Writers of Pro Football Prospectus 2008

24 Apr 2017

BackCAST 2017

by Nathan Forster

Last year, Football Outsiders introduced "BackCAST," a metric that projects the likelihood of success for running back prospects available in the NFL draft. BackCAST follows "Speed Score," which evaluates running backs based on 40-yard dash time and weight. Like Speed Score, BackCAST uses 40-yard dash time and weight in its projections, but it also includes college statistics from each prospect that correlate to NFL success. BackCAST also projects whether each running back is likely to be heavily involved in the receiving game, or is more of a "ground-and-pound" back.

As we stated last year, BackCAST was (and is) a work in progress, and there was a good chance that we would be able to improve the model in the future. We have used the last 12 months to reconfigure BackCAST in a way that we believe will make it more effective going forward.

Let's start with what stays the same. BackCAST still uses 40-yard dash time and weight, the two holdovers from Speed Score. BackCAST also uses each prospect's receiving yards per game. However, BackCAST is removing the metrics Yards Over Expected Per Game (YOE/G) and Peak Attempts in exchange for two new metrics: Adjusted Yards Per Attempt and Attempts Over Expected Per Season (AOEPS).

YOE/G created a baseline for the strength of each prospect's college team's running game and compared the performance of the prospect against that baseline. More specifically, YOE/G compared the player's yards per attempt during his entire college career to the yards per attempt of all other teammates to record rushing attempts during the player's career, as well as the year before the player started at the college. The idea behind this metric was to control for situations where the running back's system or supporting cast inflated (or deflated) his numbers.

Indeed, when we first started building BackCAST, one of our goals was to create a metric that would address the "Wisconsin running back problem," where running backs from certain schools with consistently strong offensive lines tended to be overdrafted. When that metric seemed to correlate with success, it was easy to rationalize including it in the model. It even had some good anecdotes to support it -- you may recall that the baseline was bad for super-bust Trent Richardson's projection, but great for the fabulous LaDainian Tomlinson's projection.

Having had more time to work on the model, we subjected the "baseline" theory to more scrutiny, and it did not hold up. When isolated, the baseline was just noise that was mixed with other factors that correlate with success. Specifically, what YOE/G was really doing was measuring efficiency and volume, sort of like an extremely primitive DYAR for college running backs. Put more simply, it's one thing to average 6.5 yards per carry after 30 carries, but it's quite another to average 6.5 yards over 550 carries. Similarly, most players with high YOE/G had a high yards per attempt average and a lot of carries.

The new BackCAST removes the "baseline" noise in the prior version and measures yards per carry directly with an adjustment for running backs with low attempt numbers. If the running back prospect has at least an "average" number of total carries for a drafted running back (518), BackCAST assumes that his yards per attempt is reliable and uses it for the metric. However, if the running back has less than 518 carries (for example, 450), BackCAST recalculates the yards per attempt by adding the difference between the average number of carries for a drafted running back and the prospect's running back's carries (in the example, 518 - 450 = 68) and assumes those extra carries each gained 5.36 yards (the average yards per carry of a drafted running back). These adjustments prevent a running back with good efficiency numbers but few carries from skewing the results.

After we removed YOE/G and added adjusted yards per attempt, we were able to take a much closer look at the best way to measure the relationship between total rushing attempts and NFL success. The metric that ended up working best to predict NFL success together with adjusted yards per attempt is one we're calling Attempts Over Expected Per Season (AOEPS). AOEPS gives the running back credit for soaking up a high volume of his team's attempts early in his career. Typically, the most special prospects become significant factors in the running game early in their college careers and increasingly dominate as they gain more experience. An average drafted running back has 19.4 percent of his team's carries as a freshman, 29.2 percent as a sophomore, 36.5 percent as a junior, and 42.4 percent as a senior. Thus, a running back who soaked up 29.4 percent, 39.2 percent, 46.5 percent, and 52.4 percent of his team's rushing attempts during a four-year career would have an AOEPS of 10 percent, because he averaged 10 percent more of his team's rushing attempts than would be expected for a drafted running back.

