The leader in Anthony Amico’s early running back projections for the 2018 draft will surprise you.
Building the Model
RB production is a little bit tougher to project than that of wide receivers, but we still have some indicators for a frame of reference. Kevin Cole’s model last year was comprised of 40-yard dash time, weight, final season rushing yards per game, and final season receptions per game. The great Jon Moore found that special teams production was also predictive of success.
I took 207 RBs drafted since 2001 and tested relevant variables looking to build something that could (relatively) accurately predict success. Just as with the WR model, the RB model seeks to predict a player’s best season within his first three years in the league. The resulting scores were scaled 0 to 100. After running the variables through regression testing, here were the three variables that tested most significant, and were used to build the model (listed in order of statistical significance).
- Final season adjusted all-purpose yards per game
- NFL Draft Scout rank
You’re probably wondering what the heck adjusted all-purpose yards per game are. In my initial iterations of the model, I noticed that rushing yards per game were significant (as shown by Cole), but that predictive power increased when utilizing all-purpose yards instead. However, I was not positive that all yards were created equal. In doing some testing, it appeared to me that rushing, receiving, and kick return yards were all worth relatively the same, but punt return yards were worth about five times as much. As a result, I used adjusted all-purpose yards per game in the model, which stems from a simple calculation.
Adjusted all-purpose yards = rushing yards + receiving yards + kick return yards + (punt return yards*5)1
You’ll notice that this model does not include the physical traits that Cole’s did, but that is not an indictment of either model. NFL Draft Scout rank is a strong proxy for draft position, and Cole’s model did not include such a factor. My thesis would be that the market is properly factoring in things like weight and the 40. Since we have information about the market in this model, the physical traits are rendered obsolete.
The model tested with an out of sample r-squared value of 0.3102.
Here are the results of the model when interacting with the 2018 RB class. Included are the top RBs according to NFL Draft Scout, along with some other players I examined in my NFL Draft Preview series.
The output is listed in two different ways. “Predict” is the raw total for fantasy points in the player’s best season (first three years). “Prediction Score” is the result of scaling that number from 0 to 100.
|Player||School||Age||DS Rank||ADJ APYPG||Predict||Prediction Score|
|Saquon Barkley||Penn State||20.9||1||179.2||224.2||83.8|
|Nyheim Hines||NC State||21.1||8||185.3||194.8||76.7|
|Josh Adams||Notre Dame||21.2||11||117.8||135.9||62.6|
|Jaylen Samuels||NC State||21.4||9||93.3||124.2||59.8|
|Ito Smith||Southern Mississippi||22.2||20||146.4||96.9||53.2|
|Kalen Ballage||Arizona State||22.0||15||92.1||86.2||50.7|
|Dalyn Dawkins||Colorado State||23.0||27||132.7||42.5||40.2|
|Justin Crawford||West Virginia||22.8||28||91.5||15.3||33.6|
|Demario Richard||Arizona State||21.1||37||92.1||12.1||32.8|
|Larry Rose III||New Mexico State||22.2||43||151.5||3.9||30.9|
|Diocemy Saint Juste||Hawaii||23.8||53||138.9||-78.4||11.1|
|Jamal Morrow||Washington State||22.9||57||105.7||-98.9||6.2|
And here are all of the prediction scores represented visually.
Many will be surprised to see that the model’s top ranked RB is not Saquon Barkley, but Rashaad Penny. The two scores are, of course, razor thin, but it was Penny’s ability to gain yards that vaulted him to the top. His 250.3 adjusted all-purpose YPG are by far the most in the class and 71.1 more than Barkley’s. If he can do well at the Combine and move up the scouting ranks, his score would improve a good bit. Barkley’s expected draft position inside of the top 10 – and established production against top-flight competition in the B1G – would still have him as the top RB on my board.
The next group of players contains three names people know and one major surprise. Nyheim Hines comes in as the third-best RB in the model due to his impressive blend of production in all phases of the game. Hines had over 100 yards rushing, receiving, and in both phases of the return game. His placement in the projections is in spite of losing some work to teammate Jaylen Samuels.
Nick Chubb and Sony Michel are far lower in the projections than they are in most ranking systems. They both rank in the top seven on NFL Draft Scout but fall outside of the top 10 in prediction score. They are not the first dynamic backfield we’ve seen in the draft, with Joe Mixon and Samaje Perine each drafted a year ago out of Oklahoma. However, that pairing was more productive overall, with scores of 85.4 and 67.0 respectively. Mixon in particular was able to supplement his backfield split with return game production, something the Bulldog tandem sorely lacks.
Something to keep in mind while observing the results is that they will likely change leading up to the draft. In particular, I would expect a lot of the rankings to change across the industry once the NFL Combine and various pro days take place. My plan is to release another iteration of these scores sometime after the combine and before the NFL draft occurs, then follow up with a post-draft model utilizing draft position.
- To get adjusted all-purpose yards per game, I obviously just divided by games played. (back)