Converging an AR-EGARCH model

[Nikke-]

Hi there!

I have a bit of an issue with getting my rather complex model to converge. First I estimated a model with OLS and HAC no problem there, but as the residuals were both autocorrelated and heteroskedastic (no real surprise there, I just wanted to go through the model building process) I estimated an ar(1), ar(2) process for the error term, which seemed to get rid of the autocorrelation quite nicely. On top of this I have an EGARCH specification with some variance regressors added. As I have understood basically the point estimates from my final model should equal those from the initial OLS and just the standard errors should improve. I have tried the various estimation options (as I understand it the goal is the maximize the log-likelihood value) but they do not seem to converge well by themselves so I have resorted to user supplied starting values and worked from there through trial and error. Should I basically get (close) to the initial OLS-estimates? Or is it acceptable to reach somewhat different values if they have economic interpretations and residuals are well behaving?

Thanks in advance and hope someone figured out what I was going after

[Nikke-]

ok well I did some further research on this and it seems to me that because the ar(1) and ar(2) refer to the Cochrane-Orcutt transformation for serial correction in the error term it is possible to get different results than from an initial OLS without the correction. I actually managed to ind a thread here on eviews forums about this and it seems to me that I´m doing it correctly. The only thing is: should I just go on with finding the maximum log likelihood simply by changing the starting values. It seems to me judging by the literature, that there is a possibility for multiple maximums and therefore I should proceed with experimenting. And on finding the correct "lag" I should continue applying AR-processes untill the rho values do not differ from 0 anymore... and obviously that, the autocorrelation is gone.