The Empty Spreadsheet and the Discipline of the Sports Analyst

3:47 a.m. in Busan, forty minutes after the match ended. I open the spreadsheet named after the date, nine sheets, each of them a layer of questions: patch, format, roster, region, finance, competition rules, risk, public narrative, industry transmission. Nine blank sheets. Not a number, not a team name, not a timestamp. On the phone, three messages from three different people, identical in content: “Got any numbers yet?”

I sat looking at those white cells long enough to realise something I would never have considered six years ago: an empty dataset has its own anatomy. It does not fall silent in one single way. A cell is empty because the match has not been played. A cell is empty because the provider has not opened its API. A cell is empty because that metric was never defined for that league. And a cell is empty because the analyst is lazy. Four kinds of emptiness, four different ways of handling it.

When the data says nothing, the first job of a professional is to classify the silence, not to fill it with feeling. That is the entire spirit of what I call the empty analysis frame: nine pre-built layers of questions, and the discipline to tolerate the fact that most of them will return an indeterminate result.

Foundation before floors

In 2026 I was nineteen, a second-year student in Busan. On World Cup night I fed all 23 shots taken by Germany against South Korea into an xG model I had written myself in Python. The screen returned: 1.32 xG, 0 goals, a 0-2 defeat. I checked every shot. Eighteen of the twenty-three came from outside the box, 78 percent. South Korea's goals arrived in the 90th plus 2nd minute through Kim Young-gwon and the 90th plus 6th through Son Heung-min, after Manuel Neuer pushed forward for a set piece and left his goal empty.

In the Russian night, for the first time I saw a number that could hurt.

The real lesson of that night was not the number. It was how close I came to writing something false, because the data I fed the model was missing three fields: weather, pitch condition, substitution order. None of those three appeared in any statistical table I could download. Had I drawn conclusions from that deficient table, I would have published a conclusion that was arithmetically correct and causally wrong. That kind of error does not expose itself. It only surfaces when a team wins away more often than any model predicted, and people call it character. Character is something I cannot measure. Home win rate I can.

Three questions before typing a single word

Since then, every piece I write follows a mandatory procedure: before discussing victory or defeat, I must first interrogate the numbers. Three questions, always in order. Where does this data come from, which source tier, what is its delay. How many matches are in the sample, and my minimum threshold is thirty matches or three patch cycles. If the sample is zero, what am I allowed to write.

The third question is the hardest, and it decides whether the article exists at all. When the sample is zero, I may only write about method, about limits, and about the list of signals to track in the next round. I may not grade a team, may not call a player the best, may not predict a scoreline.

Based on my experience following matches across seven consecutive seasons, I record the data extraction date next to every figure in a piece, because the same metric pulled at two different moments can diverge after a provider issues a retrospective correction. Readers have the right to know which version of a number they are reading.

Source tiers and delay

Data sources in this profession split into four tiers, each with its own error profile. Tier one is the publisher's official API, slow to update, usually within twenty-four to forty-eight hours, but with stable metric definitions. Tier two is community statistics platforms, faster, but with definitions that shift from one contributor to the next. Tier three is internal club data, the most accurate and never public. Tier four is journalism, and it carries the highest transmission error, because one wrong figure at tier four gets copied hundreds of times.

In my articles I label the source tier of every figure. When two tiers contradict each other, I present both and state plainly which one I lean toward, and why. This approach makes the work slower, and I accept that cost.

The 0.08 coefficient and the price of a broken foundation

In 2026, K League 1 became the first football league in the world to resume in front of empty stands. The xG model I built in 2026 began to drift systematically. I collected 152 matches, compared them with the 2026 season, and found the home win rate had fallen from 46.2 percent to 31.6 percent. I wrote a forty-page report concluding that every 10,000 spectators was worth 0.08 additional expected goals for the home side.

The 0.08 coefficient does not measure the silence; it measures what we lost.

