The winning formula moved the same way — the 2022 champion won on moments (shots on target, z +1.10), and the 2026 champion won on structure (completed line breaks, z +1.97). Yet both champions pressed less than the average of their own tournament. All 168 official FIFA Post-Match Summary Reports were parsed in full and checked against 336 team-matches and 5,287 player-matches; one conclusion that did not survive the check is corrected in the body text itself. The conclusions are gathered in Chapter 12.
The sides that won in 2022 were the kind of side that runs less,
and the sides that lost in 2026 lost because they pressed more.
One problem has to be settled before anything else. Put FIFA's two reports side by side and compare the numbers, and you end up reading a change in FIFA's ruler as a change in football.
Forced turnovers fell from an average of 71.1 per match to 41.5 — a 42% drop. Ball-winning cannot have nearly halved in 4 years. The quartile ranges of the two tournaments do not even overlap (2022: 38–104; 2026: 23–59). This is a change of definition.
So all 35 metrics were audited on three criteria — change in mean · change in standard deviation · quartile overlap — and sorted into three grades.
Bar length = the % change from the 2022 mean to the 2026 mean. Bar colour carries the grade — this is the only figure in the report that uses the grade colours. Metric names are exactly as they appear in the official FIFA reports (hover a bar for a plain-language description).
Each metric is converted to a z-score within the distribution of all team-matches in its own tournament, and the resulting z-profiles are what get compared.
z(team, metric) = ( team per-match mean − field mean of that tournament ) ÷ field standard deviation of that tournament
This single conversion solves three problems at once. Definition changes → the field moves with them, so they cancel out 32 teams→48 teams dilution → dividing by the width of the distribution normalises it and the scale now matches the question itself: "how does a side differ within its tournament's trend?"
All 64 of the 2022 PMSRs were checked. The line Expected Goals appears not once. FIFA did not carry xG in the official reports for the 2022 tournament.
Without xG you cannot measure who created the better chances. Shot counts alone are not enough — a shot struck 1v1 with the goalkeeper inside the box and a hopeful effort from 30 metres out both count as 1.
Filling the gap with third-party xG data was an option, but it was not used. Every provider calculates xG differently. Set provider A's 2022 xG next to FIFA's 2026 xG and you can no longer tell a tournament difference from a provider difference. So a different route was taken.
2026 has xG, and it has the other metrics too. So a computer can be asked to find "the rule that predicts xG from the other metrics". If that rule is accurate enough, the same rule applied to 2022 gives "the value xG would have shown". This is what is called the Threat Index.
You cannot just throw any metric at the computer. Three conditions did the filtering.
Condition one — it must exist in both tournaments. A metric missing from 2022 cannot be applied to 2022.
Condition two — it must be graded "stable" in Figure 1. If FIFA changed the way it counts something, applying it to 2022 returns nonsense. This condition is why receptions in the final third were dropped — an excellent measure of threat in principle, but its maximum jumps from 174 in 2022 to 426 in 2026, which points to a change of definition.
Condition three — they must not overlap with each other. This is where the first attempt went wrong. It is explained in full below.
It keeps coming up from here on. It runs from 0 to 1, and the closer to 1, the better the fit. What matters is knowing exactly what it measures.
Think of a game where you guess your friends' heights. With no information at all, the best you can do is answer "170cm, the average" every time. That is the baseline. Now let someone tell you their shoe size and you will do far better.
R² measures "by what % did you cut the error, compared with just guessing the average". R² = 0 means you did no better than guessing the average; 1 means perfect. Bigger is better.
In football data, anything in the 0.6s is fairly high. It is a sport played by people, and the remainder is territory that was never predictable in the first place.
The first model used six ingredients: shots on target · total shots · final-third phase share · crosses · completed line breaks · defensive line breaks. Accuracy was good at R² 0.687, but the crosses coefficient came out at −0.0236. That reads as "the more crosses you put in, the less threat you create".
This is not football, it is a statistical side effect. Take crosses on their own against xG and the correlation is r = +0.404 — clearly positive. Sides that crossed more also had higher xG.
The problem was that the ingredients overlapped badly. Total shots and shots on target correlate at 0.78; crosses and final-third phase share correlate at 0.78. In effect the same thing was being counted twice, or three times. In that situation a regression fights over the shared portion and shoves one coefficient into the negative (multicollinearity). It does not mean crosses subtract threat; it means the model adjusted so that crosses would not re-count the share the shots had already taken.
Statistically that is explicable, but you cannot use a metric that collides with the football common sense that "a cross is an active attacking act that delivers the ball into a dangerous area". So the ingredients were redesigned.
| What changed | Why |
|---|---|
| Total shots dropped | Correlates 0.78 with shots on target. It counted the same thing twice and was the main culprit in the fight over coefficients |
| Crosses → completed crosses | Counts only the deliveries that actually reached a team-mate. The correlation with xG also rises from 0.404 to 0.474 |
| Every coefficient fixed at 0 or above | The principle that "attacking acts do not subtract threat" is nailed into the calculation rule itself. The computer cannot use a negative even if it wants to |
The third one is the key. When you are confident a principle holds in football, constraining the data in advance so it cannot break that principle is the honest approach. In statistical terms it is called non-negative least squares.
That leaves these five ingredients.
| Ingredient | What it counts | Why it is threat | Correlation with xG |
|---|---|---|---|
| Shots on target | Shots that went inside the frame | The most direct threat | +0.75 |
| Completed line breaks | Passes played through an opposition defensive line | Volume of progression | +0.57 |
| Final-third phase share | Share of time spent in the opposition 1/3 | Time held in the danger area | +0.56 |
| Defensive line breaks | Breaking through the opposition's last defensive line | The most dangerous progression | +0.53 |
| Completed crosses | Balls delivered from wide that found a team-mate | Sending the ball into the danger area | +0.47 |
How much each of the five ingredients should count was not decided by a human. The computer looked at all 208 team-matches from 2026 and found the combination with the smallest error against actual xG. Under one constraint: every coefficient must be 0 or above.
By analogy: you know the weight, height, age and shoe size of 208 people and you are filling in the question marks in "weight = ?×height + ?×age + ?×shoe size", with one extra rule that "a question mark cannot be negative". The result looks like this.
How to read the numbers. The 0.1716 in front of shots on target means "1 more shot on target adds 0.17 of xG". 6 shots on target gives 0.17 × 6 ≈ 1.0 xG. That matches the feel of anyone who knows the game.
