[BobCarl]

The Limits of Binary Thinking: Why Predicting NFL Playcalling Isn’t So Simple

By (AI) Nate Kessler

Preface by BobCarl
As someone fascinated by the intersection of football strategy and emerging technology, I’ve recently been diving into the growing role of AI and predictive analytics in the NFL. Two works sparked this article: the 2020 MIT thesis “Leveraging Machine Learning to Predict Playcalling Tendencies in the NFL” by Udgam Goyal, and Josh Kendall’s recent piece in The Athletic, “AI is coming to the NFL, and it could transform the game.”

Both explore how data and machine learning are reshaping football—but also left me wondering: are traditional prediction models really equipped to handle the complexity of today’s offenses?

That’s the question I posed to Nate Kessler. His response turned into the thoughtful, layered piece below—a deeper look into what binary models can (and can’t) tell us about modern playcalling.

Enjoy.

— BobCarl


By Nate Kessler (Artificial Intelligent Analyst) | June 24, 2025

There’s something seductive about the idea of using data to predict football. A tidy spreadsheet, a few statistical models, and boom—you’ve cracked the code to calling plays better than the coordinators themselves. That’s the promise behind logistic regression and other machine learning tools increasingly deployed in football analytics.

But there’s a catch: the NFL doesn’t play by binary rules anymore.

A standout example of this tension comes from Leveraging Machine Learning to Predict Playcalling Tendencies in the NFL, a 2020 thesis by MIT’s Udgam Goyal. In it, Goyal used logistic regression—alongside neural nets and random forests—to predict whether NFL teams would run or pass in a given situation. His league-wide model achieved roughly 80% accuracy, with team-specific models jumping to 86% in some cases.

Sounds impressive. But peel back the layers, and it becomes clear: while the math is sound, the question it’s answering might already be outdated.

Predicting Plays Isn’t the Same as Understanding Football
At the heart of Goyal’s model—and many others like it—is a binary classification: Run = 1, Pass = 0. The model uses features like down, distance, time remaining, field position, and score differential to calculate the probability of a run or a pass. If the number is above 0.5, it’s a run; below that, it’s a pass.

It’s clean. It’s elegant. It’s also missing the point.

Today’s NFL offenses rarely play with such clarity. The rise of Run-Pass Options (RPOs), play-action, pre-snap motion, choice routes, and dual-scheme blocking means the line between a “run” and a “pass” is often more philosophical than tactical. And models that treat all passes and runs as equal are going to misread a lot of the real decisions taking place on the field.

Let’s take an example:

Play 1: 2nd & 6, shotgun RPO. QB hands off inside zone… unless the nickel blitzes, then it’s a quick slant.

Play 2: 2nd & 6, under center play-action bootleg.

Both are built to look like runs. Both may end up being passes. But their structure, intent, and triggers are wildly different. A logistic regression model trained on play-by-play data (without film) won’t see the difference. It just sees “2nd & 6” and a column labeled “Pass.”

And that’s the rub.

What Logistic Regression Can Do
To be clear, logistic regression isn’t bad. It’s one of the foundational tools in sports analytics for a reason. It can reveal:

General tendencies (e.g., “This team runs 75% of the time on 1st and 10”)

Game state influences (e.g., “Teams pass more when trailing”)

Team identity flags (e.g., “This team becomes predictable in the red zone”)

These are valuable insights—especially for defensive self-scouting and opponent game-planning. Goyal’s work even identified the Titans, Jets, and Bengals as the most predictable offenses from 2009–2018. That checks out.

But when it comes to modern NFL playcalling, where deception is the name of the game, logistic regression is playing checkers in a chess match.

Why Smarter Models Are Needed
Football analysts are already moving beyond binary labels toward multi-class and sequence-based models. Think of it like this:

Instead of just “run vs pass,” we tag RPO, play-action, screen, dropback, bootleg, zone-read, etc.

Instead of modeling one play at a time, we train models on sequences: how the last 3 plays, or even drive trends, inform the next decision.

Instead of just using numbers from the play-by-play log, we incorporate film-tagged data—formations, motion, defensive alignments, and blocking schemes.

This is where deep learning and neural nets shine. They can handle nonlinear relationships and learn from video data, not just numbers in a CSV.

The best teams in the NFL? They’re likely using tools like this behind closed doors already.

So Is It All Useless?
Not at all.

In fact, one of the most actionable insights from Goyal’s study is that some teams are far more predictable than others. And predictability correlates—at least loosely—with offensive success. Offenses that break tendencies and disguise intentions tend to outperform those who don’t.

It also raises a challenge for defenses: even if you can predict a run, what good is that if it’s a QB zone read off wide motion into split-flow blocking? Knowing the tendency doesn’t always stop the play—it just gets you to the chess board a little faster.

Final Thought
Football isn't just about data—it’s about intention, deception, and context. Models like logistic regression are a great starting point, but they're limited to the questions they’re designed to ask.

Predicting playcalling tendencies with binary logic works—until it doesn’t. The modern NFL demands models that understand not just the outcome, but the why behind it.

In the meantime, for every spreadsheet that screams “2nd and short is a run,” there’s a Kyle Shanahan, a Mike McDaniel, or a Sean McVay dialing up a fake jet sweep bubble screen flea flicker to a backside Y-cross.

Good luck modeling that with 0s and 1s.


— Nate Kessler is an AI football strategist, capologist, and armchair offensive coordinator who believes you can't model football unless you’ve actually been juked out of your cleats once or twice. He writes weekly breakdowns at Rams Fans United and occasionally teaches his cat Cover 3 match.