topic Re: dynamic regression model in SAS Forecasting and Econometrics

dynamic regression model

AndreaRossi — Wed, 23 Nov 2016 11:43:52 GMT

Hello

How to implement dynamic regression model like

(1 - u1*B-u2*B^2)*(1-u24*B^24)*(1-u168*B^168)*Yt = (vo+v1*B+v2*B^2)*Xt + Et

with Et WN?

I have just read the proc arima user guide...And I don't know to transform my model in a transfer function model....

please help me!

dyanmic regression

AndreaRossi — Wed, 23 Nov 2016 10:06:59 GMT

Hello

How to implement dynamic regression model like

(1 - u1*B-u2*B^2)*(1-u24*B^24)*(1-u168*B^168)*Yt = (vo+v1*B+v2*B^2)*Xt + Et

with Et WN?

Re: dyanmic regression

Ksharp — Wed, 23 Nov 2016 11:23:36 GMT

If it is about Time Series Analysis, plz post it at Forecasting Forum.

Proc arima can do dynamic regression.

Re: dyanmic regression

AndreaRossi — Wed, 23 Nov 2016 11:40:35 GMT

Ok..I post there my msg.

Re: dyanmic regression

AndreaRossi — Wed, 23 Nov 2016 11:47:50 GMT

But I have just read the proc arima user's guide.....It declare to write in a transfer function form..In my case is not ok!

Re: dynamic regression model

alexchien — Wed, 23 Nov 2016 15:34:44 GMT

you could try the following estimate statement in proc arima:

estimate p=((1,2)(24)(168)) input((1,2)x);

Re: dynamic regression model

AndreaRossi — Wed, 23 Nov 2016 16:11:53 GMT

thanks for the reply

I think, in this case, to write Yt = (v0 + v1*B + v2*B^2)*Xt + 1/((1-u1*B - u2*B^2)*(1-u24*B^24)*(1-u168*B^168))*Et

right or not?

Re: dynamic regression model

alexchien — Wed, 23 Nov 2016 20:43:40 GMT

i think you need to divide the backshift operator for Xt as well..

Yt = (v0 + v1*B + v2*B^2)/((1-u1*B - u2*B^2)*(1-u24*B^24)*(1-u168*B^168))*Xt + 1/((1-u1*B - u2*B^2)*(1-u24*B^24)*(1-u168*B^168))*Et

Re: dynamic regression model

AndreaRossi — Thu, 24 Nov 2016 14:31:50 GMT

Yes i think so! there is a way to impose the two denominator to be equal?

Re: dynamic regression model

AndreaRossi — Fri, 25 Nov 2016 16:26:29 GMT

Anyone have others idea?

Re: dynamic regression model

alexchien — Mon, 28 Nov 2016 14:16:02 GMT

Sorry, what do you mean by "impose the two denominator to be equal?"? One more thing, you need to add an NOINT to the ESTIMATE statement as there is no intercept term in ur model.

Re: dynamic regression model

AndreaRossi — Mon, 28 Nov 2016 15:17:07 GMT

Thank for the reply:

If I write my ARX like this

Yt = (v0 + v1*B + v2*B^2)/((1-u1*B - u2*B^2)*(1-u24*B^24)*(1-u168*B^168))*Xt + 1/((1-u1*B - u2*B^2)*(1-u24*B^24)*(1-u168*B^168))*Et

(whit NOINT option) the parameters of the two denominator are estimated different:

Yt = (v0 + v1*B + v2*B^2)/((1-u1a*B - u2b*B^2)*(1-u24a*B^24)*(1-u168a*B^168))*Xt + 1/((1-u1b*B - u2b*B^2)*(1-u24b*B^24)*(1-u168b*B^168))*Et

whit uia != uib

Re: dynamic regression model

AndreaRossi — Mon, 28 Nov 2016 15:18:07 GMT

but in my model uia ==uib

Re: dynamic regression model

rselukar — Mon, 28 Nov 2016 16:16:30 GMT

You can specify Y = (numerator poly) / (denominator poly) X + E / (AR Poly) type model for your polynomials as follows:

proc arima data=test;

identify var=y crosscorr=x;

estimate p=(1 2)(24)(168)

input=((1 2)/(1 2)(24)(168) x)

noint;

run;

It will indeed happen that the estimated denominator polynomial coefficients in the transfer function will generally be different from the estimated AR polynomial coefficients. Currently PROC ARIMA does not allow constraining these coefficients to be the same.

You seem to be dealing with hourly data and are trying to capture hour of the day and hour of the week seasonal patterns. Is this particular model very important for you or some other model might do that still does a good job?

Re: dynamic regression model

AndreaRossi — Mon, 28 Nov 2016 17:18:11 GMT

thank for the reply rselukar

you have grasped fully my problem. It is a model of electricity price (Y) related to demand (X). I need to re-execute examples of a book for my thesis. I am searching others model too.

Re: dynamic regression model

AndreaRossi — Tue, 29 Nov 2016 17:29:51 GMT

other suggestion?