Machine learning in trading: theory, models, practice and algo-trading - page 2255
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Rare activations on Exam rather means that the market has changed and what often happened on the trayn has stopped happening. And it doesn't necessarily mean that there weren't many sheet activations there either.
Yes, I agree that there is a market change effect as well.
Let's look at Train.
The situation is slightly better, but there are also leaves with a rare number of activations.
Notice how the learning happens - a tree with a large weight is built - conditionally successful, and then a set with small weights, and then again large - such a pie, and if you remove the veins with small weights, then you get a shift in probability.
Yes, I agree that there is a market-changing effect as well.
Let's look at Train.
The situation is slightly better, but there are also leaves with a sparse number of activations.
Notice how the training happens - a tree with a large weight is built - conditionally successful, and then a lot with small weights, and then again large - such a pie, and if you remove the veins with small weights, and you get a shift in probability.
I wonder what will happen if you train a new model on this diagram?
In general the idea is to train the second model on the "guts" of the first.
Why are you messing with this carburetor? You're not improving anything with it.
If you understand what the problem is, you can look for a solution. Obviously these trees have flaws.
But I agree that I can't figure out the CatBoost code to make edits to it, alas.
However, there is an opportunity to influence the model, perhaps zeroing in rare examples in the leaves will give a positive effect, but it is desirable then to recalculate the coefficients of the leaves - with this is more complicated, but globally solvable.
Take a simple neural network without leaves. It will work on new data as badly as boosting. What does this tell you?
I agree that there will be overtraining effects there too, but of a different nature - the question is which of these effects can be more accurately detected and evaluated and which is easier to deal with.
There is an excellent SHAP tool for feature selection and interpretation, but it's in python. Everything has been done for you a long time ago )
In fact, the vast majority of these methods only talk about using predictors in models, but don't do any evaluation of them themselves. You need estimates of predictors independent of the model - I'm working on that, there are modest positive results.
Of course I want to play around with ready-made solutions in Python or R, but I have doubts that I can handle new syntax.
I wonder what happens if you train a new model on this diagram?
In general, the idea is to train the second model on the "guts" of the first
This model in the example from the ancient deposits, now I have 60k leaves in the models, which is certainly a lot to form a sample. It is possible to try significantly reducing the number of trees. However, I note that I evaluated the leaves from CatBoost and they are very weak in their characteristics individually compared to the leaves from the genetic tree.
On leaves(thousands of leaves) from the genetic tree I have trained - it is possible to improve metric performance.
If you understand what the problem is, you can look for a solution. Obviously, such trees have disadvantages.
But I agree that I can't figure out the CatBoost code to make edits to it, alas.
However, there is an opportunity to influence the model, perhaps zeroing out the rare examples in the leaves will have a positive effect, but it is desirable to re-weight the leaf coefficients afterwards - this is more difficult, but globally solvable.
I agree that there will also be overtraining effects, but of a different nature - the question is which of these effects can be more accurately identified and evaluated and which is easier to deal with.
In fact, the vast majority of these methods only talk about using predictors in models, but do not make any assessment of them themselves. We need estimates of predictors independent of the model - I'm working on it, there are modest positive results.
Of course I want to spin ready-made solutions in python or R, but there are doubts that I can handle the new syntax.
There it is the impact of features on the behavior of a particular model that is evaluated
If you understand what the problem is, you can look for a solution. Obviously, such trees have disadvantages.
But I agree that I can't figure out the CatBoost code to make edits to it, alas.
However, there is an opportunity to influence the model, perhaps zeroing out the rare examples in the leaves will have a positive effect, but it is desirable to then re-weight the leaf coefficients - this is more difficult, but globally solvable.
I agree that there will also be overtraining effects, but of a different nature - the question is which of these effects can be more accurately identified and evaluated and which is easier to deal with.
In fact, the vast majority of these methods only talk about using predictors in models, but do not make any assessment of them themselves. We need estimates of predictors independent of the model - I'm working on it, there are modest positive results.
Of course I want to spin ready-made solutions in Python or R, but there are doubts that I can handle the new syntax.
Came to the conclusion that adding by 1 (or removing by 1) is the best. Here's my research. I guess you've already seen it.
It is the impact of features on the behavior of a particular model that is evaluated there.
That's what I'm saying, the evaluation goes through the resulting model.
Came to the conclusion that adding 1 at a time (or removing 1 at a time) is the best. Here's my research. You've probably already seen it.
Haven't seen it before - looked it up - in general I agree that the real effect is through removal. CatBoost has a method of removing the predictor and sort of reweighting the model without it, but I haven't dealt with it. So far I limited myself to adding and removing predictors, but not just one, but in groups.
That's what I'm talking about, the evaluation goes through the resulting model.
And that's a good thing.
you can see what signs are flawed in the new data
I don't know... maybe it's my experience or maybe it's my drinking...)
but it seems to me that you're suffering from dickishness...)