A team with 65% of the ball, eighteen shots, and the better expected goals was not necessarily the better side. The first question is what the score was. Nine times out of ten the honest answer explains the whole stat line: they were losing. Chasing a deficit is the most reliable way in football to inflate every number a casual viewer treats as proof of control. The scoreline a stat was compiled under is not a detail you check afterwards: it is the first thing you need before the stat means anything at all.
What "game state" means
Game state is the scoreline context in which a statistic was accumulated: whether a team was winning, drawing, or losing at the moment each pass, shot, or tackle happened, and often by how much and with how long left. It sounds like a footnote. It is closer to a hidden variable that shapes almost every number on the match sheet.
The reason is that the scoreline changes how both teams choose to play. A side that goes 1–0 up has its objective flipped: the most valuable thing it can now do is not concede. So it drops deeper, keeps its shape, stops committing bodies forward, and is happy to let the opponent have the ball in front of it. The team that goes 1–0 down has the opposite incentive. It must score, so it pushes players up, takes more risks, presses higher, and shoots from worse positions because a low-percentage shot beats no shot at all. The football you are watching after the first goal is partly a product of the goal, not of the teams' true strengths.
Score effects in 9,158 Premier League matches: goals per 90 by game state
Added November 6, 2026. We timed every Premier League goal ESPN records from 2002-03 to 20 September 2026 and measured how fast teams score in each game state. Read raw, the numbers run the wrong way for a score effect. A team one goal up scores 1.34 goals per 90 minutes; the team one goal down scores 1.20, barely more than the 1.18 a team scores while level. Two or more ahead, the leader scores 1.62 per 90 and the side behind 1.09. The reason is who is in each state: the side in front is usually the better team. Adjust for the two teams and the minute, and the score effect shows. A side one goal behind scores 9% more than its own attack, its opponent's defence and the clock predict; a side one goal ahead scores 8% less. Two or more behind, +12%; two or more ahead, −11%.
| Game state | Share of minutes | Goals scored per 90 | Scored vs expected | Next goal: this team | Next goal: opponent | No more goals |
|---|---|---|---|---|---|---|
| Trailing by 2 or more | 7.9% | 1.09 | +12% | 23.6% | 34.9% | 41.4% |
| Trailing by 1 | 17.4% | 1.20 | +9% | 35.3% | 38.8% | 25.9% |
| Level | 49.3% | 1.18 | +2% | 42.3% | 42.3% | 15.4% |
| Leading by 1 | 17.4% | 1.34 | −8% | 38.8% | 35.3% | 25.9% |
| Leading by 2 or more | 7.9% | 1.62 | −11% | 34.9% | 23.6% | 41.4% |
The left panel shows where the raw gap sits. From the 30th to the 60th minute a team a goal up scores 1.39 to 1.41 per 90 and a team a goal down 1.09 to 1.20; in the last half-hour the lines nearly meet (1.44 and 1.41 against 1.33 and 1.38). The right panel measures the same goals against each team's own level. The intervals of the four states with a margin all sit clear of zero; the level row's +2% is inside its interval.
Who scores next. From the moment a one-goal margin begins, the next goal is the leader's in 52.3% of the cases where one comes (38.8% of these leads against 35.3% for the side behind; 25.9% see no further goal). The edge fades with the clock. For one-goal leads that begin before the 30th minute the leader scores 52.0% of next goals; for leads that begin at 60:00 or later it is 49.5%, a coin flip, and 60.4% of those leads see no more goals at all. A lead of two or more keeps about 60% of next goals whenever it starts (59.9%, 59.0% and 60.5% in the three periods). For what a one-goal lead is worth at the final whistle, and how the longer added time since 2023-24 changed that, see how safe a one-goal lead is.
The match opens up. Both teams together score 2.37 goals per 90 while level, 2.55 at a one-goal margin and 2.71 at two or more. Part of that is the clock: the last 15 minutes carry 1.12 times the average goal rate and the first 15 only 0.83 times, and margins come later. The same fact sits behind the advice below to lean on first-half numbers: 65.4% of first-half minutes are played level against 32.1% of second-half minutes. Goals in a match are not independent events, which is one reason a plain Poisson model misses some scorelines (see do football goals follow a Poisson).
