The Explosive Forest: Predicting the Big Play
Exploring explosiveness over expectation
Big plays in football are filled with jubilation or devastation depending on which team you root for. Offensive coordinators spend dozens of hours scheming up plays in the hopes of ripping open the defense for one of these big plays. Big plays have always been important to winning football games, and explosiveness is one of ESPN’s Bill Connelly’s five factors to winning a football game.
Today we will be looking at the explosive play itself, whether passing or running yields more explosive plays, and we will be observing which teams excel at averaging higher than expected explosive plays. In order to accomplish this, we need a baseline expectation for an explosive play. We need to figure out this question: Given the game situation (down, distance etc.), if the play is successful, how explosive can we expect the play to be? To answer this question we will be building a model that attempts to answer this question.
The model of choice…. the Random Forest!
Image: Afroz Chakure
A random forest model is simply a bunch of uncorrelated decision trees predicting your dependent variable, and then all trees come together in order to give you its final prediction. Think of a random forest regression model as one of those “how many M&M’s are in this jar?” competitions. Everyone is trying to predict the amount of M&M’s while using different factors in their decision making. Some may use the weight of the jar, some may use size or shape. In the end everyone makes a prediction, and the average of all of the predictions becomes the group answer. Thats a random forest model in a nutshell! (A more detailed explanation can be found here).
The Explosive Play Model
There are numerous definitions of an “explosive” play. For the purposes of this article, an explosive play is any play where the Expected Points Added (EPA) of the play was in the 75th percentile (0.80 EPA and higher) of all plays in the last 5 years (2016-2020). Inputs into the model included down, distance, yard line, score differential, and a couple of other variables.
All things considering I believe this is good enough for what we are trying to accomplish. The model could be improved with variables such as coverage data or how many men are in the box, but unfortunately this data is not public. You’ll notice the model tends to under predict as you can see from the conglomerate of dots above the line. In a high variance game like football, we shouldn’t be too worried that our model isn’t regularly predicting super explosive, 90+ yard touchdowns. Now that we have our model, we can answer the first question: passing or rushing, what is leading us to the sweet glory of a big play?
In what appears to be a case of priors confirmed, you’re expected more explosive plays by passing then you are running the football. Interestingly enough on first and second and long (greater than 10 yards to go), rushing and passing yield nearly identical expected explosiveness numbers. You can also notice a downward trend as you get closer to the sticks. It would be interesting to pair this with coverage data in order to see if this is a case of conservative defenses bending or if its aggressive defenses getting snake bitten.
From a goal line to goal line perspective, you can clearly the higher upside in passing than rushing. The spread and RPO revolution on offense has shown teams that if you want to move the ball down the field in a hurry, it better be through the air. You certainly could not make any definitive conclusions only off of this graph, but the “tired defense” in the second half doesn’t appear to hold up here. If defenses were tired and giving up more big plays, you’d see higher predicted explosiveness numbers here, but that is not the case. More research would have to be conducted in order to make any stable conclusions on that front, but nevertheless it is one more piece to the puzzle of football!
Its cool to hit a couple big plays, but can you do it consistently? (Another excellent Bill Connelly piece on this concept). Just because you exceed expectations on a big play, does not mean your offense is an efficient machine. In order to truly be great, you need the efficiency and the explosiveness in order to put points up at will. With that in mind, lets look at the top offensives in both EPA/Play and Explosiveness over Expectation to see who comes out on top in the CFP Era.
*Note* Rankings based on Power 5 + Notre Dame only
Here are the Power 5 teams that rank top 10 in both EPA/Play and Explosiveness over Expectation. You may be wondering “Where is 2020 Alabama? No Joe Burrow LSU??”. Both teams are comfortably in the top ten in EPA/Play, but in terms of Explosiveness over Expectation that is not the case. You can see some expected faces in Oklahoma, and the Lane Train at Ole Miss. On the bottom of the list you can see Cal and Texas Tech, when their QB’s were guys named Jared Goff and Patrick Mahomes, respectively. This could be a case of fuzzy memories but yes, the 2016 Pittsburgh Panthers had a top 10 offense nationally. They finished 4th in EPA/Play, 4th in offensive SP+ and 1st in Explosiveness over Expectation.
*Note* Rankings based on Group of 5 only
The power 5 offenses that made the list were primarily pass heavy spread offense teams. On the Group of 5 side, however, its nearly flipped with more run/option heavy teams making the list. The 2016 New Mexico Lobos were explosive out of the pistol, earning them their first 9 win season since 2007, and were 2 points away from a 10 win season. As a FSU alum and fan it is very welcoming to see a team with the presumed starting QB (2018 UCF with McKenzie Milton) and a team with my head coach (2019 Memphis, Mike Norvell) both on the list.
Big plays are more likely the result of team play rather than individual QB play, but I couldn’t help myself. Leading the charge in passing explosiveness over expectation… that would be superstar Kansas City Chiefs QB Patrick Mahomes. The biggest surprise comes from Justin’s placement at the bottom of the list. For a QB that is known for their big play capabilities, it is surprising to see a negative explosiveness over expectation. I believe this is another time when offensive scheme and play calls could solve the mystery. Pairing explosiveness over expectation with the play calls/scheme can provide in interesting case study into a QB’s players abilities. As stated earlier though things like scheme and WR YAC could influence an individual players numbers to the point where it is mostly noise.
I tried to keep this relatively short but I feel there are still questions left to answer. In the future I may come back to investigate why option teams seem to be more explosive in the G5/P5, and potentially look deeper into the value of the metric on the individual level. Overall we gave “passing > rushing” more validity, and investigated which teams are both efficient and explosive on offense. I hope you enjoyed the article and hope you all have a wonderful 4th of July weekend!
If you want to dive in to the data like I do, check out @CFB_Data and @cfbfastR on Twitter, where you can learn how to get started in the world of College Football data analysis!
If you want to see more charts and one off analysis, follow my twitter page, @CFBNumbers