NS + indicators. Experiment. - page 5
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 don't have this understanding and still don't have it. Prival, I don't accept your hypothesis that it's some kind of a pre-interval (do you remember the 200 point spike on the cable, which is made with one tick?). No yu can predict it, but Fibami... I think it's quite likely...
Why, Mathemat, should a neural network be able to predict such studs? And in general, any TC, should do such a thing? That means that your requirement for a TS is to predict all ticks accurately in time and magnitude. Is that too much to ask?
to klot
I would really like to help you, as I unscrupulously use your FFT library and haven't even sent a bottle of champagne for the New Year. And to many others, who are trying to build a neural network based trading system (TS), I'd like to help too.
I would like to comment on some of my comments in this thread. Maybe my knowledge is not so outdated and may be useful. Somewhere in 1987 I was doing research on creating an algorithm that recognizes tanks. At that time there were no such beautiful software packages, and the 286 computer was a distant blessing. So the ideas, which I now see in the National Assembly, were investigated there.
One of the most striking is Fain Reader, what could really come out of those algorithms in such tasks they worked quite well. That is, at the input of the algorithm a photo (scanned text) image was divided into clusters, the edges of these spots were smoothed and the image sharpness was increased, then by the maximum correlation integral this spot was assigned to a class of tank, BMP, APC, KamAZ, etc. At Fain Reader it is the letter a, b, c, etc.
Since that time I've got the attitude towards these algorithms: "In fact, just trying different indicators is the same as trying all known combinations of indicators and see what happens. Everything is determined by the architecture, by what's inside. And the correct input is only one, it's a stream of quotes. Everything else is a transformation of this stream.
An algorithm cannot be universal for everything at once. It has to be tailored to perform a specific task. And it has one input, a photo (scanned sheet). Try to twitch the sheet during scanning and the algorithm crumbles like a house of cards. These algorithms work well with stationary objects and recognise them, i.e. assign them to a certain class a,b,c, etc.
They may well be able to detect (recognize) the state we are in now (a, b, c, etc.) but they cannot predict. It (the algorithm) does not know what I will do in the next minute with this sheet.
How this all manifests itself in forex. NS can recognize the state we are in now and will not tell us what will happen in 5 minutes. When transforming the whole data flow it finds peaks (bottoms), builds a line, waits till the price comes to this line again (recognizes the situation) and decides to play on breakout of this line or on a rebound. In the same way I think NS may well recognize that we are at a certain point (a.b.c. ...), but won't tell us what to do next.
It takes a long time to explain everything. In my opinion, the output of NS should be like this 'Random Flow Theory and FOREX'.
And if you still want to use the NS, then give it at the input of the derivative of this indicator 'Optimized version of Kaufman's adaptive moving average AMA from wellx' and I think they need a lot (should differ periodAMA nfast nslow G) until the 'curse of dimensionality' will not kill (so as not to go through all in a row, try to choose the most uncorrelated). Going back to the tanks, I do not care where it is now, I need to know where it will be when the projectile reaches it (where quotes will go up or down from the point of entry into the market). Therefore, you need to analyse the velocity vector as it behaves.
I think many of you have already come to this conclusion that it is necessary to apply the MA and the zigzag for entry. It seems to me that the MA takes the derivative of the MA and the zigzag is points where the velocity vector changes its direction. Therefore the derivative of this indicator should work.
So it goes like this. Maybe, I have misled you, and my knowledge is as outdated as that of a dinosaur. And I am mistaken that the input alone is a stream of quotes (a sheet of paper, a photo) and it should be of good quality. And using NS to recognise one class is nonsense.
But I am sincerely trying to help.
a little bit on the side of tanks. when i was a third year student, i remember doing image recognition, applying various Kuwahara filters, vectorization and pr..... and then ns. then i found a solution. the task was to recognize faces on the image. and since it was stupid to send the whole picture to the network, i limited myself to a window. and the same neural network could resize the window and move it along two axes. It looked very interesting, the whole thing at work. Life is movement. Artificial life, you know. )) Later, more advanced algorithms came to mind. But never reached the point of implementation.
Next day I will try to implement something based on filters with NS. How is that. mmm... Yes.
(x-ma(x,n))/(3*stdev(x,n)), lately I always use this formula. And, actually, go ahead for training, cross-checking and OOS. .
I see, thank you. And I took a look at your example. It turns out the normalisation to form -1 +1. I will try to experiment with your version.
I forgot to ask one more question. I understood that you are using NeuroShell DayTrader. Why are you not satisfied with NeuroShell2? I'm asking because I have NS2 version 4.0, and I'm not interested in other similar packages. May be I am mistaken? What do you personally like about DayTrader?
klot, there is such a suggestion. You only use price differences. That is, you only take into account the value of the price. Try using time intervals as well. Most indicators only work with price. And not only indicators, but most traders use price changes in the market. But price is very much related to time. The available versions of indicators for pattern search (not only Gartley) also consider mainly only price. In relation to ZZ we can propose the use of such a parameter. The number of bars during which the ZZ ray was built.
It is possible to use the Fibo Time tool. I will try to show charts using Fibo Time on Onyx in the nearest future. Closer to the New Year or after the New Year.
I agree, a confidence interval in the actual meaning of the term does not fit here. But I think Prival was referring more to an analogy. After all, indeed, the price value (why necessarily Close ?) is in this interval. But the mo here changes, imho, so there is no gain.
And getting a true confidence interval at least on the current bar, let alone the future, would be more than cool. In fact, it would be a ready-made solution to the problem of building a strategy.
I agree, a confidence interval in the actual meaning of the term does not fit here. But I think Prival was referring more to an analogy. After all, indeed, the price value (why necessarily Close ?) is in this interval. But the mo here changes, imho, so there is no gain.
And getting a true confidence interval at least on the current bar, let alone the future, would be more than cool. In fact, it would be a ready-made solution to the problem of building a strategy.
Well, it seems that my thoughts are starting to converge. Now please look from the same point of view (confidence interval) here 'Need an indicator reflecting the price in operating time'.
Before looking anywhere else, it is worth resolving the basic issues after all. In my opinion, there are two.
1. The confidence interval is, after all, an analogy. The value of the confidence interval for the price values on the given bar is not known a priori. We can make some forecasts concerning this, but it is already an element of TS and therefore must be substantiated at least ideologically. Where are they? The value of High-Low changes on every bar. So any prediction of this "confidence interval" can only be made statistically. Based on what statistics ? What are its properties ? Perhaps these statistics can be linked to the local volatility which one? Where does it come from? How to describe its dynamics? How to determine the High-Low from it? There are classical ideas, but may be there are more fruitful ideas?
2. The High-Low value on a future bar is only half of it. If there is no possibility to predict the dynamics of mo at least with the same level of reliability, then it is all vanity and catching the wind. Hence the question: how to predict mo dynamics ?