How the numbers are made. The sample is ESPN's public Premier League scoreboard: 9,170 matches after removing 357 duplicated 2009-10 rows, of which 9,158 have a goal list that adds up to the final score. A goal shown at minute 67 is placed at 66:30. ESPN records how long each half lasted only from 2016-17, so the main rates run on the regulation clock: the 2,067 goals scored in added time change the state but are not counted, and added time adds no minutes. Using real half lengths from 2016-17 (3,847 matches), the one-goal leader scores 1.38 per 90 and the trailer 1.22, the same pattern. Red cards are handled by cutting each match at its first one: 1,210 matches (13.2%) had a red, and the cut removes 4.3% of the minutes, so the state effect above is not a man-advantage effect. Counting every minute instead gives 1.35 and 1.21. Penalties and own goals count for the side credited, taken from the goal list rather than ESPN's box-score penalty field.
What this cannot show. The archive has goals and red cards with their minutes, but no shots, possession or expected goals by minute, so it measures the score effect on goals only, not the shot-volume and possession swings described below. The strength control is rough: each team's season factors include the very goals being tested, and form changes within a season. ESPN shows whole minutes, so every goal time carries up to 30 seconds of error.
Why it distorts almost everything
Once you see game state, you see its fingerprints on every aggregate stat. The leading team and the trailing team are, in effect, playing two different sports, and the box score blends them into a single misleading average.
- Possession. The clearest victim. A team protecting a lead cedes the ball on purpose; a team chasing the game hoards it because it has to. High possession is therefore as much a symptom of being behind as a sign of control. "Dominating possession" and "losing" are not opposites: they frequently arrive together.
- Shot counts. The trailing team throws bodies and shots forward; the leading team is content to defend and counter. So the team that is behind usually out-shoots the team that is ahead, which inverts the intuition that out-shooting your opponent means you were on top.
- Expected goals. Even xG, the metric built to be more honest than shots, inherits the bias, because xG is summed over shots, and the trailing team takes more of them. A side can lose the match and "win" the xG simply because it spent an hour pouring forward in search of an equaliser. xG totals are more trustworthy than raw shots, but they are not immune to the scoreline. (For what xG does and does not claim, see expected goals explained.)
- Defensive and pressing numbers. Tackles, interceptions, and pressing intensity all swing with game state too: a team defending a lead defends more, presses less, and clears more, none of which means it is a better defensive side.
The slogan that captures the whole phenomenon is simple: losing teams shoot more. It is one of the most robust regularities in football analytics, and it is the reason a raw shot count, or a raw possession figure, can tell you almost the opposite of the truth about who controlled a match.
A real example: Liverpool 1-2 Manchester United, October 2025
On 19 October 2025 Manchester United scored in the second minute away at Liverpool and led for 82.5 of the 90 regulation minutes: Liverpool equalised in the 78th, and United scored again in the 84th. Here is ESPN’s full-match box score next to the clock.
| Team | Goals (minute) | Minutes ahead | Minutes level | Minutes behind | Possession | Season possession | Shots (on target) |
|---|---|---|---|---|---|---|---|
| Manchester United | 2 (2′, 84′) | 82.5 | 7.5 | 0.0 | 36.5% | 50.9% | 12 (4) |
| Liverpool | 1 (78′) | 0.0 | 7.5 | 82.5 | 63.5% | 59.2% | 19 (6) |
Read only the full-match line and Liverpool look the dominant side: nearly two-thirds of the ball and 19 shots to 12. Read it against the clock and the story changes. Liverpool were level for 7.5 regulation minutes and never ahead, so almost every one of those numbers was compiled while they chased. United’s 36.5% was 14.4 points below their own 2025-26 average; Liverpool’s 63.5% was 4.3 points above theirs. What ESPN’s box score cannot show is the split while level, because it publishes match totals with no minute stamps on passes or shots. The early goal did not just change the score; it shaped most of the box score that appears to contradict it.
One match proves little, so we ran the same comparison over the whole 2025-26 season. Of its 674 team-matches without a red card, a team that led for more than half the regulation minutes had 3.4 points less of the ball than its own season average, and was below that average in 65.0% of those matches; a team that trailed for more than half had 2.8 points more, and was above its average in 63.8%. Shot totals barely moved on the same comparison: teams that trailed for most of a match took 0.3 fewer shots than their own average and faced 0.3 more, most likely because the matches a team trails are mostly against stronger opponents who limit its chances. Full-match totals cannot separate the two, and the ESPN archive has no minute-by-minute shot log that could.
