Machine learning in trading: theory, models, practice and algo-trading - page 2168

New comment
 
Aleksey Nikolayev:

It seems to be written everywhere that Fourier is good just for periodic signals. Or close to it - with a narrow spectrum.

By the way, for MO in trading, it seems to me, the Walsh decomposition would be more suitable, but for some reason I have not seen it mentioned on the forum.

Not just periodic, but always with the same period (time from the beginning of one oscillation to the beginning of another) and always the same shape. And in quotes both periods and the form of the curve are changing all the time. Therefore, no signal transformations in the market are appropriate.
I also studied electronics in high school, and I know what I'm talking about.

 
Igor Makanu:

Yes!

I just finished - you got it right

I have to stop, because one moron stirred up everyone with his fantasies )))

It showed a filter and 10 deals, of this type. From tails to the mean (filter lines) or something like that

As usual, when the trend starts, it'll be negative in the whole account

And it posts these stumps every time in every thread, and you all cheerfully discuss it for 10 pages for many years ))


 
Maxim Dmitrievsky:

It showed a filter and 10 trades, of this type. From the tails to the average (filter lines) or something like that

As usual, when a trend starts, it's going to be negative in the whole account

And it posts these stumps every time in every thread, and you all together have fun discussing it for 10 pages for more than 1 year ))

ok

now show the same on the mashka and compare

rookie ;)

 
Aleksey Nikolayev:

It seems to be written everywhere that Fourier is good just for periodic signals. Or close to it - with a narrow spectrum.

And there are no options there - basis functions are sines and transformation only defines for each sine its offset (phase) and period

and the intersection of all these sines (x-axis) gives us the points where we take signal values (y-axis)


UPD: here is another good explanation,https://habr.com/ru/post/196374/

the most valuable thing in this article are hand-drawn drawings ;)

 
Renat Akhtyamov:

ok

now demonstrate the same thing on the car and compare

bully ;)

there are plenty of other mashcas that look similar. It doesn't change anything.

Like HMA or something like that

if you're going for MAs, you should at least normalize them for volatility and so on.

And don't write that it's a grail. It's not even close.

 
Maxim Dmitrievsky:

there are plenty of other mash-ups that look similar. It doesn't change anything.

For example, HMA or something like that.

If you're going for MAs, at least normalize for volatility and stuff like that.

And don't write that it's a grail. It's not even close.

Show it in a screenshot, Max, not in words.

Keep your mouth shut.

 

If you need it, here is a small script that calculates the relative area under the intersection of the curves of the distribution of the value of a feature for two clusters.

cl_1 = samples.drop(samples[samples["targets"] == 0].index).reset_index(drop=True)[samples.columns[1]]
cl_2 = samples.drop(samples[samples["targets"] == 1].index).reset_index(drop=True)[samples.columns[1]]

min_x = min(cl_1.min(),cl_2.min())
max_x = max(cl_1.max(),cl_2.max())
x = np.arange(min_x,max_x,(max_x-min_x)/100)
hist_1 = np.histogram(cl_1,x)
hist_2 = np.histogram(cl_2,x)
plt.hist(cl_1,x, histtype= 'step',label='cl_1')
plt.hist(cl_2,x, histtype= 'step',label='cl_2')

area1 = np.sum(hist_1[0])/sum(hist_1[0])
area2 = np.sum(hist_2[0])/sum(hist_2[0])
ymax = np.maximum(hist_1[0],hist_2[0])
ymin = np.minimum(hist_1[0],hist_2[0])

area_overlap = sum(ymin/sum(ymax))
print(area1, area2, area_overlap)

I think this is a good metric for choosing a thinning method, target analysis, features, and who knows what else.

area_overlap is yellow in the picture

 
Renat Akhtyamov:

Show me a screenshot, Max, not just words.

You don't talk like a sack of bags.

What do you want to see? Here is the real TC on the MO from my article. Training - only 1 month, then generalization for 2 years. And it is real and it works.

 
welimorn:

If you need it, here is a small script that calculates the relative area under the intersection of the curves of the distribution of the value of a feature for two clusters.

I think this is a good metric for choosing a thinning method, target analysis, features, and who knows what else.

area_overlap is yellow in the picture.

You do ingenious things ) I'll take a look at it on Monday.

 
Everyone knows that the deviation from the average can be significant, but no one knows when it will happen.
1...216121622163216421652166216721682169217021712172217321742175...3699
New comment