Trang chủInternational FootballThe Empty Column: The Limits of Football's Data Model
International Football

The Empty Column: The Limits of Football's Data Model

**Câu trả lời cốt lõi**: Mô hình dữ liệu bóng đá thất bại vì bị buộc phải điền vào mọi ô, chứ không vì thiếu số liệu. Giá trị thật của phân tích chiến thuật nằm ở việc nhận diện ô nào phải để trắng — như Liverpool 2020-21, khi PPDA tăng từ 9,8 lên 13,4 chỉ ra khoảng trống giữa Robertson và Wijnaldum. **Sự kiện chính**: - Liverpool thua sáu trận sân nhà liên tiếp ở Premier League mùa 2020-21; PPDA tăng từ 9,8 lên 13,4, tương đương chậm gần bốn đường chuyền. - Morocco cầm bóng khoảng 23 phần trăm trước Tây Ban Nha ngày 6 tháng 12 năm 2022, hoà 0-0 sau 120 phút và thắng luân lưu 3-0. - Chelsea trả khoảng 106,8 triệu bảng cho Enzo Fernández tháng 1 năm 2023 và 115 triệu bảng cho Moisés Caicedo tháng 8 năm 2023. - Everton bị trừ mười điểm tháng 11 năm 2023, giảm còn sáu điểm tháng 2 năm 2024, trừ thêm hai điểm tháng 4 năm 2024. - Premier League cho phép thay năm người từ mùa 2022-23, biến hai mươi phút cuối thành cuộc chiến tiêu hao đội hình. **Nguồn**: Phạm Anh, phân tích chiến thuật, cập nhật ngày 13 tháng 8 năm 2026 | Cross-checked: VuaBong.vn **Hỏi đáp liên quan**: - Hỏi: Vì sao tỷ lệ cầm bóng không dùng được để đánh giá trận Morocco – Tây Ban Nha? Đáp: Vì tỷ lệ cầm bóng đo ai giữ bóng, không đo ai kiểm soát không gian mà Morocco chủ động nhường. - Hỏi: Chỉ số PPDA tăng có nghĩa gì? Đáp: Đối thủ được phép chuyền thêm trước khi bị áp sát, tức hệ thống pressing chậm lại ba đến bốn giây. - Hỏi: Vì sao xG không dự báo được các màn lội ngược dòng lớn? Đáp: xG đo chất lượng cú sút theo vị trí và tình huống, không đo được trạng thái tâm lý khi một đội bóng từ bỏ nguyên

On 10 July 2026, in Saint Petersburg, the second half of the France–Belgium semi-final reached the 51st minute when Griezmann swung in a corner from the left and Umtiti rose to head past Courtois. I wrote one line in my notebook: 1-0, minute 51. Then I kept the stopwatch running, counting every second the ball was live and every second it was not, the way I had done all summer. When the whistle went, the notebook read 54 minutes of ball in play for France and 61 for Belgium. Belgium had more of the ball, completed more passes, and went home. France reached the final, then won it.

I wrote more than six thousand words about that match and gave France's approach a name: spatial pragmatism. Controlling space matters more than controlling the ball. I counted twelve high-speed sprints from Mbappé, most of them in the second half, most of them aimed at Belgium's right flank where Vertonghen was overloaded. Belgian supporters attacked the piece. They felt I had disparaged a beautiful way of playing. They were probably right about the emotion, and I was wrong about the method.

Six years later, opening that file again, I noticed something uncomfortable: 54 and 61 do not measure control. They measure how often the referee blew, how often the ball left the pitch, how many seconds players spent on the ground, and the heat of a July evening. My measurement was elegant, tidy, and almost empty. It only meant something when placed beside a different question — where was the space during those 54 minutes.

That was when I started caring about the empty column.

In the dataset I built for the 2026-21 season, the most valuable column was not PPDA and not successful duels. The most valuable column was the one I left blank, because I did not have enough data to fill it, and I chose to leave it alone. Football analytics has taught me a great deal over twenty-two years. The hardest lesson remains the lesson in silence.

