probably because - if you assume that the probabilities of all the discrete events of a storm outcome that could occur are approximately uniformly distributed (I get that is probably a stretch), and that there are more discrete events that favor a non-snow outcome vs a snow outcome for us, then the sample space of all possible outcomes are more heavily dominated by non-snow ones. Thus the probability of selecting a non-snow outcome is higher, so if a model is all by itself in predicting one, it's more likely to be correct than a model all by itself predicting a snow-outcome. I realize that there are a ton of assumptions here that are probably not empirically grounded (independent events, etc), but it seems that might be the general logic.