Artificial neural networks. - page 3
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
The network is a tool, a kind of universal non-linear function that can be optimised (fitted) for different data (input-output). This function cannot extract any regularities.
It does :)
NS is trained on input-output data and if the patterns are universal for the general population, then the NS will work successfully outside the sample.
And isolation of regularities is an additional layer of work to analyze trained weights and transfer functions of NS, I myself have never done it personally.
The human brain is capable of learning non-linear patterns very successfully. Take, for example, calculating the trajectory of a boomerang launched by bushmen tribes somewhere in Africa. Nature has given us everything. You can purposely learn non-linear patterns in the marketplace. I personally would like to write a VBA learning program of the following kind: I will display a series of small quotes, say, 100-150 bars on a chart and I will need to guess the general direction of price in future, say, for 50 bars. I will press the up/down buttons and the program will record my answer and whether I guessed or not. After that, the chart will randomly shift on the time axis, etc. As a result, I will either learn (percentage of guessed directions will increase), or not (like, sorry, failed). This would be a good example of the brain's trainability.
The network is a tool, a kind of universal non-linear function that can be optimised (fitted) for different data (input-output). This function cannot pick up any regularities. You may as well argue that a polynomial fitted to a smooth curve finds its hidden patterns. Step out of the data range to which the function has been fitted and you know what you will get as an example of polynomial regression. By the way, number of parameters to be optimized in the network is much larger than in the decision tree as the network contains sub-connections that do not influence the correct decisions and whose weights are decreased during optimization.
It is one thing to approximate market data by a network, and quite another thing to do pattern recognition.
The second seems to be more correct - after all it is the trader's brain that recognizes images.
It is one thing to approximate market data by a network, and quite another to recognise images.
The second seems to be more correct - after all, it is the trader's brain that recognises images.
Engineers at Google fed the self-learning network (what type I don't know) screenshots from YouTube videos and the network learned to separate cats into a separate class. It is possible to estimate how much information was spinning around in the RAM. Theoretically, it is possible to send charts to the grid but it must be a huge complex network and a suitable computer. It is easier to send a price pattern normalized in the range of, say, [0;1]. And stationarity is preserved and it's easy to implement. Basically, the trader sees the price pattern, while there are traders who trade using a pure chart (without indicators). But apparently, the network has to be constantly retrained. Because our brain is also constantly updating connections and improving our understanding of the process.
When a trader "sees" a pattern, he or she does not analyze a pure price series. No human brain has such amount of RAM (by the way, it is a thinking RAM - a couple hundred bytes maximum). Therefore, signal preprocessing is mandatory.
I hadn't heard about that. OK. Well then of course it's hard to understand what aspects of CD are perceived by the brain. That's the problem with all the attempts to make something similar based on NS. The important signs of incoming information are not defined, so we have to guess.
But - personally I understand bends and spikes most of all :)
I hadn't heard about that. OK. Well then of course it's hard to understand what aspects of CD are perceived by the brain. That's the problem with all the attempts to make something similar based on NS. The important signs of incoming information are not defined, so we have to guess.
But - personally I understand bends and peaks most of all :)
I think it is very promising to clear price information from noise with the help of NS. Basically, a useful signal is described by a small number of parameters, for example, if we take a day on M1, there can be 10-20 (instead of 1440), which is quite easy for human brain to perceive. The question is how to correctly clear the signal without losing important information. This is where a network can help, IMHO. Well, and to process (to classify, cluster, regress, etc.) the rest part it is possible by any suitable method, and not the fact, by the way, that NS will be the best choice. In any case, there are many tools developed for such tasks, and not the least reason is that all of them work equally badly)))
Another thing about information actually going to the brain. A very instructive example is fractal compression of images. It just shows that actually visually to a person for perception there is enough information in thousands and even tens of thousands times less than "raw" size of image.
I wonder what ratio can be achieved by compressing a quotient using the fractal algorithm? 100? 500? more?
Alexey, you're talking directly about the Feature Selection problem, in plain language: how to select the 1% of data from the array that is most informative for a forecast like: enter long, enter short, wait. But even when applying sophisticated techniques to this problem, the pitfalls come out quickly. First, it's not clear what sauce to use to feed the selected data points: raw values are not appropriate, you need to preprocess. Maybe it would work if you took the difference between 457 and 891 bars, and between 1401 and 1300. There are a lot of variants, and I don't have enough time to bruteforce them all.
There is also the problem of interpreting the output of the algorithm. We often try to build the same network so that the output has a clear signal of what to do, or at least information that is relatively understandable (to the designer) to be converted into such a signal. But it's not certain, that it is convenient to output information in such form, maybe it would be much easier to give us not roughly one and a half bit (by-sell-stop), but, for example, 10 bits of information?