Yonathan Melamed’s NFL Offensive Clusters

NFL Offenses K Means Clustering

In this short summary, I’ll be writing about what my goals were during this research, my methodology for arriving at the NFL offensive clusters, and some takeaways. The app can be found here.

Goals

After recently viewing a simple clustering analysis on historical NBA teams, I got inspired to a do a similar analysis on NFL teams. My idea was to find some clustering method that would allow me to compare teams from different eras. A couple weeks ago, my Ravens yet again suffered an unexpected, ugly loss in the playoffs. This gave me a second motivation: after grouping teams by different clusters, we could look to see if certain types of offenses were more succesful than others, particularly in the playoffs.

Methodology and Results

I utilized data from nflfastR’s play by play dataset, which has been an awesome resource. I used regular season data only, and included all seasons 2006-2023, since some of the data is only available starting in 2006. In total, I calculated the following variables, grouping each by offensive team and season: Pass EPA/Play; Rush EPA/Play: Pass EPA/Play in pure pass situations (situations where, based on game factors such as down, distance, time, and score, the expected passing rate would be greater than 70%); Rush EPA/Play in pure pass situations (situations where the expected rushing rate would be greater than 70%); Pass rate over expected (‘PROE’); Explosive play percentage (the rate at which offensive plays gain at least either 20 passing yards or 10 rushing yards); Turnover rate; Average depth of target; YAC EPA per catch (on receptions only); Sack percentage; Percentage of a team’s passing EPA that is gained through targets to the middle of the field; and scramble rate (percentage of dropbacks in which the QB rushes). An example code for explosive play percentage can be found below:

library(nflreadr)
library(tidyverse)
#load data 2004 to 2023
pbp=load_pbp(seasons = 2006:2023)
#filter to only regular season and no aborted plays
pbp = pbp %>% 
  filter(season_type=="REG", aborted_play==0)
#create df of explosive rate
explosive = pbp %>% 
  filter(!is.na(epa), !is.na(posteam),pass==1|rush==1) %>% 
  group_by(posteam, season) %>% 
  mutate(explosive=ifelse((!is.na(rushing_yards) & rushing_yards>10)|(!is.na(passing_yards) & passing_yards>20), 1, 0)) %>% 
  summarise(explosives=sum(explosive), plays=n(), explosive_rate=explosives/plays) %>% 
  select(-explosives, -plays)

Once I created these datasets, I joined them together using leftjoins and checked for any issues with the data, confirming there were no NAs. Then, I standardized the data using R’s scale() function. I decided to standardize each column by each season separately. Doing so allows us to compare offenses from different eras. For example, The 2023 49ers and the 2006 Colts are the two best passing offenses when looking at historical Pass EPA/Play. While the 49ers offense was slightly more efficient, passing offenses in general are more efficient in the 2023 season than they were in 2006. Therefore, once we normalize the data by season, we actually get 2006 Colts as the best offense since 2006. It’s important to remember this is taken in context of the other offenses during that season.

Next, I moved on to the clustering. After experimenting with different variables, I decided on including only pass EPA/play, rush EPA/play, explosive rate, and pass rate over expected (all standardized). I found that if I included too many variables, the clustering algorithm wasn’t really showing me much. We could have 8 different clusters showing different combinations of all the variables, but if we can’t explain why it’s done and what the clusters are even showing, we’re not doing it right. I used both the silhoutte method and elbow method to arrive at 5 centers for the k means model, and I found my results to be easily interpretable. Here, I plot a PCA analysis of the clusters.

PCA Analysis

Once I got my results, I wanted to understand what type of offense each cluster represented. I found the clusters to be a good portrayal of the different types of offenses. Here, I present the mean of each standardized variable by cluster.

Starting with Cluster 2, this cluster shows offenses that we can simply describe as ‘Bad’. They tend to have lower than average pass and rush EPA and explosiveness (think Jets). Cluster 3 represents offenses that tend to excel at rushing, and do so a lot. I call them ‘Rush Centric’ offenses. This year, those offenses include the Browns, Cardinals, Titans, and others. I tend to struggle with this cluster the most, since at times it can combine good rushing teams (with little passing prowress) and teams that just run the ball a lot, but aren’t bad enough to be under the Cluster 2 umbrella. Next, Cluster 1 represents teams that tend to have OK passing offenses, but pass the ball a lot. I call them ‘Meh Offenses’ and you can think of them as teams like the 2023 Jaguars and the 2023 regular season Chiefs (very heavy emphasis on regular season). Cluster 4 represents Elite Passing attacks. They throw a whole lot and are pretty good at it (think every other Mahomes offense). I’ll show later how this cluster has had the most success in terms of Super Bowl wins. Finally, my favorite, Cluster 5, which I term “Schematic Wonder”. These attacks excel at both rushing and passing, though they tend to be a bit better at rushing (meaning they are very, very good at rushing) and do so a lot. They are also incredibly explosive. I call them Schematic Wonders both because of my intuition, that is that in order to be good in all phases, especially rushing, you have to have some good scheme attached AND because of the playcallers in this cluster. Consider the 2023 Schematic Wonders: the 49ers, Dolphins, and Lions, who have arguably the 3 brightest offensive minds in the league. Kyle Shanahan teams are frequent guests of this cluster; in fact, since 2019, the 49ers offense has been in this cluster every year but one. It’s important to stress that not many teams make this cluster, and that in general offenses aren’t exactly perfectly balanced to each cluster, although not too unbalanced to worry about the clustering itself.

Takeaways

Using my app, you can compare every variable I collected (including those not in the actual model and the 4 standardized variables) for every offense from 2006 to 2023. Additionally, I circled every Super Bowl winner. As you can see, most of the Super Bowl winners were Elite Passing offenses. This should come as no surprise, as most of us know by now that the saying “Defense Wins Championships” is a myth, and that the best offenses are those that can sling it.

In fact, since 2016, every Super Bowl winner has had an Elite Passing offense! It’s beautiful when stats agree with your priors. Nonetheless, this year that record will break, as we have the Meh Passing offense Chiefs and the Schematic Wonder 49ers, although the Chiefs obviously come with an asterix. Finally, to revisit my question about the Ravens’ offense sustainability in the playoffs, we see that the 2023 Ravens were an Elite Passing offense (although they were close to a Schematic Wonder with their rushing success, they weren’t quite as explosive as MIA and SF). I’ll leave you with a list of the 2023 offenses by cluster, thanks for reading and make sure to check out the app!