Defensive line breaks 0.0110 against completed line breaks 0.0036. A factor of 3. The data says that putting the ball behind the last defensive line is far more threatening than slipping a pass between lines in the opposition half. No human set that; it is the output of a calculation across 208 matches.
Completed crosses 0.0111, defensive line breaks 0.0110. Effectively identical. It means connecting a ball from wide into the box is worth the same as breaking the last line through the middle. That is the opposite conclusion to the first model, which treated crosses as a deduction — and it agrees with football common sense.
To see how well a formula holds, run it on a match where the answer is already known. The 2026 final has actual xG.
| Spain — actual xG 2.52 | ||
|---|---|---|
| Ingredient | Value | Contribution |
| Starting value | — | −0.154 |
| Shots on target | 12 | +2.059 |
| Completed line breaks | 182 | +0.656 |
| Final-third phase share | 22% | +0.298 |
| Defensive line breaks | 22 | +0.241 |
| Completed crosses | 4 | +0.044 |
| Total | 3.145 | |
| Argentina — actual xG 0.07 | ||
|---|---|---|
| Ingredient | Value | Contribution |
| Starting value | — | −0.154 |
| Shots on target | 0 | +0.000 |
| Completed line breaks | 87 | +0.314 |
| Final-third phase share | 5% | +0.068 |
| Defensive line breaks | 5 | +0.055 |
| Completed crosses | 1 | +0.011 |
| Total | 0.293 | |
Spain: 2.52 actual against 3.15 predicted. Argentina: 0.07 actual against 0.29 predicted. It does not hit the exact value, but it gets the order and the size of the gap right — the verdict that Spain were more than 10 times as threatening as Argentina is the same verdict actual xG gives. The next figure confirms this was not the luck of one match, across all 208.
xG is a value produced separately for each side. So every match has four bars — the home side's actual xG and prediction, the away side's actual xG and prediction. From top to bottom, matches are ordered from the smallest gap between the two sides' xG to the largest
Read it in two steps.
First, are the black bar and the blue bar for the same side similar in length. If they are, that side's xG was predicted well.
Second — and this matters more — within a single match, was the higher of the two sides identified correctly. Look at the final at the very bottom: actual xG is Spain 2.52 against Argentina 0.07, and the Threat Index is 3.15 against 0.29. The values differ a little, but the verdict "Spain were overwhelmingly on top" is the same. That verdict is what this report actually uses.
Across all 104 matches, the rate at which "which of the two sides was on top" was called correctly is 77.9%. But that rate splits sharply by type of match.
| Actual xG gap between the sides | Matches | Correct call on which side was higher |
|---|---|---|
| 0.02 – 0.49 (dead level) | 26 | 53.8% |
| 0.52 – 1.13 | 26 | 69.2% |
| 1.16 – 1.67 | 26 | 92.3% |
| 1.68 – 3.77 (one-sided) | 26 | 96.2% |
In the 59 matches where the xG gap is 1.0 or more it calls 94.9% correctly. Where the gap is under 0.5, it drops to 53.8% — coin-flip territory.
That is actually a reassuring result. What this report deals in is the large difference between sides that created threat and sides that did not, not tight matches decided in the second decimal place. Splitting into 4 quartiles and comparing top with bottom is precisely the way to look at that large difference.
| Model | R² (accuracy) | Cross-validated R² | Winner accuracy | Coefficient signs | Verdict |
|---|---|---|---|---|---|
| v1 · total shots included, unconstrained | 0.687 | 0.650 | 78.8% | crosses −0.0236 | discarded |
| alt · drop crosses only | 0.665 | 0.639 | 78.8% | final third −0.0067 | the problem just moved |
| alt · all 6 ingredients non-negative | 0.664 | 0.645 | 78.8% | no negatives | crosses and final third died at 0 |
| v2 · total shots out + completed crosses + non-negative | 0.615 | 0.591 | 81.2% | all positive | adopted |
R² fell from 0.687 to 0.615, but winner accuracy rose from 78.8% to 81.2%. Those two numbers measure different things — R² is "how precisely does it reproduce the xG value itself", winner accuracy is "how often does it get right which side was the more threatening".
The way this report uses the Threat Index is the latter. It splits sides into 4 quartiles and rules on which of two sides was on top. The ordering matters more than the exact decimal. So the trade taken was to give up a little R² in return for ordering accuracy and football validity.
Why cross-validation is needed. The computer built the formula by looking at those 208 cases, so of course it fits those 208 well. It is like writing your answers after seeing the exam paper.
So this was done instead. The 208 cases were split into five pieces; the computer is shown only four pieces to build the formula, then asked to predict the piece it was not shown. Rotate the pieces and repeat five times. The result, 0.591, barely drops from the original 0.615 — which means it learned a real rule rather than memorising.
The threshold set before any calculation began was "discard it if R² is under 0.55 or cross-validated R² is under 0.50". The result, 0.615 / 0.591, passed, so it is used. Had it fallen short, the index would have been thrown away and shots on target alone would have carried the work — and that failure would have been written up right here too.
And one thing to hold on to alongside it — this is not xG, it is an approximation of xG. Winner accuracy is 5.0 percentage points below actual xG. Chapter 11 checks which way this approximation distorts the conclusions (the short answer: it comes out more conservative than reality).
Anyone who watched Qatar 2022 shares an impression. Nobody pressed furiously high up the pitch, and nobody sat huddled in front of their own goal either. Everyone stood in the middle. That impression survives intact in the data.
Split the out-of-possession phases into high block, mid-block and low block, and calculate each side's share within those three, and this is what comes out.
| Out-of-possession phase | Qatar 2022 | North America 2026 | Change |
|---|---|---|---|
| High block / press | 10.4% | 10.1% | almost none |
| Mid-block | 48.9% | 43.1% | −5.8%p |
| Low block | 40.8% | 46.8% | +6.0%p |
In 2022 the mid-block is 48.9% and the low block 40.8%. In 2026 that reverses: mid-block 43.1%, low block 46.8%. The high-block share sits near 10% in both tournaments, essentially unchanged — what changed was not "how high you press" but "whether you stand in the middle or drop off."
A high share is only a trend in itself. The question is what happened to the sides that followed it. Sides were split into thirds by their tournament-average mid-block share, and the mean rounds reached was calculated for each group.
