Machine learning in trading: theory, models, practice and algo-trading - page 2076

[Maxim Dmitrievsky]

mytarmailS:

you should write a random enumeration of everything with everything and train on it, select what is trainable + genetics to shorten the search, and leave the computer for a couple of months....

I'm about to write an article on random sampling, but in python

to have it )

https://www.mql5.com/ru/articles/8642

I plan to compare it with RNN, CNN, etc. But first make some more modifications on catbust.

[Evgeniy Chumakov]

mytarmailS:

I think it is better not to look for a pattern after the pattern, but to look for the rebound price related to this pattern, it is much more difficult to formalize, but it seems to me

It is possible to calculate the percentage of price growth on history at a given pattern (arithmetic mean, root-mean-square), for example.

[mytarmailS]

Evgeniy Chumakov:

You can calculate on the history by how many percent the price increase was for a given pattern (arithmetic mean, root mean square) for example.

You can simply normalize the pattern to the range 0-1, for example, and then normalize the space relative to this pattern, and that's it.

see what I mean?

[Evgeniy Chumakov]

mytarmailS:

you can simply normalize the pattern to the range 0-1, for example, and then normalize the space relative to that pattern, and that's it

see what I mean ?

I don't get it.

Here I have the current pattern:

Pattern 'long' = X ,

The 'short' pattern = B,

Past event = L.

I can use history to find when it was the same.

Then normalize the space relative to this pattern - by the type of function in the fix range (or there 25 % below the maximum, etc.)

[mytarmailS]

Evgeniy Chumakov:

I don't get it.

Look, it's simpler, I also puzzled for years...

we have the x1 patren

x1 <- rnorm(40)

we have two other patterns x2 and x3 they are the same as x1 but with different volatility

x2 <- x1*3
x3 <- x1*6

we consider the first 5 points to be patterns, all the following will be considered the space relatively to the pattern

1) bring the patterns into a single range, say 0-1

the usual normalization

r01 <- function(x)    (x-min(x))  /  ( max(x) - min(x))

all three patterns will be the same

plot(   r01(x3[1:5])   ,t="l",lwd=50) 
lines(   r01(x2[1:5])   ,col=2,lwd=20) 
lines(    r01(x3[1:5])   ,col=3,lwd=10)

now it is necessary to normalize the space relative to the pattern

this is done by the function r02

r02 <- function(x,y)    (y-min(x))  /  ( max(x) - min(x))

where " y " is space and "x" is not a normalized pattern

Then we connect the results of transformations from functions r01 and r02 in one row (vector)

and we obtain a pattern normalized in the range 0-1 and the space is normalized relative to this pattern

plot(c(r01(x3[1:5]),r02(x = x3[1:5],x3[-c(1:5)]))  ,t="l",lwd=10) 
lines(c(r01(x2[1:5]),r02(x = x2[1:5],x2[-c(1:5)])),col=2,lwd=5) 
lines(c(r01(x1[1:5]),r02(x = x1[1:5],x1[-c(1:5)])),col=3,lwd=2)

similarities should be searched not by correlation but by Euclidean metric

eucliden.distance <- function(x1, x2) sqrt(sum((x1 - x2) ^ 2))

the smaller is the value, the closer the patterns are to each other

[Evgeniy Chumakov]

I see that it's different.

Your pattern describes the price, and my pattern describes the segments (shorter, longer). If the pattern is equal to the letter B , then the identical ones are only the same B.

But the space around the pattern is naturally different because of volatility, but I don't need it because I care about formality (shorter or longer).

[mytarmailS]

Evgeniy Chumakov:

I see that it's different.

Your pattern describes the price, and my pattern describes the segments (shorter, longer). If the pattern is equal to the letter B, then the identical ones are only the same B.

What difference does it make what you're comparing?

Your pattern is 10-20-10 and 20-40-20 is one and the same pattern or not?

[Evgeniy Chumakov]

mytarmailS:

What difference does it make what to compare?

Is your pattern size 10-20-10 and 20-40-20 the same pattern or not?

What size? I have several clearly described patterns they are all different as for example 0110 and 1001

[mytarmailS]

Evgeniy Chumakov:

What size? I have several clearly described patterns, they are all different, like 0110 and 1001

Well, the sections are yours.

 у меня отрезки (короче,длиннее)

10 candles - 20 candles - 10 candles

Don't slow down)

[Evgeniy Chumakov]

mytarmailS:

Well, your segments are

10 candles - 20 candles - 10 candles

don't slow down)

What candles 10-20-10 ?