Although these changes to the model may seem dramatic, they don't significantly change the projections for 90 percent of individual prospects. Last year's model generally liked big, fast, efficient running backs who dominated their backfields in college, and this year's model also likes big, fast, efficient running backs who dominated in their backfields in college. However, the differences do cause big shifts for some running back prospects whom BackCAST significantly downgraded or upgraded due to a team baseline or unusual usage patterns in college. One of those running backs happens to be last year's standout rookie Ezekiel Elliott, whom BackCAST graded as a good prospect but probably overrated as a top-ten pick. The baseline hurt Elliott's projection significantly, and he also had an unusual usage pattern in college -- he was used extensively as a freshman, but did not earn an unusually large share of Ohio State's carries over the latter years of his college career. The new BackCAST (correctly, it seems) grades Elliott as an elite prospect.

Here are the top 20 projections from 1998 to 2016 based on the new BackCAST:

Top 20 BackCAST Projections, 1998-2016
Name Year Rnd Pick College/Univ BackCAST
Ricky Williams 1999 1 5 Texas +185.7%
T.J. Duckett 2002 1 18 Michigan St. +160.6%
Ron Dayne 2000 1 11 Wisconsin +150.8%
LaDainian Tomlinson 2001 1 5 TCU +134.2%
Darren McFadden 2008 1 4 Arkansas +130.3%
Ronnie Hillman 2012 3 67 San Diego St. +126.3%
DeAngelo Williams 2006 1 27 Memphis +115.7%
Todd Gurley 2015 1 10 Georgia +115.2%
Ezekiel Elliott 2016 1 4 Ohio St. +114.7%
Rudi Johnson 2001 4 100 Auburn +113.1%
Luke Staley 2002 7 214 BYU +112.8%
Willis McGahee 2003 1 23 Miami (FL) +111.5%
Edgerrin James 1999 1 4 Miami (FL) +111.2%
Steven Jackson 2004 1 24 Oregon St. +109.8%
Curtis Enis 1998 1 5 Penn St. +109.5%
Reggie Bush 2006 1 2 USC +108.6%
LenDale White 2006 2 45 USC +105.7%
Kevin Jones 2004 1 30 Virginia Tech +105.6%
Derrick Henry 2016 2 45 Alabama +99.6%
Knowshon Moreno 2009 1 12 Georgia +97.9%

Certainly, not every prospect in the above chart was a winner. However, where BackCAST really shines is in its list of the first- and second-round picks with the lowest BackCAST projections:

Bottom 20 BackCAST Projections, 1st- and 2nd-Round Picks, 1998-2016
Name Year Rnd Pick College/Univ BackCAST
Joe Montgomery 1999 2 49 Ohio St. -77.8%
Brandon Jackson 2007 2 63 Nebraska -59.0%
Kenny Irons 2007 2 49 Auburn -56.2%
John Avery 1998 1 29 Mississippi -38.4%
Chris Henry 2007 2 50 Arizona -37.0%
Travis Henry 2001 2 58 Tennessee -34.5%
Julius Jones 2004 2 43 Notre Dame -30.5%
Christine Michael 2013 2 62 Texas A&M -24.2%
Montario Hardesty 2010 2 59 Tennessee -23.6%
Ameer Abdullah 2015 2 54 Nebraska -22.2%
Mike Cloud 1999 2 54 Boston College -12.8%
Carlos Hyde 2014 2 57 Ohio St. -11.3%
Montee Ball 2013 2 58 Wisconsin -9.3%
DeShaun Foster 2002 2 34 UCLA -6.7%
Joseph Addai 2006 1 30 LSU -5.5%
Chris Perry 2004 1 26 Michigan -5.2%
David Wilson 2012 1 32 Virginia Tech -4.7%
Robert Edwards 1998 1 18 Georgia +2.1%
Isaiah Pead 2012 2 50 Cincinnati +4.5%
J.J. Johnson 1999 2 39 Mississippi St. +7.7%

One other change that we have not made this year, but expect to make in the near future, is to switch the independent variables being projected from each running back's total rushing and receiving yards for his first five years in the NFL to each running back's total rushing and receiving DYAR for his first five years in the NFL. However, we do not expect the switch to result in dramatic changes to the model.