Rounds reached: 0 = group-stage exit … 5 = reached the final. 2022 is a perfectly monotone staircase; in 2026 the staircase disappears
Splitting sides into those whose mid-block share exceeds their low-block share and those where the reverse holds gives the same picture. In 2022 the 22 mid-block-dominant sides reached a mean of 1.77 rounds, against 0.70 for the 10 low-block-dominant sides. In 2026 it is 1.53 against 1.14 — the gap shrinks to less than half.
| 2022 semi-finalists | Mid-block | Low block | High press | Verdict |
|---|---|---|---|---|
| Morocco | 33.6% | 22.7% | 3.1% | mid-dominant |
| France | 31.6% | 16.6% | 3.6% | mid-dominant |
| Argentina | 27.0% | 15.9% | 5.4% | mid-dominant |
| Croatia | 22.4% | 21.3% | 5.6% | mid-dominant |
| 2026 semi-finalists | Mid-block | Low block | High press | Verdict |
|---|---|---|---|---|
| France | 25.1% | 11.6% | 6.0% | mid-dominant |
| Spain | 16.1% | 12.4% | 9.2% | mid-dominant |
| Argentina | 20.9% | 25.6% | 5.4% | low-dominant |
| England | 16.1% | 25.0% | 7.5% | low-dominant |
In the 2022 semi-finals all four sides spent more time in a mid-block than in a low block. In 2026 half of them are flipped. The FIFA Technical Study Group (TSG) reached the same conclusion — "all four semi-finalists used the mid-block as their principal defensive phase", "sides holding 33% possession or less recorded 14 wins and draws against 5 defeats."
Whether sides could stand in the middle because they were strong, or became strong by standing in the middle, cannot be separated with this data. Weak sides are pushed into a low block from the start. The staircase in Figure 3 must not be read as "use a mid-block and you reach the semi-finals."
One thing is certain, though — that staircase existed in 2022 and does not exist in 2026. If the same endogeneity operates in both tournaments and the staircase appears in only one of them, the difference comes from the character of the tournament, not from team strength.
To explain why the mid-block won, we have to look at the other side of it. The opposite of the mid-block is not the low block but excess. What happened to the sides that ran more and pressed more.
Here the hypothesis splits in two. "Running a lot" and "pressing a lot" are usually talked about as one thing, but in the data they behave completely differently.
Bias one — matches played. A side that reaches the final plays 7-8 matches; a side that goes out in the group stage plays only 3 of them. Counting "which sides turn up in this band" means the sides that survived longest automatically appear more often. That is not evidence of quality, it is arithmetic.
Bias two — extra time. 5 matches in 2022 and 9 in 2026 went to extra time. Team distance covered in those matches averaged 145km, more than 30% above matches decided in normal time (112km). They ran 30 minutes longer, so that is only to be expected.
So two things were done. ① Extra-time matches were converted to a 90-minute basis (120 minutes × 3/4), and ② the unit is the team, not the team-match, with that side's per-match average as the figure. This way every side is counted exactly once.
Every side in each tournament was ranked by distance covered per match, converted to 90 minutes, and split into 4 quartiles. Q1 is the sides that ran least, Q4 the sides that ran most. The vertical axis is that group's mean rounds reached — higher means further (0 = group-stage exit, 5 = the final)
In 2022 the half that ran less overwhelms the rest. The 16 sides in the lower half of distance covered have a mean of 2.00 rounds reached, the 16 in the upper half 0.88. That is a 2.3-fold gap.
More decisive is the team list below. All 4 of the 2022 semi-finalists sit in Q1 and Q2, that is, in the half that ran less. Q1 has Argentina and France; Q2 has Morocco and Croatia. Q3 and Q4 hold none of the 4 semi-finalists.
2026 is different. The 4 semi-finalists are scattered across Q1·Q2·Q3 — and the gap is far narrower, 1.46 for the lower half against 1.12 for the upper.
Comparing the 4 quartiles mixes side strength together. The cleanest test is comparing the two sides inside the same match. Of two sides that ran under identical conditions for 90 minutes (or 120), did the one that ran less win?
Each line is one match. Left is the side that ran less, right the side that ran more. Bold line = the side that ran less won
The tight rounds admit no exceptions. France ran less than the opponent and won 4 matches in a row — 106.4km to 107.8km against Poland (3-1), 109.2km to 112.8km against England (2-1), 106.9km to 112.5km against Morocco (2-0). In the semi-final Argentina won 3-0 while running 200m less than Croatia.
Apply the same calculation to the 2026 knockouts and it is 12 of the 28 matches that produced a winner (42.9%) — a coin toss. The effect disappeared completely in 4 years.
This is where the received wisdom breaks. The sentence "the sides that ran a lot and pressed a lot lost" sounds like a description of 2022 — but where pressing is concerned, 2022 shows no relationship at all.
The same method as Figure 4 above. Sides were split into 4 quartiles by pressures per match, converted to 90 minutes. Q1 is the sides that pressed least, Q4 the sides that pressed most
In 2022 the volume of pressing barely separated results. The lower-pressing half sits at 1.69, the upper-pressing half at 1.19. And the 4 semi-finalists are scattered across Q1·Q2·Q4 — Morocco made the last 4 while sitting in the group that pressed most in the tournament.
2026 is another matter entirely. 1.92 for the half that pressed less, 0.67 for the half that pressed more. Almost a 3-fold gap. And all 4 semi-finalists are in Q1 and Q2 — Argentina in Q1, Spain, France and England in Q2 — Q3 and Q4 hold none of the 4 semi-finalists.
Q4 holds 12 sides: Australia, Czechia, Cabo Verde, Ghana, Iran, Austria, DR Congo, Uzbekistan, Paraguay, Curaçao, Saudi Arabia and Jordan. Mean rounds reached 0.58 — most of them went out in the group stage.
In 2022 the volume of pressing did not separate winners from losers. It was random, somewhere between 34% and 44%. In 2026 the steepest slope in this entire dataset appears — from 55.8% in the least-pressing of the 4 quartiles down to 11.5% in the most-pressing 4th.
In other words, the 2022 lesson was "run less" and the 2026 lesson is "press less and win it back more". They point the same way, but the layer being measured is different. In 2022 the question was still "how much do you run"; by 2026 it had become "what do you spend that running on".
— But whether "run less" is a prescription or a by-product cannot be settled from this section alone. The first check in Chapter 11 answers that question.
