hi !

i said MA (or ARMA) will be better to model colored noise because this is 
their purpose (large band spectrum modeling).
AR model perform less well for large band spectrum than MA.
ARMA try to model in the same time : sharp frequency peaks and large band 
frequencies.

As e(n) is not a true noise (because not a true noise, mean not egal to 0 
etc...), maybe yes it can work with AR.
You have to check with the classic FFT (or Short-FFT) if AR or MA will be 
better before using the model.
If you can model e(n) with another AR modeling, so that is that previous AR 
modeling on the data didn't make good his job (ar modeling of Y), no ?

I think you have to check :
- spectrum FFT of e(n)
- error variance or the modeling AR of e(n) (should be not signifinat 
compared to e(n)... if it is, this is that e(n) cannot be modeled by AR).

cheers,
mich.



----- Original Message -----
From: Paul Ho
To: [email protected]
Sent: Friday, December 01, 2006 5:56 AM
Subject: RE: [amibroker-ts] confidence level asymptoe of Predicted AR


Ok just being a bit more rigorous this morning

Let e0(n) = y1(n) – y0(n) y1(n) is the first approximation and y0(n) is the 
true value;

* e1(n+1) = y2(n+1) – y0(n+1) e2(n+2) = y3(n+2) – y0(n+2)

* y1(n) = sum[ y(n-k).a(k)]

* y2(n+1) = y1(n). a(1) + sum[y(n-k).a(k)] k = 2….m

* y3(n+2) = y2(n+1).a(1) + y1(n).a(2) + sum[y(n-k).a(k)] k = 3….m

* e1(n+1) = e0(n+1) + e0(n).a(1)

* e2(n+2) = e1(n+2) + e1(n+1).a(1) + e0(n).a(2)

* e3(n+3) = e2(n+3) + e2(n+2).a(1) + e1(n+1).a(2) + e0(n).a(3)

* ek(n+k) = ek-1(n+k) + sum[e(k-j)(n+k-j). a(j)] j = 1…..k

Unfortunately e(k-1)(n+k) cannot be determined using recursive means and 
e*(k-1)(n+k) may provide a good enough fit but I suspect it may also 
increase the errors instead. Noise is not white as mean is not zero. By the 
way, what the difference between modelling AR of Y and AR of e using burg 
and ARMA which means to model the same thing.

Cheers

Paul.

_____

From: [email protected] [mailto:[EMAIL PROTECTED] On 
Behalf Of Tom Tom
Sent: Friday, 1 December 2006 7:39 AM
To: [email protected]
Subject: RE: [amibroker-ts] confidence level asymptoe of Predicted AR

You calcul are ok, i found same too : e'(n+1) = e(n+1) + Sum[ e(n-k) .a(k) ]
with your (2).
But how do you find :
(3): e'(n+1) = e(n+1) + Sum[ e’(n-k) .a(k) ]…… ? What are "……" ?

For the other calcul ok, if (3) is good, we can use : e*(n+1) = Sum[ e(n-k).
A(k)] .
But problem come from using AR to model white noise. It is in contradiction
because purpose of AR modeling is to do :
Y(n) = Sum[ Y(n-k).a(k) ] + e(n)
With e(n) a white noise (if AR model is good to model our data signal).
So if a good model for the white noise Y(n) is a model with all the coeff
a(k) = 0... and Y(n) = e(n).

But e(n+1) can be predicted from e(n) by AR if e(n) is not true white noise
(and i think it is the case for trading data, because they don't follow AR
model exactly...).
So if we check that e(n) is not white noise, not why use AR... ? Because
e(n) will be so like a colored noise... and AR is bad to model large
spectrum signal. AR is for sharp spectral signal (sinus ray...). Better to
model e(n) with an MA model for sure, or ARMA maybe.
Instead of those if you cannot so use a long AR (long AR = MA). Use an AR
model of high order.

Just my thought.
But i will look for some source about interval confidence of prediction. If
i find i will send them.

Cheers,
Mich.

----- Original Message -----
From: Paul Ho
To: amibroker-ts@ ts%40yahoogroups.com> yahoogroups.com
Sent: Thursday, November 30, 2006 3:58 PM
Subject: RE: [amibroker-ts] confidence level asymptoe of Predicted AR

Mich,

I think we are on the same page, except that I want to get from Y”n to Yn so
I define

(2) : Y''(n+1) = Y(n+1) + e'(n+1) = Sum[ Y'(n-k).a(k) ]

So the end result is

e'(n+1) = e(n+1) + Sum[ e(n-k) .a(k) ], but in the next recursion it will be

e'(n+1) = e(n+1) + Sum[ e’(n-k) .a(k) ]……

But e(n+1) is not known….. unless I do an AR on e(k) such that e*(n+1) =
Sum[ e(n-k). A(k)] and let a*(n+1) ~= a(n+1), but there are errors in there
also.