How analysts adjust
The fix is not to throw the stats away but to condition them on the scoreline. A few standard techniques do most of the work.
Split stats by game state. The simplest and most powerful move: report numbers separately for when a team was winning, level, and losing, instead of mashing them together. The splits often tell three completely different stories, and the contrast between them is itself informative: a team whose performance collapses when ahead has a coaching problem the full-match average would hide.
Lean on the "while level" and early-game numbers. Performance recorded when the score is level is the closest thing to a neutral baseline, because neither team has yet had its incentives distorted. For the same reason, first-half numbers tend to be less score-affected than second-half ones, since more of the first half is typically played at 0–0. Many analysts treat "while level" xG as a cleaner read on which side was actually better than the full-time total.
Use score-state-adjusted models. The most rigorous approach builds the scoreline directly into the model: weighting or correcting each event for the game state it occurred in, so that a shot taken while 2–0 down is not counted the same as one taken at 0–0. This is closely related to possession-adjusting defensive stats so that volume of opportunity does not masquerade as quality; we cover that machinery in possession-adjusted stats.
When all you have is the box score
This section condenses a separate article first published on 11 June 2026.
Most of the time the data you can actually get hold of is a full-match box score with no timestamps in it. Not every number on that sheet is equally contaminated, and a rough ordering, worst first, is the most useful thing to carry in your head.
- Possession share: worst. Close to a pure function of who is chasing; treat a raw possession percentage as evidence about the score, not about control.
- Shot count: contaminated, less than you would think. Trailing sides do push for shots, but in the Premier League of 2025-26 it did not reliably flip the count: of the 125 sides that conceded first inside 20 minutes, 76 ended with more of the ball and only 50 out-shot their opponent.
- Defensive volume: badly, and in the flattering direction. Tackles, interceptions, clearances and blocks all rise for the side defending a lead.
- Pressing rates: badly, in both directions at once. A full-match PPDA averages a desperate late press with a comfortable early stand-off.
- Expected goals: contaminated, but less. It discounts the low-value attempts score effects generate most of, but it is summed over shots, so it is not immune.
- Per-shot quality: the most robust thing on the sheet. xG per shot, shot distance and the share of attempts from inside the box are ratios; they drift, but they do not swing by a factor of two the way volume does.
The ordering follows from one fact: score effects are overwhelmingly a volume distortion. The closer a metric sits to a raw count, the less you should trust it; the closer it sits to a per-event rate, the more you can.
Without event data you can still spot the distortion. Cover the stat line, look at the result and the goal times, and predict the box score; if your prediction lands, the box score has told you nothing the scoreline had not. Use the goal times as a crude split: a match whose first goal arrives in the 84th minute is nearly uncontaminated, one settled in the ninth minute is almost entirely score effect. Prefer first-half splits where they are published. And watch for the tell-tale mismatch: high possession with low xG per shot is the signature of a side that was behind and shooting from wherever it could; strong xG per shot on modest volume is the signature of a side that was ahead and picking its moments.
Season totals are contaminated too, because game states are not randomly distributed: good teams lead a lot and bad teams trail a lot. The table below splits the 2025-26 seasons of the sides that finished top and bottom, Arsenal (85 points) and Wolverhampton Wanderers (20), by the state each regulation minute was played in.
| Game state | Arsenal: share of minutes | Arsenal: scored per 90 | Arsenal: conceded per 90 | Wolves: share of minutes | Wolves: scored per 90 | Wolves: conceded per 90 |
|---|---|---|---|---|---|---|
| While ahead | 42.7% | 1.36 | 0.55 | 9.0% | 0.58 | 1.17 |
| While level | 48.8% | 1.67 | 0.70 | 48.1% | 0.44 | 1.70 |
| While behind | 8.5% | 2.49 | 0.00 | 42.9% | 0.74 | 1.53 |
| Season, all minutes | 100% | 1.61 | 0.58 | 100% | 0.58 | 1.58 |
Arsenal spent 42.7% of their minutes ahead and 8.5% behind; Wolves 9.0% ahead and 42.9% behind. So the two season totals lean on different kinds of football: about half of each is level play, but Arsenal’s other half is mostly spent leading and Wolves’ mostly trailing. The goal rates move the way the score effect predicts, on small samples: Arsenal scored 1.36 per 90 while ahead and 2.49 in the 289.5 minutes they spent behind (8 goals); Wolves scored 0.44 per 90 while level and 0.74 while behind. The box scores show the same lean in possession: Wolves averaged 44.6% of the ball in the 15 red-card-free matches they spent mostly behind and 40.6% in the 16 where neither side led for most of the match, against 42.3% across the season. That produces three common misreadings: the unlucky-dominance story ("65% of the ball and eighteen shots and lost"), the style misattribution (a team labelled possession-based on totals that mostly record how often it was behind), and the improvement mirage (a side starts winning, chases less, and its falling volume looks like decline).