Half supported. The "running a lot" half is strongly confirmed in 2022 (81.8% in the knockouts); the "pressing" half is disconfirmed in 2022 and confirmed in 2026 — the original hypothesis had folded the phenomena of two tournaments into one.
Chapter 11 revises the conclusion of this section. Hold side strength constant and compare within the same side, and the direction of distance covered flips (mean points-per-match gap −0.30). Pressing survives the within-side comparison as well (+0.82). → Chapter 11 · Checks
Now that the character of the tournament is known, the question narrows. Within that trend, how did the two sides that reached the final differ?
Start with the point where Argentina stood out from the tournament. Convert each of the 16 metrics to a tournament-relative z-score and only one axis rises clearly.
How many standard deviations each side sits from the mean of its own tournament. 0 = tournament mean. The checkboxes switch sides on and off
Argentina 2022 were not a side that overwhelmed the tournament on structural metrics. Line breaks +0.73, possession +0.65, final-third phase share +0.65 — all of it "decent enough". The one place where they were clearly ahead of the field was shots on target. It is the profile of an ordinarily good side that spiked at the finish.
France are another matter again. The metric on which France deviated most is mid-block share (+1.06). Shots on target is only +0.39 by comparison. France were the side that executed the tournament's dominant trend most faithfully.
Follow possession in match order and the same pattern comes out of both finalists.
In match order from group match 1 through to the final. A line moving down means the side had less of the ball. Point size is that match's Threat Index — the bigger, the more threatening
| Round | Argentina 2022 (winners) | France 2022 (runners-up) | ||||
|---|---|---|---|---|---|---|
| Opponent | Possession | Result | Opponent | Possession | Result | |
| Group 1 | Saudi Arabia | 62.0% | 1-2 loss | Australia | 56.2% | 4-1 win |
| Group 2 | Mexico | 51.0% | 2-0 win | Denmark | 44.1% | 2-1 win |
| Group 3 | Poland | 67.4% | 2-0 win | Tunisia | 56.4% | 0-1 loss |
| Last 16 | Australia | 53.2% | 2-1 win | Poland | 48.7% | 3-1 win |
| Last 8 | Netherlands | 44.7% | shoot-out win | England | 36.1% | 2-1 win |
| Last 4 | Croatia | 34.4% | 3-0 win | Morocco | 34.1% | 2-0 win |
| Final | France | 46.5% | shoot-out win | Argentina | 40.6% | shoot-out loss |
Argentina came down from 67.4% against Poland to 34.4% against Croatia, and that 34.4% match was a 3-0 win. France came down from 56.4% against Tunisia (a defeat) to 34.1% against Morocco, and that match was a 2-0 win.
For both, the match with the lowest possession was the match they won most convincingly. And both lost matches in which they had more of the ball — Argentina lost to Saudi Arabia with 62%, France lost to Tunisia with 56.4%.
This is not coincidence but the logic of a mid-block tournament. When the opponent stands in the middle, the ball comes easily but no threat travels with it. In 2022 having the ball was not an advantage but a task.
The formations tell the same story. Argentina 2022 moved between 4-4-2 · 4-1-2-3 · 3-5-2 across 7 matches (1 match not recorded in the FIFA original), and France alternated between the two shapes 4-1-2-3 ↔ 4-2-3-1.
Yet 4 years later Spain played 7 of their 8 matches in a single 4-1-2-3 and won it. Conversely, Argentina 2026 used four different shapes and finished as runners-up.
The side that changed most won, and the side that changed least won. Formation variation is a cheap way of keeping the opponent from knowing what to prepare for the next match; it is not a condition of winning. That is why this item is recorded as disconfirmed in the Chapter 9 scoring.
The tournament, expanded to 48 sides, wore a different face. The mid-block staircase had gone, and another staircase appeared in its place.
The low-block share in 2026 is 23.2%, higher than the 19.1% of 2022. But that figure has the 48-side expansion effect mixed into it and must not be read at face value (metric-definition audit grade). What is certain is the relationship within a tournament.
| Low-block share | Q1 (lowest) | Q2 | Q3 | Q4 (highest) |
|---|---|---|---|---|
| 2022 win rate | 31.2% | 31.2% | 46.9% | 43.8% |
| 2026 win rate | 48.1% | 48.1% | 42.3% | 15.4% |
In 2022 the top quartile by low-block share (mean 32.5%) won 43.8% of its matches — higher than the bottom quartile's 31.2%. Dropping off was a valid strategy. In 2026 the group in the same position collapses to 15.4%.
And then there is pressing. Looking again at Figure 6 from Chapter 3, the 2026 pressing-quartile staircase (55.8 → 55.8 → 30.8 → 11.5%) is the steepest gradient in this entire dataset. Converted to pressing efficiency, it falls from 59.6% to 9.6%.
Comparing win rates by Threat Index quartile across the two tournaments produces an interesting difference.
2022: 21.9 → 34.4 → 40.6 → 56.2% · 2026: 7.7 → 30.8 → 46.2 → 69.2%
In 2026 the group that generated the most threat wins 69.2% of the time. In 2022 the same group stops at 56.2%. The other end behaves the same way: the group that generated the least threat won 21.9% in 2022 but only 7.7% in 2026.
In 2022 it was hard to win even when you created good chances, and possible to survive even when you did not. The Threat Index also picks the winner 69.4% of the time in 2022 against 81.2% in 2026. A tournament of the mid-block turns out to mean a tournament with few decisive moments at either end.
The 2026 champions stood in exactly the opposite place from the 2022 champions.
Return to the z-profiles in Figure 7 once more. Every point at which Spain 2026 dominated the tournament is a structural metric.
| Metric | Argentina 2022 | Spain 2026 | Difference |
|---|---|---|---|
| Completed line breaks | +0.73 | +1.97 | +1.24 |
| High-press share | +0.14 | +1.50 | +1.36 |
| Forced turnovers | −0.03 | +1.10 | +1.13 |
| Defensive line breaks | +0.69 | +1.23 | +0.55 |
| Possession | +0.65 | +1.09 | +0.44 |
| Shots on target | +1.10 | +0.86 | −0.23 |
| Pressures | −0.60 | −0.27 | +0.33 |
| Mid-block share | +0.50 | −0.71 | −1.21 |
The only metric on which Argentina 2022 stood out in their tournament was shots on target. Spain 2026 stood out in line breaks, high press and forced turnovers, and their shots-on-target z is in fact lower than Argentina's.
