To check that they were too noisy, we graphed out the distribution of "up up", "up down" "down up", etc. in a contour plot, where "darker" means more density, and "lighter" means less density. A "-1" means down, a "0" means same, and "1" means up. Therefore, "Same Up happens 5% of the time" would correspond to "x=0 y=1 z=.05". This is displayed graphically below (click to enlarge). As a reader can see, most of the time approval ratings stay the same. But the point is that we have skew--rarely do we get "up up" or "down down", the upper right and lower left corners, respectively.
The main point here is that for reasons inexplicable to Corrections, a "down" tick is more likely to be met by an "up" tick than a noisy process would predict. We posit this warrants further analysis. Suggestions as to causes are welcome.
Update: Upon further reflection, Corrections suggests this is likely due to the absence of a unit root with noise. That is, every movement we see is one of two types. The first is a "real" movement. The second is a "noisy" movement. Noisy movements are transitory and likely to disappear. Real movements are permanent, in expectation (letting opinions be a random walk, for example). Any movement due to noise will likely go away next period--up noise will be met with down, and down will be met with up. This explanation doesn't completely satisfy Corrections, but it seems relevant.