Well, I felt as if I'd been called out in class ;-)
Given:
Aj, Bj, Cj..... Nj at each round of measurements, calculate all the pair wise products and add to the accumulated correlation coefficient.
Phi(A,B) += Aj*Bj is the running estimate of the crosscorrelation coefficient of A & B. If I were doing this I would put all the correlation coefficients in a file as ASCII text:
number pair
sort -k 1n <file >sorted
and then plot sorted using gnuplot. In general one gets a plot of the cumulative distribution function. As most data are quasi-Gaussian this takes on a sigmoidal shape with the outliers at the right and left.
The optimal measurement length is as long as possible, as often as possible. However, daily over a few years is adequate. The lower the drift, the longer the series needed to characterize it.
I should like to note that the measurement drift of the DVM is correlated across devices. That introduces an error in the correlation coefficient estimate. I'm still contemplating the particulars, but a sparse L1 pursuit should be able to separate the DVM drift from the reference drift if there are multiple references. The part I've not determined is how the number of references in excess of two affects the accuracy of the DVM drift estimate. I'm also a bit unclear if it's possible using Wiener's L2 methods. Likely one can, but the procedure may be rather involved.
If there is interest from TiN, cellularmitosis, Andreas and others who are measuring and comparing multiple references I'll write a bespoke program to do the calculations and generate files for use with gnuplot.
I know little about the Allan deviation as I have never coded it, just looked at plots of clocks. So I'll need to read the definition closely and work out the arithmetic required. The article I checked suggested it was far simpler than I thought.
If you are at all interested in problems such as posed by Zia, you really should get a copy of "Random Data" by Bendat and Piersol. It is very readable and describes in clear language how to solve most problems. There are a few things that sparse L1 pursuits can do that can't be done with classical L2 methods, but pursuits are not a replacement for classical methods. Allan Piersol passed away, so the 4th ed is the last. Octave/MATLAB will handle all the numerical chores.