The two sides do have one thing in common. Both pressed less than the mean of their own tournament (−0.60, −0.27). What changed is what they did with the energy they saved — the forced-turnover z flips from −0.03 to +1.10.
Forced turnovers is a metric where a definition change is suspected. No raw comparison was made; only each side's relative position inside its own tournament was compared. A z-score normalises mean and variance at the same time, so even if the definition changed, "where they sat inside the tournament" still holds.
Compare the defensive blocks of the 2 champions using the measured values from the FIFA diagrams and they are strikingly alike. Low-block line height is 18.9m for Argentina 2022 and 16.6m for Spain 2026 — a gap of only 2.3m. The mid-blocks are 36.1m against 39.2m.
Where the block was set did not separate these champions. Only what came out of that block did.
Everything up to here has been about sides. The last layer is the individual. The FIFA PMSR carries, for each player, pressures (direct and indirect) and distance covered by speed zone (Zone 1–5). All 5,287 player-matches from the two tournaments were pulled.
Zone 1 is 0–7 km/h — in practice, walking. Calculate walking as a share of total distance covered and you see exactly how a player budgeted his 90 minutes.
In the 2022 final Messi covered 12,109m in total. Of that, 7,128m was walking. In the same match Enzo Fernández walked only 4,905m of his 15,055m.
Every outfield player with 4 matches or more. x-axis = walking share (%), y-axis = attacking output per match (shots + goals×2 + take-ons + offers in behind). The top right is the "walks a lot, creates a lot" zone
The first step was to find each side's primary finisher from the data. The criterion is the sum of two things — that player's share of all his side's shots, and his share of all his side's runs in behind. It measures who carried the job of shooting and running into the space behind.
The answer from the four sides is this.
| Side | Primary finisher | Of side's shots | Of side's runs in behind | Total = finishing share | Of side's dribbles past |
|---|---|---|---|---|---|
| Croatia | Perišić | 37.8% | 40.1% | 77.9 | 32.3% |
| France | Mbappé | 33.7% | 29.1% | 62.8 | 61.5% |
| Argentina | Messi | 45.1% | 11.0% | 56.0 | 42.9% |
| Morocco | En-Nesyri | 23.3% | 32.1% | 55.3 | 6.4% |
All 4 carried more than half of their side's attack on their own. Messi took 45.1% of Argentina's shots, Mbappé 61.5% of France's dribbles past opponents. Perišić accounts for 40.1% of Croatia's runs in behind. In a match played by eleven men an even share for one man is 9%, and these four took 3 to 7 times that.
In loading the attacking role onto one man, the four sides of the last 4 were exactly alike. Up to this point there is no exception.
Next came the pressing load of the same four players. Position adjustment is essential — without it you get an artefact in which the isolated lone striker of a weak side comes out on top (the first attempt returned Saudi Arabia, Serbia and Cameroon in places 1 to 3 and was discarded). The median pressures for a forward at the 2022 tournament is 33.0 per match. That value is set to 1.00 for the comparison.
| Side | Primary finisher | Pressures / match | Vs forward median | Walking share | Defensive load | Semi-final result |
|---|---|---|---|---|---|---|
| France | Mbappé | 22.2 | 0.67× | 51.7% | 33% lighter | 2-0 v Morocco won |
| Argentina | Messi | 27.6 | 0.84× | 57.9% | 16% lighter | 3-0 v Croatia won |
| Croatia | Perišić | 32.0 | 0.97× | 40.3% | at the median | lost 0-3 v Argentina |
| Morocco | En-Nesyri | 46.3 | 1.40× | 30.7% | 40% heavier | lost 0-2 v France |
France gave Mbappé 63% of the attack and asked him for only 67% of the forward median in pressures. Argentina likewise gave Messi 45% of the shots and cut his pressing to 84%.
Croatia handed Perišić 40% of the runs in behind and still asked him to press at the median (0.97×). Morocco gave En-Nesyri both the shooting and the running in behind, and made him press 40% more.
And in the semi-finals, the two sides that lightened the load beat the two that did not, 3-0 and 2-0 respectively.
The last 4 split precisely along that line. Argentina (0.84) against Croatia (0.97), France (0.67) against Morocco (1.40). In both matches the side with the lower pressing ratio for its star won without conceding.
This result breaks the original observation into two pieces. "Load the attack onto the star" was a condition common to the last 4; "and relieve him of defending in return" was the fork in the road to the final.
The four players named at the outset were Messi, Mbappé, Modrić and Ziyech. Checked against the data, Croatia's primary finisher was not Modrić but Perišić, and Morocco's was not Ziyech but En-Nesyri.
And Modrić (pressing ratio 1.19×) and Ziyech (1.38×) both pressed more than the median for their own position. The two of them were not conserving energy; they ran more, not less. On this point the original observation is disconfirmed.
What holds is that Messi and Mbappé were exactly right. And their two sides met in the final.
| Tournament | Finalist | Star | Walking share | Tournament rank | Positional median |
|---|---|---|---|---|---|
| 2022 | Argentina (champions) | Messi | 57.9% | 1st / 100 | FW 35.8% |
| 2022 | France (runners-up) | Mbappé | 51.7% | 2nd / 100 | FW 35.8% |
| 2026 | Spain (champions) | Yamal | 48.4% | 13th / 195 | FW 39.9% |
| 2026 | Argentina (runners-up) | Messi | 64.7% | 1st / 195 | FW 39.9% |
The 10 top attacking players with the lowest pressing ratios in 2022 include Kane (0.64) · Saka (0.66) · Son Heung-min (0.78) · Lewandowski (0.80) · Olmo (0.80) · Gakpo (0.81). England went out in the last 8; Korea Republic, Poland, Spain and the Netherlands went out before that.
At side level the correlation is rounds reached × how far the star is exempted from pressing, r = −0.08. That is 0 for all practical purposes.
Taking the star out of the defending is neither a necessary nor a sufficient condition for the last 4 — what is true is that all 4 cases of reaching a final, across both tournaments, were of that type without exception. This should be read not as "do this and you reach the final" but as "the sides that reached the final were doing this".