When no split is available, switch to a ratio, aggregate over enough matches that the game states vary, and state the contamination out loud: "They out-shot the opponent 18–9, though they trailed from the ninth minute" is a complete and honest sentence. The correction has limits of its own. Splitting by state shrinks every sample, so a "while level" figure from one match can rest on twenty minutes of football; game state is entangled with team strength, so "while ahead" numbers are disproportionately strong teams' numbers; and one goal up with eighty minutes left is not the same instruction as one goal up with three left, which is why the more careful models weight by time remaining as well as by margin.
Why this matters beyond a single match
Game state is not only a single-match trap. It feeds straight into the models that try to project a whole season. If a model ingests raw shot or possession totals without accounting for the scorelines they were built under, it will systematically misjudge teams: over-rating sides that spend a lot of time behind and pile up empty volume, under-rating efficient front-runners who win early and then defend. Differences in exactly how models handle this kind of context are one of the reasons two reputable forecasts of the same league can disagree, as we explore in why league projection models disagree.
The practical takeaway is a habit rather than a formula. Before you let any match statistic settle an argument, ask the first question: what was the score while this number was being made? A possession figure, a shot count, even an xG total means something quite different at 0–0 than it does at 2–0 down with twenty minutes left. The scoreline is not context you can add later. It is the lens the whole box score was shot through.
Reproduce it
Every figure in the Premier League section (added November 6, 2026) comes from our analysis script. It reads data_layer/epl_goal_times_2002_2027.json (ESPN's public Premier League scoreboard, pulled 5 October 2026) and, for the real-minutes check, data_layer/epl_added_time_2016_2027.json (ESPN match summaries, pulled 5 October 2026), through our archive loader, which drops the 357 duplicated 2009-10 rows. It rebuilds each match minute by minute, cuts it at the first red card, splits the minutes and goals by game state, fits the expected rates, counts who scored next, writes every number to data/stats/game-state-and-score-effects-epl.json and draws the chart. No network, nothing hand-entered.
The two 2025-26 examples (added November 7, 2026) come from a second analysis script. It reads the same goal-times file and data_layer/epl_2025_26_boxscores.json (ESPN match summaries, pulled 7 September 2026), picks the Liverpool and Manchester United match by a stated rule (the 2025-26 match without a red card, involving a top-three side, in which a team scored first inside 15 minutes, never trailed, won, and had the smallest share of the ball), and writes its figures to data/stats/game-state-and-score-effects-examples.json.
Sources, notes & further reading
- Premier League section (added November 6, 2026): goal and red-card minutes for every completed Premier League match from 2002-03 to 20 September 2026 (9,158 of the 9,170 in the file have a usable goal list), /data_layer/epl_goal_times_2002_2027.json, ESPN public scoreboard (soccer/eng.1), pulled 5 October 2026; half lengths from 2016-17, /data_layer/epl_added_time_2016_2027.json, ESPN match summaries, pulled 5 October 2026. Computed by our analysis script.
- The 2025-26 examples (added November 7, 2026): goal minutes from /data_layer/epl_goal_times_2002_2027.json and full-match possession and shots from /data_layer/epl_2025_26_boxscores.json, ESPN public scoreboard and match summaries (soccer/eng.1), pulled 5 October 2026 and 7 September 2026. Computed by our analysis script.
- StatsBomb: analysis and methodology on score effects and game-state splits.
- StatsBomb open data: event data timestamped and scored so events can be bucketed by game state.
- Understat: match and season xG data for the major European leagues.
- FBref: match logs and advanced stats useful for splitting performance by scoreline.
Published by SoccerAnalytics from public data and the sources named in this piece. How this site is made.