Machine learning in trading: theory, models, practice and algo-trading - page 785

 
Aleksey Vyazmikin:

Okay, you don't want to talk about that, let's talk about something else.

Why do you take a delta of 10 bars and not the delta between the opening and the maximum/minimum for those 10 bars?

The ten bars are chosen at random, just a nice number. With my variable, I get a number that tells me that in ten bars the difference between the claws will be so-and-so. And what does the data you suggest tell me. What is the physical meaning in them.

In my TF the point is: if NS shows that the difference in ten bars between the current and future indexes will be 300 pips, I open, if less, I do not. And what will you do when you get a forecast? How do you use it? I just do not understand you a bit.

 
Mihail Marchukajtes:

Where do these percentages come from. There are tests that have been conducted. We are still unable to decide on the initial data, and you are already talking about the result .....

If i'm wrong, i can't see the difference between one bar and 10 bars. And yes, i've done lots of tests to predict the number of points of the next bar, but even the number is guessed in 53% of cases, and the direction together is 50/50.

 
Anatolii Zainchkovskii:

If you do not know how to do this, then you may be wrong, but I do not see the difference between one bar and 10 bars. And yes, I have repeatedly made tests to predict the number of points of the next bar, and even only the number is guessed in 53% of cases, and the direction together is 50/50.

No problem. I just picked 10. Let's predict for one bar or two. I'm asking people. What's the interest? Let's define unambiguously what we want to get from TC and continue...

Unfortunately Rashit killed his article about BOO, so I predict the color of a candle in 1, 2 or 3 bars ahead. But I do it with the help of classification. That's not fair. Let's build a system that uses regression.

By the way my Boshs are predicted correctly on average 8 out of 10... So...

 
Mihail Marchukajtes:

No problem. 10 was just my choice. Let's predict by one bar or two. That's what I'm asking people. What's the interest? Let's define unambiguously what we want to get from TC and continue...

Unfortunately Rashit killed his article about BOO, so I predict the color of a candle in 1, 2 or 3 bars ahead. But I do it with the help of classification. That's not fair. Let's build a system that uses regression.

By the way my Boshs are predicted correctly on average 8 out of 10... So...

If you want to check the exact position of the indicator, you'll have to choose one of the two possible ways, like 1-60 bars for 1 minute or 2-1bars for 1 hour.

 
Anatolii Zainchkovskii:

The final result, of course, it is interesting to get at least in 60% of cases to get to those announced 300 points.

Well, no problem. There is interest in predicting the color and size of the next bar. There is no time now, I propose to wait for MEMBERS until evening, gather together and discuss how it is. There and prepare the target at your request there and continue. The only thing let's determine when we will make this prediction, every bar or at certain times. As an example, I'm working on the BOO only when the window is forming on the basic TS. When green dots appear then predictions. as an example...

 
I really tell you that with the help of Reshetov's classification I predict 8 out of 10 candles correctly on the OOS. Approximately, but I do it only during the formation of the window. I'm off, see you tonight...
 
Mihail Marchukajtes:

Well, no question. There is interest in predicting the color of the next bar and its size. Now there is no time, I propose to wait for MATERS until the evening, gather all and wipe what and how. There and prepare the target at your request there and continue. The only thing let's determine when we will make this prediction, every bar or at certain times. As an example, I work in BO only when the window is forming on the basic TS. When the green dots appear then the prediction. as an example...

I think i can try to learn on each bar, then look at the result of trained forward prediction and if there is a pattern at a certain time the forecast works better, then continue to use only this time range. i tried the number of bars in a day, the result has not improved.

 
Mihail Marchukajtes:
I'm really telling you that with the help of Reshetov's classification I predict 8 out of 10 candlesticks correctly in OOS. Approximately, but I do it only during window formation. I'm off, see you tonight...

Ok, see you tonight.

 
Vizard_:

I'm more about ideology - it's desirable that the state has the property of markoviability, as much as possible include information for
to predict the next states, etc... About the network I don't really understand and have no desire, but the implementation can be anything, you can plug in anything,
combine, + validation (I have not tried), + number of iterations you can probably try to limit, etc., as long as the response on the OOS was.
Compare it to knn, KX, etc. with the teacher.
Try to put only time in for interest. Example for clock - xH = 1,2,3,4,5...... sin(xH), cos(xH) and, or so forth, similarly for min...

Okay, I'll put all sorts of stuff in it and try it out... I'll let you know what I get later.

 
Mihail Marchukajtes:

The ten bars are chosen at random, just a nice number. With my variable I get a number that tells me that in ten bars the difference between the claws will be so-and-so. And what does the data you suggested say? What is the physical meaning in them.

In my TF the point is: if NS says that the difference in ten bars between current and future indexes will be 300 pips, then I open. And what will you do when you get a forecast? How do you use it? I just don't understand you.

What is important is not what will be in X bars but what was in 10 bars, i.e. if in 10 bars X points have been reached, then we open.