Machine learning in trading: theory, models, practice and algo-trading - page 2976

New comment
 
Aleksey Nikolayev #:

As far as I understand, "quantisation" (histograms) is used in bousting to speed up, so that there are fewer variants for splits. If so, the solution is good for its universality, but may be bad in a particular case - the real boundary may be lost.

Yes, that's correct. It speeds up and can be attributed to regularisation. But it also loses the exact split.

 
Forester #:

Maybe, but I don't see any fish there. I don't use quantisation at all. I prefer to explore float data.

I'm sorry you don't believe me.

I can demonstrate the effectiveness on your sample, compare the learning curve.

 
Forester #:

Yes, that's right. It speeds up and can be attributed to regularisation. But it also loses the exact split.

Accurate split on history. If the nature of predictor values distribution is known, quantisation can pick out exactly the range with characteristic stable behaviour. For trading it is just relevant.

 
Aleksey Vyazmikin #:

Accurate split on history. If the nature of the distribution of predictor values is known, quantisation can pick out exactly the range with characteristic stable behaviour. For trading it is just relevant.

The task of searching for ranges/splits is solved by tree learning. There is at least some meaningful formula separating the rows with respect to the target.
In quantisation it is just a counter + skipping of doubles. Quantisation occurs without any check on the target function.
.

If a tree with training on the target does not give stability (or gives very weak stability), how will a counter that has no relation to the target give it? Only random and sometimes randomly good segments, which in time will cease to be such.

 
Forester #:
The task of searching for ranges/splits is solved by a tree during training. At least there is some meaningful formula separating the rows with respect to the target.
In quantisation it is just a counter + skipping of doubles. Quantisation occurs without any check on the target function.
.

If a tree with training on the target does not give stability (or gives very weak stability), how will a counter that has no relation to the target give it? Only random and sometimes randomly good segments, which in time will cease to be such.

Quant tables have to be selected for each predictor. Assuming a lucky random is hit, that's what I want to identify. Random or not. Not with 100% reliability, but even by eliminating 30% of randomness you can improve the quality of the trained model.

I am developing my split estimation function (algorithm), which should reduce the disadvantage of trees - greed.

It is strange of course, I have been working on this topic for years, I have done a lot of experiments with different samples, I have statistics on the effectiveness of the approach, I say that the method works, but I meet with distrust.

 
Aleksey Vyazmikin #:

Quantum tables should be selected for each predictor. Let's say a lucky randoms hit - that's what I want to detect. Random or not.

How can it be NOT random with respect to the target if the target is not involved in choosing the quantisation point? Only random.

 
Forester #:

How can it be NOT random with respect to the target, if the target is not involved in choosing the quantisation point? Only random.

It is random, but the pattern is not random. I.e. it will persist in the future. The estimation takes into account the same target.

On the other hand, no one prevents to more accurately immediately split the predictor into quantum segments taking into account the target.
 
Aleksey Vyazmikin #:
On the other hand, nobody prevents to split the predictor more accurately at once into quantum segments taking into account the target.

It is the task of the tree to find the best splitting point, so that the purity of the classes of the target right and left sections is maximised.

Do you want to estimate purity during quantisation? Essentially you want to do the same thing the tree will do later. Turn off quantisation and you get what you want. The tree will pick the best split point given the target.

 
Forester #:

It is the task of the tree to find the best split point, so that the purity of the target right and left classes is maximised.

You want to estimate purity during quantisation? Essentially you want to do the same thing that the tree will do later. Turn quantisation off and you get what you want. The tree will pick the best split point given the target.

Tired of explaining that "best" is often not the best choice.

Instead of questions - statements - as if we are doing religion....

 
Aleksey Vyazmikin #:

Tired of explaining that "best" is often not the best choice.

Instead of questions - statements - as if we are dealing with religion....

You have a specific approach with specific terminology, people are not ready to sacrifice space on their hard drive for such information, without understanding the result. I see the way out is either to make a detailed article to find enthusiasts who will (most likely) start pushing this in the marketplace right away 😀, or pay someone a little bit as a journeyman. Or a student for a stick of sausage.

Here's an example of my last article about meta models even. How many people wrote with suggestions and improvements, ideas? Zero :) and there's a ready-made TC. How many have taken it and are using it? At least several, without feedback.

It's a very difficult question in general, when you do something and then you're afraid to put it in the public domain, so that you don't turn out to be a sucker. But at the same time you want help. Here there are 1000 kites for one Mother Teresa, they will just eat you up. This is such a specific field of activity. Then you're on your own.)
1...296929702971297229732974297529762977297829792980298129822983...3595
New comment