Re: How to interpret coefficients from ARMA in R (r-project.org)?

From: *** Startz (richardstartz1_at_comcast.net)
Date: 06/12/04 

Date: Sat, 12 Jun 2004 20:40:48 GMT

The problem is that you are estimating an ARIMA(0,1,0) model as
ARMA(2,2). The model you are estimating is equivalent to

(y-y(-1)) = epsilon.

This model can also be represented as an ARMA(2,2) where the first AR
coefficient is 1+alpha, the second AR and the first MA coefficient is
minus alpha and the second MA coefficient equals zero, for arbitrary
alpha.

By the way, when the AR coefficients add to one the model is said to
be nonstationary, not non-invertible.

Out of curiosity, is the original data an asset price?

-*** Startz

On 12 Jun 2004 09:56:58 -0700, dave@autobox.com (David Reilly) wrote:

>dakshing64@yahoo.com (Bush will disarm all workers next) wrote in message news:<f26c87f0.0406110727.67535f6@posting.google.com>...
>> Ran summary.arma(arma(inp, order=c(2,2))) on a timeseries data set and
>> generated these results:
>>
>>
>> Call:
>> arma(x = inp, order = c(2, 2))
>>
>> Model:
>> ARMA(2,2)
>>
>> Residuals:
>> Min 1Q Median 3Q Max
>> -0.40833 -0.37282 -0.33614 -0.06072 60.63911
>>
>> Coefficient(s):
>> Estimate Std. Error t value Pr(>|t|)
>> ar1 1.9484320 NA NA NA
>> ar2 -0.9486320 NA NA NA
>> ma1 -0.9577766 NA NA NA
>> ma2 -0.0009986 0.0503448 -0.02 0.984
>> intercept 0.0382860 NA NA NA
>>
>> Fit:
>> sigma^2 estimated as 10.01, Conditional Sum-of-Squares = 3945.79,
>> AIC = 2051.33
>>
>> How are the coefficients interpreted? Is there a way to
>> predict/forecast?
>>
>> Thanks
>> Dakshin
>
>
>Dakshin,
>
>There is no need to interpret thos model as it is non-invertible and
>thus an invalid model
>
>
>By inspection the two ar coefficients when added together are just
>about 1.0 which indicates non-invertibilty.
>
> ar1 1.9484320 NA NA NA
>> ar2 -0.9486320 NA NA NA
>
>
>Older programs just simply fitting senseless AR/MA structures are to
>be avoided like the plague. Programs that simply fit higher and higher
>orders that don't test for invertibility are to be studiously avoided.
>
>Perhaps you might want to get some freeware , which actually verbally
>describes the equation ....http://www.autobox.com/freef.exe ..The
>program is called FreeFore . It is so much fun to play with that many
>users have suggested the name ForePlay ! (LOL !)
>
>
>hope this helps ..
>
>David P. Reilly
>AUTOMATIC FORECASTING SYSTEMS
>
>215-675-0652
>
>P.S. For some details on FreeFore click on
>http://www.autobox.com/ffpro.doc
>
>for other info please see
>http://www.autobox.com/pdfs/whytheyswitched.pdf

----------------------
Richard Startz RichardStartz1@comcast.net
Lundberg Startz Associates