3

I've fitted a model with auto arima, with independent variables with the below codes:

Regressors<- cbind(GDPr_chg,r)
arima1 <-Arima(RVD_all_chg, order=c(2,0,0),seasonal=c(0,0,1), xreg=Regressors)

Turned out the model results are as below:

ARIMA(2,0,0)(0,0,1)[4] with non-zero mean
Coefficients:
        ar1      ar2     sma1  intercept  GDPr_chg        r
      1.5285  -0.5668  -0.7920     0.0990    0.4459  -0.7613
s.e.  0.0766   0.0744   0.0724     0.0274    0.1948   0.3437

I know that xreg will fit an ARIMA model for the errors, so the model should be:

y = 0.099 + 0.4459*GDPr_chg -0.7613*r + 1.5285*u(t-1) -0.5668*u(t-2)-0.7920*e(t-4) + "random error

But I couldnt obtain the same fitted values by using the above formula. Can anyone give me a hint?

Also, how does R fit the error terms for the first few forecasts when u(t-1) and u(t-2) are still not available [as u(t) should be the regression residual in time=t, right?]

Great thanks in advance everyone.

Heywood
  • 71
  • 1
  • 6
  • 1
    This will not answer your question, but still is a great resource to double-check if you got things right; it is Rob J Hyndman's blog post ["The ARIMAX model muddle"](http://robjhyndman.com/hyndsight/arimax/). – Richard Hardy Mar 10 '15 at 10:03
  • Your AR polynomial is close to being invertible you may be using a poor ARIMA identification scheme like AIC or BIC . It appears that your MA structure is redundant to your AR structure. – IrishStat Mar 10 '15 at 12:22
  • For your 2nd question about error terms, you may want to see the answer given in [here](http://stats.stackexchange.com/questions/98989/fitted-values-of-arma-model/99004#99004). – Stat Mar 10 '15 at 12:25
  • Once I remove the Seasonal MA term, the residuals appeared to be seasonally related, cannot pass the port test... – Heywood Mar 11 '15 at 02:29

1 Answers1

5

The autoregressive coefficients are weights for the observed series $y_t$ not for the residuals $u_t$. The correct representation of the model for the output that you show is:

\begin{eqnarray} \begin{array}{ll} \eta_t = y_t - 0.099 - 0.4459\, \hbox{GDPr_chg}_t + 0.7613 \, \hbox{r}_t \\ y_t = 0.099 + 1.5285\,(y_{t-1} - \eta_{t-1}) - 0.5668\,(y_{t-2} - \eta_{t-2}) + u_t - 0.7920\,u_{t-4} \end{array} \end{eqnarray}

As the model contains a MA part, there isn't a closed-form expression expression to obtain the fitted values. An iterative procedure is required. stats::arima uses the Kalman filter. See also the post linked by @Stat in the comments above.

For illustration of a more simple case with no MA part, you may be interested in this post and this post. The latter gives a numerical example to obtain out-of-sample forecasts in an AR model with a external regressor (which is the same as for the fitted values, observed values are replaced by forecasts).

Note: there are other definitions of the ARIMA model. The signs of the coefficients and the definition of the mean/intercept may vary depending on the software implementation you are using. This answer and related posts linked here are based on the R function stats::arima.

Community
  • 1
javlacalle
  • 11,184
  • 27
  • 53
  • Sorry , didnt know i would cancel the question when I put a green tick there @IrishStat – Heywood Mar 11 '15 at 01:40
  • I tried the above calculation, seems the "nt" in the 1st equation is simply finding the regression residual (actual y - modelled y); but if this is ture, then in the 2nd equation, the difference of (y-n) is simply the y minus the residuals obtained in the first equation. actual y minus regression residuals = modelled y again... am i right? – Heywood Mar 11 '15 at 01:50
  • I have fixed the index in the seasonal moving-agerage, $u_{t-4}$. – javlacalle Mar 11 '15 at 08:01
  • The first equation is not the residuals of the model, it is the observed series minus the deterministic terms. In a model with no external regressors it would simply be the series $y$ minus its mean. The example in the second link that I gave may be helpful to see this. – javlacalle Mar 11 '15 at 08:05