Machine learning in trading: theory, models, practice and algo-trading - page 2018
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
I think the article wanted to... Sketch the essence of the approach. I still don't understand what you are doing :D
I adhere 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, as in NLP (neural language processing) a neural network itself determines the context in word sequences, i.e. connection between counts of time series.
The genetic tree and CatBoost are weakly connected, the article I plan to write about CatBoost. Postponed to write for the reason that I identified my shortcomings in the stability of the predictors and threw all my forces to fix it, at the same time and the new predictors made. By the end of the week I plan to start the computational process (it's frustrating when the servers are idle) and I will have time for the article - I'll try to write the first part by the end of the month. The article will be about my kitchen of model making on CatBoost.
With genetic trees everything is more complicated, there won't be an article about it yet, but the approach is that we select leaves from trees which stably classify a part of data on history - in fact 0.5%-3% of responses from all sample, the more such leaves the better, now there are about 1000 to buy and sell, in addition I search leaves which also filter selected leaves, i.e. I carry out additional training, which increases their accuracy. Leaves are grouped according to their similarity (there is still some work to be done), then their responses are weighted within each group on the history and the threshold at which the signal from a group of leaves is generated is determined. An additional filter here is the genetic tree built from the responses of all leaves or only groups. This approach allows to significantly increase the completeness of classification in an unbalanced sample, in my case with 3 targets, where the target "0" is about 65%.
Work on the criteria for selecting leaves and the methodology for combining them has great potential for improvement, and hence models can be of higher quality.
What does this have to do with predictors?
I must be confused, thinking about my own)
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, as in NLP (neural language processing) neural network itself determines the context in word sequences, i.e. connection between samples of time series.
I agree about architecture, we need a completely different architecture, we need a set of networks:
1. which identifies images
Determining the spatial ordering of images.
Looking for patterns in the images placed in space.
Now I solve for 1 and 2 networks with my brain - by composing predictors, and the third task is handled by CatBoost. It would be difficult to combine these networks into one, maybe to try to work with each direction separately, and then to combine these networks?
Now I solve for 1 and 2 networks with my brain - making up predictors, and the third task is handled by CatBoost. It will be difficult to combine these networks into one, maybe try to work with each direction separately, and then combine these networks?
It is necessary to watch for innovations, they are constantly improving. Modern grids have exactly the same task, to do everything at once.
Searching for predictors manually is the last century, like a pickaxe on a rock. And, as everyone has seen, it almost doesn't work.You have to watch the innovations, they are constantly improving. Modern grids have exactly the same task, to do everything at once.
Searching for predictors manually is like using a pickaxe on a rock. And, as everyone is convinced, it almost doesn't work.It has to be a very complex architecture to do everything at once, and the more complex the architecture, the more processing power is needed.
However, if there is a need in capacities (there are old servers and GPUs), I am ready to provide them for the idea ;)
The more complex the architecture, the more processing power is needed.
However, if there is a need for power (there are old servers and GPUs), then I'm ready to provide them for the idea ;)
not complicated, you just need to understand
power is not needed at all. I have LSTM on my laptop learning in a few minutes without any graphics cards. About power is a myth.
The more complex the architecture, the more processing power is needed.
However, if there is a need in capacities (there are old servers and GPUs), then I am ready to provide them for the idea ;).
Ready to float an idea, in person?
not complicated, you just have to figure it out
You don't need any power at all. I have LSTM on my laptop learning in a few minutes without any video cards. It's a myth about power.
Not complex architectures don't work, as you said above. Uncomplicated ones need stationarity... cycles.
Wow, a couple of minutes is cool, and what network topology is it, how many layers, neurons?
Ready to voice ideas, in person?
You can do that in person.
Non-complex architectures do not work, you said it yourself. Uncomplicated ones need stationarity... cycles.
Wow, a couple of minutes is cool, and what network topology is that, how many layers, neurons?
Man... it's not complicated in the sense that you can understand
Usually a couple of layers are enough, you don't need much depth in forex
There are more advanced networks for BP, cooler than lstm. There may be profits from there, haven't checked yet. All "classics" like boostings and perseptrons are not suitable for BP.