Testing real-time forecasting systems - page 27
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
Sorry :-) Here's the prediction at the beginning:
gpwr apparently cooks Fourier differently, my method has no parameters.
:о)))
And how is GRNN different from Fourier? And what is this GRNN anyway? I just don't know. :о(
No, no, no,
I'm posting a forecast at the beginning, not showing it at the end.
Also, if there are no parameters, how does gpwr get another line?
GRNN can be coded in different ways. I used the simplest code with fixed sigma (cluster size). The length of the pattern is another parameter. I optimized it using past data and got 140 bars. Prices without smoothing were used as input. By the way, the 3rd method (non-linear autoregression) gave similar results.
According to the new data these two methods give the following predictions
GRNN:
non-linear AR:
grasn, GRNN is a type of nerve net. Here's a link to the definition. Or here 's a bit more clever.
Yes, yes - already figured that out, thanks...for participating :o))))))
Addendum: I was confused by the phrase "lack of parameters". Any NS has "parameters always".
:о)))
How is GRNN different from Fourier? And what is this GRNN anyway? I just don't know. :о(
Mathematically, GRNN (general regression neural network) is most simple but very efficient neural network which was proposed by Specht in 1991. See link here
http://people.cecs.ucf.edu/georgiopoulos/eel6812/papers/general_regression_network.pdf
It has nothing to do with Fourier. GRNN refers to probabilistic neural networks such as nearest neighbours. It takes all past patterns and calculates Euclidean distance from current to past patterns this way
D[n] = SUM( (Open[i] - Open[n+i])^2, i=0...PatternLength )
A weighted average type prediction from past "future" prices is then calculated
Open[-1] = SUM( Open[n-1]*exp(D[n]/(2*Sigma), n=0...AllPastPatterns) / SUM( exp(D[n]/(2*Sigma)), n=0...AllPastPatterns)
In nearest-neighbors, after calculating Euclidean distances to past patterns, the closest pattern is selected and its "future" values are used as predictions for the current pattern. This is in the simple version, which is rarely used. Usually the nearest neighbours are found and their "future" values are averaged or weighted to find predictions for the current pattern.
Yes, I understand,
Just to beclear, the question was posed to neoclassic - just to remind you of the content of his post:
gpwr видимо, по другому готовит Фурье, у моего метода нет параметров
That's what I asked neoclassic:o))))) What is the difference, because his forecast a bit like Fourier's, remotely so.
to gpwr
thanks for the synopsis.
to Mathemat
I already said thanks but I'm always ready to repeat it :o)))
Grasn, see 'Dynamic extrapolator based on Fourier transforms'.
Here you can see the working principle and the indicator itself :-)
Grasn, see 'Dynamic extrapolator based on Fourier transforms'
Here you can see the principle of work and the indicator itself :-)
and then:
It looks like GRNN has hit the jackpot :-)
Or have you decided that the day will actually be a lost day if you don't confuse me unnecessarily? :о)))))
In no way meant to confuse you :-)
Похоже GRNN сорвала куш
I said I meant that the gpwr GRNN forecast was the most accurate, and my picture was just a follow-up.
In no way meant to confuse you :-)
I said, meaning that the gpwr GRNN forecast turned out to be the most accurate, and I gave my picture just to develop the topic.
>> all clear :o))))))
PS: Except for one thing - GRNN gave just one of the worst predictions. But that's mine, IMHO. I mean, it's obvious.