Football Outsiders is developing a new RB prospect rating system which they're calling
BackCAST.
Their old system, Speed Score, looked at 40 time & weight. It was basically a linear combination of those 2 variables, where each 0.10 seconds of 40 time was worth 20 pounds. (Though they used a weird formula.)
Their new system uses 5 variables:
- 40 time
- weight
- yards per carry relative to teammates (which they call "yards over expected per game")
- market share of rushing attempts
- receiving yards per game
I include things similar to each of these in my RB projection system, except I haven't gathered data on teammates' rushing stats so I don't look at rushing efficiency relative to teammates (if I had that data then I would include something like it).
First their top 10 RBs, then some thoughts on their system vs. mine.
+63.1% Derrick Henry
+46.2% Ezekiel Elliott
+23.4% Devontae Booker
+18.2% C.J. Prosise
+18.1% Jordan Howard
+16.6% Kenneth Dixon
+15.2% Daniel Lasco
+14.2% DeAndre Washington
+8.6% Tyler Ervin
+8.1% Paul Perkins
The percentage rating is "how much better than the average draftable RB rospect is this guy?" This class is not very good, by FO's metric. The top 25 RB prospects since 1998 range from +182% to +85%. They basically agree with conventional wisdom that Henry is an early 2nd round prospect. They're just a lot lower on Elliott.
(One note - there may be some errors in the data in that they're using to calculate these, especially in weights and 40 time. Or it's possible that the errors are just in the table that got posted to their website.)
On to methodology. There are methodological differences between FO's approach and mine, as I talked about
last year in relation to their WR rating system.
To take one example, it seems reasonable to include some variable related to rushing attempts in a RB prospect rating formula. It seems plausible that RBs who carried the load in college are more likely turn into NFL successes, and historical data seems to bear that out. But there are a few different ways that you might do this. You might just look at number of rushing attempts in a RB's best season - if he had 300 attempts, then mark him down as 300 attempts. Or, you might look at his market share relative to his teammates - if he had 300 attempts and his team had 600 total rushing attempts, then mark him down for 50% of the attempts. Or, you might look at his market share relative to the RBs on his team - if he had 300 attempts, the other RBs combined for 200 attempts, and QBs/WRs had 100 attempts, then mark him down for 60% of the RB attempts.
Football Outsiders' approach is to run a separate regression with each of these predictor variables and then pick the one which has historically been most correlated with NFL success. They did this and picked the 2nd option - market share relative to all teammates - even though that includes sacks as rushing attempts. I suspect that this is a case of overfitting - it's probably just a coincidence that guys whose QB had lots of rushing attempts did worse in the NFL - and this choice hurts the rating of guys like Ezekiel Elliott whose QB had lots of rushing attempts.
My approach is to either pick the one which seems most plausible to me, or to choose "all of the above" and average them together. In the case of attempts, I picked total number of rushing attempts, but averaged together the player's most recent season and the season with the largest workload, and I capped the number with a floor of 80 and a ceiling of 225. If there was enough data to figure out which variation on the "number of attempts" variable was most predictive then my ad hoc approach would be missing out on that information, but I think that we don't have enough data to reach anything more than the crude conclusion that "something like number of attempts is predictive of NFL success". So I rely on intuition about what's plausible, and averaging together a few similar variables to smooth out the rough spots for guys who have a weird profile, in order to put my system together.
Similarly, they pick one rushing efficiency metric (YOE/G) while I average together several, and they pick one athleticism measure (40 time) while I take a weighted average of several (primarily 40, vert, and broad).