When the opponent has the ball, do not look at the ball — look at the space they leave behind.

I first wrote that line in 2026 and I still believe it, but I now understand it differently. Space is not a poetic metaphor. It is a geometric zone with coordinates, a timestamp, and a person responsible for it. And what I have learned after all those years is this: most football data models fail not because they lack numbers, but because they are forced to fill every cell.

A NEW INDUSTRY OF SURGEONS

Twenty years ago a match left an analyst roughly twenty lines of statistics. Today a Premier League fixture generates around three thousand event records with coordinates, plus positional data at ten to twenty-five frames per second for twenty-two players and the ball. A mid-table club's season can produce tens of millions of rows. Nobody reads them all. Nobody can.

The industry's problem has reversed direction. Once the question was how to get more data. Now the question is how to know which of those millions of rows is answering the question you actually asked.

Companies such as StatsBomb and Stats Perform turned event logging into an industry with its own standards. Expected goals emerged from a simple need: a penalty is worth more than a shot from a tight angle, so counting shots equally is meaningless. From there came an entire ecosystem — xG, xGA, PPDA, packing, field tilt, progressive passes, spatial density. Each metric is a way of slicing a match.

Each metric is also a promise. And every promise carries fine print.

I have sat in meeting rooms in Manchester where a fitness coach opened a positional dataset and talked about distances between running lines. I have watched a recruitment director present a dossier on a young Belgian midfielder with twelve charts, and the first question from the room was whether he could handle a derby. No chart answered that. And that is precisely where I want to pause.

The transfer window is where the industry shows its nature most clearly. When a club must commit a hundred million pounds inside four weeks, they do not lack data. They lack the time to verify it. What grows in that gap has a name: the panic premium. A defender signed on 30 January usually costs more than the same defender signed on 30 June, even though his professional value has not changed. The market is not pricing the player. It is pricing impatience.

That is why I begin any deal analysis with contract structure rather than headline fee. A hundred-million-pound fee paid across five years, plus wages, plus agent fees, plus amortisation, creates a different financial obligation from the same fee paid once. Spreadsheets understand that. Models often forget.

WINTER 2026 AND THE GAP BETWEEN ROBERTSON AND WIJNALDUM

In early 2026 European football was played inside bubbles. Liverpool lost six consecutive home league games, something that had not happened in sixty years. I was in Manchester, unable to enter a stadium, and to fight the helplessness I went into the data. Over three nights I reconstructed every Liverpool turnover from public event data.

Their PPDA — passes allowed per defensive action — rose from 9.8 to 13.4. After losing the ball, they were slower by roughly four passes before closing down. Four passes at elite level is three or four seconds. Three or four seconds is enough for a shape to advance and a gap to open.

The easy explanation was Van Dijk's injury. I almost wrote it. But when I separated turnovers by zone, a pattern appeared that I had not gone looking for: most of the delay sat on the left, in the space between Robertson and Wijnaldum. Robertson pushed high by instinct, as a full-back in a high-pressing system does. Wijnaldum dropped deeper to cover a defence missing a centre-back. Between them a vertical strip opened, fifteen to twenty metres long, and opponents learned to switch the ball into it in two passes.

Liverpool did not collapse because of a storm of injuries. Their machine had forgotten the language it operated in.

I wrote 2,400 words on it, and that piece changed how I work. Since then each analysis tracks a single break point rather than listing ten problems. But it also produced a new error that took me two years to see: I became too confident that I had found the right break point.

Event data told me where the ball went. It did not tell me what Robertson was thinking when he chose to advance. Perhaps he was following instructions and Wijnaldum was not. Perhaps both were following instructions that had gone stale after a substitution. My model pointed at a zone. It did not point at who was responsible inside that zone.

That is the first limit, and it is not small.

QATAR: 23 PERCENT OF THE BALL AND A SPATIAL TRAP

In December 2026 Morocco met Spain in the round of sixteen. I watched from Manchester at midnight and logged every time Hakimi advanced. Morocco had about 23 percent of the ball. Spain completed over a thousand passes. The score after 120 minutes was 0-0, and Morocco won the shootout 3-0.

