The Stoppage Time That Never Gets Played - Michael Atkin

Michael Atkin

SubscribeSign in

The Stoppage Time That Never Gets Played<br>If stoppage time were measured correctly, 21% of matches would finish with a different scoreline, and 11% with a different result.

Michael Atkin<br>Jun 29, 2026

Share

Stoppage time is a sham.<br>If stoppage time were measured correctly, I estimate that 21.4% of matches in the last six major international tournaments would have finished with a different scoreline [95% CI: 18.2%, 25.2%] — and 10.8% with a different result [9.1%, 12.8%] (World Cups ‘18 & ‘22, Euros ‘20 & ‘24, Copa ‘24, AFCON ‘23).<br>Why I built the model

The motivation for this model comes from two observations about football matches:<br>1. Stoppage time is systematically under-awarded. During the 2018 World Cup, Nate Silver hand-measured every match in the group stage and found that referees under-award stoppage time in 97% of matches, by an average of 6 minutes. I pulled the stoppage time data across five major tournaments since. The direction and magnitude of error is striking; 98% of matches ended too early, by an average of 7 minutes.

2. Teams are much more productive during stoppage time. The goal-scoring rate in 2nd half stoppage time is 1.9× the average minute.

The data

All of the match data in this project comes from StatsBomb, one of the sport’s gold-standard providers. StatsBomb codes each match into a timestamped log of on-ball events: pass, shot, tackle, etc. Each match carries roughly 3,000+ logged events, over 30 per minute. The coding is done by a combination of computer vision and human intervention. Professional clubs buy this data to analyze opponents, among them Club Brugge. Since 2018, StatsBomb has released part of that professional database free for research. This project runs entirely on the free release. I use all 314 match logs from the last six major international tournaments.<br>I also use World Football Elo rankings to stress-test the model by factoring in team quality (see Objection #5).<br>Methodology

The core idea is simple. Take all 314 matches as played. For each, append the stoppage the referee should have added but didn’t, and ask: what’s the probability of a goal in the appended minutes? Average that probability across the 314 matches. That average is the headline.<br>Step 1 — Measure stoppage played and stoppage owed , and validate against Nate Silver’s data. First, stoppage owed. The Laws of the Game (Law 7) tell the referee to add all time lost to substitutions, injuries, celebrations, cards, VAR checks, and time-wasting. But there is no stated limit on how long a routine throw-in or goal kick may take before the delay counts as lost time. To measure owed stoppage, you have to supply the thresholds the Laws omit. Nate Silver’s 2018 stopwatch study supplied them, and I adopt his table unchanged. (His allowances are generous. For example, a throw in gets 20 seconds before any of it counts as owed):

A throw-in that takes 50 seconds owes 30; one that takes 15 owes nothing. Genuine stoppages (celebrations, subs, cards, injuries) are credited separately.<br>I calibrate the estimator against the one independent ground truth that exists: Nate Silver’s by-hand measurement of all 32 World Cup 2018 matches, where he recorded both the stoppage that should have been added and the stoppage that was.

Stoppage played is the easy half — whistle-to-whistle added time, read straight off the event clock — and my played clock matches his almost exactly (r = 0.992). That near-exact match is what makes the owed estimate believable: Nate, with a stopwatch and no access to this pipeline, found the same shortfall (r = 0.836, average error 1.54 minutes). Across the 314 matches, the owed stoppage average (mean) is 16.2 minutes and the omitted average is 7.3 minutes, positive in 98% of matches.<br>Important Context - Changes to the stoppage time rules in 2022: FIFA directed referees to count stoppages more honestly. The biggest change was to add all of a goal celebration to stoppage, rather than the perceived “excess” of one. (I update my stoppage time estimate for post-directive tournaments to count 100% of goal-celebration stoppage, rather than the ~25% I calibrated to Nate’s data.)<br>Step 2 — Reconstruct the live football, and check it against outside numbers. Dead time is the gap between two timestamps: a pass goes out of play, a throw-in is received, and the seconds in between are dead. Summed across a match, those gaps give ball-in-play. I validate it against the best external anchors there are. For the 2022 World Cup my reconstruction reads 57:40 against Opta’s 58:04 (−24s); for 2018, 56:00 against Opta’s 54:50 (+70s), with FiveThirtyEight’s 55:18 in between. (Opta, now Stats Perform, publishes the industry-standard ball-in-play figures.)<br>Some gaps have no restart to explain them: the ball goes out of play, then a pass arrives, with no throw-in logged in between. The matches are coded by hand off the broadcast...