Nobody commissioned that report. I did it because I knew that when the foundation is wrong, every analytical floor built on top of it is wrong too, and that error does not confess itself. In the report I placed the limitations section before the conclusions, because the 2026 season also involved a compressed schedule, more substitutions, and different accumulated fitness. I was able to isolate the crowd variable to a degree, not absolutely. Correlation is not causation, and I wrote that sentence down instead of leaving readers to guess it.

Layer one: patch and meta

Every meta update is a confession from the publisher. It states that the previous balance state is no longer what they want, and usually because one group of champions or one playstyle is holding a win rate that is too high in professional play.

When I analyse a match without patch notes, every tactical judgement is worthless, even when the numbers look handsome. A team winning with a roster that has just been buffed says nothing about their ability in the next patch. I check three things first: when the patch took effect, which positions the changes touched, and whether average game length shifted as a result. If the patch landed after match day, I state clearly in the article that the data belongs to a previous cycle.

The trap in this layer is time. Readers see a team playing well for two weeks and assume it is their peak form. Two weeks is seven matches. Seven matches is too small a sample to name anything, let alone establish a trend.

Layer two: format and schedule density

Format is not back-office trivia. Best-of-three and best-of-five produce two different probability distributions, and a team can be strong in a Bo3 and weak in a Bo5 without changing a single thing about how they play. I always build a schedule density table: matches within ten days, flight hours, rest days between rounds. A one-day gap is negligible. A three-day gap is significant, especially for a team with a thin roster, and especially in qualifiers staged across multiple cities.

The path into a tournament affects data quality. Teams seeded directly into the group stage have a smaller official-match sample than teams that came through qualifiers, so all their metrics carry wider error bars early on. I write that down, instead of comparing two teams using samples of different sizes and declaring one better than the other.

Layer three: the roster

On 8 June 2026, I was the first to report a loan deal with a 2.8 million euro purchase option, involving a Korean midfielder at a mid-table club. The data source I used came from an analytics company in Lisbon.

A transfer fee does not measure talent; it measures the hunger of the buyer.

What I actually want to say in the roster layer is not the transfer story. It is a division problem. That player logged 564 minutes the previous season, while his contract recorded 1,200, meaning actual minutes fell 41 percent against the committed figure. A number like that can come from injury, from a change of system, from a new head coach, or from all three. To separate those three causes I need the injury list, a substitution timeline by minute, and role changes on the pitch. Without those three, I am permitted one sentence about minutes and not one sentence about form.

There is another dimension to a roster: depth. I count the players who passed one thousand official minutes in a season, then weigh that against the number of matches to be played. A team with only twelve players past that threshold will break during a compressed stretch. This is the kind of prediction I am willing to make, because it rests on arithmetic rather than a feeling about morale.

Layer four: the regional map

Regional strength is measured by three indicators: international results over the last three years, academy output into the talent pool, and the quality of practice servers. The third is usually ignored, yet it determines how fast young professionals develop. A region with strong practice servers produces players faster than its international results yet reflect.

The sample in this layer is at least three years. One year is noise. I have never personally seen a regional ranking built on a single season survive into the following one.

Layer five: the money

In the finance layer I check four lines: sponsorship revenue, distributions from the league or publisher, salary expenditure, and capital injection. The transfer race among the giants is a brand arms race. The biggest spender is usually the club with the largest sponsorship revenue, not the club with the best win rate afterwards. I do not turn that into a declaration. I choose case studies so it surfaces on its own, and let readers verify it themselves.

A club paying triple a rival's wages for identical results is burning money to buy a position on the media ranking table. There is nothing wrong with that. It is simply brand business, not sporting business. The genuinely valuable contracts usually sit at small clubs: buy the right gap, sell at the right price, and never pay for a name.

Layer six: rules and governance

The rules framework of a competition covers competitive integrity, transfer and registration rules, contract compliance, protection of minors, and disputes between publishers and organisations. This layer is the only one where a blank cell can be good news: no violation means no data.

But a blank cell here is also the most dangerous, because violations only surface after the ruling is complete. I track it by reading past rulings to derive the average sanction by violation type, then asking what sanction would be reasonable if a similar situation occurred today. With no rulings to read, I mark insufficient information and stop, rather than speculate about the intentions of the parties involved.

