On Wed, 15 Sep 2010, Allin Cottrell wrote: > On Wed, 15 Sep 2010, Mike Pfeiff wrote: > > > Is there any obvious reason why using the ARIMA feature (without any AR > > or MA terms) with a seasonal difference=1 would yield different > > coefficients that using OLS and manually seasonally differencing the > > independent and dependent variables? > > No. Can you post an example script that shows differences in > results? We have some test cases on file where we get identical > results via the two methods, but maybe we're overlooking > something...
Sorry, what I said there was not correct. The test cases I had in mind, where we got identical results, were in fact comparisons of (a) ARIMA estimation and (b) plain ARMA estimation, with the data manually differenced first. As for "ARIMA" with no AR or MA terms versus OLS on differenced data, you would expect the results to be close but not necessarily identical, because the calculation method is quite different (using the Kalman filter and BFGS for ARIMA). Or I should say "was quite different", because I've now changed this in CVS: if we detect that an ARIMA model has no ARMA content (and no missing values) we now just apply OLS (= MLE), there being no point in using heavy artillery when a pea-shooter will do the job. Allin Cottrell
