PSA - NFL.COM's Rushing Yards Over Expectation Is Not A Comparative Statistic (1 Viewer)

[rockaction]

Any of you into advanced statistics or that see guys on Twitter use this statistic (PROVIDED BY NFL NEXT GEN STATS - PFF'S AND OTHER PROPRIETARY ONES ARE DIFFERENT) to compare backs should be totally aware that the model that they use uses that particular RB's MPH and acceleration at the point of the handoff as an input to calculate his expected rushing yards. See the first bolded passage in the quote below. In addition, any blocking after the RB gets the ball is unaccounted for (this is less problematic for comparing backs, but it is problematic in claiming the Next Gen is a fully-realized and realistic isolated RB stat).

I have no idea where or why it came into use to compare guys, but I've read articles saying it, Sumer Sports says it, any data guy who understands it will tell you that.

DO NOT THINK that it uses a baseline, objective, average running back and then calculates how far he would have gotten.

This is supremely important in how you look at the statistic. It's essentially calculating the running back against himself or others that have the same MPH and acceleration with similar defensive coordinates at the time of handoff. I do not know why stat guys like it at all for comparison purposes. Anybody who wants a PM can have it broken down more thoroughly and illustratively by AI.

This is from NFL.com

NGS | NFL Next Gen Stats

NFL’s Next Gen Stats captures real time location data, speed and acceleration for every player, every play on every inch of the field. Discover Next Gen Stats News, Charts, and Statistics.

nextgenstats.nfl.com

How the model works​

The team from Austria admittedly had NO exposure to American football prior to the Big Data Bowl. It didn't matter. Singer and Gordeev's understanding of the problem -- coupled with their expertise in machine learning -- allowed them to "think outside the box," as Gordeev put it in his post-competition summary of their winning solution.

Singer and Gordeev built a 2D convolutional neural network based on the relative location, speed and acceleration features of every player on the field at the moment of handoff. Singer and Gordeev's elegant solution was rooted in the fusion of their advanced understanding of deep learning and the simplification of a complex problem. Here is how Gordeev explained it on the 2020 Big Data Bowl discussion board on Kaggle, the web-based community of data scientists that hosted the competition with the NFL:

"If we focus on the rusher and remove other [offensive] team players, it looks like a simple game where one player tries to run away and 11 others try to catch him. We assume that as soon as the rushing play starts, every defender, regardless of the position, will focus on stopping the rusher ASAP, and every defender has a chance to do it. The chances of a defender to tackle the rusher (as well as estimated location of the tackle) depend on their relative location, speed and direction of movements."

 

[Pickle Rody]

I do not consider it a statistic at all. What? Yards Over Expectation? Who's expectation?

If some folks find it helpful (for what I've no idea) more power to them. To myself, it is useless dribble.

 

[Chaka]

I do not consider it a statistic at all. What? Yards Over Expectation? Who's expectation?

If some folks find it helpful (for what I've no idea) more power to them. To myself, it is useless dribble.

Agreed. "Over expectation" "statistics" are a load of rich creamery butter.

 

[Bri]

I do not consider it a statistic at all. What? Yards Over Expectation? Who's expectation?

If some folks find it helpful (for what I've no idea) more power to them. To myself, it is useless dribble.

I've argued ad nauseum as to how they can consider the stat when the expectation could be wrong.

Rotoballers? Someone about a decade ago used yards over projection/expectation for the whole game. That was great and a major oversight by the entire industry to not be calculating it but yet constantly discussing the topic.
I've got this guy projected to be 80ypg and he keeps getting me 60- adjust. Simple. We all are off and ..again should have been doing this forever.

Per carry or per reception or per pass is absurd

 

[rockaction]

I do not consider it a statistic at all. What? Yards Over Expectation? Who's expectation?

If some folks find it helpful (for what I've no idea) more power to them. To myself, it is useless dribble.

I've argued ad nauseum as to how they can consider the stat when the expectation could be wrong.

Rotoballers? Someone about a decade ago used yards over projection/expectation for the whole game. That was great and a major oversight by the entire industry to not be calculating it but yet constantly discussing the topic.
I've got this guy projected to be 80ypg and he keeps getting me 60- adjust. Simple. We all are off and ..again should have been doing this forever.

Per carry or per reception or per pass is absurd

I disagree. I think it'll only get better. This was created because you can put microchips in guys' pads that tell you ten times a second where he is and can tell you meters/second and acceleration data. Then they get a picture at the handoff from those chips where everybody was on the field, their speed and acceleration/deceleration, and their angles and direction, and then map it onto a computer learning machine where they have an index of a ridiculous number of actual rushing plays using the speed and locations of guys where the computer can calculate by lanes and the open field how far the guy will get or should be expected to get.

Now, it does two things that hurt it as a comparative model. It inputs the runner's speed and acceleration at the time of the handoff and uses that to project the expectation, which means you're already using his abilities as a given in the expectation, and therefore it is inextricably linked to that runner's abilities and his abilities (or runners with very similar abilities) only. So if you have a guy going, say, 10 MPH and his acceleration says he will hit about 20 MPH and therefore he will go x amount of distance in the direction he is heading and likely to go before the mapped defenders at each defender's speed, acceleration, and direction are able to get close enough to be EXPECTED to corral him and take him down, you can really say you're comparing his actual result with his theoretical result. It's a "within-player" comparison. And that's my main beef with it and how some fantasy analysts are using it. It does not allow us a player comparison of two guys with differing natural speeds and differing capacities to accelerate. You only get the expected result from the inputs of speed and acceleration you're modeling for that particular player, and we know how players vary. The model only compares similarities in everything.

That's my understanding of it.

There's also the issue where the handoff happens and all of the offensive players disappear from the picture. There's no more blocking going on. That's the last instant and then everything is based on unfettered pursuit of the ball carrier. This, as you have already quickly probably noted, doesn't allow for downfield blocking or even any real blocking on the edges.

That said, there are a ton of virtues to the model. It does not claim to measure a back and whether that back is elite, good, average, lousy. It doesn't claim that. It looks at geographical location and speed and compares it to similar situations of location and speed and direction. And that's up front if you dig a little into it. If you look at it on the NFL site and just roll with it (I once did) then you won't know that.

But yeah, I think you can get some feel of how a play is going on a particular team or how a particular team's back is going given the context of the play if you feed it a decent sample size. I think it's very useful.

I actually think it's wicked powerful for teams to use. I mean, if you have this constant flow of what you should be getting you will know which types of plays a running back runs well based upon the expectation of similar contexts. Then, by removing the blockers you get an actual virtue in that it isolates the runner from his teammates' blocking prowess or failure downfield and it helps tell you whether it is the scheme or the runner. I also think that despite the inability to compare directly and completely authoritatively, it can backdoor a comparison—and hang with me here—if you are indeed consistently above expectation and it is all the time, you are likely forcing them to miss tackles or you are breaking tackles, but you cannot say for sure or really compare with other backs (but you can get a sense of it). Those are interesting applications of it, and it's an interesting and useful measure, in my opinion.