So, in sum, BackCAST (as revised) includes the following factors:

  • The prospect's weight at the NFL combine;
  • The prospect's forty-yard dash at the NFL combine. If he did not run at the combine, BackCAST uses his pro day time;
  • The prospect's yards per attempt, with an adjustment for running backs who had fewer career carries than an average drafted running back
  • The prospect's "AOEPS," which measures how much, on average, the prospect's team used him in the running game during his career relative to the usage of an average drafted running back during the same year of eligibility; and
  • The prospect's receiving yards per game in his college career.

BackCAST is expressed in terms of the percentage that the running back is projected to over-perform or under-perform the average running back prospect. For example, a player who has a +50.0% BackCAST score is expected to be 1.5 times as productive as the average drafted running back. Conversely, a player with a BackCAST score of -50.0% is expected to be only half as productive as the average drafted running back.

BackCAST also includes "RecIndex," which measures whether the player is likely to be a ground-and-pound two-down back, a player who catches passes out of the backfield more often than he takes handoffs, or something in between. In short, RecIndex measures the likelihood that the player records a disproportionately high or low number of receiving yards versus his rushing yards. The two factors that are significant in predicting RecIndex are receiving yards per game in college and weight, as smaller players are more likely to be receiving backs.

Our new and improved BackCAST suggests that we have an amazing class of running backs in the 2017 NFL draft. This class was so good that I had to double-check things after the first time I ran the numbers, as I was sure that some sort of error in my spreadsheets caused BackCAST to rate this class too highly. Indeed, if you compare the prospects below to the top-20 historical BackCAST projections above, you will notice that not one, not two, but three running back prospects would make the list of the top-ten prospects in BackCAST's database.

Leonard Fournette, LSU

BackCAST Score: +142.2%
RecIndex: -0.01

Leonard Fournette has an amazing size/speed combination. The average drafted running back is 216 pounds and runs a 4.55 40-yard dash. Fournette, by contrast, is faster than the average running back at 4.51 seconds and nearly a full standard deviation heavier at 240 pounds. Fournette dropped to 228 pounds at his pro day, but unfortunately for football analytics enthusiasts, he did not run the 40-yard dash again.

Fournette was also no slouch when it came to production. Fournette had an AOEPS of +11.76%, even after missing a good third of his junior season to injury (and also, perhaps, due to a desire to preserve his health for the NFL draft). Fournette also averaged more than 6.2 yards per attempt during his career. Fournette does not absolutely blow you away in the receiving game, but his RecIndex is greater than the typical running back (the median running back has a negative RecIndex, which is balanced by a smaller group of third-down specialist backs).

Teams are right to hesitate before using a first-round pick on a running back, but Fournette's numbers are so good that it's hard to argue that teams at the top of the draft with a need at running back should pass on him.

Dalvin Cook, Florida State

BackCAST Score: +136.0%
RecIndex: +0.61

Talk about a consolation prize. A team that misses out on BackCAST's top running back in this draft need only settle for one of the best projected running backs of all time.

Dalvin's Cook's projection is so high because he sports an unusual combination of usage and efficiency. When a college running back is the focal point of his team's offense, his opponents will scheme against him. Although the running back can still be efficient, it would take a unique talent to post peak efficiency numbers while carrying an unusually large load.

There have been plenty of running backs who have been used as heavily as Cook or have been as efficient as Cook, but only one running back ever who has been used as heavily and been as efficient as Cook. Ricky Williams is the only other running back prospect to ever score an AOEPS above 20 percent and to average more than 6.0 yards per attempt.

You could make a case for drafting Cook over top-ranked prospect Fournette. Although Fournette has an awesome projection, his particular profile somewhat matches other running backs who boasted great projections but ultimately disappointed. BackCAST was also high on Ron Dayne and T.J. Duckett, for example, who scored highly in large part due to unusual size/speed combinations. It is a bit harder to find a bust whose high projection, like Cook's, had more to do with college production. Also, Cook has an excellent RecIndex, which means that even if he does not succeed as a traditional running back, he likely has some value as a Reggie Bush-type back who can catch passes out of the backfield.

Joe Mixon, Oklahoma

BackCAST Score: +115.7%
RecIndex: +0.90

Of course, we condemn the off-the-field actions of Joe Mixon in the strongest terms. BackCAST, however, is a mathematical construct that does not make any moral judgments and has no way to factor in the effect of a prospect's off-the-field actions. If nothing else, BackCAST's positive outlook on Mixon's talent is a testament to the potential that Mixon has squandered.

Mixon was not quite as heavily used as Fournette or Cook, but he was more efficient, averaging more than 6.7 yards per carry -- the highest of any running back in this year's class. Mixon was also a prolific receiver, and he therefore also has the second-highest RecIndex of any running back in this class.

Brian Hill, Wyoming

BackCAST Score: +84.6%
RecIndex: -0.07

Brian Hill is a major BackCAST sleeper who could potentially be had in the fourth round. Indeed, Hill's strong BackCAST projection would theoretically put him on top of a weaker draft class. Hill's 4.54-second 40-yard dash is "average" for a drafted running back, but Hill is a big back at 219 pounds, which gives him a good size/speed combination. The Cowboys used Hill heavily, and his +20.6% AOEPS is better than any other running back in this class save for Dalvin Cook. Hill's biggest drawback is his lower average of 5.53 yards per attempt, although that is not unusually low for a running back with a workload as heavy as Hill's.

Marlon Mack, South Florida

BackCAST Score: +80.8%
RecIndex: +0.10

Although not quite the sleeper that Hill is, Mack could be a pleasant surprise for a team that spends a mid-round pick on him. Due to a strong freshman season, Mack posts a +11.5% AOEPS. It is hard to say whether AOEPS under- or overrates him. As his career progressed, Mack ceded carries to running quarterback Quinton Flowers, but he also lost carries to fellow running back D'Ernest Johnson. Mack also posted a strong 6.2 yards per carry, which, combined with his good AOEPS, makes him a sort of discounted Dalvin Cook.

Christian McCaffrey, Stanford

BackCAST Score: +76.2%
RecIndex: +0.99

Typically, when a first-round running back prospect comes in sixth on BackCAST's list, it would be time for an explanation about why the prospect is overrated. McCaffrey, however, is simply a victim of being a fairly "average" first-round prospect in a draft full of historically great prospects and strong sleepers. McCaffrey's projection suffers a bit because he had few carries as a freshman and lost carries to Bryce Love as a junior. Overall, McCaffrey is a strikingly similar prospect to Marlon Mack. McCaffrey has much better receiving numbers, but Mack edges him slightly in AOEPS and size/speed combination. The best reason to draft McCaffrey is his receiving potential, as he has the highest RecIndex in this class.

Samaje Perine, Oklahoma

BackCAST Score: +72.9%
RecIndex: -0.27

Samaje Perine is a good sleeper for a team that is looking to flat-out pound the ball. Perine is rather slow with a 4.65-second 40-yard dash, but he is an absolute load at 233 pounds. Perine somehow managed to post a nice AOEPS despite the fact that he had to share the backfield with Joe Mixon during his sophomore and junior seasons. That freshman season, however, was particularly impressive: Perine ran the ball 263 times for 1,713 yards, working out to an average of 6.51 yards per attempt. One could only imagine what Perine's college numbers may have looked like if he had not had to share the load with Mixon. Perine is not particularly explosive in the passing game. He never had more than 108 yards receiving in any single season, hence the low RecIndex.

POTENTIAL BUST ALERT

Alvin Kamara, Tennessee

BackCAST Score: -30.4%
RecIndex: +0.71

Alvin Kamara is not a hopeless case, but it is unusual for a player with this many holes in his BackCAST projection to be rated as a second-round pick by scouts. Kamara had trouble getting carries for the Tennessee Volunteers, posting a -14.2% AOEPS. It would be one thing if Kamara had been stuck behind a particularly talented teammate, but the available evidence does not suggest that Kamara was permanently locked in some sort of Thurman Thomas/Barry Sanders situation (both played for Oklahoma State at the same time). Kamara was the clear second running back behind Jalen Hurd as a sophomore (though, in all fairness, Hurd was considered a possible future high-round draft pick at the time). During Kamara's junior year, Hurd battled injuries and his production cratered, resulting in an abrupt decision to transfer from Tennessee. Even with Hurd out of the picture, Kamara barely edged out sophomore running back and former three-star recruit John Kelly for carries. It certainly could be the case that Tennessee's coaches failed to realize what they had in Kamara and that his NFL career will prove them wrong for failing to give him the rock enough times to shine. However, Kelly was actually more productive than Kamara on a per-carry basis, so Tennessee's coaches were not clearly wrong to platoon their backs.