The standard reading is a fairy tale: a North African side defending heroically against a powerhouse. That reading is not wrong, but it discards the most interesting part.

Morocco did not come to Qatar to tell a fairy tale; they came to prove that defending is also a language of poetry.

What Regragui built was not a plain low block. It was a deliberate spatial trap. When Spain circulated to Morocco's right, Hakimi pushed almost to a wing-back's height, producing two effects at once. He forced the opposing full-back to choose between following him and holding position. And, more importantly, he opened space behind himself — but that space sat inside a zone the two centre-backs and the holding midfielder had been drilled to seal within two seconds.

The elegance of the design is that it invites the opponent to throw the ball into a hole with a floor.

Spain passed a great deal and advanced very little into dangerous areas. They controlled the ball in the region Morocco was happy to let them control. That is why I do not use possession share to judge this match. Possession measures who holds the ball. It does not measure who holds the space.

The second half of the story is Bounou. In the shootout he saved two and Sarabia struck the post. I wrote about his dive tendencies in one-on-one situations, and the piece travelled, quoted by a Bayern Munich assistant coach. But I have to be honest about one thing: my sample was tiny. For a goalkeeper facing ten penalties across three seasons, every percentage is fragile. One save shifts the number from 70 to 80.

I presented it as a rule. It was only a probabilistic tendency.

The difference between those two sentences is the entire content of this article.

EURO 2026: THE TWELVE-METRE MAP AND A SOURCE I COULD NOT VERIFY

On 14 July 2026 Spain beat England 2-1 in Berlin. Yamal had turned seventeen one day before the final, and he had been the tournament's central figure since his curled finish against France in the semi-final on 9 July.

After the tournament a man describing himself as a data analyst for the Spanish federation contacted me privately. He said the federation had built a forbidden-zone map for Yamal: receiving on the right half-space, within the final twelve metres from the touchline, using a technique called spatial density to measure which zones favoured a left-footed player operating on the right.

I could not verify that source. No name, no document, no second confirmation. I wrote the piece and flagged it in the third paragraph, and it drew 180,000 views in three days. But the more I reread it, the clearer it became that the most-shared section was not the caveat. The most-shared section was the map.

What I could observe with my own eyes, needing no source at all, was how Spain pulled their full-backs inside. Carvajal or Cucurella repeatedly drifted into the half-space, leaving the flank to Yamal or Nico Williams. The opposing full-back faced a choice with no good answer: follow the man inside and vacate the wing, or hold the wing and leave a midfielder sprinting back. Either option created a receiving zone for Yamal exactly where he wanted it.

That is geometry, and it is observable without positional data. What positional data adds is quantified evidence of frequency and efficiency. What it does not add is an answer to the central question: does Yamal receive there because the system pushes him there, or because he is Yamal and has been finding that spot since he was twelve?

After rewatching Barcelona's youth matches I lean towards the second. But I do not have enough data to assert it, so I say I do not know.

A model does not replace reality. It only narrows the range of things you can say without being wrong.

THREE MATCHES WHERE xG SURRENDERED

There is a category of match where every probabilistic model fails, and these tend to be the ones people remember.

On 25 May 2026 in Istanbul, Milan led Liverpool 3-0 at half-time. Inside the first six minutes of the second half Liverpool were level. No first-half xG model forecast that, because xG measures shot quality from location, shot type and the creating situation. It cannot measure the psychology of a side three goals down with nothing left to lose. Recklessness changes decision quality. Decision quality changes shot quality, in ways a model trained on ordinary matches does not capture.

On 8 March 2026 Barcelona needed three goals against PSG in the last twenty minutes and scored three, the last in the 95th. Across those twenty minutes they played several times their season average of passes into the box. The pre-match model gave them a very low chance of progressing. It was not wrong about probability. It simply could not describe what happens when a team abandons every principle of risk control.

On 4 May 2026 at the Bernabéu, Real Madrid equalised in the 90th minute and led in the 91st against Manchester City, then won in extra time. Rodrygo came on and scored twice in two minutes. I have rewatched that match four times. What I see is not a miracle. I see a team concentrating every available body into one zone, accepting that the rest of the pitch goes unguarded, and an opponent unable to restructure its defence inside so short a window.

All three share one feature worth bolding: a data model describes the normal operating state of a system, while the biggest results come from the moments a system stops operating normally.

That is not a rejection of data. It is a reminder of its scope.

THE TRANSFER MARKET: WHERE MODELS GET BILLED

In the transfer window every error comes with an invoice.

In January 2026 Chelsea paid around £106.8m for Enzo Fernández from Benfica — the highest fee ever paid for a player in a mid-season window in England at that point — after he was named best young player at the 2026 World Cup. In August of the same year they paid £115m for Moisés Caicedo, whom Brighton had signed for roughly £4.5m from Independiente del Valle two years earlier.

Two deals, two entirely different valuation stories. Brighton bought a nineteen-year-old with data on ball recovery and progressive passing, then sold him for more than twenty-five times the price. Chelsea bought a proven Premier League player at the market rate of a club that needed to solve a problem immediately.

Both were rational. Only one was a model-driven investment. The other was buying certainty, and certainty always carries a surcharge.

Brentford is the example I use most when talking to young scouts. They signed Ivan Toney from Peterborough for around £5m after their model suggested his scoring output in League One would translate to the Championship better than the market price implied. That is a model with explicit assumptions, verification, and self-awareness about its limits. They did not buy Toney because he was the best player available. They bought because the ratio between price and probability of success tilted their way.

Based on my experience watching matches, I have noticed something about expensive failed deals: they rarely fail because the player is bad. They fail because the player was bought to solve a problem the club described incorrectly. Antony joined Manchester United for €95m plus add-ons, and the problem was framed as a lack of a one-v-one right winger. The squad's actual problem was a missing tempo-setter in midfield and no structure out of possession. A good winger cannot fix a system with no rhythm.

This is the most common modelling error in recruitment: judging a player by his own metrics instead of judging him inside the system that will surround him.

The fix is not technically complex, only time-consuming. Before signing, reconstruct the player's last ten matches and count where he receives, under how many opponents, and how long his team keeps the ball after he receives. If those numbers match the new team's structure, the deal has a basis. If they do not, every beautiful chart is meaningless.

PSR: WHEN THE SPREADSHEET WAS RIGHT AND THE MODEL WAS WRONG

The Empty Column: The Limits of Football's Data Model

In November 2026 Everton were docked ten points for breaching the Premier League's profitability and sustainability rules. In February 2026 the deduction was reduced to six on appeal. In April 2026 they received a further two-point deduction for a separate period. Nottingham Forest were docked four points the same month. And the 115 charges against Manchester City, published in February 2026, ran across multiple seasons.

I read these rulings as a lesson in modelling, not in club ethics.

The rules were built on the assumption that revenue is forecastable, transfer amortisation is schedulable, and player value is recorded stably. All three are true on paper and all three wobble in practice. Broadcasting revenue can collapse when a contract is restructured. A player signed for £80m can lose book value in one season through injury. And a loss recorded under regulation may not reflect a club's real health, in either direction.

The striking thing is that the docked clubs did not lack data. They had numbers, accountants and advisers. What they lacked was a forecast humble enough to price in scenarios the spreadsheet had no cell for.

In my own work I always ask one question when reviewing a club's finances: if revenue falls fifteen percent for two consecutive seasons, does this wage structure still stand. Most answers are no. And most of those answers appear in no annual report.

That is why I treat financial rules as a modelling problem rather than an accounting one. Accounting records what happened. A model must speak to what could happen. A good model is one with a reserved cell for the possibility that it is wrong.

THE FIVE-SUB RULE AND THE FINAL TWENTY MINUTES

From 2026-23 the Premier League permitted five substitutions across three windows. I supported the change and still do, though for a different reason than most.

The popular justification is player welfare under a congested calendar. That justification is correct and sufficient. The tactical consequences are more complicated.

