With practice, it's easy to make meaningful predictions down to that granularity. For example, Tetlock finds that with the superforecasters in GJP, their exact percentages are important enough that merely rounding to the nearest 5% or so makes their overall performance significantly worse. (This isn't so mysterious if you think of it in terms of frequencies: predicting a NK nuke test at 5% rather than 10% may seem specious at first, but surely there is a big difference subjectively between 'every 20 years' and 'every 10 years'?)