The official FIFA physical analysis points the same way. Dr Paul Bradley records that "Messi, Mbappé and Lewandowski all three sit in the low-volume, low-explosiveness quadrant", and gives the interpretation as "outstanding game intelligence makes them extremely selective about when to run at high intensity". In the individual case study he writes that "Messi is below the role average on almost every metric, and beats the average only in on-ball share and quality".
The two tournaments share one uncommon natural control. Argentina won 2022 and finished 2026 as runner-up. Same country, same star. Team strength is, by design, partly held constant.
Where Argentina stood inside each tournament's field. A bar running to the left means it became relatively weaker
| Metric | ARG 2022 z | ARG 2026 z | Change |
|---|---|---|---|
| Defensive line breaks | +0.69 | −0.05 | −0.74 |
| Ball progression | +0.46 | −0.24 | −0.70 |
| Shots on target | +1.10 | +0.52 | −0.58 |
| Mid-block share | +0.50 | −0.07 | −0.57 |
| Second balls | +0.22 | −0.32 | −0.54 |
| Pressures | −0.60 | −0.07 | +0.53 |
| Low-block share | −0.34 | +0.19 | +0.53 |
| Forced turnovers | −0.03 | +0.31 | +0.34 |
It moved from a mid-block to a low block, from a side that pressed less than the field to a side that pressed about as much as everyone else, from a side that stood out on shots on target to an ordinary one. The result was a runner-up finish, and in the final Argentina's xG was 0.07 — the lowest value among 2026's 32 knockout matches.
Take Messi alone and the picture reverses. His walking share went up further, 57.9% → 64.7%. That was the choice of a 38-year-old. The side moved toward running more; the star moved toward conserving more. In 2022 side and star pointed the same way; in 2026 they split.
Comparing the two tournament means across the 16 metrics graded stable in the metric-definition audit, the largest movements are shots (+10.3%) and direct pressures (−11.0%). Most sit within 10%.
Within that, the 3 block-geometry measures are almost completely still — high-block line height 47.8m → 48.5m, low block 17.8m → 18.2m, and mid-block line height goes 37.55m → 37.56m, a difference of 1cm.
Different tournaments, a different 48 sides, values measured 4 years apart, and they agree to this degree: it means the consensus on where to set the defensive block is already settled. What changed is not the position but what a side does from that block.
I write down where the data supported my thinking and where it contradicted it, together. Drop the items that were contradicted and this report stops being a check and becomes promotion.
In 2022 the amount of pressing had nothing to do with results: the 4 quartiles ran 34–44%. In 2026 the least-pressing quartile won 55.8% and the most-pressing quartile 11.5%, the steepest slope across the 4 quartiles anywhere in this dataset. Both champions pressed less than their own tournament mean (−0.60, −0.27), and the champions' forced-turnover z flipped from −0.03 to +1.10 between the two tournaments. Over 4 years this claim became stronger.
And this claim is the only conclusion in this report that passes the team-strength control — compared within the same side, 23 of 32 sides took more points in their "less pressing" matches, with a mean points-per-match gap of +0.82 (Chapter 11). Running volume did not pass.
The 2022 champion stood out on a single axis, shots on target (+1.10); the 2026 champion stood out on structural metrics (its shots-on-target z is in fact lower). A title that leans on individual moments is observed in 2022; in 2026 it weakens. But a new piece of evidence appeared, walking share — in all 4 cases of a side reaching the final, the star ranked at the top of the tournament for walking. To run the individual variable you have to bank the energy for it: for the first time that is confirmed with physical data.
Win rate rises monotonically across the 4 Threat Index quartiles in both tournaments (2022 25→56.2%, 2026 11.5→71.2%). Yet the two sides that reached the 2022 final won most convincingly in the matches where they had the least of the ball (Argentina beat Croatia 3-0, on 34.4% possession, and France beat Morocco 2-0 on 34.1%). Proactivity is not having the ball, it is creating threat.
Argentina 2022 won with 3 shapes, and Spain 2026 won with what was effectively 1 shape (4-1-2-3 in 7 of 8 matches). Argentina 2026 used 4 shapes and finished runner-up. The side that changed most won, and the side that changed least won. Variation is a cheap way to leave an opponent unsure what to prepare for; it is not a condition of winning.
The attempt to use defenders' share of ball progression as a proxy for role fluidity failed in the 2026 analysis (the top of the 4 quartiles won 23.1%). That metric was measuring not role fluidity but a side whose progression is blocked. This time again I found no alternative. It stays unresolved.
Mid-block share, low-block share, distance covered and the star player's pressing volume are all symptoms of a side's strength as much as anything else. Weak sides sit deep, sides being pushed back run more, and only strong sides can afford to spare their star. The staircase in Figure 3 must not be read as causation.
The nearest thing to a defence in this report is the head-to-head within the same match (Figure 6) and the 4-year comparison within the same country (Chapter 8). Both hold side strength partly constant. Even so, it is not a clean separation.
The 81.8% in the 2022 knockouts rests on the 11 matches that were decided. It leaves out the 5 draws (penalty shoot-outs), and n is small. Only 2 sides reach the final per tournament, so the figure "4/4" is 4 cases. It makes a strong impression, but the statistical power is low.
8 of the 35 metrics are graded definition change suspected (forced turnovers, total offers to receive, take-ons, step-ins, receptions in the final third, share of offers in behind, transition passes, Zone 4 running). Zone 4 was dropped entirely, and the rest were used as z-scores only. Even so, any sentence in which these metrics appear is less reliable than the sentences around it.
The Threat Index is an approximation with an R² of 0.615 behind it. Its winner hit rate of 81.2% is 5.0pp below real xG's 86.2%. Every conclusion computed with this index could come out slightly differently with real xG. Check 5 in Chapter 11 establishes the direction of that distortion — it comes out more conservative than reality.
This is the most important limit. PMSR is match-total data, so the time structure inside a sequence is invisible. There is no way to know how many seconds apart the line breaks happened, or how long it took to reach the final third after the defensive line was broken. Nor how many repetitions it took before Lure and Shock worked.
Next step: from event-level logs, measure (a) the standard deviation of the interval between line breaks and (b) the time from a defensive line break to entry into the final third. That is the point at which the analogy becomes a number.
Whether the star not defending is what took the side to the final, or whether being a side bound for the final is what made sparing the star affordable, cannot be separated with this data. Chapter 7 is the observation "the sides that reached the final were doing it this way", not a prescription.
Chapter 10 listed the limits. Listing them is not enough. Here each conclusion is actually run through a check. And one of them did not pass.
