[cnvogel]

Part of if might be a constant offset/day.

26 seconds over 43 years (1972-2015) is 0.60 seconds/year. If this matches the average offset from 1820...1972 (which would add up to about 92 seconds), it means that our average is wrong by ~0.6 seconds, but the time length doesn't gain additional seconds due to "spinning down" (caused by tidal interaction with the moon, or whatever other effects there might be).

That's actually also what the graph suggests: It's havnig a short-term daviation of +/- 1ms over the course of (judging by eye) weeks, and a long-term deviation of +/- 2ms over the course of decades. The average deviation of 0.6seconds/year (number of leap seconds inserted) translates to an average of 1.65ms/day from 1972-now.