Methodology

Tendency, not optimal play

Most public 4th down tools tell you the mathematically optimal call. This project answers a different question: what will a specific coach actually do, based on how they've really called these decisions in the past. The two often disagree, and that gap is the whole point.

Why this matters

Coaches don't always make the textbook-correct call. Their real tendency reflects risk tolerance, how much they trust their offense, and years of instinct built up over a career. That's worth modeling directly, instead of assuming every coach plays a mathematically optimal game.

How it works

Every 4th down decision an NFL head coach has made since 2010 (going for it, punting, or attempting a field goal) is pulled from real play by play data, then matched to the coach who called it and the exact situation: down and distance, field position, score, time remaining, timeouts, and how much head coaching experience they had at that point.

Two machine learning models are trained side by side on that history: a decision tree ensemble and a neural network, each learning its own picture of how a coach's tendencies shift with the situation. Averaging their two predictions together consistently performs better than either model alone.

A coach's very first 4th down decision comes with no track record to go on. Early in their career, the model's prediction leans on how similar coaches have handled similar situations, then gradually shifts toward that coach's own real behavior as more of their decisions come in, the same way a scouting read sharpens over a coach's first season.

A third model answers a different question: if a team does go for it, how likely is the attempt to actually succeed? That model looks only at the situation, not at who's calling the play, because whether a conversion works is a property of the down, distance, and field position, not of who's coaching. Its estimate appears next to every prediction, and it's what drives the conversion odds chart on Findings.

The data

Play by play and schedule data comes from nflverse, a widely used, freely available source for real NFL statistics. It refreshes automatically every week during the season, so coaching changes and new tendencies show up without anyone doing manual work.

The code

Full source, the model training pipeline, and the real accuracy numbers behind these predictions are on GitHub.