Machine learning in trading: theory, models, practice and algo-trading - page 2699

New comment
 
Aleksey Nikolayev #:
Cyclic time (hour number, etc.) is easy to use, for example, in KNN, if the metric is written correctly. Or in some development of this method, such as local regression.
They write such things in textbooks, but in fact cyclic time is already embedded in increments, and there is more useful information in them. They are needed for the sake of time, and for price levels we need to add something else.
 
elibrarius #:
I see a familiarity, I've seen it 3-4 times now in your posts.
2 times 0.5 per turn.)))))))

2 times, it's like this:

2 times 0.5 is in the centre :-) An average of two, which suddenly suddenly describes tics well

 
Is this it? Is the discussion of automatic feature search over?
Or was that the point?
 
Ahahahaha, you had a lot of pressure, but when we got down to business, we just got deflated..
The main thing is to trindle, trindle, and then you'll see your grandchildren grow up...
 
mytarmailS #:
0) yes I am...)

1) I haven't deployed the whole thing yet,
1. there are problems with the curse of dimensionality and combinatorial explosion, but this is solvable in theory, in favour of accuracy....
2. There is a problem with the fact that the search algorithm is slow, a lot of things need to be written in C or C++, and I don't know how to do it.
3. Even an optimised algorithm will not be able to search for patterns in a large date, we need to search for patterns locally.....
But in general, if it doesn't work, nothing works...

2) Yes.


By the way, you can replace the word "event" with the word "rule".


My method fixes the number of spaces in which the regularity is searched and limits the step of coordinates in these spaces, so there should be no explosions. Plus there are ideas how to immediately reduce the number of combinations to be searched by analysing the spaces beforehand.

I will do the search in MQL5 through the "mathematical computation" mode, the advantage here is the debugged system of agent support, which will allow to manage parallelised computation tasks. I have quite a lot of weak cores on my servers, so it is important to me.

A rule is an analogue of a tree leaf, if I remember your research correctly. The leaf contains the conditions that describe the pattern, and the Event is the source for finding the pattern.

The Event is perhaps the stump of the tree, which will be built up by interacting with other predictors.

Building up, even it is possible to tell growth, if to use representation about a tree, - it is the second already stage, to realise which it is possible either through the algorithm (while sketches on a paper) or on R through genetic trees (it is simply already worked out methodology, to you threw the script), or as you do - but working already with a small in general table - searching relative regularities, and it is possible to think up something else. And at this stage CatBoost can already digest data with joy, as an intermediate solution. It is possible to pull out leaves and rules from it, but they are usually weak there.

 
Maxim Kuznetsov #:

the probability of overcoming any line by the price (and triggering of indicator signals) depends on the time of day and day of the week.

It is necessary to add cyclic time to NN and DL. The simplest way is a sine wave. Dependencies are non-linear, so it is simply squared, taking into account the sign. There are two additional inputs that are responsible for time references. Midnight/midday is different everywhere, so it is better to calculate and give the phase in advance. This is the connection of the model with the real world and its time

If they are not explicitly given, then IMHO you will either get a pumpkin or the whole thing will try to get and output them by itself.

Yes, time is one of the most important scales, and of course I use it.

How is the issue of transition to summer/winter time solved, do you think any correction is required?

Let's say we trade Euro/Ruble - on the history we have different moments of transition to winter/summer time, and then the absence of transition for the ruble, but the presence of the euro, let's say planned news events are important, but with the time shift they will be on the chart at different times, and how to be? Maybe it makes sense to use the time scales of two currencies at once, and maybe more?

 
Aleksey Vyazmikin #:

Yes, time is one of the most important scales, and of course I use it.

How is the issue of the transition to summer/winter time solved, do you think any correction is required?

Let's say we trade Euro/Ruble - on the history we have different moments of transition to winter/summer time, and then the absence of transition for the ruble, but the presence of the euro, let's say planned news events are important, but with the time shift they will be on the chart at different times, and how to be? Maybe it makes sense to use the time scales of two currencies at once, and maybe more?

this is a well-known b@##... and it constantly confuses everything, no matter what we trade:-) in two major centres - USA and England, the hands of the clock are moved on different days. Up to more than 1 week apart. The intervals between the most important events change and two or three weeks in six months can be thrown out of the analysis. And our people are making a mess of things, "we change the clocks, we don't change the clocks".

I don't know a universal or even more or less successful solution to this problem. Either just ignore these "critical days" or teach winter/summer time separately. The latter seems more reasonable, but we're already critically short of data as it is

 
Aleksey Vyazmikin #:

My method fixes the number of spaces in which a pattern is searched and limits the step of coordinates in these spaces, so there should be no explosions. Plus there are ideas how to immediately reduce the number of combinations to be explored by analysing the spaces beforehand.

I will do the search in MQL5 through the "mathematical computation" mode, the advantage here is the debugged system of agent support, which will allow to manage parallelised computation tasks. I have quite a lot of weak cores on my servers, so this is important to me.

A rule is an analogue of a tree leaf, if I remember your research correctly. The leaf contains conditions describing the pattern, and the Event is the source for finding the pattern.

The Event is perhaps the stump of the tree, which will be grown by interacting with other predictors.

Building up, even it is possible to tell growth, if to use representation about a tree, - it is the second already stage, to realise which it is possible either through the algorithm (while sketches on a paper) or on R through genetic trees (it is simply already worked out methodology, to you threw the script), or as you do - but working already with a small in general table - searching relative regularities, and it is possible to think up something else. And at this stage CatBoost can already digest data with joy, as an intermediate solution. It is possible to pull out leaves and rules from it, but they are usually weak there.

Are there any tools in your approach to take into account data invariance?

https://en.wikipedia.org/wiki/Affine_transformation
 
mytarmailS #:

Any tools in your approach to account for data invariance?

https:// en.wikipedia.org/wiki/Affine_transformation

Maybe it is relevant for multiple points, for example to find similar patterns, but in my case there is essentially one point at the first stage. The point is converted/normalised into different relative measurement systems - time scale and price, plus a third space - any discrete predictor continuously describing the market. You get 3 dimensions in the initial representation. Each has its own quantum table.

 

Glad the topic is in the top!

The more adherents - the more I will swoop)))))

1...269226932694269526962697269826992700270127022703270427052706...3496
New comment
Reason: