[Gretl-users] acf and pacf output

I recently got an email from Alvaro Novo, who pointed out that the ACF and PACF figures produced by gretl's "corrgm" command were not in agreement with other packages. I have now tested this myself (against gnu R and Eviews), and Alvaro is right. I believe the gretl figures are asymptotically equivalent to those given by the other programs, but since the gretl figures are "non-standard" I have fixed this for the next release: in gretl 1.2.5 the ACF and PACF output will be in agreement with R. In case anyone is interested in the details: 1. Gretl's figure for the k-order autocorrelation coefficient was calculated as the correlation between y_t and y_{t-k}, which is OK, except that the formula used treated y_t and y_{t-k} as if they were distinct variables with (possibly) distinct means. The correct formula uses a common value for the mean of y. In this case the old gretl result would be asymptotically correct for a series that is stationary in mean. 2. Gretl's PACF numbers used the estimator that is set out in several standard sources (Greene, Hamilton, Davidson and MacKinnon): namely, the coefficient for lag k is the last coefficient in an OLS regression of y_t on [1, y_{t-1}, y_{t-2}, ..., y_{t-k}]. However, it appears that the procedure used by R and Eviews is the Durbin-Levinson algorithm, which computes the sample PACF recursively from the sample ACF values. I have now switched to Durbin-Levinson. (From my reading on this topic, I'm not quite clear on why the results differ: any insights into that?) -- Allin Cottrell Department of Economics Wake Forest University, NC