College football data is entering a new age, my friends. Thanks to some hard work from collegefootballdata.com and my new online friend @903124, we have a reliable, useful, and pretty darn consistent measure of Expected Points Added (technical details here).

This is great news! Now, we can start to talk about plays in context, and fine tune a metrical evaluation and comparison of teams. First, I’ll explain EPA and highlight how it is informative. Then, I’ll look at some TCU performances in 2018 according to EPA, and in doing so, I’ll lay out an analytical framework for the upcoming season using EPA.

Expected Points Added, by the Numbers

An expected points metric allows us to quantitatively evaluate a play, attaching a specific magnitude and direction to each play’s effect on a game. The EPA metric is derived using a model that breaks plays into down, distance, and yard line bins, and then estimates the actual points expected from each state, using some fancy math.

Expected points actually explains pretty well, at face value, what happened on a play. For example:

Top five plays by EPA, 2018:

1. UAB vs Rice, 2nd Quarter. 3rd and 6 from their own five, up 21 points: UAB throws a 95 yard touchdown pass. Expected Points: -2.19. Actual points: 7. EPA: 9.79.
2. Notre Dame vs Va Tech, 3rd Quarter, 2nd and 12 from their own 3, up 8 points: Notre Dame scores a 97 yard rushing touchdown. Expected Points: -2.64. Actual Points: 7. EPA: 9.26.
3. Colorado vs Colorado St, 3rd Quarter, 3rd and 14 from their own 11, up 25: Colorado scores an 89 yard touchdown pass. Expected points: -2.18. Actual Points: 7. EPA: 9.18.
4. Minnesota vs Illinois, 4th Quarter, 3rd and 19 from their own 14, down 24. Minnesota throws an 86 yard touchdown pass. Expected Points: -1.95. Actual Points: 7. EPA: 8.95.
5. UTEP vs. NMSU, 3rd Quarter, 3rd and 18 from their own 24, tie game: UTEP Throws a 76 yard touchdown pass. Expected Points: -1.79. Actual Points: 7. EPA: 8.75.

Top five TCU plays by EPA, 2018

1. vs Ohio State, 2nd quarter, 2nd and 10 from the 7 yard line, down 4: Darius Anderson runs for a 93 yard touchdown. EP: -0.8 AP: 7 EPA: 7.8
2. vs Baylor, 2nd quarter, 3rd and 6 from the 35: the Mule hits Reagor, who runs for 65 yard touchdown. EP: -0.6 AP: 7. EPA: 7.06.
3. vs Oklahoma State, 3rd quarter, 1st and 10 from the 17, up 18: Jalen Reagor runs 83 yards for a touchdown. EP: .107. AP: 7. EPA: 6.89
4. vs Kansas State, 3rd quarter, 2nd and 10 from the 33: Collins hits Reagor for a 67 yard touchdown. EP: .392 AP: 7. EPA: 6.6.
5. vs Ohio State, 3rd quarter, 2nd and 11 from the 49: Robinson hits Hights for a 51 yard touchdown. EP: .888. AP: 7. EPA: 6.111.

Most of these were situations where a team wasn’t expected to score very often, and they did. Hence a high EPA. This can be fun to play around with, especially when trying to consider context - the EPA doesn’t account for when the margin is close, or when the game is late, but it does allow to compare by summing up across those situations. For example:

TCU EPA total 2018 (-80.406).
TCU EPA second half of games 2018: (-56.65)
TCU EPA first half of games 2018: (-23.754)

Wow - TCU was really bad! They were especially bad in the second half! There’s a wealth of breakdowns we could do - and that I will do - but for now, this is enough to introduce the general idea and kind of check the validity.

Using EPA to Evaluate College Football Teams

The NFL Analytics Dark Web (yes, it’s a thing), loves EPA! It’s a measure of success, of explosiveness, and of player execution, all in one convenient metric. Of course, that adds a bit of murkiness in interpretation, but on the whole, a “mean EPA” stat really tells you a lot about a team, and helps you to compare a team.

