Machine learning in trading: theory, models, practice and algo-trading - page 698

New comment
 
Such a small error value usually indicates that the prediction is implicitly looking into the future.
 
Maxim Dmitrievsky:

What "of" the value of the next bar? What are you all laughing at?

By which the SMA is built
 
sibirqk:
By which the SMA is built

It seems to be calculated incorrectly, there are no such predictions

although you can adjust it on the history, but it won't work on the OOS anyway

 
Maxim Dmitrievsky:

It seems to be calculated incorrectly, there are no such forecasts

Although you can adjust it on the history, but it won't work on the OOS anyway

What does the EPO have to do with it?

The result of the window movement is shown at 41 bars. It was predicted to be 42 bars. Then this predicted 42 bar has been compared with the moved fact.

 
SanSanych Fomenko:

What does this have to do with OOS?

The result of a 41 bar window movement is shown. 42 bar was predicted. Then this predicted 42 bar was compared to the shifted fact.

Did the movement occur in the training sample or in the test one?

 
Maxim Dmitrievsky:

It seems to be calculated incorrectly, there are no such predictions

Although you can adjust it on the history, but it won't work on the OOS anyway

The question is why he makes such a big fuss over people. People will not sleep now - think suddenly someone has already baked a philosopher's stone.

 
Maxim Dmitrievsky:

was the movement in the training sample or in the test sample?

Maxim! Well, what is the training sample of the wrecker? A bar appeared - calculate the scale for it, it is always the same.

 
sibirqk:

I have to ask SanSanych why he is worrying people at night. People will not sleep if someone has already created a Philosopher's Stone.

It's like shoeing a flea: an extraordinary skill, but completely useless.

Here I am looking at the graph with this very wand. What good will it do if I know its value one step ahead? The price has a life of its own.

 
sibirqk:

f_ma=(sum(Cn(1:11))+ftr_cn)/12 => ftr_cn=12*f_ma-sum(Cn(1:11))

f_ma is future value of SMA(12) you have predicted, ftr_cn is future bar price value, sum(Cn(1:11)) - sum of values of prices of the previous 11 bars.


SanSanych Fomenko:

This variant does not pass - look at the size of the error, moreover there are sections where the fact is completely the same as the prediction by 1-step-ahead. Count the error in your variant.

Should fit, this is the formula for how to get a prediction of the next price from the predicted SMA and past prices.

#формула как найти SMA для самого нового бара
SMA = (open[11]+open[10]+open[9]+open[8]+open[7]+open[6]+open[5]+open[4]+open[3]+open[2]+open[1]+open[0])/12

#SMA  для следующего бара выглядело бы так:
SMA_next = (open[10]+open[9]+open[8]+open[7]+open[6]+open[5]+open[4]+open[3]+open[2]+open[1]+open[0]+open[-1])/12
#open[-1] это цена открытия следующего бара, и мы её не знаем, это будущее.
#SMA_next вы можете спрогнозировать и узнать вашим способом. Дальше используя её можно выразить open[-1]:
open[-1] = 12*SMA_next - (open[10]+open[9]+open[8]+open[7]+open[6]+open[5]+open[4]+open[3]+open[2]+open[1]+open[0])
open[-1] == грааль
 
SanSanych Fomenko:

Maxim! Well, what is the training sample of the dash? A bar appeared - counted for it, it is always the same.

))))

Did you build some kind of model and then make a comparison of the predicted waveform with the actual one in the area where the model was built or in the new data?

Because it's either a very high quality overfit, or it looks into the future... because it's unrealistic to predict like that in forex

1...691692693694695696697698699700701702703704705...3495
New comment
Reason: