When the Model Returns Nothing: A Lesson on Data Integrity in Professional Sport

The clock on my second monitor in Los Angeles read 3:12 a.m. on 16 May 2026. I was waiting for the data table from the first Bundesliga round after football froze. The table returned no rows. It was not a connection error, not an account error, not a server failure. It returned exactly what it had: a blank. After eighteen years of reading numbers, I realised something no textbook teaches: the hardest part of sports data analysis is not reading the number, it is recognising when there is no number to read.

That blank was not a mere technical glitch. It was a test of professional integrity. With an empty table, I had two choices: impute season averages so the model still produced a plausible-looking output, or keep the blank and write the line nobody in the industry wants to write — insufficient information to assess. I chose the second. That night I lost a small contract. But I kept something far harder to build: the ability to tell a correct number apart from a number that merely looks correct.

Eighteen years ago I started as an esports competitor and then a tournament organiser in Vietnam. Back then we scored by hand on paper, copying from video tapes, where one mistake in a single map ruined the whole sheet. Because I once wrote numbers with my own pen, I carry a reflex many younger colleagues lack: whenever I see a figure, my first instinct is to trace it backwards and ask who produced it, how, and what was dropped along the way.

Before trusting a number, ask where it was born. I say this to every intern on day one. Not to frighten them, but to explain that sports data is an industrial supply chain with many joints that can break, not a tablet on which truth has been carved.

Start at the lowest layer. In major football leagues, event data is produced by two different groups: optical camera systems recording ball and player coordinates, and human annotators labelling every pass, duel and shot from a screen. Cameras measure position but do not understand intent. Annotators understand intent but depend on camera angles, typing speed and an internal rulebook hundreds of pages long. A through ball slightly touched by a defender may be logged as a failed pass by one provider and a completed pass by another. Neither is lying. Neither is complete.

Higher up, every metric carries its own definition. For the same shot from the edge of the box, one system may assign a scoring probability of 0.08 and another 0.14, depending on whether it accounts for defensive pressure, strong foot, or the type of preceding pass. No authority arbitrates. So when two outlets publish different xG values for the same match, readers often conclude one is lying. Most of the time, both are telling the truth according to their own definitions.

Professional basketball followed a similar path, later. From the 2026-14 season, the top North American league installed ceiling-mounted motion tracking across every arena, recording ball and player positions twenty-five times per second. That is what produced concepts like actual distance travelled, overtime sprint speed, or shooting efficiency when guarded inside one metre. But that system only measures what happens on the floor, not how much a knee hurts. So when a team announces a player is resting for load management, motion tracking can neither confirm nor deny it. It simply stays silent.

Esports makes the problem harder in one specific way. There are two entirely separate data layers. The first is public data released by the publisher after each update: pick rates, win rates by rank bracket, ban rates. The second is internal scrim data, which nobody publishes and nobody can verify. When a team suddenly changes its composition in a knockout match, outside analysts can only explain it with two words: new strategy. Nobody has a denominator to prove it. That is the most dangerous kind of blank, because it looks like a conclusion.

Taken together, every number on a sports dashboard is the endpoint of a long chain of assumptions: about camera angle, about definitions, about labelling rules, about how missing data was handled. I am not saying this to dismiss data analysis. I am saying its value depends on whether the reader is willing to trace it backwards.

On 27 August 2026 I sat in front of a screen watching Liverpool host Arsenal at Anfield. I was then a mid-level analyst at a sports data firm in Los Angeles, and I had just been granted access to a new column I had never used, called xG. The match ended 4-0 to Liverpool. Traditional statistics showed the shots were not far apart: Liverpool 18, Arsenal 9. Looking only at that column, I would have concluded it was a somewhat fortunate home win.

The new column told a completely different story. Liverpool recorded 3.6 xG; Arsenal just 0.3. The gap in chance quality was many times larger than the gap in shot count. Being the kind of person who trusts process over instinct, I did not believe it immediately. I printed the entire match dataset, bound it, and cross-checked it against the next ten rounds for both clubs. The result forced me to rewrite how I worked for years afterwards: the xG model predicted direction correctly in roughly 80 percent of cases in my verification sample.

That Liverpool shock did not make me afraid of data; it made me afraid of confidence. If I had trusted a simple scoreline for years, there was no reason to think I would not repeat the same mistake with a new tool — except this time the mistake would look far more convincing.

On 27 June 2026, in Kazan, my model failed most clearly. Germany held 74 percent possession, took 26 shots and generated 1.8 xG against South Korea. Everything in the dataset said a goal was only a matter of time. South Korea took 4 shots for 0.8 xG. The final score was 0-2, with both goals arriving in the 90th and 96th minutes.

The important point is that my model did not miscalculate the scoring probability of individual shots. It miscalculated something else entirely: the probability of generating a shot at all. When a team is forced to chase, the quality of chances it creates falls while the number it concedes rises sharply, because it pushes defenders forward. My old model measured the tip of the iceberg and ignored what lay beneath. After that match I made two variables mandatory in every report: the opponent's PPDA — passes allowed per defensive action — and the real intensity of the match measured by time the ball was in play.

