Machine learning in trading: theory, models, practice and algo-trading - page 1334

 
Aleksey Vyazmikin:

I looked up MSUA, I don't know which book it refers to specifically, but it's not searchable with that name. As I understand it, this thing is used in CatBoost

--l2-leaf-reg.

l2-leaf-regularizer.

L2 regularization coefficient. Used for leaf value calculation.

Any positive values are allowed.

3

CPU and GPU


Or is it about something else? This method can also be used when creating predictors, for example to describe patterns on certain sections.

Well, this is regularization by Tikhonov, and where is the bagging temperature?

 
Maxim Dmitrievsky:

Well, this is Tikhonov's regularization, and where is the temperature bagging?

But the meaning seems to be the same, no? I just don't know what algorithm is inside there...

--bagging-temperature

Defines the settings of the Bayesian bootstrap. It is used by default in classification and regression modes.

Use the Bayesian bootstrap to assign random weights to objects.

The weights are sampled from exponential distribution if the value of this parameter is set to"1". All weights are equal to 1 if the value of this parameter is set to"0".

Possible values are in the range. The higher the value the more aggressive the bagging is.

1
 
Aleksey Vyazmikin:

But the meaning seems to be the same, no? I just don't know what algorithm is inside there...

--bagging-temperature

Defines the settings of the Bayesian bootstrap. It is used by default in classification and regression modes.

Use the Bayesian bootstrap to assign random weights to objects.

The weights are sampled from exponential distribution if the value of this parameter is set to"1". All weights are equal to 1 if the value of this parameter is set to"0".

Possible values are in the range. The higher the value the more aggressive the bagging is.

1

it's different of course.

kind of useful when you have a lot of features, I guess.

will change the model a little bit, pure fine tuning no more

the details need to read, in general it is clear but not to the end

 

By the way, I found the lectures I mentioned earlier, with examples in python, for anyone who needs to use XGboost mostly. There, or in the next lectures regularization is also discussed.


 
Maxim Dmitrievsky:

That's different, of course.

well, it's kind of useful when you have a lot of features, I guess.

will change the model a little bit, purely subtle tuning no more

Let's see what will be the spread - today or tomorrow will be the next 100k models, I'll decide whether to apply this parameter in the overshoot...

 
Aleksey Vyazmikin:

Let's see what will be the variation - today or tomorrow there will be another 100k models, I will decide whether to apply this parameter in the overshoot...

I don't have a normal manual on the parameters? I don't use catb yet, I'm reading about other things

 
Maxim Dmitrievsky:

Why is there no normal manual on the parameters? I'm not yet using a katb, I read about other things

Well, all that is settings and a brief description, plus a well-known roller with explanations.

 
Aleksey Vyazmikin:

If you look carefully, you can see that financial results of models in one sample can vary greatly - from 5000 to 1500, i.e. significantly, which means that Seed does affect the models. I will assume that it is the selected models that are similar (I'll check), and they have slightly different profit margins, but almost all models are flat in the middle, which is surprising - they make mistakes on the same margins (an anomaly in the new data?).

You have a box with some kind of very hilly landscape created inside it. We throw a lot of balls in there (that's the sids), and our job is to make sure that most of the balls hit the deepest hollows. This will be the learning, and this is the principle by which learning in MO is structured.

1. If we jiggle the box slightly, most of the balls will not be able to leave the hollows where they originally hit - learning will not happen.

2. If we shake the box hard, some of the balls have a chance to get and stay only in the deepest hollows, but the shallower ones will remain unfilled, because the balls will pop out of there. Full learning will not happen.

3. If we shake the box with medium force, only the deepest and middle troughs will be filled, but the rest of the balls won't find anything and will keep randomly bouncing around the box. The training is better than in 1 and 2, but it's also not ace.

There are always settings in the learning methods - exactly how and when to shake the box to get the most effective learning.

If the different "sids" don't add up, then either something is wrong with the learning algorithm - you shake it wrong, or there are no deep troughs to latch on to in our box.

 
Yuriy Asaulenko:

You have a box, inside which a kind of very hilly landscape is created. We throw a lot of balls in there (that's what sids are), and our job is to make sure that most of the balls hit the deepest hollows. This will be the learning, and this is the principle by which learning in MO is arranged.

1. If we jiggle the box slightly, most of the balls will not be able to leave the hollows where they originally hit - learning will not happen.

2. If we shake the box hard, some of the balls have a chance to get and stay only in the deepest hollows, but the shallower ones will remain unfilled, because the balls will pop out of there. Full learning will not happen.

3. If we shake the box with medium force, only the deepest and middle troughs will be filled, but the rest of the balls will not find anything and will continue to randomly bounce around the box. The training is better than in 1 and 2, but not ace either.

There are always settings in the learning methods - exactly how and when to shake the box to get the most effective learning.

If the different "sids" don't add up, then either something is wrong with the learning algorithm - you shake it the wrong way, or there are no deep hollows in our box to get a grip on.

Or a shoebox )

The balls are a good explanation.

and a good box shakes itself.

 
Yuriy Asaulenko:

You have a box, inside which a kind of very hilly landscape is created. We throw a lot of balls in there (that's what sids are), and our job is to make sure that most of the balls hit the deepest hollows. This will be the learning, and this is the principle by which learning in MO is structured.

1. If we jiggle the box slightly, most of the balls will not be able to leave the hollows where they originally hit - learning will not happen.

2. If we shake the box hard, some of the balls have a chance to get and stay only in the deepest hollows, but the shallower ones will remain unfilled, because the balls will pop out of there. Full learning will not happen.

3. If we shake the box with medium force, only the deepest and middle troughs will be filled, but the rest of the balls will not find anything and will continue to randomly bounce around the box. The training is better than in 1 and 2, but it's also not ace.

There are always settings in the learning methods - exactly how and when to shake the box to get the most effective learning.

If the different "sids" do not converge, then either something is wrong with the learning algorithm - you shake it the wrong way, or there are no deep troughs in our box to latch on to.

Good abstraction, if by deep troughs we mean answers with minimal error on validation, for which there is a learning stop, then this can also explain the fact that he got better results with increasing the size of the validation sample, and this may be the result of a formal increase in the abstract terrain size and the number of troughs accordingly.