Win projections and regression to the mean

I get a kick out of win projections regardless of sport; not to pick on Peter King, but this is just the most recent in a long, long, long-standing tradition.

King picks no team in the NFL to have a record better than 12-4 and no team to have a record worse record than 3-13. At face value that may not seem like anything bizarre, but that would be a truly historic level of parity. Not since 2003 have either of those predictions come true-- no team finished below 3-13 during that season-- and the last time both of those conditions were met was in 1993. Basically, he's predicting a level of parity not seen in professional football in about twenty years.

When computer models run these things, you tend to find most likely outcomes. The problem is, when you've got 30+ samples of teams, something fairly unlikely is probably going to happen to somebody. There was a projection for baseball wins a few years back that predicted something like 8 teams to have exactly .500 records. Not only had that never happened (or even close), but the total number of exactly .500 teams was something around 8 over the previous 10 years.

I can understand why someone would predict this-- someone's more likely to remember a significant error (you picked my team to go 1-15 and they made the playoffs!) than a pretty-close prediction (boy, they sure showed you-- 11-5 instead of 9-7!). Nonetheless, I'd love to see someone step out on a limb-- and I'd have no problem with them trumpeting their prediction at the end of the season if it ended up correct.