Layer seven: the risk profile

Risk splits into six categories: competitive, financial, personnel, regulatory, public opinion, systemic. For each I record three fields: level, probability, impact. When information is insufficient, the level field must stay blank. A less experienced writer will fill it with a mid-point average to make the table look tidy. I leave it blank and note why it is blank.

A risk profile with three honest blanks is worth more than a profile with three blanks filled by guesswork, because a blank still preserves information about where our ignorance lies, while a filled blank erases that location entirely.

Layer eight: public narrative and expectations

Public narrative has a heat cycle. A story with a statistical foundation lives for weeks. A story made purely of emotion lives for days. I measure the gap between market expectation and objective assessment along three fields: team results, individual form, transfer activity. The wider the gap, the higher the probability of reversal.

This is the layer I work through slowest. Most surges begin with a match whose metrics drift outside the normal band and nobody checks the error bars. A hat-trick from three shots is three pieces of luck, and it will repeat at a different rate if the sample is large enough. Every shot that hits the post is a world not yet born, and also a goal that never existed in the data.

Layer nine: industry transmission

Transmission runs in one direction: publishers, then clubs, competitions and streaming platforms, then sponsorship and derivative markets. I record each block with a direction, a magnitude, a time horizon. A change in the first block needs six to eighteen months to reach the last.

In this layer I separate two things that are habitually mixed: the short-term interest of the sponsor and the long-term health of the ecosystem. A large sponsorship deal can drain next season's payroll. The balance sheet of a small club is where the truth shows up first.

Vocabulary is an analytical decision

I strip vague words out of my work. Loss of form becomes minutes played down 41 percent against last season. Being pinned back becomes choosing to sit deep, because the first phrase describes the viewer's feeling and the second describes the team's decision. A match for the ages becomes a match in which two teams generated a combined 4.02 xG and scored nothing from open play.

Every time I swap an emotional phrase for a measurable one, I lose a little allure and regain a little verifiability. In my profession, that exchange rate is profitable.

Counterintuitive: a beautiful sample is not evidence

There is a temptation that runs opposite to leaving cells blank. Once you have spent the effort building a model, collecting data, constructing tables, you want your numbers to say more than they have the right to say. A tidy metric, a handsome chart, a high fit coefficient, and suddenly you speak as though you have discovered a law.

PPDA 25.1 — sitting deep is not a concession, it is stretching the game.

In December 2026 I compiled Morocco's three knockout matches. They conceded 71.6 percent of possession, conceded one goal, while their opponents generated a combined 4.02 xG. PPDA 25.1, nearly double the league average of 13.2. The media called it luck. That framing ignores how Morocco deliberately allowed passes in harmless areas and maintained the structure of their defensive block through extra time. I replaced the phrase being pinned back with choosing to sit deep when describing teams like that, and I stated explicitly the error margin of my xG model across those three matches.

The second danger is the reverse reaction: using insufficient information as a shield never to conclude anything. That is laziness disguised as scholarship. If I reach the end of a piece without a judgement in the present tense, made actively, open to verification, then I have not done my job.

The third danger sits elsewhere. Data analysts are walking into the dressing room. We draw conclusions about things we do not live beside, do not witness at four in the morning after a cancelled flight, do not hear in the technical meeting room. Our conclusions often detach from a team's real rhythm. I hold one rule: at most one predictive conclusion per piece, and that conclusion must carry explicit trigger conditions and a date on which readers can come back and check.

What remains

Back to that night in Busan. After classifying the four kinds of emptiness, I answered the three messages with one sentence: no numbers yet, and the absence of numbers is also a number. Then I wrote at the top of the file the list of signals to track for the next round, each with a trigger threshold and a data source.

I do not write about football. I write about the light that data illuminates.

If you are following a team this season, try one thing: take the three metrics you trust most, record the extraction date and the number of matches in the sample, then read them again in four weeks. Most of us are not wrong in our conclusions. We are wrong about when we reached them.

I still keep that nine-sheet spreadsheet. The day a blank sheet fills one cell with correct data is the day I start writing.