Machine learning in trading: theory, models, practice and algo-trading - page 63

New comment
 
Alexey Burnakov:

I didn't even say anything about overbidding, don't twist my words.

I see, it means the time of transaction from signal to signal. Well, the rest is clear.

The same method, I've never tried it, so I'll shut up. The question is whether the signal system itself is adequate. I.e. why in exactly those places where there is a signal we start watching and not in other places.

Any TS, no matter how complex it is, leads to a fait accompli. The event has happened and it cannot be cancelled. I mean the formed bars. I'm not counting the zero bar. So, when the event occurred and there is a signal, this is the moment we save indicator values. The rest of the market does not interest us. Only at the moment when the signal appears. But we can save the indicators at this moment, the moment when the signal appears. We can save indicators for any period, for example, it is floating because TS gives a signal with floating bars. That is why I take the moment for 5 bars in one signal, and for 8 bars in another, and for 3 bars in the third one, etc. A certain point of adaptation, if you look at my screen you will see the green dots, so their number (and it is always different) is the period of calculation for the indicators. But in general, we take only the moments of the signal appearance and therefore the Predictor is trained only on the signals and we are not interested in the rest of the market. Although we can build a model between the signals, but this is so...... to the word.....
 
Now here comes the fun part. Just now there was a signal... And I must say that all day binary model and trinary model corresponded to each other, but now the signal of binary model "False Sell" and trinary model "Dash" to what would it be???? So I keep buying... Sitting in BU.... :-)
 
Mihail Marchukajtes:
Any TS, no matter how complicated it is, leads to a fait accompli. The event occurred and it cannot be cancelled. I mean the formed bars. I'm not counting the zero bar. So, when the event occurred and there is a signal, this is the moment we save indicator values. The rest of the market does not interest us. Only at the moment when the signal appears. But we can save the indicators at this moment, the moment when the signal appears. We can save indicators for any period, for example, it is floating because TS gives a signal with floating bars. That is why I take the moment for 5 bars in one signal, and for 8 bars in the other, and for 3 bars in the third one, etc. A certain point of adaptation, if you look at my screen you will see the green dots, so their number (and it is always different) is the period of calculation for the indicators. But in general, we take only the moments of the signal appearance and therefore the Predictor is trained only on the signals and we are not interested in the rest of the market. Although we can build a model between the signals, but this is so...... to the word.....

" That's all understood.

How is Demark's choice better than, for example, such points, after which, for example, within a day movement of more than 100 points? In other words, they are selected according to the future movement, but not according to some signal system?

 

Today is a perfectly acceptable day, so I can't complain.... And you say Predictor is bullshit. What am I talking about.... What a predictor - CLASSIFICAIOR :-)

 
Alexey Burnakov:

" That's all understood.

How is Demark's choice better than, for example, such points, after which, for example, within a day movement of more than 100 points? That is chosen based on the future movement and not on some signal system?

If you want to choose such points, you have to know the future, and it is not known.
 
Mihail Marchukajtes:
To choose such points, you need to know the future, and it is not known, so I do not quite imagine how you can choose such points??? in what way???
How. We make a training set, in which after each point on the chart we gather the information that will be in several hours. (max, min, marginal difference, etc.). We sift out the points with less than n points modulo, as you have. and make a classifier that says 1 - buy, -1 sell, 0 - do nothing. Just do everything without signal system.
 
For example, you can make a cross of the wands and specify in the target function those signals, after which the rate went up by 100 points, then train the network and it will find only those signals, after which the price flew up by 100 points. But do not forget that you still need to train the network with a sufficient level of generalization, so that its work was real..... And even after the Classifier says that yes, this is the signal, it's not the fact that the rate will fly by 100 points, the main thing is that it will go in the right direction, and no one knows how much it will go, it's the fact, as they say....
 
Alexey Burnakov:
How. We make a training set, in which after each point on the chart we collect the information about what will happen within a few hours. (max, min, marginal difference etc.). We sift out the points with less than n points modulo, as you have. and make a classifier that says 1 - buy, -1 sell, 0 - do nothing. Just do everything without a signal system.
What criterion will be the point. What is the condition for the appearance of this point?????
 
If the criterion for the appearance of a point is future movement, then it is pure redrawing. That is, there is no point, the movement began, the point appeared in the past. What should we classify then? We need a fait accompli, not an event that depends on the future.....
 
Mihail Marchukajtes:
What will be the criterion for the point. What is the condition for the appearance of this point?????

I will repeat. In m hours the price will break 100 points up or down. And thus the trade with a take profit of 100 pips is simulated.

Or in m hours the price will be at least 100 points above/below. Then it simulates closing of a position in m hours. This is how I do it.

1...565758596061626364656667686970...3495
New comment
Reason: