Every forecast starts from how often a lab actually ships: the median gap between releases on that product line. Then public claims adjust it, weighted by who made them: a date from the lab itself counts roughly five times what a repost of that date counts, and a claim naming a day counts far more than one naming a season.
A model that has already missed a stated date gets pushed later and its window widened, because dates that move once tend to move again. Betting markets pricing a release below 40% are treated as evidence against that date. Open any card to see every claim and exactly what it moved.