Machine learning in trading: theory, models, practice and algo-trading - page 2017
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
Has anyone tried using "round levels" as signs?
Or as a way to handle prices ?
You can mark prices with circular values for example...
You can remove values that are the same in a row ...
It is a good information compression, plus filtering ... Maybe it will be easier to look for patterns on such a chart for a model ...
I plan to make predictors, the logic here is that there are option strikes at levels, so it may be useful for Moex.
The charts are interesting, if we could get a fast algorithm in MQL...About jamming - perhaps we need to change the way the error correction.
Well, why a black box, if there are only 2-3 layers, it is quite realistic to unscramble by coefficients. Small coefficients here can be coarsened and zeroed out, which will reduce the number of inputs to the neuron.
What do you mean by "maybe you should change"? Learn the math and how activation f-ions work. Or the network developers are so dumb they wouldn't have known
Why get into all this stuff and reinvent the wheel without any specific training or mathematics? It is a stupid waste of time. There are technologies and it is written how they should be used.
What do you mean "probably need to change" learn the math and how activation f-ions work. Or network developers are so dumb that they would not have guessed
Because it's nothing to nothing, just assumptions and zero specifics. They trained autoencoder, put fiches into boosting or NS, showed results. That's all. You don't have to untwist anything. Deep architectures are not built to be parsed, but to reduce the analytical routine.
Why even get into all this and reinvent the wheel without a degree or mathematical background? It's a stupid waste of time. There are technologies and it is written how they should be used, that's all. Lots of people are working on this.
If ready-made solutions out of the box to solve the problems I have set before them, then there was no need to invent anything, but, alas.
I'm preparing a big sample now and will train a lot of models on CatBoost, have some ideas how to estimate the quality of the model for further successful real-time application - I'll share the results of my research.
If ready-made solutions out of the box would solve the tasks that I set before them, then it was not necessary to invent anything, but, alas.
Now I'm preparing a large sample and will train a lot of models on CatBoost, have ideas how to assess the quality of the model in order to further their successful application in real time - I'll share the results of the study.
I don't think CatBoost is good for forecasting of time series, it doesn't work with sequences.
You can play with the classification, but it will be uselessCatbust is not suitable for time series forecasting, it does not work with sequences
you can purely play with classification, but it will be uselessAnd how do you determine whether it works or not?
I have models that are profitable for a year (trained about a year ago) in the tester - are you suggesting to consider them a fluke?
Yes, CatBoost is inferior to a genetic tree with post-processing leaves, but it is very fast to train.
And what works - NS?
And how do you determine whether it works or not?
I have models, which for a year (trained about a year ago) show a profit in the tester - you suggest to consider them a fluke?
Yes, CatBoost is inferior to a genetic tree with post-processing leaves, but it is very fast to train.
And what works - NS?
I didn't define it, it's the architecture itself for other tasks
yes, it's all a fluke
nothing works yet )
Predictors are not prices in naked form - many relative points that may be similar...
I'm not sure that screening by correlation would be effective...
Why not try it? A negative result is also a result (in the sense of food for further thought).
I once even suggested the following formula for the correlation coefficient: C = (n1 - n2)/n, where n is the number of bars where at least one of the two systems gives a trade signal, n1 is the number of bars where the signals are given by both systems simultaneously and in the same direction and n2 is the number of bars where the signals are given by both systems simultaneously and in the opposite directions.
The matrix of these ratios can be used for clustering, thinning and portfolio forming.
I didn't determine it, but the architecture itself is for other tasks
Yes, it's all a fluke.
so far nothing works )
Of course, there's no time series sharpening here, so the predictors should contain information about the X coordinate, not just the Y coordinate.
If you can learn to identify such random patterns, you'll be a pro.
I have more than 60% of the leaves sampled in previous years working, which is very, and in my opinion, validates the idea of a badly classified data processing approach. If more people worked on this idea, the result would be better, but everyone has their own bling.
Of course, there is no sharpening for time series, so predictors should contain information about X coordinate, not only Y coordinate.
If one learns how to identify such random patterns, one will be profitable.
I have more than 60% of the leaves sampled in previous years working, which is very, and in my opinion, validates the idea of a badly classified data processing approach. If more people were working on this idea, the results would be better, but everyone has their bling.
Like, the article wanted to... Sketch out the gist of the approach. I still do not understand what you do :D
I stick to the view that signs should be extracted automatically by the model itself from the time series (if there are any). And there is no need to do anything manually. The increments are enough. The question is the architecture. For example like in NLP (neural language processing), the neural network itself determines the context in word sequences, i.e. connection between time series references.
Why not try it? A negative result is also a result (in the sense of food for further thought).
I believe I once even suggested the formula for the correlation coefficient: C = (n1 - n2)/n, where n is the number of bars where at least one of the two systems gives a signal for trading, n1 is the number of bars where the signals are given by both systems simultaneously and in the same direction and n2 is the number of bars where the signals are given by both systems simultaneously and in the opposite directions.
The matrix of these ratios can be used for clustering, thinning and portfolio forming.
What does this have to do with predictors?
I do something similar for leaf selection, but there is a catch in the fact that the number of leaf responses is different in the sample and you need to take into account that leaves with similar responses but different length can belong to the same group.