When I began this series last offseason, I was frustrated that I couldn’t watch more backs, and in turn have a larger sample to speak from. I only got to watch over 200 carries for 10 players, all of which were in the good-to-great running back tier. As a data addict, it bugged me that I didn’t have enough recorded to make more inferences about specific running backs, and the position as a whole. But ultimately, it’s been for the best.
With time to think about how my numbers work during the season, I came up with a few nice adaptations. The series set out to separate running back play from surrounding variables, and these tweaks go a great additional distance toward doing just that. They are what I’ll be talking about today. Let’s get into it.
Before reading on, you’ll be lost if you haven’t read the introductory installments of this series (parts one and two, and perhaps zero). It would probably also help you get a feel for the series if you read the following editions about individual players. Here’s an archive of everything to date.
Stuff Generated Yards vs. Hit Generated Yards
My original method for reconstructing Yards per Carry was to form a “Supercomposite” of four observation-based metrics: Offensive Line-Generated Yards per Carry (OL-GY), Hole Rate (HR), Success Rate (SR), and Running Back-Generated Yards per Hole Hit (RB-GY/Hit). (In order,) These stats attempted to measure the baseline amount of yardage a blocking unit provided on each carry, how often that unit provided opportunities, how often a back successfully made it to the hole, and what the back did once he got to that hole.
(The Supercomposite is meant to be looked at as a way to reconstruct the running situation, where we separate distinct variables from one another. In a few words, it measures what percentage of rushing production can be chalked up to a skill or set of skills–such as vision for the back. Read part two to get a better grasp of its significance and see a fuller picture.)
Three of those measures are pretty dang good when contextualized, but the final stat needed work. Why?
RB-GY/Hit took the amount of yards a back generated on every carry and divided it by the amount of times he hit a hole. This is because players create much more when hitting a hole than not, so RB-GY/Carry would be relatively unfair for players behind bad lines. The problem? Players still generate without a hole, which means that over many carries, RB-GY is unfair to players behind good lines.
The solution? Separate the two types of runs. It’s pretty logical, but I couldn’t do it with how I recorded these stats at the time (I effectively had to switch from a per-game basis to a per-carry basis of data entry to make this switch). Now, there are three metrics that measure creativity (in other words, what the player makes up on his own) for three different situations:
- Hit Generated Yards (HGY) are recorded when a back has successfully hit a hole.
- Stuff Generated Yards (SGY) are recorded when there is no hole to be hit.
- Failed Generated Yards (FGY) are recorded when a back misses a hole for whatever reason. I basically let this stat go untouched, because the nature of missed holes leads to a huge variance in generated yards thereafter, which don’t really help in understanding a back any better.
HGY and SGY are the stats to pay attention to here. They ensure that blocking plays less of a role (it is flatout impossible to entirely separate rushing and blocking, unfortunately) in representing how creative a back is. I hadn’t thought about it at the time, but a terrific by-product of this change is that we separate two very distinct skill sets.
It doesn’t take a football expert to understand that there are differences between what makes a good back in the trenches (strength, size, etc.) and in daylight (elusiveness, long speed, etc.). This is clearly displayed in the (lack of a) direct relationship between SGY per Stuffed Carry (SGY/C) and HGY per Hole Hit (HGY/C, I know): The two stats hold a .05 R-squared coefficient in my current set of data.
After splitting the Supercomposite into five categories, this change is probably the most significant for now. But there are two additional tweaks to be discussed–both of which should be just as prominent down the line.
It shouldn’t be surprising that (a) it’s harder to block seven people than six and (b) it’s harder to run past seven people than six. One of the most alarming properties of my stats last year was that the Cowboys were placed firmly below-average in both Supercomposite blocking figures.
My response was, for the future, to take note of the compared counts of box defenders and blockers. Even if a relative numbers disadvantage in the box wasn’t the sole cause of such an anomaly, it’s a necessary additional step to separate RB play from other variables.
This offseason, I started to also track a team’s blocker-to-defender (dis)advantage in the box. The equation here is pretty simple:
Box disadvantage (at the snap) = Box defenders - (Offensive box players-2)
The minus-2 is there because the ball carrier and quarterback won’t be adding any blocking to the equation, and we use “disadvantage” because in most cases, there are more box defenders than blockers in a given play.
For now, I’ll apply box disadvantage by filtering runs so that only those with a disadvantage of one (effectively one-high coverage) or zero (effectively two-high coverage) are counted–any more, and the run is likely in a short yardage situation; any less, and the run is probably coming on a third-and-long where the defense is content with conceding yards on the ground. Down the line, I’ll probably see what the difference is between zero and one, too.
