Machine learning in trading: theory, models, practice and algo-trading - page 1534
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
Indicators from the sample on which the training was conducted.
And on the sample outside of training, what? The next six months, for example?
So your logloss is ~1, which means that the model is in awe of what's going on.
I'm blind, where is the negative logloss? I specified the logloss for the whole model.
You want to find a separate logloss for an individual input? That is, the amount of information between the values of the sheet and the target?
I'm not sure what logloss is, I want to find quality information, and logloss is looking for rather the quantity of the whole model, maybe the quantity will do - I do not know. Any idea how to do this?
The "do nothing" class is clearly unbalanced relative to the others, which is what I wrote above that may be such a problem because of the 3-class model. The rest of the metrics are depressing.
Yes, it's not balanced, but what can you do... You can of course do two samples - breaking the sample into vectors, but again, some of the information that might be useful would be lost. In general, I agree that the sample is not balanced, but how to do it without losing information, I do not know. I tried duplicating strings with small number of targets - the effect is negative.
I'm blind, where is the negative logloss? I specified a logloss for the whole model.
I'm not sure what logloss is, I want to find quality information, and logloss is looking for rather the quantity of the whole model, maybe the quantity will do - I don't know. Any idea how to do that?
Yes, it's not balanced, but what can you do... You can of course do two samples - by splitting the sample into vectors, but again, some of the information that might be useful would be lost. In general, I agree that the sample is not balanced, but how to do it without losing information, I do not know. I tried to duplicate rows with a small number of targets - the effect is negative.
Not negative, but tends to unity, i.e. maximum. It should tend to zero.
Logloss shows correlation between your features and your target. Roughly speaking, 0 means full correlation, i.e. they describe the target well. 1 is no description at all, i.e. they are completely uninformative. This is a good indicator in the sense that it says whether there is any correlation at all. It turns out that you have no correlation.
I do not know, do the normal 2 class and not 3.
And on a sample outside of training, what? The next six months, for example?
And sampling outside of training is kind of gone... I didn't, I trained for about March 1, 2019, decided that I should use all the information from 2014.
But, decided to see for myself what these trees provide, stuck it in the EA and did three passes with 3 different trees from March 1, 2019 to September 15, 2019.
1. First split step
2. Twentieth split step
3. Forty-eighth step of splitting
And, in general, I myself am surprised that the results are positive for all three models!
Interestingly, the 2nd model is smoother, and the third model is already crawling into the plus side with difficulty.
Interestingly, the accuracy of the models is not much worse, and sometimes even improved, below is a table with percentage changes relative to the training sample
I can't say about completeness and logloss - you have to take the numbers and do the sampling.
And yes, most models will loop when training to zero, since there really is a preponderance of that class relative to the other two, and then zeros are easier to find - that's what messes things up.
is not negative but tends to unity, i.e. maximum. It should tend to zero.
logloss shows how much your features correlate with the target, roughly speaking, i.e. 0 is full correlation, i.e. they describe the target well. 1 is no description at all, i.e. they are completely uninformative. This is a good indicator in the sense that it says whether there is any correlation at all. It turns out that you have no correlation.
I don't know, make it a normal grade 2 instead of a grade 3.
I'm not sure that Logloss in multiclassification is equal to one... In general, I can't figure out how to implement the formula myself - I don't understand these ciphers from public sources. And I would like to see Logloss not the final one, but for the whole sample, how it changes and where it sags. And as I understand it, it's more correct with a balanced sample...
I'm not sure that Logloss for multiclassification equals one... In general, I can't figure out how to implement the formula myself - I don't understand these ciphers from public sources. And I would like to see Logloss not the final one, but for the entire sample, how it changes and where it sags. And as I understand it, it is correct in a balanced sample to a greater extent...
I don't want to rack our brains with this... big companies like Yandex are doing things. Said: do it this way and you'll be fine. Just do it and don't do it on your own. Otherwise you will drown in wording and different approaches.
It shows the change as it goes along the gradient, building up treesAnd sampling outside of training is kind of gone... I didn't, I trained for about March 1, 2019, decided that I should use all the information from 2014.
But, decided to see for myself what these trees provide, stuck it in the EA and did three passes with 3 different trees from March 1, 2019 to September 15, 2019.
1. First split step
2. Twentieth split step
3. Forty-eighth step of splitting
And, in general, I myself am surprised that the results are positive for all three models!
Interestingly, the 2nd model is smoother, and the third model is already crawling into the plus side with difficulty.
Interestingly, the accuracy of the models is not much worse, and sometimes even improved, below is a table with percentage changes relative to the training sample
I can't say about completeness and logloss - you have to take the readings and do the sampling.
And yes, most models will loop when training to zero, since there really is a preponderance of that class relative to the other two, and then zeros are easier to find - that's what messes things up.
This is what I get. Just moved all bot logic to python, forest replaced by boost. I can't find the error, there seems to be no peeks. Depending on the settings, acuras can be raised to 0.7-0.8, while reducing the number of transactions.
For wood, the error range is about the same, but it is not acuras, it is classification error. It behaves similarly in trayne, even better. But it is much worse in the test.
training:
The OOS is 10 times bigger than the training.
What's the input? Pure prices?
increments
I will finish the connector this weekend and put it to the tests. I plan to load the model in the cloud and retrieve signals from the terminal. I will also put it to the cloud and use it to collect signals from the terminal. I can send you the client on MT5 laterA simple and interesting approach to describing patterns for MO
https://www.quanttrader.com/index.php/a-simple-algorithm-to-detect-complex-chart-patterns/KahlerPhilipp2019