Other signs that Kamara could be a transcendent talent stuck in a bad situation are just not there. Kamara ran a 4.56-second 40-yard dash at 214 pounds -- those are firmly average numbers. Kamara averaged more than 6 yards per carry, which is good, but not unusual for a back with relatively low attempts, nor is it a number that makes Kamara stand out amongst the running back prospects in this class. Kamara has strong receiving numbers, but that should only get him so far. Would teams spend a second-round pick on a player who was going to be the next Travaris Cadet?

The following table provides the BackCAST and RecIndex numbers for all of the halfback prospects invited to this year's NFL combine.

BackCAST Projections, 2017 Combine Invitees
Name School Weight 40 AOEPS Adj Y/A Rcyds/G BackCAST RecIndex
Leonard Fournette LSU 240 4.51 11.8% 6.22 16.4 +142.2% -0.01
Dalvin Cook Florida St. 210 4.49 20.7% 6.50 24.6 +136.0% 0.61
Joe Mixon Oklahoma 228 4.45 1.9% 6.16 35.8 +115.7% 0.90
Brian Hill Wyoming 219 4.54 20.6% 5.53 10.9 +84.6% -0.07
Marlon Mack South Florida 213 4.50 11.5% 6.16 13.8 +80.8% 0.10
Christian McCaffrey Stanford 202 4.48 9.1% 6.21 32.8 +76.2% 0.99
Samaje Perine Oklahoma 233 4.65 11.9% 6.02 8.9 +72.9% -0.27
Jeremy McNichols Boise St. 214 4.49 8.3% 5.61 32.0 +69.5% 0.86
Elijah McGuire Louisiana-Lafayette 214 4.53 1.9% 6.06 27.3 +50.3% 0.66
D'Onta Foreman Texas 233 4.58 -3.4% 6.26 5.4 +43.9% -0.42
Elijah Hood North Carolina 232 4.59 2.0% 5.89 6.4 +43.5% -0.37
Kareem Hunt Toledo 216 4.62 6.0% 6.32 12.6 +40.6% 0.03
Stanley Williams Kentucky 212 4.51 -2.9% 6.41 8.8 +30.1% -0.10
Aaron Jones UTEP 208 4.56 1.7% 6.25 18.5 +27.8% 0.34
Donnel Pumphrey San Diego St. 176 4.48 12.8% 6.05 19.2 +12.0% 0.63
Name School Weight 40 AOEPS Adj Y/A Rcyds/G BackCAST RecIndex
James Conner Pittsburgh 233 4.65 -1.2% 5.59 10.6 +9.1% -0.20
Wayne Gallman Clemson 215 4.60 8.5% 5.06 11.6 -3.1% -0.01
Matt Dayes North Carolina St. 205 4.47 -5.2% 5.19 20.7 -22.6% 0.46
Jamaal Williams BYU 210 4.59 0.1% 5.37 13.2 -26.0% 0.10
Alvin Kamara Tennessee 214 4.56 -14.2% 5.69 28.5 -30.4% 0.71
T.J. Logan North Carolina 196 4.37 -10.5% 5.42 13.5 -31.6% 0.23
Joe Williams Utah 190 4.41 -11.9% 5.75 10.1 -49.8% 0.13
Justin Davis USC 208 4.53 -9.8% 5.35 9.3 -54.3% -0.05
De'Veon Smith Michigan 223 4.57 -7.2% 4.55 6.1 -61.7% -0.30
Corey Clement Wisconsin 220 4.68 -8.9% 5.37 7.2 -67.7% -0.24
Christopher Carson Oklahoma St. 218 4.58 -16.9% 5.23 14.2 -71.5% 0.08
Jahad Thomas Temple 190 4.62 1.2% 4.62 21.7 -97.0% 0.62
Rushel Shell West Virginia 227 4.74 -10.4% 4.58 9.7 -100.0% -0.18
Dare Ogunbowale Wisconsin 213 4.65 -17.0% 4.99 12.7 -100.0% 0.05
Devine Redding Indiana 205 4.76 1.9% 4.45 7.4 -100.0% -0.10

Here's what BackCAST had to say about running back prospects last year:

(Portions of this article appeared previously on ESPN Insider.)

Posted by: Nathan Forster on 24 Apr 2017

12 comments, Last at 09 May 2017, 7:21am by Mr Skinner

Comments

1
by justanothersteve :: Mon, 04/24/2017 - 12:41pm

I fully expect Joe Mixon to be drafted by the end of Round 2. He's definitely a talent. My main worry is that any time I hear about a player who has put his past behind him, I always end up thinking this.

2
by poboy :: Mon, 04/24/2017 - 4:54pm

How do you folks test these models to check for issues with overfitting, p-hacking, etc.? My concerned with these sorts of closed models is it's hard to evaluate how predictive they'll be without some sort of sense of the diagnostics that you've run.

3
by Denaina :: Mon, 04/24/2017 - 6:19pm

Still seems to be a work in progress compared to Sackseer, QBase, and Playmaker Score

4
by ChrisS :: Tue, 04/25/2017 - 1:25pm

You mention Mixon's troubles, but according to Mike Tanier Dalvin Cook is also a problem child. "Cook was charged with misdemeanor battery for allegedly striking a woman outside a bar in 2015. There was also a criminal mischief charge involving a BB gun in 2014. And a robbery charge as a juvenile in 2009. And an animal abuse citation involving three mistreated puppies. And more.

Cook was found not guilty of the battery charge. He was forced to pay a fine for the animal abuse citation. The other charges were dropped or minimized."

5
by Mr Skinner :: Wed, 04/26/2017 - 3:39am

Any chance of getting the fully updated new backcast scores for 2016 and other years?

6
by Mr Skinner :: Wed, 04/26/2017 - 10:01am

Also, is playmaker coming up soon?

7
by agauntpanda :: Wed, 04/26/2017 - 11:24am
12
by Mr Skinner :: Tue, 05/09/2017 - 7:21am

Thank you. Looking better than last year, thankfully.

8
by gofastjoey :: Mon, 05/01/2017 - 6:33pm

I had a question about QBs influencing the running stats here & I see someone had the same question after last season's article, but I cannot find that it was ever answered.

Re: Introducing BackCAST
by Dan :: Tue, 04/26/2016 - 2:50pm
Does YOE/G also include quarterbacks? It seems like that could be misleading - QBs who take a lot of sacks would have a low (or negative) YPC, and QBs who are great scramblers would have a high YPC.
reply
16
Re: Introducing BackCAST
by Dan :: Sun, 05/01/2016 - 8:43pm
For example, Devontae Booker has a 4.95 career YPC (2014-15). Non-Booker Utes had 3.97 YPC from 2013-15, giving Booker a 0.98 YPC edge over his teammates. But non-Booker RBs on Utah had 4.41 YPC, giving Booker only a 0.55 YPC edge over his fellow RBs. Utah's QBs had only 3.48 YPC (including sacks), and accounted for 43% of Utah's non-Booker carries.

9
by LionInAZ :: Mon, 05/01/2017 - 11:58pm

I'm pretty sure you would be told that QB rushing is treated separately from RB rushing, just as it is for NFL players.

10
by Karl Cuba :: Wed, 05/03/2017 - 12:49pm

At the combine Joe Williams was listed at 210 lbs and they must have weighed him, I assume this would upgrade his score a little?

11
by gofastjoey :: Fri, 05/05/2017 - 10:29am

something still didn't feel right with your LT/TCU stats. I just went on sports-reference.com and according to their stats, non-LT TCU players had 1151 attempts for 4303 yards during the 4 seasons from 1997-2000 that LT played for the team.

please explain how you came up with 1078 attempts for 2146 yards; basil mitchell alone had 1830 yards on 351 carries in LT's first 2 seasons, 97-98.

for the record, i live in san diego (never a charger fan) and LT was one of the most amazing players i have ever been lucky enough to watch on a weekly basis. i do not attempt to disparage the player, i am only trying to understand the context of some of your numbers.

"stats can be used to prove anything. 80% of all people know that." - homer (simpson)