Machine learning in trading: theory, models, practice and algo-trading - page 2003

New comment
 
Evgeniy Chumakov:

1) Here you are (if I may call you that, correct me if I do)

2) The future. She only sees numbers. Or is that something mystical?

3) We look at the graph and don't know reliably what will be on the right in the future. And if we look at the same graph backwards we know the past (what's on the left side of the graph) because-we can look, but if we hide part of the graph on the left can we reliably tell what should be there?

1) Yes, as you like.

2) yes, it's hardly mystical

3) really interesting thing, it turns out that the past we have not seen is easier to predict than the future that we have not seen.

 
mytarmailS:

3) really interesting point, it turns out that the past that we have not seen is easier to predict than the future that we have not seen


So it turns out that the future weakly depends on the past (50-60%) in the time series, but the preceding past strongly depends on the future?

For example - the current candle goes up, because the previous one was going down, but it is not reliable, and the previous one went down, because the current one is going up. How can this be?



 
Evgeniy Chumakov:


So what happens, the future in the time series weakly depends on the past (50 - 60%), but the preceding past strongly depends on the future?

For example - the current candle grows because the previous candle was falling, but it is not reliable, and the previous candle was falling, because the current one is growing. How can this be?



It's not the candles you have. It's something else that you never told us. Perhaps this pre-processing of yours is laying high predictive power for the past. Zigzag will do that.


And purely by candlestick, I think that it should predict both unknown candlestick from the future and unknown candlestick from the past in approximately the same way.

 
elibrarius:


And purely on the candles, I think it should predict about the same for an unknown candle from the future as for an unknown candle from the past.

Try it, if you're not too lazy. I think there will be a surprise. I don't know how widely known it is, maybe it just passed you by somehow. At least for me it's a very long known thing, without any machine learning it's much easier to build a good MTS on "inverse" quotes than on "straight" ones. There is no "symmetry" there.

I don't know exactly why, although I can guess.

 
Wizard2018:

Try it, if you're not too lazy. I think there will be a surprise. I don't know how widely known it is, maybe it just passed you by somehow. At least for me it is a very long known thing, without any machine learning it is much easier to build a good MTS on "inverse" quotes than on "straight" ones. There is no "symmetry" there.

I don't know why exactly, though I'm guessing.

Laziness. And it's a pity to waste time on something that will never come in handy.
 
Wizard2018:


I don't know exactly why that is, although I'm guessing.


Can I hear a suggestion?

On the transformed series it is understandable why such good results on the reverse course, but if the performance is better on the normal price, then it is a paradox, it turns out that the past depends on the future.

 
Evgeniy Chumakov:


Can I hear a suggestion?

On the transformed series it is clear why such good results in the reverse course, but if at the normal price the indicators are better, then it is a paradox, it turns out that the past depends on the future.

Analysis of transformations and several rows of different ones and then it is possible to draw conclusions.

Scientifically it should not depend if the transformations are the same for all members of the series. If not, anything is possible. If we transform the current value from the past values by one or 2 bars for example. So the conversion from 0 to the end of the series will be different than from the end to zero. The transformation should be the same whether it is forward or backward))) then it will be correct.

 
Evgeniy Chumakov:

If you inverted the series, then the signs should also be inverted. Multiply by -1

 
Maxim Dmitrievsky:

If you have inverted the row, you must also invert the signs. Multiply by -1.


According to those logs I got 50/50, so it will not improve.

That series that I immediately posted the last "price" at the top of the file, and they say so wrong.

I posted two files here https://www.mql5.com/ru/forum/86386/page2002#comment_18198406

Original - last value at the top

reverse - last value at the end of the file

 
Evgeniy Chumakov:


According to those logs I got 50/50, so it will not improve.

The series that I immediately posted the last "price" at the top of the file, and they say so wrong.

I posted two files here https://www.mql5.com/ru/forum/86386/page2002#comment_18198406

Original - last value at the top

reverse - at the end of the file is the last value

If I reverse the second one, the error is also bad. It should be like in the first one. So incorrectly prepared

reversed the first - the error remains the same (good)

>>> regr.score(X_train, y_train)
1.0
>>> regr.score(X_test, y_test)
0.9958279845956355

The direction of the row does not play a role

No, stop. Yes, this row does not work in the opposite direction.
1...199619971998199920002001200220032004200520062007200820092010...3495
New comment
Reason: