Fourth downs: when going for it wins, and what kicking costs
From 2020 through 2025, NFL teams made 19,358 fourth-down decisions outside the last two minutes of a half. In 65% of them going for it and kicking were too close to call. In the 3,772 where going was clearly better, teams went 45% of the time, and 98% of the win probability given up on clear calls came from kicking, not from going when they shouldn’t have.
- The cost of fourth downs is almost all on one side. Teams rarely go when kicking is clearly better (3% of those calls). The losses come from punts and field goals on plays where going was the better bet.
- Most fourth downs aren’t worth an argument. A decision counts here only when the whole 90% interval favors one choice. That leaves 65% as toss-ups, where neither call can be called a mistake.
- A season of clear calls is worth about 0.4 wins to a typical team, the sum of the win probability it gave up; the most cautious team-seasons gave up 1.0.
When should a team go?
Who goes when they should?
The costliest calls
How a fourth down is valued
Each choice is worth the offense’s chance to win the game afterward, averaged over what can happen next. Going for it either converts, and the offense keeps the ball somewhere past the line to gain, or fails, and the opponent takes over near the spot. A field goal is good or it isn’t. A punt nets some number of yards, lands in the end zone, or is blocked. Each of those outcomes leads to a new game situation, and a win probability model says how often teams have won from it.
Four models feed the calculation: win probability from any game situation, the chance a fourth-down try converts, the chance a field goal is good, and how far a punt nets. All four are fit only on seasons before the one being judged, so the 2025 decisions were valued by models that had never seen 2025.
Why the end of a half is left out
A fourth down with a minute left is a different problem: the clock, the timeouts and the kneel-down decide what each choice is worth. The win probability model here is accurate over a game as a whole but not in those moments. It rates a team with the ball and a one-score lead in the final 40 seconds at 95.9% when such teams won 99.8% of the time. An error of a few points is larger than the difference between going and kicking, so decisions in the last 2 minutes of either half and in overtime, 15% of all fourth downs, are not judged at all.
Why most calls are toss-ups
The gap between going and kicking is usually small: a point or two of win probability. The models behind it are estimates, so the gap has an interval around it. Refitting every model 50 times on resampled games gives that interval, and a decision counts as clear only when the whole interval sits on one side of zero.
By that standard 65% of judged fourth downs since 2020 are toss-ups. That share is a statement about this model’s precision as much as about football: a steadier win probability model would call more of them. It also means the clear calls are the ones worth arguing about, and on those the pattern is one-sided. Teams went on 45% of clear go calls and on 3% of clear kick calls.
Are teams getting bolder?
In 2020 the league went for it on 43% of clear go calls; in 2025 it was 53%. That rise is more than chance alone would produce (90% interval +6 to +15 points). The seasons aren’t judged by the same model (each uses every earlier season), so small moves from year to year can come from the model as well as from the coaches.
Method, data, and limits
Data. nflverse play-by-play, 2016–2026, regular season and playoffs. A decision is a fourth-down run or pass, field-goal try, or punt in a game with a winner, plays wiped out by penalty excluded: 19,358 judged decisions in 2020–2025, plus 2026 through week 10.
What isn’t judged. Decisions in the last 2 minutes of either half and in overtime (3,327, 15% of all fourth downs) are left out. There the clock decides what a choice is worth, and the win probability model isn’t reliable: on seasons it hadn’t seen, teams with the ball and a one-score lead in the final 40 seconds won 99.8% of the time (617 plays) and the model gave them 95.9%. Its calibration error is 0.016 over the rest of the game, 0.022 in the last minute and 0.057 in overtime. Errors that size are bigger than the gaps between going and kicking. The end of the first half is left out for a different reason: the calculation treats the second-half kickoff simply. These are the fourth downs people argue about most, and this page has nothing to say about them yet.
The four models, and how each did on seasons it hadn’t seen (2020–2025).
- Win probability: gradient-boosted trees on score, time, down, distance, field position, timeouts, home field, the second-half kickoff, and team strength (point margin so far). Brier score 0.158 against 0.170 for a score-and-time baseline. NFL win probability v1.0.1 (validated in the backtest harness).
- Conversion: the conversion rate of fourth-down tries by yards to go. Calibration error 0.017, inside what chance alone produces (0.023). NFL fourth-down conversion v1.0.0 (validated in the backtest harness).
- Net punt: gradient-boosted trees on field position, roof, wind and season. Off by 6.52 yards on average against 6.60 for a field-position table. NFL net punt yards v1.0.1 (validated in the backtest harness).
- Field goals: kick distance and its square, fit on the last three seasons. Calibration error 0.016 against a chance level of 0.013: it did not pass, and it still predicts slightly fewer makes than happen. It replaced a first version that missed by more. NFL field-goal make probability v2.0.0 (experimental: chosen after version 1.0.0 failed its calibration gate on 2020-2025 (error 0.026 against a noise bar of 0.014: too few makes predicted from 50+ yards). Those seasons were read before this version was specified, so they are not a clean test of it; the 2026 season is, scored in February 2027).
Putting them together. Yards gained on a converted or failed try come from past fourth-down plays with a similar distance to go; punt outcomes are the model’s prediction plus the spread of past punts from that part of the field, with touchbacks and blocked punts (0.5% of punts). After a change of possession the opponent’s win probability comes from the same model with the situation flipped. A kick is offered only up to 61 yards, the longest distance teams try with any regularity; a punt only from 30 or more yards out.
Intervals. Every model and table is refit 50 times on games drawn with replacement, and each decision’s go-minus-kick gap is recomputed each time. The interval is the middle 90% of those. It covers how uncertain the models are; it does not cover what they leave out.
What it can’t tell you. The models don’t know who is hurt, who the kicker or quarterback is, the matchup, or the weather beyond wind and roof. Team strength is a season-to-date point margin. A touchdown counts 7; returns for scores, safeties and penalties aren’t modeled; overtime is valued as the model sees it. Long field goals are tried mostly by strong kickers in good conditions, so their make rates flatter an average kicker. A clear call here is clear to this model, and “win probability given up” is the model’s estimate, not a fact about the game.
Code: nfl/fourth/ · Data: nflverse play-by-play (CC-BY 4.0). Team numbers: team efficiency · all teams · offense vs. defense. Sparvi is not affiliated with the NFL or any club.