Machine learning in trading: theory, models, practice and algo-trading - page 613
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
Surprised. What kind of model is it that counts for more than an hour?
It should be 1-2 minutes at the most.
88-50-20-2 network, 86400 lines of training data on 88 predictors. + 28800 on validation data and on the test site (however, they are counted quickly).
Teacher, I'm embarrassed to ask. How much should a model cost?
Well, that certainly does not bring pennies? Maybe there is a formula (price-quality)?)).
Well obviously not two kopecks, Trickster.....You should understand it, you are not a little one.....
The search for a generalizing model in multidimensional data space is not fast from the optimization point of view, because increasing requirements and introducing more and more strict rules for model building in order to reduce the effect of overlearning leads to an increase in time to find such a model. That is, the data must be shaken thoroughly.....
Why the same file on AWS and Reshetny different amount of time? AWS 2-5 minutes, Reshetov 3 hours and his model is twice better than AWS models/ Why?
88-50-20-2 network, 86,400 lines of training data on 88 predictors. + 28,800 for validation data and test plot (however, they count quickly)
With such a set of Reshetov would count forever :-)
Let me tell you a forex secret. Alpha in the data can be only on a very short segment. With my data I have not yet been able to increase this parameter more than 50 rows. That is, I have 100 columns and 50 rows. This covers about two weeks of the market. That is, if I start to increase the training interval, the quality of the model falls below 75% and the quality of the CB becomes such that it is impossible to work on it, at your own risk. So I don't understand what you want to do with these thousands of records. Build a model for the entire market with an acceptable level of training quality you can not, the larger the learning curve, the worse the model. And if it shows a good result in such an area, it has nothing to do with the concept of generalization..... You know what I mean......
Do you know why you can't build models on a long plot with good performance??? Because such data... inputs.... that could build such a model does NOT exist in nature in principle..... There is no such data or it would have been used everywhere a long time ago. I mean publicly available data....... not insider or whatever.... So... I don't understand why so many lines????
With such a set of Reshetov would count forever :-)
Let me tell you a forex secret. Alpha in the data can be only on a very short segment. With my data I have not yet been able to increase this parameter more than 50 rows. That is, I have 100 columns and 50 rows.
You can't build such models, sampling length should be at least 5 times the number of features, and you have more features than sampling and the curse of dimensionality
You show ignorance instead of revealing the secret of Forex.
So I don't understand what you want to do with these thousands of records? You won't be able to build a model for the whole market with an acceptable level of training quality. And if it shows a good result in such an area, it has nothing to do with the concept of generalization..... You know what I mean......
This is M1, only 60 days. So it's not for the whole market, but for the last 3 months.
When the duration doubles, the model no longer builds.... Optimize the duration of course, but I have not got to that yet. This is where the number of layers needs to be figured out first.
If I build at least 10 models, their calculation will take 8-10 hours(((
There are 3 formulas to calculate, find min and max, count them, then 2-3 between them, and 2-3 outside. And then from all this to choose the best models, well, if they are already counted - in the ensemble of them.
PS Hm. This is with 2 hidden layers, and the same amount with the 1st layer.
In general, 24 hours will determine the model.
In general, 24 hours will determine the model.
I'd like to see a good video card for this kind of data satanism, it's not a hoax anymore :)
Miners freaks jacked up the price tag 3 times over on all the normal cards
This is M1, only 60 days. So it's not for the whole market, but for the last 3 months.
When doubling the duration, the model was no longer built.... Optimize the duration, of course, but I haven't gotten to that point yet. I would like to figure out the number of layers first.
If I build at least 10 models, their calculation will take 8-10 hours(((
There are 3 formulas to calculate, find min and max, count them, then 2-3 between them, and 2-3 outside. And then from all this to choose the best models, well, if they are already counted - in the ensemble of them.
PS Hm. This is with 2 hidden layers, and the same amount with the 1st layer.
In general, 24 hours will determine the model.
In general, I can suggest you a dedicated server for i7 3-4Ghz and ssd for $7-8 per month. Counts ok, and the computer will not be busy as much.
At the moment I keep thinking that it is possible to input (and, probably, output) distributions of quotes at a certain depth, or moments of distributions
You will get a certain smoothing and probabilistic picture and, perhaps, a limited number of certain variants, which is important. But I haven't done it yet - I need to pump up my analysis of variance for that.
If we take into account fractals, then mb. the relations of distributions between different fractals. But the topic needs to be seriously worked out, to draw a scheme
Searching for a generalizing model in the multidimensional data space is not a quick task from the optimization point of view, because increasing requirements and introducing more and more stringent rules for model building in order to reduce the effect of overlearning entails increasing the time for searching for such a model. That is, the data must be shaken thoroughly.....
Why the same file on AWS and Reshetny different amount of time? AWS 2-5 minutes, Reshetov 3 hours and his model is twice better than AWS models.
Reshetov's model is not a benchmark. For example the search for a set of predictors in it is done by trying different variants - the model takes a random set of predictors, it is trained, and it remembers the result. This is repeated in a loop a huge number of times, eventually the best result is used as the final model. This process can be noticeably accelerated if you first do the selection of predictors by a special algorithm, and then train the Reshetov model just once on this particular set. And you get Reshetov model quality at a speed comparable to AWS. The "cost" of such a model will drop significantly, but the quality will remain the same.
I think you should try to build a model on the principle of fractal analysis with such data. When several timeframes are allowed to enter. Maxim showed a good video on the world fractality.
In general, I can suggest you a dedicated server for i7 3-4Ghz and ssd for $7-8 per month. Counts ok, and the computer won't be busy as much.