Machine learning in trading: theory, models, practice and algo-trading - page 3406
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
Well, for graphs to speak better than words, it would be good to explain what they mean in general, what is on them, how counted, etc....
Here I already wrote that this is the financial result of 100 trained models CatBoost. Models with different seed values - and so the same hyperparameters.
Here take the selection of linearly dependent features, the essence of the method is to throw out those features that essentially duplicate others.
You had 2410 features, after the selection there are 500 features left, the model works the same or even better, the method worked 100%.
How is this far from the optimum?
There are many different methods. And there is a sense in the proposed method, as tests have shown.
However, using a different approach, you can potentially get a better result, as the graph shows.
In general, I think that it is not quite correct to draw final conclusions from a single sample.
It seems that you still don't understand what you were doing at all.
This is a sign of megalomania :))))))
It's a sign of megalomania :))))))
Alexei,
you are contrasting cleaning the sample from unnecessary features with a method to improve the quality of classification.
These are different tasks like classification and clustering like warm and soft.
And if you're doing it, and you are, it's not my megalomania, noooooo))))
You're the one who's frankly stupid, no offence.
But thanks for the experiment.
Alexei,
you're contrasting cleaning the sample from unnecessary features with a method to improve the quality of classification.
These are different tasks as classification and clustering are like warm and soft.
And if you do it, and you do it, it's not my megalomania, noooooo))))
You're the one who's frankly stupid, no offence.
In fact, the predictors that matter are those with which the model has shown/can show the best result. And the goal is to select such predictors without looking into the future. How this will be done is another question. The result for me would be to reduce the variance of the model's financial results - it is not.
I agree that ideal conditions are needed to evaluate the effectiveness of a method. Why the model may train better on linearly dependent predictors - because the model appears to actually factor in the gain of certain predictors in importance by choosing the same information to split in different trees.
And, drop the habit of throwing sand around in the sandbox - it can get in your eyes.....
Here I already wrote that this is the fin result of 100 trained CatBoost models. Models with different seed values - and so the same hyperparameters.
And why only sids are different? Why not on different samples?
In Boost you can also set class balance, have you done it?
Why only different sids? Why not different samples?
Is that sarcasm? Seed affects the random number generator, which here affects the selection of a trait when it is chosen after efficiency calculation. I.e. the best one will not always be selected.
You can also set class balance in the boost, have you done that?
Of course.
Is that sarcasm? Seed affects the random number generator, which here affects the selection of the trait when it is chosen after the efficiency calculation. So the best trait will not always be selected.
What do you see as sarcasm?
Why is it not sarcasm to run the same model 100 times on three identical samples?
but making 500 samples and running a model is sarcasm?
Or do you think that different sids somehow change the sample, maybe it's sarcasm?
What did you see as sarcasm?
In the absurdity of the sentence.
why running the same model 100 times on three identical samples is not sarcasm.
The model is not one model, we create different models due to different seed.
Using comparable training data allows the result from that training to be comparable as well.
and taking 500 samples and running the model is sarcasm?
For what purpose? Describe the goals and objectives of setting up such an experiment and what it might show.
or do you think that different sids change the sample in some way, maybe that's sarcasm?
The predictors used to build the model and their order of selection change.
Accordingly, we assume that initially there are unnecessary/unnecessary predictors that worsen the result, so the task is to exclude them from training.
The model is not just one model, we create different models through different seed.
different models or the same model on different sids?
one model on the same sample with different sids is not different models.
For what purpose? Describe the goals and objectives of setting up such an experiment and what it can show.
For a more realistic evaluation, like crosvalidation or the same casual.
different models or the same model on different sids ?
One model, on the same sample with different sids is not different models.
How can there be one model with different sids - I obviously don't understand you.
You build a model and then change the seed when you apply it? CatBoost doesn't even have this functionality.
For a more realistic assessment, like crosvalidation or the same casual.
So you describe specifically - general phrases don't make the approach clearer.
Like:
1. selected predictors on the total sample.
2. Made 500 subsamples, fixing them.
3. trained 100 models on each subsample with different methods
4. Got 500*100*4=200,000 (two hundred thousand) models down a bunch of time
5. Now we will evaluate them in a tricky way - how?
6. Let's conclude.
How can there be one model with different sids - I obviously don't understand you.
You build a model and then change the seed when you apply it? CatBoost doesn't even have this functionality.
So you describe specifically - general phrases do not make the approach clearer.
Like:
1. selected predictors from the total sample.
2. Made 500 subsamples by fixing them.
3. trained 100 models on each subsample with different methods
4. Got 500*100*4=200,000 (two hundred thousand) models down a cloud of time
5. Now we will evaluate them in a clever way - how?
6. Let's conclude.
you have a TS on moving averages.
you select parameters (periods of averages) to this TS on the same plot but with different sids.
The TS is a catboost model.
Is the analogy clear?
On the second question, read Max's article about casual, because I'm tired of explaining, I'm already tired to be honest.