Now, there’s a difference between “what” and “how well.” In this article, I want to use a mathematical tool called principal component analysis to take our extensive play-by-play data and measure similarities between schemes on the offensive side of the ball. The reason for using this approach is that if you’re not careful, extensive data can turn into misleading data in no time if its dimension isn't reduced.
Here’s one example: play-action rate and eleven personnel usage on neutral downs. While the two variables are not perfectly correlated, there is a correlation, suggesting that instead of having two unique pieces of information, we really have less than two:
PCA takes the linear relationships between all of the variables in a data set pertaining to what offenses are trying to do, which are how much time the quarterback spends in the pocket, how frequently they utilize play-action concepts, which routes they throw, which run concepts they employ, which personnel groupings they use, and much more. Once those are squished together, we can look at the first two principal components, which are the subject of many tweets during the NFL season:
It’s interesting that, given the richness of the data, two principal components only capture roughly 15% of the variability in the data. However, a few themes emerge.
Distance from the origin (0,0) — at least as it pertains to the first two principal components — is a measure of the scheme’s uniqueness. There is actually a more robust way to measure scheme uniqueness, which is to take each of the principal components, multiply them by the amount of variance they contribute and add them up. If this is done, these are the offensive play-callers in 2021 in order of uniqueness, which actually can be shown to contribute to betting prices on a weekly basis:
For completeness, here are the offensive coordinators with the least unique schemes:
The heavy, run-first-but-not-Shanahan (MIN, TEN, NE, CLE) teams do appear to cluster together. All four of these teams used 11 personnel less than 40% of the time in neutral situations in 2021, but Minnesota was the only one of the bunch that used play-action on less than 30% of neutral situations.
The Kyle Shanahan-Matt LaFleur-Mike LaFleur offenses are in relative close proximity to each other in two dimensions, but variations in the passing game — with the 49ers throwing much shorter passes on average relative to the teams ran by LaFleurs — made them not clustered together quite as much as one would assume. Additionally, I’ve noted that linebackers do not flow at all against the 49ers, but they flow voraciously against the Green Bay Packers and New York Jets in the running game, which might speak to Shanahan’s ability to deceive:
working through some stuff for a paper right now
GCOE is "ground covered over expected" – change in Euclidean distance between RB and LB/SS after two seconds, adjusted for situation. Against these offenses in 2021: pic.twitter.com/GzeGE8Smpk
— Eric Eager ???????? (@PFF_Eric) May 11, 2022
The Chiefs are in a league of their own, as they are the only team in their cluster when we do cluster analysis based on these PCAs. This is how a team can host four consecutive AFC Championship Games despite the league spending a great deal of its time trying to stop them specifically.
Teams with running quarterbacks — BAL, ARZ, PHI — are able to achieve uniquenesses that are likely less a product of the actual offensive play-caller and more due to what is available to them in the run game that isn’t available to other teams.
The 2021 Washington Commanders left a lot on the table offensively from an execution standpoint. Only Kansas City faced fewer men in the box on average on neutral downs than Washington, and as a result, the Commanders blocked up a lot of running plays well. However, there were running backs in the NFL who performed better overall in terms of yards per carry than Antonio Gibson did on perfectly-blocked plays.
using team- and play-level data, here is some expected and actual yards before contact during the 2021 NFL season. I included penalized plays here because expectation was taken pre-play:
Chiefs OL + scheme make an easy game for RB
Antonio Gibson woof pic.twitter.com/H8rdC6xZj5
— Eric Eager ???????? (@PFF_Eric) April 21, 2022
Additionally, only Baltimore used more play-action on neutral downs than Washington, but when the Commanders were not able to use deception, quarterback Taylor Heinicke averaged less than six yards per pass attempt and threw 14 interceptions to just nine touchdown passes with 21 turnover-worthy plays to just eight big-time throws. His 9.1 yards per attempt on play-action concepts ranked eighth in the NFL.
Lastly, Washington was a top-five team in motion rate used in neutral situations, which is usually a recipe for success. While Carson Wentz is not a superstar quarterback, he might actually be above the line of competence that can succeed in a system run by Scott Turner that is clearly doing the right things.