Market etiquette or good manners in a minefield - page 33
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 can do it with words...
I formulated a new reality paradigm not so long ago... -:) One that can neither be proved nor disproved.
Here it is:
Regularity is a way of existence of chance (i.e. a particular case). The converse, in general, is not true.
From here:
1. Perceived reality is a stream of randomness in the space of (local fluctuations of) regularities, or a stream of regularities (intentions of sentient entities) in the space of randomness.
2. Thus, the first claim is that reality (the universe), as a phenomenon of perception, contains neither randomness, nor regularity as such. Both are manifestations of a nature beyond human perception.
So, we can "zastruyat"...-:)
Here's some more thought on your g*th(w), where g=0.005
It's an interesting process... In fact, with this operator you're pulling all the weights into the vicinity of +/-0.005 zero. From where they will start to "run up" again in the learning process on the next cycle.
You get a kind of "learning impulse" that happens once per countdown.
Of course, there is no point in subjecting weights to such influence every epoch - the grid simply does not have time to learn properly, since one epoch is most likely not enough for normal learning even on a smooth vector (like that five-membered sine you gave me for the grid test). I propose to call the yuga the optimum number of epochs needed to train the grid. Since the yuga(N epochs) in your system happens once before each forecast, then at the end of each yuga(after weather forecast) it makes sense to try... "Continuity" of knowledge, supposedly, should be preserved and will be lost very gradually, since the new vector differs from the old one by only one datum.
And there is one more thought. It concerns the topic of the limits of the optimal range of weights. In my opinion, we should try +/-ln(D) as a limiting range:
Before(before starting conversations with you) I spent very many hours(and even days) racing single layer perseptrons in genetics. In the course of this fascinating but pointless exercise, I managed to notice that the weights in successfully sharpened models, rarely go beyond +/-(2.5 : 3.0), and the inputs in those perseptrons were at most 8. Then a second, or alternative, point for applying g*th(w), would be for one of the weights to reach the permissible range of +/-ln(D)
Here, Fedor, there is an important trick: If we have a trained NS on some input vectors, then applying operator th() to all its set weights does not destroy its knowledge, but only compresses the area in which its weights are defined. This is an important point that allows to get rid of the "saturation" effect, which saves the computational power of NS and exploits possible quasi-stationarity of market processes.
Regarding the rest of what you said - I need time to think about it.
I'm learning to use Matcad... a good tool. Sergey, I'm curious, how do you look at your grid results in Matcad? Do you draw charts?
And one more important question - how to cram quotes from MT4 into Matcad?
Well, yes - I make charts. Very handy!
As for exporting data to Matkad format, there's nothing easier. Go to your quotes archive and click the export button on the quote you need. Select the "ASCII Text (*.prn)" option from the context menu, specify the path to save the file, and you're done. This is a native Matkadian format. In Matcad you read from the file: Open=READPRN("FileName.prn")<2>. Two means (together with brackets in upper index of command, see dashboard), that you read only second column corresponding to opening prices on chosen TF (take minutes) from the file.
That's not what I meant...
Well, okay. If you're interested, let me know and I'll tell you all about it.
"Continuity" of knowledge is supposed to be preserved and will be lost very gradually, because the new vector differs from the old one by only one count.
Here's something that comes to mind on the subject.
Not so long ago I played with "exact" training, one single neuron, training vector with length equal to the number of neuron inputs (without constant offset - it was absent) P=w. I was doing it just for fun. It is clear, in this formulation, the grid can be trained as precisely as I wish on the training sample (number of adjustable parameters equal to the number of linear equations), so I don't bother with ORO, and get the exact values of weights by solving a system of linear algebraic equations via Newton's method in a fraction of a second. The effect is tremendous - we take a perseptron with 1000 inputs and in a second we get the values of weights, which for 1000 training vectors, each 1000 samples long, gives us the error of training "0"! Can you imagine? - A matrix of 1000x1000 and not a single error! It would seem that you add only one small element to this matrix - a new sample (try to predict one step ahead) and nothing extraordinary should happen. The grid will still show +1 or -1 well, maybe not guess in the extreme case... However, the result is discouraging. If this grid had hit the bull's-eye every time on 1000000 previous counts, here, right away - into Cosmos - instead of +/-1 - 4872365695. How about that! And you say "just one count"...
It's all a consequence of the perseptron's wild overlearning.
Glitches in the forum!!!
I'm learning to use Matcad... a good tool. Sergey, I'm curious, how do you look at your grid results in Matcad? Do you draw charts?
And one more important question - how to cram quotes from MT4 into Matkad?
Here is an example of how I work. I transfer data to Matkad. This is more convenient.
In the archive there is a script that sends information in a required format
(Time-Open-High-Low-Close-Hour-Min-Day-Month-Year-Day of the Week)
You just need to attach it to the needed chart and specify the date of the needed history (the history has to be already uploaded through the archive of quotations) .
We transfer the obtained files (GBPUSD_4.prn in this example and EURUSD_4.prn) to the directory where the Matcad file is located and work there.
If you carry out multi-currency analysis, don't forget about the holes. I showed you how to synchronize data in the matcad file.
Matcad version 14. Everything is in the archive.