ARIMA issue

[noetsi]

After deciding you had no non-stationarity I ran this model suggested by MINIC and SCAN (a ma and ar of 1)

proc arima data= tsdata;
identify var=spend scan esacf minic;
ESTIMATE p=(1) q=(1);
run;
ods graphics off;

when I ran this I got the following warning which I have never encountered in ARIMA.

Warning: The model defined by the new estimates is unstable. The iteration process has been terminated.

Warning: Estimates may not have converged.

Any suggestions? I have 50 data points if that matters.

 

[hlsmith]

No idea, not my area - but perhaps if you want to force it to run you can change the convergence criteria. Results may be dubious, but you can see what it came up with.

 

[noetsi]

Yeah. I have never run into this.

How come no time series people around this board

Ok this is likely the issue.

"This warning typically indicates that one or more of the model parameter estimates is approaching the invertibility or stationarity boundary. You may want to look at
your table of parameter estimates to see if any of them are approaching 1 in absolute value. If they are, then that is an indication of model misspecification. If you have an AR parameter that is nearly 1, then that may be an indication that you need to difference your data. If you have an MA parameter that is nearly 1, then you may have over-differenced your data.
If you think that your model is correctly specified, then you can try adding the NOSTABLE option to the ESTIMATE statement. This allows the procedure's optimization algorithm to iterate outside of the stationarity and invertibility region. In some cases, the final parameter estimates will satisfy the stationarity and invertibility conditions. In this case, you may be able to use the model for forecasting. However, if the NOSTABLE option is used and the final estimates fall outside the stationarity and invertibility region, then the forecasts from the resulting model could become explosive.

I looked at the formal tests, the data etc and decided the data was stationary. But sas does not agree. Strangely getting rid of non-stationarity did not eliminate the problem. I had to also specify an AR 1 term which makes no sense to me.