Discussion of article "Practical application of neural networks in trading (Part 2). Computer vision" - page 2
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
Converged networks are not suitable for trading.
In trading it is necessary to determine whether to buy or sell now, i.e. at the 0th bar or at the rightmost border of the picture.
The convolutional network is not tied to the 0 bar. It answers the question - is there a certain pattern (for example, a kitten) in the picture. In any place, not pinned to the right side. And as long as there is a kitty in the picture (on the right, in the middle, on the left) it will signal that it is there.
For example she found a pattern for Buy, at the 50 bar and at the 100 bar. Obviously, at 50 bar and at 100 bar it is too late to open a Buy trade. The rare times when the Buy pattern will be on the right, the answer will be right to trade right now.
Obviously the patterns for Sell will be in every (or almost every) picture as well. For example at 0 bar Buy, and 50 bars ago you could Sell. And 80 bars ago also Buy, etc. I.e. each picture will most likely contain several Buy and several Sell patterns. Averaging them, the forecast probabilities will be about 50% +-10%. I.e. the answer would be deciphered as in this picture there are 3 places for Buy and 4 places for Sell. The probability for Sell = 4/7 = 57%. But it cannot be taken as a decision to open a deal right now on the 0th bar.
But in general, the fact of working with pictures is interesting.Converged networks are not suitable for trading.
In trading it is necessary to determine whether to buy or sell now, i.e. at the 0th bar or at the rightmost border of the picture.
The convolutional network is not tied to the 0 bar. It answers the question - is there a certain pattern (for example, a kitten) in the picture. In any place, not pinned to the right side. And as long as there is a kitty in the picture (on the right, in the middle, on the left) it will signal that it is there.
For example she found a pattern for Buy, at the 50 bar and at the 100 bar. Obviously, at 50 bar and at 100 bar it is too late to open a Buy trade. The rare times when the Buy pattern will be on the right, the answer will be right to trade right now.
Obviously the patterns for Sell will be in every (or almost every) picture as well. For example at 0 bar Buy, and 50 bars ago you could Sell. And 80 bars ago also Buy, etc. I.e. each picture will most likely contain several Buy and several Sell patterns. By averaging them, the forecast probabilities will be about 50%.
For this purpose, convolution and lstm are combined
For this purpose, convolution and lstm are combined
For example, the combination is sequential: trend up-> reversal pattern-> trend down-> reversal pattern-> buy signal.
and without lstm you can do it in an interesting way by building a special architecture so that reversal patterns are searched on the edges of the window and the trend in the middle.
For example, the combination is sequential: trend up-> reversal pattern-> trend down-> reversal pattern-> buy signal
and without lstm you can do it in an interesting way by building a special architecture so that reversal patterns are searched on the edges of the window and the trend in the middle.
Not considering that I think that teaching a time series through pictures is a high-tech overkill, you can probably add additional symbols to the pictures, specifying what you want from a given picture.
Very good article! Thank you!
I am currently working in and try to get the maximum performance out of it. Should I get meaningful results, I will report about it.
Interesting work.
When saving/displaying pictures it uses, as I understand, autoscaling enabled in the terminal, which destroys the information about motion force. Maybe make a single size - to a set number of points in height?
Folks, does anyone have sources of trading robot on python?
Interested in the fish itself, and thinking will be my neuronka with reinforcements....
Just lazy to write everything from scratch =)