A new toy

I have spent a decent amount of time over the past few months building up a list of season-by-season player performance that ties to salary cap numbers. Due to this (and laziness, and having and wanting to keep my job) I have not produced as many posts this fall as I normally do. In the next few weeks I will start extracting (hopefully) interesting posts from it while also continuing to refine the data as errors become clear.

The data set is neither perfect nor comprehensive. It covers the key players for a period in which I could find good data on salaries via the USA Today database. The relatively manual nature of matching individuals who fail out of the automated linking led me to prioritize matching those with significant playing time over some who may have drifted into the league for a game or two.

Out of 44,866 “units” of Approximate Value from 2003 through 2009, all but 488 are tied to players who have a salary linked. On the salary side $18,017,691,274 in cap numbers are accounted for out of a total of $19,144,980,923. Finally, all but 29 of the 5442 player seasons as a starter have a matching salary.

Within the “matched” data there are sure to be errors at the individual player level – the USA Today salary database is only so accurate – but the data will provide opportunities for a wide variety of analyses along dimensions of age, position and tenure in the league.

Some news is worse than no news: The NFL Scouting Combine and projecting draft picks

In general it is a pretty good assumption that adding information to a decision making process improves the outcome. In specific instances, however, the added information can be extremely disruptive. With respect to the NFL Scouting Combine, does it help? It depends…

The method

While we do not have visibility into teams’ draft boards before and after the combine, we do have access to one consistent record of pre- and post-combine sentiment from a consistent (and remarkably well-coiffed) source: Melvin A. Kiper, Jr.

By pulling several years of draft predictions (2004 to 2008) and evaluating the change in quality of predictions on either side of the combine, we can evaluate whether they improved. For each player, the absolute difference between the value they have delivered and the value of the player picked in that position will be the quality of the prediction: the lower the better. If, before the combine in 2004, Kiper projected Larry Fitzgerald to be the 1^st pick, then the quality of that pick would be the value provided by Philip Rivers as the actual most valuable draftee (64% as ranked by Career Salary Cap Value) less the value provided by Fitzgerald (52%). The average of all those individual values is the aggregate quality of the prediction[1]. A prediction with 0 error will have all players ranked in order of their actual career value to date.

Remarkably (or maybe not remarkably, he’s been doing this for a long time), Kiper shows an improvement at the aggregate level from before the combine to after. Across 152 individual players the average improvement in the prediction per player is 0.03% of the salary cap closer to the actual value of the players. Of course this 0.03% means that the average prediction improved by $38,000 of football value at the current cap number. The real action is looking at the distribution of increased (or decreased) accuracy for each position.