The biggest risk in this report is endogeneity. If weak sides run a lot because they are being pushed back, and press a lot because they are being pushed back, then "the sides that ran more lost" is just another way of saying "the weak sides lost".
The cleanest way to separate the two is to compare within the same side. Split one side's matches into halves by that side's own running volume (or pressing volume), and compare points per match in the half where it did less against the half where it did more. Side strength becomes a constant by design.
Sides with 4 or more matches only. Each line is one side. Left dot = mean points from the matches where that side did less, right dot = mean points from the matches where it did more
On 2022 distance covered, of the 16 sides that played 4 or more matches, only 4 took more points in "the matches where that side ran less". 10 were the other way round, and the mean points-per-match gap is −0.30. Within the same side, that is, results were actually better in the matches where they ran more.
This collides head-on with the between-side comparison in Chapter 3 (Figures 4 and 5). Put the two results together and the reading narrows to this — "the sides that ran less won", not "running less makes you win". The sides that did well in 2022 were structurally low-running sides (mid-block sides), and those sides won. But a side running less on a given day did not win more that day.
The sentence in Chapter 3 is corrected: not "the sides that ran more lost" but "the sides that won in 2022 were the low-running kind of side" is the accurate version. Running volume is most likely not a cause but a by-product of a cause called the mid-block. And that reading meshes exactly with the Chapter 2 mid-block staircase.
Run the same calculation on 2026 pressing volume and 23 of 32 sides took more points in "the matches where they pressed less", with a mean points-per-match gap of +0.82. A gap of 0.82 points per match is large.
It is the only claim whose direction holds when side strength is held constant. The conclusion in this report that comes closest to causation is pressing. Not running volume.
The 2022 knockout figure of 81.8% rests on 11 matches. Whether that number could be chance was checked two ways.
Bootstrap — put the 11 results you have into a bag, draw one out, write it down, put it back, and repeat 11 times to build a fake tournament. Do that 20,000 times. What you get is a picture of "if this tournament were played again 20,000 times, roughly where would the result land".
| Tournament | Matches | Lower-running side wins | 95% confidence interval | Probability of chance | Verdict |
|---|---|---|---|---|---|
| 2022 knockouts | 11 | 81.8% | [54.5%, 100%] | p = 0.033 | not chance |
| 2026 knockouts | 28 | 42.9% | [25.0%, 60.7%] | — | no relationship |
The confidence interval is the range that holds the middle 95% of those 20,000 results. 2022 comes out at [54.5%, 100%], which does not contain 50% — not chance. 2026 comes out at [25.0%, 60.7%], which does contain 50% — meaning there is no relationship.
p = 0.033 means that "the probability of getting a result this strong purely by luck, when there is no relationship at all" is 3.3%. Below 5% is normally taken as not chance.
That said, a lower bound of 54.5% also means it cannot tell "almost always" apart from "slightly favoured". The honest wording is this — that the side which ran less was favoured in the 2022 knockouts is confirmed, but the size of that edge cannot be measured precisely with this sample.
The staircase in Figure 3 is what came out of splitting 32 sides into 3 groups. With only 32 sides, that shape could appear by chance.
It was checked with a permutation test. Shuffle the sides' rounds reached as if drawing lots, stick them onto sides at random, then compute the correlation. Repeat that 20,000 times and you build "the pile of values you get when there is no relationship at all". If the actual observed value sits outside that pile, it is not chance.
| Tournament | Sides | Mid-block × rounds reached | Probability of this value at random | Verdict |
|---|---|---|---|---|
| 2022 | 32 | r = +0.443 | p = 0.011 | 1 time in 100 — not chance |
| 2026 | 48 | r = −0.165 | p = 0.263 | middle of the pile — no relationship |
r (the correlation coefficient) is a number between −1 and +1 showing how far two values move together. In football data, +0.443 is fairly strong. That the staircase is there in 2022 and not in 2026 is confirmed statistically.
The moment a researcher picks the number of splits, they gain the freedom to pick whichever result suits them. Splits into 2 / 3 / 4 / 5 groups were all computed, and all of them are published here. Below is the mean rounds reached going from the top mid-block group → the bottom group in each split.
| Split | 2022 (top → bottom) | 2026 (top → bottom) |
|---|---|---|
| 2 groups | 2.00 → 0.88 | 1.04 → 1.54 |
| 3 groups (used in the text) | 2.40 → 1.27 → 0.73 | 1.19 → 1.12 → 1.56 |
| 4 groups | 3.00 → 1.00 → 1.12 → 0.62 | 1.25 → 0.83 → 1.33 → 1.75 |
| 5 groups | 2.83 → 1.50 → 1.29 → 1.00 → 0.71 | 1.22 → 0.90 → 0.78 → 1.60 → 1.90 |
In 2022 the top group goes furthest in all four splits, and in 2026 all four go the other way. However the number of groups is chosen, the conclusion does not change.
In 2026 both are available, so they can be compared directly. Is the staircase built from real xG the same shape as the staircase built from the Threat Index.
Taking the same data and splitting it into 4 by real xG against splitting it into 4 by Threat Index, is the win rate in each band the same shape. Q1 is the 25% with the least threat, Q4 the 25% with the most
The 2026 pressing conclusion leans on "pressing efficiency = pressures ÷ forced turnovers", and forced turnovers is graded definition change suspected. That metric was dropped completely and the calculation was re-run on pressures alone.
Pressures only — 2026: 55.8 / 55.8 / 30.8 / 11.5% · 2022: 34.4 / 37.5 / 43.8 / 37.5%
A single metric whose definition did not change (pressures is graded stable, change −10.2%) reproduces the 2026 staircase as it was, and the 2022 non-relationship as it was. This conclusion is independent of the metric-definition changes.
| Claim | Side strength controlled | Sample test | Threshold sensitivity | Independent of definition drift | Overall |
|---|---|---|---|---|---|
| 2026 — sides that press more lose | +0.82 | n=208 | passes | passes | most robust |
| 2022 — mid-block sides go further | side-level metric | p=0.011 | 2–5 groups | passes | robust |
| Threat Index as an xG substitute | — | CV R² 0.65 | — | stable metrics only | conservative estimate |
| The finalists' star player walking | cannot be controlled | n=4 | — | passes | observation (not a prescription) |
| 2022 — sides that run more lose | −0.30 | CI [54.5, 100] | — | passes | reinterpreted as a by-product |
The sentence this report began with was "the sides that lost in 2022 are the sides that ran more, and the sides that lost in 2026 are the sides that pressed more". After the checks, the sentence changes to this.
