Machine learning in trading: theory, models, practice and algo-trading - page 1798
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
It is better to direct the energy to the generation of new signs... IMHO...
If signs are good then we can even try to predict extremums, I have about 400 and 10 models, but I see that I need tens of times more, training samples can reach > 100 GB , no way )))) I don't have such power...
When there are a lot of predictors - overtraining starts, especially if it's a boosting - there are leaves that are activated on a sample of only 0.01% - it's clearly garbage.
So it's important to work in two directions - look for ways to improve models through their structure and pre/post data processing, as well as by adding new predictors.
Nothing worked with the data I previously sent you for experiments?
I put that EA on a real account now - and I see that there are other problems - for example, the limiter is not filled completely, thin market - large slippages...
When there are many predictors, overtraining begins,
It doesn't, there's crossvalidation + additional sampling to check...
And overtraining is just because you have signs that have 5% of useful information, and you want to squeeze 70% out of them, and it's just not there... there's no useful information...
That's what you need a lot of signs for.
There are leaves that are activated on a sample of only 0.01% - it's obviously garbage.
For example, when you trade with hands, you look at the charts, at different TFs, see some patterns and take decisions, then you go to a small TF and look for an entry point, then wait for the right moment and enter... What was that? It's a compression of information.
1) different TFs and patterns in them - compression
2) shallow TF , entry point - compression
3) The moment of entering - compression
As a result it won't even be 0.01% or even 0.001% relative to the sample, but you don't consider it garbage, do you?
Nothing came up with the data I sent you earlier for experiments?
what data? i missed it...
not start , there is crossvalidation + additional sampling for checking...
You can check, but it does not affect the training.
And overtraining just because you have signs that carry 5% of useful information, and you want to squeeze 70% out of them, and it just does not exist... there is no useful information...
If each predictor carries 5% useful information, that's already good, or how much do you expect? How do you determine usefulness in general - I'm basing it on the deviation from the average of all the targets in the sample.
When you trade with hands, for example, you look at the charts, at different TFs, see some patterns and make decisions, then you go to a small TF and look for an entry point, then you wait for the right moment and enter... What was that? It's a compression of information.
1) different TFs and patterns in them - compression
2) shallow TF , entry point - compression
3) The moment of entering - compression
As a result concerning the sample it will not even be 0.01% or even 0.001%, but you don't consider it garbage, do you?
No, it will be randomness, intuition, but not the system.
With what data? I missed something...
Link.
Yeah, sure )))) I'd like to see how they use "if, then, else" to recognize pictures or generate speech, that would be hardcore)) But seriously, you're talking nonsense!
Finerider recognizes without ns.
I once read an interview with one of their employees, where he called neural networks "discriminant analysis for the poor".
I once read an interview with one of their employees, in which he called neural networks "discriminant analysis for the poor
Link.
Ahhhh I didn't see it, I thought you never redid anything... No I didn't.
Probably misunderstood. Finerider recognizes without ns. Work on mice brain activity determined what the mouse would do half a second before the action in the 90s. When ns won chess and then with MOs comp became invincible these events were not covered in this way, though more significant. Work passport recognition I met about 10 years ago. Recognized it fine. Now the news of passport recognition with the NS will be installed on the railway ticket offices.... There is more hype than work.
This is all from misunderstanding .... There is a lot of hype, because we need a haip, we need a haip because there are no staff, it is necessary to arouse interest in people, or rather in the useless biomass that does not even go to the toilet without a smartphone.
How can you compare the challenges of the 90s with today's tasks? How can you compare a passport and driving a car with popular recognition of objects? Or video subject recognition on YouTube? Or human speech generation? 30 years have passed, the tasks have become thousands of times more complicated and they are solved, and you are talking about some lame passport from the 90s, but it is still there in the 90s.
How can you compare the tasks of the 90's with today's tasks is also stupid, how can you compare a passport recognition with driving a car with incidental object recognition? or recognizing the subjects of YouTube videos? or human speech generation? 30 years have passed, the tasks have become thousands of times more complicated and they are solved, but you are talking about some poor passport from the 90s, it is still there in the 90s
I do not agree, the problems must be understood from the outset and the history of their emergence and solutions helps in this. The MO algorithms were mostly formulated and implemented before the 90's, back in the 77 Fortran, the same Cat Boost. Today the tools, power and software have become more available. Recognition tasks follow the same algorithm, i.e., set bibliography, cataloging, comparison. Only quantity and speed change (so do catalog search methods, but it's not crucial). Galileo invented the first hash)))
Previously 10K was big today 100Gb can be implemented at home. If today's capacities allowed to do full search and full search passes to optimize GA would not be needed)))))
About Passport) https://open-dubna.ru/ekonomika/9057-razrabotka-rezidenta-oez-dubna-sokrashchaet-ocheredi-v-kassakh-rzhd
Solutions used to cost (sold) around $500. How much is it with AI? I don't think it's cheaper. The solution is more than 15 years old. The main problem there by the way to remove the protective grid))). On a homogeneous background and on a grid are different tasks.)
Ahhhh I didn't see it, I thought you never redid anything... No, I didn't.
Are you going to watch it now?
Will you watch it now?
I looked...
the current file with the balance does not contain the prices, the prices that you sent me earlier did not agree in size with the current balance
UPD==============
If i think about it, this idea is a failure. i should not analyze and forecast the balance graph, but look for good/bad entry points