First, a graph:

The intuition behind expected points is clear, but the consistent estimation is a bit tricky. This graph demonstrates a pretty smooth expected points trend, with some movement around a straight line. Also, there’s an exponential shape here - expected points increase by a lot (more than linear) the closer you get to your opponents’ yard line. This provides an immediate solution to one problem in college football analysis: all yards are not created equal.

Using EPA allows us to compare performance across context, punishing teams more for making mistakes in high leverage areas and rewarding them more for excelling in those high leverage areas.

Let’s look at TCU’s season with EPA. First, the offense:

What we are looking at here is Expected Points Added by Yard Line for all of TCU’s offensive plays. Remember, yard line 0 corresponds to your opponents’ end zone, and note that the red line is just a demarcation of 0 EPA.

You’ll notice immediately a couple of things. First, you’ll see that there are a whole dang lot (technical term) of negative EPA plays, particularly in scoring opportunities. In fact, if we define a scoring opportunity as “inside your opponent’s 40”, a loose definition, we see ELEVEN plays with -5 EPA or more. TCU ranked 124th last year in points per scoring opportunity, and in the bottom third of FBS in success rate inside the 30. This graph bears that out. Let’s look at some of those plays in detail to see what happened and how EPA corresponds to on-field activity.

TCU Worst Plays, 2018:

1. Butt-Fumble: vs. Kansas, 4th quarter, first and goal from the nine. Darius Anderson fumbles as TCU sets up game-clinching field goal. EPA: -11.4.
2. Not Kicking a Field Goal: vs. Kansas, 1st quarter, fourth and goal from the 1. Instead of kicking a field goal like rational people, the Frogs don’t convert on the fourth and goal try. EPA: -11.05.
3. We Passed it to the TE: vs. Iowa State, 3rd quarter, second and ten from the 14. Pass completed to Artayvious Lynn, who fumbles. EPA: -9.687
4. Shawn Robinson (x3 tie): Interception vs Texas Tech, Interception vs Texas, Interception vs SMU. EPA: -9.36, -8.8, -7.75. These were all bad, no need to pile on.

These are pretty much the plays you’d pick off the top of your head! EPA just quantifies that! Four of these plays were effectively giving away a win (Kansas x2, Tech, Texas), and the other two made games a lot more competitive than they perhaps should’ve been. The key takeaway: TCU was abysmal in scoring opportunities last year. Their EPA Per Play for downs 1-3 inside the opponent forty yard line was -.349!

You heard that right! On average, the value of a TCU play in opponents’ territory was negative. Even if you filter out outliers (plays with an EPA < -5), the mean is -.096. That’s... not good.

Breaking it down by rush and pass isn’t much better: TCU averaged -.114 EPA on rushes last season (8th in the Big 12) and -.082 on passes (also 8th in the Big 12).

The Frogs were the best defensive team in the Big 12 by a full tenth of a point, allowing barely above zero EPA for the season (.009).

Since it’s defense, remember that TCU’s goal line is the 0, and that negative EPA is good for defense. Here, you see a couple big plays (10 by my count), but also some pretty strong swings - TCU defense had 7 plays of at least -2.5 plays, and as they got closer to their goal line, they really kept teams from exploiting scoring opps.

Lastly, let’s examine the offense one more time, in the context of the down by down preview series I did earlier this summer. I’ll filter out possessions with more than a two score margin, and let’s see how TCU did on rushing and passing by down in a neutral game script.

TCU Offense, Neutral Game Script, 2018

Down Total EPA (Rush/Pass)

1. -.1079 (-.104/ -.113)

2. -.0256 (-.165/ +.129)

3. -.2442 (-.057/ -.335)

Third downs killed TCU! That’s especially painful considering TCU faced the 8th shortest third down situation last year, on average. The only plus situation for TCU was passing on second down. (For reference, OU’s second down passing attack was +.412). Passing on third downs was basically a non-starter, a little bit of an effect of the short third down distance, but that highlights a real deficiency in TCU’s offense.

Conclusion

I could keep going with different EPA breakdowns (and I will!) but the above is a good introduction to the concept. As we go through the season, we can quantify exactly how good and bad a play was, and start to attach EPA numbers to specific players and offenses. I’m excited to use EPA throughout the season as a better way to evaluate teams in different contexts.