Hello, folks. Welcome to Analysis Week. For five of the 100 days until kickoff, FTRS will be bringing you college football analysis content, developed by our friends at cfbfastR and Statsbomb.
Yesterday, we looked at the philosophy behind analytical work and built our own expected points model, based on vibes rather than advanced math. As a friend of mine from the football industry once said on a podcast (paraphrasing): when we teach someone something we find in data, we’re not teaching them advanced math, we’re teaching them football. That’s the idea: no human can watch every game, but a computer can. Why not see what it can “learn” and see how we can use that to play better?
But one thing we don’t really dig into is the “why” of the expected points. Yes, we build a model, and yes, we understand how to interpret its results, but what does it all mean? What’s all this about? We build a pile of rocks out of a pyramid of sticks and figure out a way to turn it on, but what do we do next?
Let’s go back to the reasons why we analyzed the expected points in the first place, specifically the last one:
If our ultimate goal with the broad world of “analytics” is to find out what helps teams win games, then it makes sense to look at the relationships between specific statistics and wins.
Expected points and their related transformations are what you might call “underlying numbers” — meaning: while they’re not explicit points on a scoreboard or yards gained on a play, being good at accumulating them can help you generate those more explicit outcomes toward your overall goal: winning games. This concept is beautifully (if somewhat implicitly) explained in Moneyball (the origin story of every analytics staff member):
With the help of Paul DePodesta (reimagined as Peter Brand for the film), Billy Beane focuses on on-base percentage for player evaluation. Why? According to FanGraphs:
On-base percentage (OBP) measures the most important thing a hitter can do at the plate: not make an out. Since a team only makes 27 outs per game, making outs at a high rate is not a good thing—that is, if a team wants to win. Players with a high on-base percentage avoid making outs and get on base at a high rate, which prolongs games and gives their team more opportunities to score.
Beane’s scouts are trying to optimize runs batted in (RBI), which is certainly a measure of player skill, but one that’s highly dependent on situational context (i.e., someone being in scoring position when a hit is made). OBP, on the other hand, is more self-constructed — a player’s ability to get on base is more dependent on his own actions (i.e., reactions to pitches). Since we know that OBP gives teams more opportunities to score, it follows that getting on base more often leads to more runs. More runs obviously leads to more wins. So rather than looking at an actual good outcome in and of itself and taking it at face value (it’s very hard to argue that racking up runs via RBI is inherently bad, because, you know, it’s runs on the board), Beane is more interested in how to generate good outcomes more often with OBP.
Given how we’ve defined and demonstrated the concept of expected points, we’re largely doing the same analysis as Beane and DePodesta: consistently accumulating expected points (via expected points added, or EPA) means that the offense is moving from situations with lower chances of generating the next score in the half to situations with higher chances of generating that score. Generating the next score in the half (obviously) adds points. Generating the next score in the half more frequently (by accumulating more EPA) leads to more points. Of course, more points lead to more wins, so we can draw this logical straight line from expected points to wins.
Obviously, we’d like to test our logic against the data to make sure we’re on the right track. If what we’re saying is true (that teams that generate more EPA typically generate more points), then we’d expect to see a strong positive correlation between EPA differential and point differential in our dataset (reminder: ~1.2M plays in ~7,000 FBS vs. FBS games over 10 full seasons of college football). Here’s what we get when we graph these two values:
There is clear evidence here of a linear relationship, and given the number of games we are evaluating this relationship in, we can say with some confidence that our logic has merit.
If generating more expected points leads to more points (and therefore more wins), we’ll want to identify teams that are very good at generating expected points in a single game. Taking it a step further, we could say that these teams are “effective” at achieving their goals. Traditionally, we would evaluate effective offenses using yards per game; in keeping with that same theme, let’s rank the 10 best and worst teams of the 2023 season using EPA per game:
This makes sense as a simple barometer of “good” versus “bad” offenses: just as with yards per game, we’re capturing high-performing offenses, and generally high-performing offenses are “good” offenses.
But critically, just like with yards per play, we’re also capturing high-opportunity offenses (as seen in the number of plays per game), whether it’s because of opponent tempo, turnovers, or 3:30 kickoffs on KeynoteUSA that turn into four-hour games. These factors are like needing runners on base to generate an RBI: While you can use EPA per play and yards per play to describe an offense’s performance in a single game, the factors outside (the offense) built into them cloud your measurement of an offense’s “true” capability.
We need to find our version of baseball’s plate appearance (a single unit of action in which the offense controls its performance) to smooth out these kinds of confounders. With this unit in hand, we can measure offensive performance like OBP: how effective was an offense at generating positive outcomes when accounting for its opportunities?
The obvious candidate unit here is the play, and the efficiency metrics that logically follow are therefore yards and EPA per play. Let’s reorder the 10 best and worst offenses of 2023 with that in mind:
And does our saying about EPA differential and point spread still apply here if we use EPA per game spread instead?
For sure yes!
If you’ve been reading closely, you’ll have noticed that we left out yards per play in our analysis. Why? Hint: it has to do with a lot of the things we talked about yesterday. Take a few minutes to read that article again.
Let’s focus on this part from yesterday:
In different situations, we feel different things depending on whether we think someone is going to score soon. The concept of expected points is just one way to quantify that feeling.
In the context of yards per play, every yard gained is equal: a yard is a yard is a yard. But deep down, we know that there is inherent value to every combination of situational factors (yard line, down, distance, etc.) for a single play, and we know that value can change from down to down, yard to yard, etc. Therefore, yards gained or lost on a play are not truly equal units.
Using expected points gained and lost (i.e., EPA) helps us explain that unit disparity. An offense with a high EPA per play is effective on a play-by-play basis, even when you factor in the situational context of each play: it’s generating outcomes that increase its chances of scoring in the next half by a significant margin. In other words (well, words), this example offense is efficient at taking advantage of its opportunities, and if we go back to our logic about expected points and wins:
- If you create more expected points, you will increase your chances of scoring the next goal in the first half.
- If you create more expected points more frequently (i.e. you are efficient at creating expected points), you will increase your chances of scoring in the next half more often.
- Scoring more frequently in the next half (i.e. scoring more often) leads to more points.
- More points lead to more wins.
If you only learn five things from this post about expected points, EPA, and how to use them, here’s what you need to know:
As a parting gift, here are some general rules for EPA per play when following CFB games online (hopefully via Game on Paper?):
We’ll be back tomorrow with more!
Keynote USA
For the Latest Sports News, Follow Keynote USA Sports on Twitter.