1. What the model considers

Every 4th down gets judged on the whole situation, not just down and distance:

• Down & distance: how many yards are needed for a first down.
• Field position: how far the offense is from the end zone.
• Score margin: leading, trailing, or tied, as of right before the snap.
• Time: the quarter and clock, and how much game is left.
• Timeouts: how many each team has left.
• Team strength: the pregame point spread and over/under.
• Possession: which team gets the ball to start the second half.
• Venue: indoor or outdoor, which matters for field goals.

2. How the numbers are calculated

The engine behind it is a win-probability model trained on thousands of real college games. Given a game state, it estimates each team's chance of winning. For a 4th down, it plays out all three choices:

• Go for it: the odds of converting (based on the distance, and counting first downs drawn by a defensive penalty), then the win probability whether they get it or not.
• Field goal: the odds of making it (based on distance and the roof), then the win probability of a make versus a miss.
• Punt: where the other team is likely to take over (blocks and returns included), then the win probability from there.

Each option's Win % is the team's chance of winning if they pick it. Success % is the chance of converting the 4th down or making the kick. The recommendation is just the option with the best Win %.

The leaderboard ranks teams by Win Probability Lost: on each 4th down, how much win probability a team threw away by not taking the best option, added up over the season. A perfect call costs nothing. It's weighted too, so a small misjudgment barely moves the needle while a real blunder hurts a lot.

Model: cfb4th (Jared Lee, Sebastian Carl, Ben Baldwin), built on Ben Baldwin's nfl4th work and cfbfastR's college-trained win-probability model. Play-by-play data from CollegeFootballData.com.

3. Limitations

A few things worth knowing when you read the numbers:

• It uses league averages, not a team's actual personnel. The go, punt, and kick numbers are league-wide, so the model has no idea whether a team has an All-American kicker or a walk-on (beyond what the spread implies).
• It ignores weather and altitude. It only knows indoor versus outdoor, not wind, cold, or elevation, all of which change how far a kicker can reach.
• It doesn't account for onside kicks. It assumes a normal kickoff after a score, so it can sell a field goal short for a team trailing late, when an onside try is part of the real plan.
• Fakes and botched snaps fool it. When a punt or field goal turns into a run or pass (a fake, or a bad snap the kicker has to scramble on), the data logs it as a regular offensive play, so the model treats it as going for it instead of the punt or kick it was actually meant to be.
• It's shakiest at the extremes. In blowouts or near-hopeless spots the win-probability estimates get noisy, and the gap between options can look bigger than it really is. (Clearly impossible results are filtered off this site.)
• It grades the decision, not the result. A call is judged on the odds at that moment, not on whether it happened to work out. A gamble that paid off can still have been the wrong call.
• The data isn't flawless. Play-by-play comes from a third party and sometimes has errors. The clearly broken rows get filtered, but not every quirk does.