Machine learning in trading: theory, models, practice and algo-trading - page 3069
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
You must have a million signs in there. My approach is set up to be fully automatic. You can just tell me which signs are good in your opinion, a few pieces or the same one with different parameters, I can run on them. And it will pick the targets itself. Because from your dataset only signs will remain in any case, the rest he will change everything else.
It's not a million, it's 6,000 traits. How many do you have on average? In general, CB can move them easily. The fact that the targets can change - let it. I don't have a big sample - 4k rows for train + test dumps for validation (I understand you have a fixed number of trees for each model).
No, it's not a million. It's 6,000 signs. How many do you have on average? Generally, CB is easy to move them around. The fact that the targets can change - let them. The sample itself is not large - 4k rows for train + test dumps for validation (I understand that you have a fixed number of trees for each model).
.
One should realise that this is a peculiar thing with its own tricks. It's not ordinary learning.
10-20 features are enough. Any of them to choose from, so that the formulas can be simply loaded into a ready-made lib. So as not to change anything. After reading the file with prices, I generate the necessary attributes, I don't read ready-made ones. I don't need a large number of sparse ones either. The more features, the more difficult it is to find stable relationships.
Isn't it easier to read from an array and work with the calculated data than to write some formulas?
I don't think you can change the code at all. I don't insist on the experiment.
Still learning
R2: 0.9806482223765112
Learn 2204 model of 3000
Isn't it easier to read from an array and work with the calculated data than to write some formulas?
I think it will not be possible to not change the code at all. I don't insist on the experiment.
You have not understood the main thing I wanted to tell you, try to run your piplan on SB (random walk), for starters, and maybe (for sure) you will have close results, what can it tell you?
8% - errors is ridiculous, on properly prepared chips and targets it can't happen in principle, you are forecasting the past mixed with the future and your forecast effectively searches out this past.
SR - Sharpe ratio normalised by the root of the number of observations, it is a standard measure of strategy performance. SR is a function of akurasi, and even more so of the correlation of the forecast return with the realised return, akurasi 60%+ gives a double digit SR, this is smooth exponential (when reinvested) equity.
The formulas need to be transferred to the terminal. I'll send you a ready-made bot. All you need from me are the features. The names of indicators, if you can not in the formulas.
You can just give me the binary models separately. I understand that in the end there are two of them? This approach will allow you to work with any data.
Can you just give me the binary models separately. I take it there are two of them in the end? This approach will allow you to work with any data.
I will not bargain. You asked me to train - give me the signs, I will train and test. If it turns out to be good, I'll give you the source code.
If there are normal signs, there can't be a lot of them. I don't need datasets with 6k signs, I don't have time for that
Otherwise I'll do other things.
I see that passions have heated up a bit in the last few pages. I ask everyone to write on the merits, not to try to undermine the opponent.
I understand that everyone thinks they are right, but try to respect the opinion of others too.
I suggest to leave arguments and division by packages (Python/R). Nobody will prove anything to anybody anyway.
I'm not haggling. You asked me to teach - give me the signs, I will teach and test. If it turns out to be good, I'll give you the source code.
If there are normal signs, there cannot be a lot of them. I don't need datasets with 6k signs, I don't have time for that
Otherwise I'll do other things.
I don't have signs in one line - you'll spend more time to reproduce them in python. It's more logical to test the effectiveness of the approach on my data and then decide whether to implement the predictor calculation code or not.
If I had very "good" predictors, relative to others, I would not be in a hurry to make them publicly available :) You can do this - take me a model with an acceptable result and from there pull out 20 predictors by importance (according to one of the ways of its definition) in the model.
In addition, I am also interested in the effectiveness of the proposed method on binary predictors - which are quantum segments of predictors, and this technology is not so fast to reproduce, so an array would be preferable - but here I am interested in the result with a large volume of predictors.
If something will be interesting, then already then we can spend time and effort on getting into the logic of calculating predictors and their implementation.