Machine learning in trading: theory, models, practice and algo-trading - page 3355
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
Somehow you are not paying attention to my posts, focusing on probabilities. It doesn't matter what the probability is called, what matters is that if it doesn't improve, the model is overtrained, into the bin. The prediction error on OOV, OOS, and VNE should be about the same.
The model out of the box does not give the correct probabilities, any of them. That's the story. You may have predicted labels that completely match, but the probabilities don't.
Added my post. Any model gives correct probabilities in the sense that the classification error will not fluctuate.
Somehow you are not paying attention to my posts, focusing on probabilities. It doesn't matter what the probability is called, what matters is that if it doesn't improve, the model is overtrained, into the bin. The prediction error on OOV, OOS and VNU should be about the same.
Here's another histogram
Different algorithm - different histogram, although the labels and predictors are the same. If you are looking for some kind of theoretical probability, implying that different classification algorithms will produce the same histograms ... that doesn't occur to me, since you have to work with specific algorithms and they will predict and they have to be evaluated, not some theoretical ideal. The main evaluation here is the overfitting of the model, not the proximity of the probabilities to some theoretical ideal.
Give up? Google classification probability calibration, it should be in R.
.
We're talking about different things.
I am writing about the result, and you are writing about the ideal of intermediate data.
For me it is obvious that probability values of specific labels given by RF and ada will be different, but predictions of specific labels are almost the same. I'm not interested in the probability values, I'm interested in the prediction error
If you theorise, it is most likely impossible to obtain the class probability in your sense, since you have to prove that your probability satisfies the limit theorem, which is very doubtful.
We're talking about different things.
I am writing about the result, and you are writing about the ideal of intermediate data.
The class probability values given by RF and ada will be different, but the predictions of specific labels are almost the same. I'm not interested in the probability values, I'm interested in the prediction error.
If you theorise, it is most likely impossible to obtain the class probability in your sense, since you have to prove that your probability satisfies the limit theorem, which is very doubtful.
Still, the original question was there, no one answered.
Why? If in the sense of a thesis....
Why? If in the sense of a thesis....
I was hoping someone would at least google the tip.
It shows the model's outcome on "probability" ranges with 0.05 step. CatBoost puts the class separation at 0.5 quite accurately (magnetta is 1, aqua is 0).
You can see that the fin outcome is positive starting at 0.35 - the green curve rises above the red curve.
Is this what you want to calibrate - shifting the point of class separation to the point of revenue generation?