The model was not wrong; the world simply changed while I was not looking. Raw data measures how many chances a team creates, but not paralysis, not how thoroughly a back line has read the opponent, and not the psychology of a group that knows only ten minutes remain.

In May 2026, when football returned to empty stadiums, every home-advantage coefficient in my model broke badly. I took all 157 Bundesliga matches played after the shutdown and found the home win rate fell from roughly 43 percent to roughly 36 percent — far beyond the normal variance of a league.

At first I did not believe it. I split the dataset by month, by the home team's league position, by whether that team still had something to play for. The trend held in every slice I tried. Only after confirming it did I add a new variable called crowd, and reduce the weight of home advantage in every match I analysed.

That same year, the North American basketball market provided a rare natural experiment. The entire closing stretch was staged in a single location, with no team playing at home in any meaningful sense. Home advantage did not shrink; it went to exactly zero. For a modeller this is a gift: one variable isolated from all others, allowing its effect to be measured on its own. The result forced me to rewrite almost the entire home-advantage component of my system.

On 11 July 2026 the European Championship final was played at Wembley. Italy and England drew 1-1 after 120 minutes; Italy won the shootout 3-2. On xG alone, England edged it 1.9 to 1.1. On the scoreboard, Italy were champions. Throughout that tournament, what made me trust Italy was not their attack but the lowest defensive xG conceded in qualifying, around 0.6 per match.

xG is not the truth; it is only a mirror — but a mirror does not know how to lie. That mirror showed me Italy did not need to create more chances than opponents to win a long tournament. They only needed to stop opponents creating any. But the same mirror cannot explain why goalkeeper Gianluigi Donnarumma saved two penalties, or why three young England players missed in the same shootout. Some things sit outside every model, and admitting that does not weaken the model; it makes it more honest.

In esports, the moment a model goes stale comes not from injury or crowds, but from a patch. An update can cut the power of a champion or a weapon overnight, and every model built on the previous patch's data instantly becomes history. The irony is that the market keeps pricing according to old habits for the first few weeks, opening a gap between price and reality. That gap is not free money, because to exploit it you must be certain the patch truly changed the landscape rather than a few numbers on paper.

Worse, the public pick rates publishers release are collected from ranked games across the whole player base, most of whom do not compete at professional level. Those rates are useful for measuring popularity, not for measuring true strength in an organised match. The scrim data of professional teams — the thing that actually decides results — is never published. So when a team wins with a composition nobody predicted, outside analysts have only one honest option: admit they lack the data to explain it. Very few choose that option.

This is where I want to pause, because it is the part of sports analysis where the industry lies to itself most.

When a risk matrix is empty — no row filled in — the correct conclusion is not that there are no risks. The correct conclusion is that there is insufficient information to assess risk. Those two sentences differ completely in meaning and completely in consequence. In my trade, confusing them has cost many people money and, worse, has convinced many people they were analysing when they were in fact guessing.

The Bundesliga lesson of 2026 is also a lesson that correlation is not causation. The fall in home wins coincided with empty stadiums, but also with a compressed schedule, with five substitutions permitted per match, and with abnormal player fitness after weeks of inactivity. I reduced the home-advantage weight, but I would not claim crowds were the only cause. An honest analyst must be able to say that, even though it yields no tidy conclusion to post online.

The greatest temptation for a data professional is to fill blanks. Missing a team's defensive metrics, use the league average. Missing injury data, assume the strongest line-up starts. Each time we fill like that, we create no new information — only new confidence. And false confidence with a heavy weight in a model is more dangerous than a simple model that knows its own limits.

From being taught by data, I built a rule I call the blank index. It measures the share of missing data across a report's entire input, and it must be published next to the conclusion, not buried in an appendix. If a report's blank index exceeds a set threshold, its conclusions must be labelled provisional. It sounds simple, yet very few sports analysis reports I have read include it. Nobody wants readers to know that a third of the evidence behind a call does not exist.

Betting markets follow the same logic. The price of a line is not an objective figure about two teams' strength. It is the aggregate of thousands of individual decisions, including decisions made by models with different assumptions and different data. When a price moves sharply, that is a notable signal, but what it signals must still be investigated. It may be genuinely new information, or merely a chain reaction among people reading the same source.

Small data is what big data always exposes. An error in the labelling layer, a definition changed without notice, a table broken for three weeks — all can be hidden inside a beautiful aggregate. They surface only when you are willing to split the dataset and look at each slice.

I read the footnote column while everyone else reads the scoreboard. That is why I still print reports, mark every source, and write in the margin the places where I am unsure. It is the habit of an eighteen-year veteran. It looks slow, and it has saved me many times from publishing a conclusion I could not myself verify.

On the night of 16 May 2026, that empty table was eventually restored about forty minutes later via a backup feed. I had enough data to finish the report. But I still recorded the event in my notebook, and to this day it is the entry I reread most. Those forty blank minutes taught me something a full season of complete data could not: the value of an analyst lies not in how much available data he can process, but in how honest he can be when the data is absent.

The next round will bring new numbers, and someone will again draw a firm conclusion from too small a sample. The only way to avoid being that person is to ask, before every judgement: if the data I am missing were published, what would my conclusion look like? Whoever dares answer that honestly is already a step ahead of the market — not because they hold more data, but because they understand the limits of what they hold.