Machine learning in trading: theory, models, practice and algo-trading - page 3248

New comment
 
Maxim Dmitrievsky #:

there in the loop for each example the correlation with the others is filled in by the already calculated matrix, so there is a lot of

25K correlations for each example. It would be interesting to see the distribution of correlation from -1 to +1.

 
Maxim Dmitrievsky #:

I'll try to speed it up some more.

I think we can apply the sieve principle.

  1. We calculated 25K correlations for one pattern. For example, MathAbs(corr) > 0.9 for 500. We took them into account and threw them out.
  2. Now the number of patterns decreased by 500 and the number of required correlation calculations decreased by 500.
I.e. at each step we remove similar places.
 
fxsaber #:

I guess you could apply the sieve principle.

  1. We calculated 25K correlations for one pattern. For example, MathAbs(corr) > 0.9 for 500. We took them into account and threw them out.
  2. Now the number of patterns decreased by 500 and the number of required correlation calculations decreased by 500.
I.e. at each step we remove similar places.

There is already a simple selection by known indices of the correlation matrix from the original dataset, for each pattern, so it is fast. The rightmost line on the screen just shows how many examples of each pattern there are. Other data like future prices are filled in and statistics are considered. You can do it another way, yes.

 
fxsaber #:

pattern entry, exit in n hours.

Exit like this?
 
fxsaber #:
The output is?

on a fixed TP

 
Maxim Dmitrievsky #:

fixed

While mining?

 
fxsaber #:

While mining?

I can see from the statistics, let's say the future 10 bars, I output all curves of each found instance of the pattern in the future (like a forecast).

Then the average of all the curves.

Like this pattern to sell, on average, and how many pips can be seen.

 
Maxim Dmitrievsky #:

pattern reference is saved, in the tester we look for correlation of current values with the reference, open trades according to the selected logic

I got it, why there are a lot of samples. If you plot the correlation on a 1d array, you will get something remotely similar to this.


Above the upper blue line, the correlation is super-high. You can see that it changes smoothly at the tops.


It means that we saw 0.9 in the tester - good, we open. On the next bar 0.91, then 0.92, ...., 0.95, 0.94, ...., 0.9. The length of such super-high consecutive values is the longer the pattern itself. Purely mathematical.


Perhaps, that is why there was a layering of a large number of trades.

Forum on trading, automated trading systems and testing trading strategies.

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

Maxim Dmitrievsky, 2023.09.21 15:07

The bot opens several trades at the same time, if the signal persists, with the same volume, from this it seems that martin


Therefore, probably, when mining it makes sense to consider in a series of consecutive samples only the first one.

 
fxsaber #:

I figured out why there are so many samples. If you plot correlation on a 1d-array, you will get something remotely similar to this.


Above the upper blue line, the correlation is super-high. You can see that it changes smoothly at the vertices.


It means that we saw 0.9 in the tester - good, we open. On the next bar 0.91, then 0.92, ...., 0.95, 0.94, ...., 0.9. The length of such super-high consecutive values is the longer the pattern itself. Purely mathematical.


Perhaps, that is why there was layering of a large number of trades.


That's why it probably makes sense to consider only the first one in a series of consecutive samples when mining.

Yeah, it does

 
Maxim Dmitrievsky #:

the statistics shows, let's say the future 10 bars, I output all curves of each found instance of the pattern in the future (like a forecast)

Then the average of all curves

Like this pattern is for sale, on average, and how many pips can be seen.

I saw something similar here.

It looks similar on the screen, but the length of the pattern is very high, because M1. It will probably show something interesting on the hour-markers, since it was found.


And this problem

it probably makes sense to consider only the first in a series of consecutive samples when mining.

seems to be solved.

  • MinStep is a conditional step assignment(MinStep * Max(|KK|)), which allows to take into account only the best of the consecutive similar plots according to theLimit condition.
But it's not mining (overshooting) there, of course. Although it's not far away.
1...324132423243324432453246324732483249325032513252325332543255...3577
New comment