The sides that won in 2022 were the low-running kind of side, and the sides that lost in 2026 lost because they pressed more.
The first half is correlation and the second half is close to causation. Leaving that difference standing inside the sentence is the most this data allows.
Eleven chapters, reduced to one. Mix conclusions of different confidence into a single sentence and the weakest eats away at the credibility of the strongest. So how confident I am is attached first, and they are written out as three separate things.
The least-pressing quartile won 55.8% of its matches. The most-pressing quartile won 11.5%. The steepest gradient in the whole dataset. All 4 semi-finalists are in Q1 and Q2 — Q3 and Q4 hold none of them.
And this is the only conclusion that passes all three layers of checks — compared within the same side, 23 of 32 teams took more points in the matches where they pressed less (mean +0.82), and if forced turnovers, whose definition changed, are removed entirely and the sum is run on pressures alone, the staircase still comes out. It is not explained by team strength, and it is not explained by FIFA changing its ruler.
Mean rounds reached for the top 1/3 in mid-block 2.40 against 0.73 for the bottom 1/3 (permutation test p=0.011, 2·3·4·5-way splits all in the same direction). All 4 of the last 4 were mid-block dominant. In the knockouts the side that ran less won 9 of the 11 matches that were decided (81.8%).
But within the same side the direction reverses (mean points-per-match gap −0.30). So the accurate sentence is not "run less and you win" but "the sides that won in 2022 were structurally the kind of side that runs less — that is, mid-block sides". Distance covered is not the cause; it is a by-product of a cause called the mid-block.
Messi (57.9%, 1st) · Mbappé (51.7%, 2nd) · Yamal (48.4%, 13th) · Messi (64.7%, 1st). But the correlation at team level is r = −0.08, effectively 0 — and many of the sides that took their star out of defensive work went out in the group stage. It is not a condition for reaching the last 4 but a condition observed only at the summit.
Sharper than that is the line that divided the last 4 itself. All four sides funnelled the attack through one player; that was the same. What divided them was whether they lightened his defensive work in exchange — Mbappé 0.67× · Messi 0.84× beat Perišić 0.97× · En-Nesyri 1.40×, 2-0, 3-0 respectively.
All three conclusions speak to "what changed". Yet the most striking number in this report is on the side that did not change. Different tournaments, 48 different teams, measurements taken 4 years apart — and the position of the defensive block is almost completely still.
| Block line height | Qatar 2022 | North America 2026 | Difference |
|---|---|---|---|
| High block / press | 47.8 m | 48.5 m | 0.7 m |
| Mid-block | 37.55 m | 37.56 m | 0.01 m |
| Low block | 17.8 m | 18.2 m | 0.5 m |
The agreement on where to stand is already settled. It holds between the two champions as well — the low-block line height of Argentina 2022 and Spain 2026 differs by only 2.3m. So the difference cannot come from position; it can only come from what comes out of that block.
And the two champions did stand in the same place and make different things. The single metric on which Argentina 2022 led the tournament was shots on target (+1.10); where Spain 2026 led was completed line breaks (+1.97) · high press (+1.50) · forced turnovers (+1.10). Spain's z for shots on target is in fact lower than Argentina's (+0.86). The 2022 champion won with moments; the 2026 champion won with structure. And both pressed less than the mean of their own tournament (−0.60, −0.27).
That is what the data said; below is the shape it takes when applied to a side. ① can be used as a prescription; ② and ③ are material for judgement only.
The moment total pressing becomes a training target, you enter 2026's Q4 (win rate 11.5%). The number to manage is not pressures but pressures ÷ forced turnovers — how many times you squeezed to win the ball once. In 2026 the win rate across the 4 quartiles of this metric ran from 59.6% at the top to 9.6% at the bottom. Where the same force is spent has replaced how much of it is spent.
Where the block is set has not moved 1cm in 4 years, and between the two champions the gap is only 2.3m. Arguing about how many metres up to set the block is reopening a question that already has an answer. What has to be designed is the output that comes out of that block — how many times the line was broken after a recovery, and whether that line break carried through to the final third.
In the 2022 final Messi walked 7,128m of his 12,109m. If the attack is going to be funnelled through one player, his defensive load has to be adjusted with it — that was the fork between the 4 semi-finalists. But this is an observation, not a prescription — this data cannot rule out the possibility that these were strong sides because they could afford to spare the star, and more sides made the same choice and still went out in the group stage.
The scoring in Chapter 9 reduces to one line. One supported, one conditional, one disconfirmed, one unresolved.
| Concept | 4-year verdict | Basis |
|---|---|---|
| Defence is structure, not effort | More valid as time went on | The only conclusion that passed the team-strength control (+0.82) |
| The individual variable | Conditionally valid | Won 2022 on shots on target, weakened in 2026; in exchange it gained a new physical basis in walking share |
| Proactivity | Valid — but the axis is not possession | Both finalists won most convincingly in the match where their possession was lowest (34.4% 3-0, 34.1% 2-0) |
| Formation variation | Disconfirmed — not a necessary condition | Won with 3 shapes (ARG22) and won with effectively 1 shape (ESP26). ARG26 used 4 shapes and finished runner-up |
| Multi-functionality | No way to measure it | The defender ball-progression share proxy failed again. Left unresolved |
Not deleting the disconfirmed item is a condition of this report. That formation variation is not a condition for winning is a result that differs from what I thought, and it stays as it is. Variation is a cheap way of keeping the opponent from knowing what to prepare for the next match; it was not itself the reason for winning.
PMSR is match-total data. So none of the conclusions in this report has a time axis. How many seconds apart the line breaks came, how many seconds it took from breaking the defensive line to the final third, how many low-threat repetitions stacked up before a high-threat moment arrived — none of it can be seen in this data.
So the next piece of work is not tournament data but event-level logs. The two things to measure are already fixed — (a) the standard deviation of the interval between line breaks, (b) the time from breaking the defensive line to entry into the final third. If this report got as far as "what was done and how much", the next report starts from "when it was done".
Where to set the block did not move 1cm in 4 years.
What changed was what came out of that block,
and the sides that lost tried to make up the difference with volume of pressing.