In the end, I want to have a standard deviation of e(k) and plot them as
bands. I’ll do some work and post results

In the meantime, it would be good if you can find those theories…..

Cheers

Paul.

_____

From: amibroker-ts@ ts%40yahoogroups.com> yahoogroups.com 
[mailto:amibroker-ts@ ts%40yahoogroups.com> yahoogroups.com] On
Behalf Of Tom Tom
Sent: Thursday, 30 November 2006 5:23 AM
To: amibroker-ts@ ts%40yahoogroups.com> yahoogroups.com
Subject: Re: [amibroker-ts] confidence level asymptoe of Predicted AR

Hi !

Hum... i am not sure but :

(1) : Y'(n+1) = Y(n+1) + e(n+1) = Sum[ Y(n-k).a(k) ]
(2) : Y''(n+1) = Y'(n+1) + e'(n+1) = Sum[ Y'(n-k).a(k) ]

If this is your equation / notation so :
>From (2), we have -> e'(n+1) = Y''(n+1) - Y'(n+1)
>From (1), we obtain -> e'(n+1) = Sum[ Y'(n-k).a(k) ] - Sum[ Y(n-k)a(k) ]
So -> e'(n+1) = Sum[ Y'(n-k).a(k) - Y(n-k).a(k) ]
So -> e'(n+1) = Sum[ ( Y'(n-k) - Y(n-k) ) .a(k) ]
>From (1) -> e'(n+1) = Sum[ e(n-k) .a(k) ]

There is a recursion to compute e'(n+1) from the k past values from e(n),
with the AR coeffs.
I cannot go further until the result you give (maybe he is good, maybe not,
it don't know)... But if you find a Sum[ e'(n-k).a(k) ] you sould have used
Y''' (and not only Y, Y' and Y'').
Someone can help maybe.
Paul can you send the calcul you made so we can have a look if it is ok ?

I think (and i have read somewhere but i don't know where... :( ) that for
testing accuracy of the forecast, algorythm use in/out know samples (back
horizon data is divided in 2 part). So it compute models from in know
samples, and calculate deviation from out know samples. And it predict the
accuracy from the deviation compute from the out know samples. Something
like that...

If i have time i will have a look on this where i find those theory...

Cheers,
Mich

----- Original Message -----
From: Paul Ho
To: amibroker-ts@ ts%40yahoogroups.com> yahoogroups.com
Sent: Wednesday, November 29, 2006 1:00 PM
Subject: RE: [amibroker-ts] confidence level asymptoe of Predicted AR

Pictures gone missing;(

The errors of each Yn such that Yfn+1 = Yn+1 + en+1 Yf is the
approximation of Y, and Yfn+1 = ‡” Yn-k . a k

The errors associated with Yhn+1 which is the second approximation using
Yfk is

efn+1 = en+1 + ‡” efn-k . a k

am I correct?

_____

From: amibroker-ts@ ts%40yahoogroups.com> yahoogroups.com
[mailto:amibroker-ts@ ts%40yahoogroups.com> yahoogroups.com] On
Behalf Of Paul Ho
Sent: Wednesday, 29 November 2006 12:17 AM
To: amibroker-ts@ ts%40yahoogroups.com> yahoogroups.com
Subject: [amibroker-ts] confidence level asymptoe of Predicted AR

Hi

I want to get the confidence level asymptoes of the predicted AR similarly
to that shown in the picture

Can anybody help?

[Non-text portions of this message have been removed]

[Non-text portions of this message have been removed]

__________________________________________________________
Windows Live Messenger sur i-mode™ : dialoguez avec vos amis depuis votre
mobile comme sur PC ! http://mobile.
<http://mobile. <http://mobile.live.fr/messenger/bouygues/> 
live.fr/messenger/bouygues/> live.fr/messenger/bouygues/

[Non-text portions of this message have been removed]

__________________________________________________________
Live.com: Le nouveau moteur de recherche par Messenger!
http://www.windowsl <http://www.windowslive.fr/livecom/> ive.fr/livecom/

[Non-text portions of this message have been removed]




_________________________________________________________________
Découvrez le blog Eragon sur Windows Live Spaces! 
http://eragon-heroic-fantasy.spaces.live.com/

Reply via email to