A (hopefully) readable explanation of the Football DNA system and why it’s more than just a pretty chart.

One of the mistakes, in my opinion, of modern sports analytics is the desire to boil everything down to a single number. In baseball, this is usually a player’s Wins Above Replacement (WAR). In football, Expected Points Added (EPA) has become all the rage.

I love EPA. It is the biggest leap in football analytics in my lifetime. However, one of the things that differentiates my approach to sports analytics is that I like data that describes the player or team, rather than boiling everything down to a single number. This philosophy is the foundation of the Baseball DNA system and the Football DNA system.

All Plays are Not Created Equal

Team A plays a ground-and-pound style, with low risk but no explosive plays. Team B plays an up-tempo passing offense, who get long touchdowns, but leads the league in turnovers. Those two approaches couldn’t be farther apart, but they both may net out to the same average EPA.

Further, some plays might tell you more about a team than others. Think about a third and short when the team is down three. Their performance on that play tells you much more about who that team is compared to a first and 10 when the game is already out of hand.

This approach isn’t just window dressing, it gives more information that a predictive model can use. Imagine you were trying to predict the outcome of a game. You could have two numbers; Team A’s EPA-per-play and Team B’s EPA-per-play, or you could have a matrix of data on each team, with their EPA-per-play by down, distance, run plays, pass plays, ahead, behind, etc.? Humans may prefer the former, but feeding the latter into a machine learning model will yield much better results because ML models can pick up on hidden predictors and relationships that are hard for humans to identify.

Creating the Football DNA

The Football DNA system is build at the play level, meaning every play that happens in a game is an observation that gets fed into the model. The overall concept is to feed a model hundreds of thousands of plays and then have it predict a team’s performance given a few specific situations.

This process is similar to how Large Language Models (LLMs) like ChatGPT work. LLMs are fed millions of sentences and it learns the relationships between words. Then, you are able to ask it to finish a sentence based on the most likely words that come next.

In the football context, the sentences we are feeding the model are plays. We will tell the model the game situation (down, distance, yard line, score, time remaining), some identifying information (the teams, coaches, and quarterbacks), and the outcome of the play (three yard gain). Once it is trained on all of those plays, we can then essentially ask it, “What do you predict would happen if the Dallas Cowboys are playing a league average defense, and have a third and short?”

Rather than a single answer, we want to know the odds that one of four outcomes will happen; an explosive gain, a first down, a negative play, or a turnover. Those four outcomes across six play situations give us a matrix of 24 data points. Do it again for defense and you get the 48 “chromosomes” of the Football DNA:

Down the left hand side are the six down and distance situations. First and second down are considered “early,” third and fourth are “late.” For distance to go, three or less yards is “short,” four to seven yards is “medium,” and eight or more is “long.” Grouping the data into buckets like this makes it easier for us to see the results and gives the model more predictive power because it gets more observations-per-bucket.

Let’s dive into a row; Late & Short. This is third or fourth down, three or less yards to go. The Cowboys have a -0.5 for an explosive gain, meaning on every 100 Early & Short plays, they will have about one less explosive play than an average team would. However, they have a +6.3 for first down, meaning they will get six more first downs for every 100 plays in this situation than an average team would. In fact, the Cowboys currently have an excellent late-down offense, especially on medium and long distances.

Continuing on the Late & Short row for the defensive side, you see a -0.8 for explosive gains. Our brains usually see negatives as “bad,” but remember that this is defense. The model thinks the Cowboys defense would give up about one fewer explosive play per 100 plays than an average defense would. You can see the Cowboys have all greens on the explosive gain column, meaning they are better than average at limiting those big plays.

What Football DNAs Tell Us

Here are a few interesting examples that show how the Football DNA chart can describe a team.

Offensively, the Steelers can move the ball, with a few green squares on the first down column. They are also better than average at limiting turnovers, with all greens there. However, they have deep reds for explosive gains and negative plays. This is not a good offense.

