I avoid some complications in the early data by ignoring 1998. This shouldn't matter since we are looking for a general persistent trend. For your info, I remove 3 days of the data in Aug 2002. This is so the data is commensurate over the period of one week. (You will recall there are already 4 days Aug 5-8, 2002 that are removed. Taking out 3 more allows the calendar week to synchronize over the whole database)

Amplitudes of the Fourier transform of the network variance. The plot is consistent with a lack of periodicity in the data. Fourier amplitudes are calculated for the time-ordered sample variances of network Z-scores. The continuous data extends from Jan. 1, 1999 to Sept. 7, 2001. The plot shows the first 5000 amplitudes of the series. Marks on the horizontal axis indicate the amplitudes corresponding to periods of 7, 1 and 0.5 days, respectively. The bottom curve shows the plot after smoothing over 20 consecutive amplitudes. The red bars at 1 and 7 days are a guide for the eye.

I have looked at the autocorrelation and it is also without periodicity. The device variance has the same story, but is a bit more complicated to represent since the degrees of freedom vary with the number of regs online. This is not the case for the network variance so its a less complicated time series.

I am still poking around. I've been trying different smoothings, etc, but nothing shows structure in the amplitudes.