is also based solely on flips, but the flip signal comes more than 10 times more often. What prevents you from training the neural network for more frequent flips? Or does the learning process become much more complicated?
And where to get so many trends? The reversal should be at the end of the trend, not just anywhere.

I am not teaching the network, I am teaching the strategy tester. It selects the weighting coefficients in such a way that the balance is maximal at the end.

If you are not satisfied with the current situation, ask MT4 developers to add "Maximum number of deals" to the testing parameters. Then you and other pipsers will be happy!

Another option is to switch to smaller timeframes and optimize the MTS for them.
 
Well, the pipsqueak thing is a separate topic. It does not touch upon what I have shown. The main thing is to define for oneself what it is. The length of time frame does not say much. You can talk about the system of 5 transactions a year. And it's easy to show such a system by optimization to have 10 deals over the previous two years and all of them profitable. You can sort the results by any parameter in the optimization results and the most interesting results will be those where there are more deals. In this case we can already suppose some stability. Also, I am not criticizing your system in any way, I mean the general case.
 
getch:
Well, the pips are a separate topic. It does not touch upon what I have shown. And trends may be 10,000, the main thing is to define for oneself what it is. The length of time frame does not say much. You can talk about the system of 5 transactions a year. And it's easy to show such a system by optimization to have 10 deals over the previous two years and all of them profitable. You can sort the results by any parameter in the optimization results and the most interesting results will be those where there are more deals. And I'm not criticizing your system, I mean the general case.
I honestly do not give a shit whether someone criticizes the system or not. I post on the principle: "Maybe someone will find it useful, and maybe someone will tell me how to do it better". Everything else is unconstructive. And I'm not going to modify a working and tested code according to the whims and lusts of all kinds of fate dissatisfied forum users. One of them wants the number of trades 10 times greater, the other one is allergic to neural networks and the third one wants something else out of the blue. If you don't like it, you may modify it for your own whims.

For example, 44 trades will not be enough for you, so take the quotes history since the Bronze Age, optimize it and run the backtest. You'll get a lot of trades. Why the hell should I do that for you?
 
I support your principle of deciding what to do on your own. In the end, it's every man for himself. In a constructive conversation, good ideas can emerge and the unnecessary can be discarded. That conversation didn't work out.
 
getch:
I support your principle of deciding what to do on your own. In the end, it's every man for himself. In a constructive conversation, good ideas can emerge and the unnecessary can be discarded. Such a conversation did not work.
The person is a herd animal and not all decisions can make himself. Sometimes he has to co-ordinate them with others.

I do not need the support of principles. There is simply a considerable difference between criticism, whimsy and constructiveness.

For example, I've opened a thread, published a code and after some time, during which it's almost impossible to make any kind of code evaluation, not to mention optimization, a local forum user comes to me and starts fretting about shit, like all neurons are full of shit... ...and you can't get anything out of them but a tweak. And so on and so forth. It's a true criticism, because it is built on bare emotion, as they say, without looking.

After a day or so, another forum user wrote in the same thread but with gratitude to the author. He informs us that he has run forward tests and found his results reassuring. It's not constructivism but at least the person preferred to spend time on trying to understand the code. That's why it took him about twenty-four hours while he was optimizing it and testing it. And when the results became evident, he voiced his opinion.

When somebody starts to pester the author with various tricks like little trades and big profit. I have never tried to do that. I have no doubts that you are a crooked, slanted, handicapped, or a brainless moron. Can't you download quotes and make a lot of deals on tests to check the system for lousiness? Why are you trying to interfere with the author's whims?

And finally, constructivism. For example, someone tried MTS on a demo account and noticed that the system corrupts orders. They have found it and inform the author. Together they have analyzed the problem area. The author fixed it and uploaded the working version. This approach is really constructive because sometimes you cannot take into account all the details and an error in the code might occur in the most unexpected places. It is almost impossible to test the program in all the variations by yourself.
 
Reshetov писал (а):
... "Maybe someone will find it useful and can tell you how to do it better?
"Classics of the genre" suggest using "time lags" as neural network inputs, i.e. essentially recognizing a "time pattern" (what is now in ArtificialIntelligence.mq4). IMHO, sometimes it turns out to be interesting to recognize AND.... let's say "situational pattern", i.e. to input values of several "indicators" (in quotes!!!!) on the last bar (for example, Fourier spectrum or "what do you call it", again "Arbitrage" ;). I'm a 3rd grader in a parochial school myself, so I'm not trained in terms :) ). The rest is like the "classics" - committees, etc.
Let's hope, that getch will not declare arithmetic (as (open-high+Low)^volume does not give stat advantages) as pseudoscience, and we will discuss not a "tool" - neural networks, but applications - inputs, architectures, etc. Even if it's in a neighboring topic.
Regarding the "development environment" .... klot was also inspired by the idea of writing neural networks in mql(http://www.fxexpert.ru/forum/index.php?topic=656.0), but then, luckily :), switched to a more productive way of using ready-made "components".