On the other side of the ball, the Steelers are a great example of a “bend but don’t break” defense. They have red scores on most of the “first down” column, including a rough +10.2 on Late & Medium. However, they have green scores all the way down the “explosive gain” column and several on the “turnover” column. They are willing to let you put a plodding drive together, but they will not give up big plays and will get more turnovers than average.

The opposite defensive philosophy is the Cincinnati Bengals:

Starting on the defensive side, the Bengals are stingy with first downs and explosive plays. However, they have below average rates of negative plays and turnovers, especially on early downs.

On offense, the Bengals are excellent. All greens on the first down column, including deep greens on short distances.

Football DNA Picks

Data is only as useful as it is predictive. To that end, we can feed two teams’ Football DNAs into yet another machine learning model and have it predict the outcome of the game.

Going this road gives the model dozens more points of articulation to use rather than only the blunt force instrument of Team A’s EPA vs. Team B’s EPA.

I’ve been using this model for a few years with some moderate success at picking games.

FAQs

What team variables are you including in the model?
For a given play, there are situational variables as described above and then five team variables; the offensive team, QB, and coach, and the defensive team and coach. I’ve narrowed in on these five through a combination of testing and data availability. Does the offensive coordinator matter more than the coach for some teams? Maybe! As the saying goes, “All models are wrong, some are useful.”

What if a team’s #1 receiver is hurt or their All-Pro defensive end?
The QB is the only player-specific variable I include in the model. Yes, other positions matter, but the gap between the importance of the QB and the second most important player is enormous. For evidence of this, look at the gambling lines when there are injuries. A QB injury moves the line 3-6 points. Take any non-QB superstar player in the league — Tyreek Hill, Justin Jefferson, Saquon Barkley, Max Crosby, anyone — when they are ruled out of a game, the line moves one point, if at all.

Why don’t you show runs vs. passes?
In earlier versions of this system, I did include runs and passes separately. Three reasons why I took the play type out of the model:

• First, the classification on some plays are gray areas; think about screens, touch passes, and QB scrambles.
• Second, an offense’s decision to pass or run is often dictated by the defense’s alignment. If the defense puts eight men in the box, the offense may decide to audible to a pass — if they score a TD, shouldn’t the run game get some credit since the earlier running success is why the defense decided to overloaded the box?
• Third, it just didn’t help. In fact, by removing the run/pass variable in the model, I was able to shave of a few percentage points of predictive error.

Why show the predicted success in each situation instead of the team’s actual performance of each situation?
It would be fairly straightforward to show, for example, how many actual first downs the Rams got on late and short situations rather than what the model thinks the Rams will get going forward. However, how many of those were late in blowouts? How many happened against bad defenses? How many happened when the Rams had a different QB under center? Aggregate stats are nice, but the model has the ability to cut through a lot of noise and get closer to the signal.

What other variables go into your picks?
In addition to the Football DNA data that shows up in the data visualization, I also include as many situational factors as possible, including time zone differences, days of rest, penalties, and special teams performance. In full transparency, most of these have very little effect, but I would rather account for them than not.

Do you only use these numbers to make picks?
No. I know there is a lot that goes into a game that the model does not account for. If the model spits out a -6 line for a game, I just consider that a spot to start from. One situation that happens often is a lack of data. For example, I won’t be picking for or against the Bears early this season because the model hasn’t had a chance to train on any Caleb Williams data.

Are your picks any good?
I’ve had moderate success over the last couple of years, but I’m not quitting my day job to become a professional gambler. Last year, I was number 30 out of 800 in Lee Sharp’s prediction game, and will keep using that as sort of a public ledger of performance. I am putting my picks out there mostly for fun. Use my data and picks alongside any other sources that you like or your own intuition. If I can help you validate a hunch or avoid a trap game, then that’s a win for both of us.

Do you use this for any fantasy football projections?
Not at this point and probably not for a while. Unlike baseball, where I release weekly player-level projections, I find that player-level projections for football have much more to do with injuries and opportunity. I think there are a lot of folks out there already doing a great job of chasing news. I only want to put stuff out there where I have a unique angle, and I frankly don’t have one for fantasy football, despite the fact that I’m an avid player.