Of course, I can't just link the model code (even though I'd love to!), but it's not just for reasons of secrecy that we use an "opaque magic number". All semi-

Of course, I can't just link the model code (even though I'd love to!), but it's not just for reasons of secrecy that we use an "opaque magic number". All semi-sophisticated ranking systems wind up being opaque magic numbers. We can't, for example, express a player's impact in terms of raw odds, because those odds are also determined by the level of their teammates and the opposition players. For the same reason, we can't express this number in terms of goals or points or any other easily-mapped statistic. So we need to use an abstract points system -- just as Elo and Fargo and other player rating systems use -- as a means of tracking and ranking player impact over time. However, right before kickoff -- when all the players are known from both teams -- we can distill those ratings back down to odds/goal/point advantages. This is, in fact, how we wagered with the model.