Generated Yard rates
I’d been toying with the idea for awhile, but I really realized I needed a new tool to measure creativity when charting Leonard Fournette (spoiler alert: my first focus this offseason is on the 2017 RB class). Fournette was struggling to generate anything, putting up terrible numbers in both SGY and HGY through his first five games. But, his final play of that fifth game against the Steelers was a 90-yard touchdown run. The next play? A 75-yard touchdown against the Chargers. These runs completely transformed his numbers, throwing him from one pole of the creativity spectrum to the other.
(Note: Fournette had a gaping hole to run through on the Steelers play, but he earned a big share of credit for the play with his speed and burst. Refer to part zero if you’d like to know more about my methodology when charting.)
I’m uncomfortable with just two runs having such an effect on my numbers. Of course, these numbers will largely balance out over a season’s worth of play–it should be said that the numbers from his first five games were unsustainably poor. However, those two plays told me that I needed more than a simple average to sum up a back’s creativity.
There are a couple methods and statistics to measure that creativity, in addition to an average–such as medians, standard deviations, 25th- and 75th-percentile gains, etc. But for now, I’ll be going with rates: What percentage of the time does Fournette generate 10-plus yards? Four-to-nine yards? Zero-to-three? Less than zero?
I think this is a pretty good foundation to work from. These rates model four stereotypical outcomes–a negative effect (<0), an okay effort (0-3), a chunk gain (4-9), and a gash (10). Of course, you can tweak those ranges, but the point is to get intervals that are meaningful to have separated. Once you get to a gash, there isn’t a ton separating 30 yarders from 60 yarders aside from where the run started. Then, the differences in generated yards largely come down to where the run started–not the back’s actual ability. (I’d say, again, that we’re doing this to isolate RB play from outside variables, but I don’t think I need to bash you over the head with this series’ objective.)
Quickly, an alternative way to measure creativity that fascinates me the most comes with generated points-per-play, per carry. Points-per-play takes the difficulty of gaining yards on one part of the field into account (example: it’s harder to score at the opposing five yard line than gain five yards at your own 40), so starting point similarly doesn’t paint the back’s ability in one light or another. However, that’d both take more legwork for me to record (taking note of starting point for each run is more work than it sounds like, given the other variables I already track) and make it harder for readers to interpret. Nonetheless, I do see points per play factoring in at some point down the line.
What can we do with this stuff?
It’s boring to read a bunch of theorizing and data explanation without any good tidbits to gain. So let’s get to tidbitting.
I’m protecting the world from some GY/Hit hot takes
Joe Mixon better than Alvin Kamara? Yeah, Mixon actually out-generated Kamara by a tiny margin (0.06 yards per hole hit) by GY/Hit measure. But if you adjust these numbers to remove that stat’s bias, the picture changes:
Kamara, still, was not particularly great at creating without any gaps, but he made up for it by being wonderful when he did find daylight (and good news for him, the Saints posted an incredible Hole Rate to get him those looks). Also, don’t freak out about Mixon just yet…
The kids are alright
Despite some disappointing Yards per Carry marks, Mixon and Christian McCaffrey put forth solid performances in their rookie seasons. Both placed near-average (among a sample of generally-good players) at both HGY/C and SGY/C, while their pretty-bad Success Rates (nearly two standard deviations below average apiece) should progress with a full season of NFL experience .
Box disadvantage does not provide many interesting tidbits (on its own) thus far
(The stat might turn into more of a quality control thing than anything else, in the future.)
The kids are pretty alright
Both Mixon (41.5%, 12%) and McCaffrey (43%, 10.7%) posted better 4-to-9-yard- and 10-plus-yard-generated rates when hitting holes than Todd Gurley managed (34%, 10%). Please don’t let people tell you that those two aren’t good.
Adjustments, for now
As things stand, the bigger picture (in other words, how I look at the rushing relationship as a whole) hasn’t changed too much. “Creativity” now accounts for 40% of the Supercomposite model (split evenly between SGY/C and HGY/C), compared to the 50% that it comprised before. Blocking rises from 35% to 50% of the Supercomposite (now that our creativity measure overlaps with blocking less). Those are significant changes, but still fairly incremental.
As for the second two changes, I’m not dealing with a good enough sample size (in terms of individual runners) to change my mind about anything. This isn’t to say that they won’t, they just don’t have enough data points to do so yet.
The arrivals of these changes are bittersweet for me. On one hand, it gets harder for me to pull data together and complete players’ statistical profiles. Furthermore, it’s harder for people to interpret my numbers at a glance: I’m adding more stats to the table for readers to wrap their heads around (however self-explanatory they may be)–and with sentences like…
“Both Mixon (41.5%, 12%) and McCaffrey (43%, 10.7%) posted better 4-to-9-yard- and 10-plus-yard-generated rates when hitting holes than Todd Gurley managed (34%, 10%)”
…I have a hell of a ways to go with making names, labels, etc. more accessible.
Still, each change is a huge step toward further removing situation from a running back’s performance–which is exactly what we set out to do in the first place. Amazingly, increased understanding of a subject demands increased complexity. We’ll deal with it.