When a team can make five changes, the starting eleven is no longer the only tactical unit. A match splits into two or three different versions of the same side. That means first-half data forecasts only the first half. Models built on full-match data are blending different systems into one cell.

There is a further consequence I have observed across several seasons: the final twenty minutes have become a war of attrition. A deep squad can introduce three attackers on 65 minutes and change the entire structure of the game. A thin squad endures. The gap between rich and poor now shows not only in the starting eleven but on the bench, and the bench carries more weight than it used to.

This is a case where data is abundant and the causal story is thin. We know goals between the 75th and 90th minutes have risen. We do not know how much is substitutions, how much accumulated fatigue, how much coaches becoming better at reading games from the stand. Those three causes coexist and cannot be separated with pure event data.

I once wrote a piece asserting substitutions were the main cause. I was wrong to assert. I was only right to say they are one of three plausible causes.

THE UNWRITTEN PART IS THE MOST VALUABLE PART

Here is what I consider the centre of this whole story.

Football analytics has a clear incentive structure: the analyst who delivers confident conclusions gets attention, the one who says I do not know gets ignored. A piece with ten firm conclusions travels further than a piece with three conclusions and four blanks. That mechanism is not born of audience stupidity. It comes from how information markets work: certainty is easier to consume than caution.

The consequence is pressure to fill the blank. When there is no data on a player's capacity to handle psychological pressure, people substitute successful duels as a proxy. When there is no data on dressing-room leadership, they substitute years of service. Each time, a real variable is replaced by an approximation, and the approximation enters the model with no footnote attached.

That is why I started naming the empty cells in my own datasets. Capacity to handle pressure: not measurable through event data, requires at least five live viewings under high pressure. Decision quality inside the box: partly measurable through xG, does not capture positioning before receiving. Dressing-room influence: not measurable through public data at all.

Naming an empty cell does not make my analysis better immediately. It makes me wrong less often and, more importantly, it tells me where I am wrong.

There is one more dimension I regard as the industry's biggest blind spot, and it concerns the human variable. I spent years viewing players as nodes in a diagram. That view is useful for understanding systems, but it erases what decides most results: a twenty-three-year-old who has just lost his mother, a defender booed by his own supporters for three weeks, a midfielder negotiating a contract and knowing one injury erases everything. None of those variables sit in any model. They sit in every result.

I have no technical solution to that. I have one principle: when a player behaves in a way that contradicts his historical data, study the person before concluding anything about the system.

It is also why I rewrote my Euro 2026 piece in a shorter version, with clearer doubt and no beautiful map. Fewer people read it. But it is the version I am willing to sign without further explanation.

One more note on the transfer market, where I am working most right now. When I read a rumour I do not ask whether the club wants the player. I ask who benefits from the rumour existing. An agent benefits when his player's value rises. A selling club benefits when a second buyer appears. A buying club sometimes benefits when supporters believe it is trying. Those three motives produce three very different grades of rumour, and none can be verified by reading ten more articles from the same source.

The only method I know is structural: release clauses, remaining contract length, current wage bill, and the selling club's financial pressure. When those four data points align, a deal has a basis. When they do not, the rumour is just a good story.

A TEST FOR THE NEXT MATCH

I am not writing this to say data is useless. I make my living from it and I believe in it. I am writing to say that an analyst's value lies in knowing where he stands on the map of what can be known.

When the opponent has the ball, do not look at the ball — look at the space they leave behind. But when you look at that space, ask one more question: am I seeing it through data, through my eyes, or through my own expectation. Those three answers produce three very different articles, and only one of them is true.

Next match I will do something simpler than any spreadsheet I have built. I will pick one full-back, count how often he advances in the first fifteen minutes, then count the space behind him over the same period. If the two counts do not match what I thought I saw, I will write in my notebook that I was wrong. And I will leave that cell empty until there is enough data to fill it.

Perhaps that is the only thing an analyst can promise without an extra clause: no invention in the places where I do not know. The rest of the match, as always, will be decided by the people on the pitch.