Machine learning in trading: theory, models, practice and algo-trading - page 3556

New comment
 
Aleksey Vyazmikin #:

You can always make one, just thought there was ready-made functionality - was interested in the selection criteria.

And so - in essence - if you need exactly clustering, you can use any, on which you can then apply new data, just pull out the thresholds on clusters and record in the quantum table, which can already be submitted to CatBoost. This will speed up the process as you don't have to re-count the clusters when experimenting.

The main problem is the partitioning of the transactions. If the partitioning is rubbish, discretisation will give something, but not much.

You may need a different markup for each bin.

You can take several markup templates and stretch them on each bin, if you don't have any idea how it should look like. This is all easily done programmatically, no need to explain here. The concept itself is important.
 
Maxim Dmitrievsky #:

The main problem is the partitioning of trades. If the markup is rubbish, discretisation will give something, but not much.

A different markup may be needed for each bin.

If the sampling is formed on each bar, you can think in this direction. Like different states - different TS....

But, the main thing is to work with stable (effective) quantum segments (bins), otherwise everything will fall on new data like a house of cards.

 
Aleksey Vyazmikin #:

If the sample is formed on each bar, we can think in this direction. Like different states - different TS....

But, the main thing is to work with stable (effective) quantum segments (bins), otherwise everything will fall on new data like a house of cards.

Often it is enough to look at distributions inside bins, they should be standard. Because the MO theory works only with such.
And among 10 bins/clusters 2-3 of them will be such. All that remains is to mark them up correctly.
 
Maxim Dmitrievsky #:
Often it is enough to look at the distributions within bins, they must be standard. Because the MO theory works only with such.
And among 10 bins/clusters 2-3 will be such. All that remains is to mark them correctly.

Often it is a question of sample representativeness. In theory it may make some sense, but in fact I don't see any dependence in my experiments.

Besides, the normal distribution should be completely available for probabilistic estimation, and when you take only a piece of it, the meaningfulness of the condition is lost.

But, of course, you try it yourself, don't trust anyone.

 
Aleksey Vyazmikin #:

Often it is a question of representativeness of the sample. It may make sense in theory, but in fact, I don't see any dependence in my experiments.

Besides, the normal distribution should be completely available for probabilistic estimation, and when you take only a piece of it, the meaningfulness of the condition is lost.

But, of course, you try it yourself, don't trust anyone.

There is a dependence and it works. It's matstat.
 
Maxim Dmitrievsky #:
There is a correlation and it works. It is matstat.

The article just talks about better learning on normally distributed data, plus it makes the caveat that it only matters for some types of model building methods.

And I agree that it can be useful for training in some way. But that's not what this is about. You take a subsample from this distribution, and you think there's a relationship between which distribution the subsample is from, but there isn't, all the remaining predictors in it have lost samples from their distributions differently - not necessarily proportionally - almost randomly. Thus, what you choose the bin from does not matter, the most important thing is the stability of the probability bias throughout history. Well, if you make new labels, this is what you need to control for correct markup.

 
Aleksey Vyazmikin #:

The paper only talks about better learning on normally distributed data, plus the caveat that this only matters for some types of model building methods.

And I agree that it can be useful for training in some way. But that's not what this is about. You take a subsample from this distribution, and you think there's a relationship between which distribution the subsample is from, but there isn't, all the remaining predictors in it have lost samples from their distributions differently - not necessarily proportionally - almost randomly. Thus, what you choose the bin from does not matter, the most important thing is the stability of the probability bias throughout history. Well, if you make new labels, this is what you need to control for correct markup.

I don't get it, new definitions again.
It is about what it is about. Not about anything else.
Or about the central limit theorem.
I guess there's no point in continuing to argue. I don't really have anything else to do.
It's enough to get it right and see that it's good. Compare distributions, for example, during clustering and during validation.
 
Maxim Dmitrievsky #:
It's enough to do it right and see that it's a good thing.

That's the most important thing.

 

Few people think about it, but quantisation can increase the number of predictors, focusing on different information from the general flow.

As an option, here we take deviations from MA (data stream) with equal step, let's say 100, and we get a channel - the data on the channel number at different moments of time will tell us about the current price position, volatility strength, current trend, trend volatility (for the time of deviation from the average price).

Now let's change the quantisation table and look only at what is happening near the levels.

Such data can already tell us about the situational price behaviour, especially if we assume that the levels are significant for the market, we will be able to evaluate the attempts to break through the level, its resistance to them, and in general - whether the level is significant or not. They can also be signal points for making a decision to open a position.

If the tables are built purely on the scale of indicators for the whole sample, it can be of little use, although the initial data will be valuable, but they will be incorrectly processed.

 
Aleksey Vyazmikin #:

we take deviations from MA (data stream) with equal step, let's say 100, and get a channel - the data on the channel number at different moments of time will tell us about the current price position, volatility strength, current trend, trend volatility (for the time of deviation from the average price).

Original. I will take it to the topic of what to feed to the input of neural network

1...354935503551355235533554355535563557355835593560356135623563...3625
New comment