Malcolm Gladwell made the following observation at his blog a few days ago:
Is the Wages of Wins algorithm an improvement over the things like the NBA Efficiency system? To make the case for their system, the authors “fit” their algorithm to the real world. For the 2003-04 season, they add up the number of wins predicted by their algorithm for every player on every team, and compare that number to the team’s actual win total. Their average error? 1.67 wins. In other words, if you give them the statistics for every player on a given team, they can tell you how many wins that team got that season, with a margin of error under two wins. That’s pretty good.”
And I noted in this forum: The accuracy story is the same if we expand our sample. If we look at the past 10 seasons the average error is 2.3 wins. Again, our methods link wins to all the statistics the NBA tracks for its players. And these statistics do allow us to measure quite accurately the number of wins each player produces.
A question has been raised in the comments at Malcolm Gladwell’s blog. Can you predict the future with Wins Produced?
If you are reading the book – and if not, why did you order it (oh wait, maybe you didn’t order yet, so go here now) – you may have missed this observation. Well, in all likelihood you missed it because it is tucked away in the second to last end note of the book.
“… Lee and Berri (2004) examined the relationship between the number of wins a team achieved and the productivity of its players in the previous season. This work indicates that between 65% and 75% of current wins can be explained by what a team’s players did in the prior season. In other words, knowing past productivity allows one to predict better than one could if all you knew was the current salary players were scheduled to be paid.”
What does this mean? If you measure the productivity of the players on a team in the past you can know something about the team's future performance. Now is 70% a “good” forecast? I have not seen another model try this, so it is not clear if this is good or bad. After all, as we emphasize in the book, good and bad are relative terms that require a reference point.
I will say this, you don’t want a model that can take player performance in the past and forecast the future with 100% accuracy. If you had that, then why play the games at all? If we can forecast the future perfectly then we can just all gather around the computer screen and see the outcome before the games are played. So if any math wizards are figuring out how to beat our work, please keep in mind that you might spoil the whole reason we are watching these games in the first place.