The Empty Cell in the Results Sheet: When Missing Data Gets Read as Zero

The results file from a short-course meet in Brisbane last March had 212 rows. Forty-seven rows were missing split times in the 100m freestyle. Nobody on the pricing desk noticed. The model ran anyway, still spat out a win probability for every lane, and the output sheet carried no warning line at all. That is the frightening part: the system does not crash when data is missing. It goes quiet, then returns a number that looks entirely ordinary.

I re-checked that file in the evening, after the final had ended. The winning team landed exactly where the model said it would. But the indicator I cared about — the back-half speed of two lanes competing for medals — was off by as much as 0.4 seconds over 50m, enough to flip the placings if set against World Aquatics data. The cause was not the algorithm. It was 47 empty cells read as zero.

When an empty cell gets read as zero

Across five years of covering swimming for the Australian market, I have learned that swimming data comes in three layers of very different quality, and almost no documentation distinguishes between them when training a model.

The first layer is data that is measured and published: finish times, reaction times captured by the starting block, splits at every 50m, stroke counts. This is the solid part, with error measured in hundredths of a second.

The second layer is ambiguous data: indicators that exist but have no common standard across meets. Is stroke rate counted across the whole race or only over the sprint segment? Is distance per stroke measured in metres or as a percentage of body length? When two meets use two different definitions, the comparison still runs — it is just comparing two things that are not the same thing.

The third layer is the unmeasurable zone: the feel for the water after a taper week, the tension before a final, the decision to change breathing rhythm at the 150m mark that a coach never publishes. Here the numbers are not wrong. They are simply absent.

The most common error in this industry is to mix all three layers into one table, then handle empty cells in one of two ways: fill in the meet average, or leave the zero untouched. The first erases variance — it turns a wildly inconsistent swimmer into a steady one. The second is worse: it invents a swimmer who never existed, one who covers the opening 50m in zero seconds, and every derived metric is dragged downward with it.

There is a more concrete example I run into constantly. In swimming, the underwater segment after the start and after each turn often decides short-course results. But public data almost never carries that indicator. Poolside cameras record it, coaching staff watch it, and the database released to the public does not. The consequence is that every prediction model for the 50m and 100m is running on a dataset missing its single most important variable, and no empty cell in the sheet tells you so.

My point is not that data is dirty. It is that an empty cell carries information, and that information is usually worth exactly as much as the information in a filled cell.

A map of limits

I call my method drawing a map of limits, and it has three clearly marked zones. The confirmed zone: where data is thick enough to conclude, for example an injury frequency for a young athlete across three consecutive seasons. The ambiguous zone: where data exists but the sample is too thin, for example two finals swum in two different pools. The sensory zone: where you have to rely on the eyes of someone who watched the race, and I have to state plainly which zone I am standing in.

In 2026, assessing the Daniel Arzani transfer for a betting firm in Brisbane, I presented three figures. An average of 8.2 km covered per match, below the 10.1 km benchmark for the Celtic forward group. A dribble frequency of 2.1 per match. And two ACL ruptures in his career. The sporting director pushed back, saying I looked at human beings like machinery. Two seasons later, Arzani had played a total of 20 minutes for Celtic.

The numbers did not decide that. The reading of the numbers did. Injury data is not a random variable independent of playing capacity — it is missing systematically, and that is precisely the class of missing data every regression model mishandles if the person building it does not deliberately separate it from the rest.

Another example, and this one sits on the environment side rather than the human side. In 2026, when European football returned to empty stadiums, I rebuilt the data from historical leagues and found home win rates had fallen by as much as 21% against the five-year average before it. Bookmakers adjusted their handicaps almost two rounds behind the data. The variable that changed was not team quality, but something that had never existed in any database: crowd noise.

The counterintuitive angle

This industry suffers from an illness that is the mirror image of the one I have just described. People fear empty cells more than they fear wrong numbers. An analytical sheet that returns all-unknown gets treated as a failure, while a sheet returning twelve indicators interpolated from three real data points gets treated as professional.

A null result is still a result. When a data file contains no information points at all, the correct answer is not to reason on behalf of the missing data. The correct answer is to say plainly: not enough basis. But saying that sells no tickets, generates no page views, retains no bettors. So the market rewards the fabricator and punishes the honest.

Kazan is the day I learned that a 99% probability can still die on the betting desk. Germany against South Korea in 2026: Germany held 74% of possession but played only 11 passes into the box and reached an xG of 0.7 — lower than their opponent's 0.9. A week later, FIFA released official data confirming every one of those numbers. The data was right. The people reading it were not.

Numbers have no gender. A number has no gender, but the people reading them do. I sat in a press room at Suncorp Stadium in 2026 and predicted Melbourne Victory would come back from 1-0 down at half-time, on the basis of xG of 2.4 against 0.6 and distance covered of 112 km against 98 km. A male commentator laughed and said football is not mathematics. Melbourne won 2-1. But my lesson was not that I was right. It was that I almost did not dare say it, because my sheet had four empty cells in the sprint-distance column, and it took me twenty minutes to decide those four blanks did not collapse the conclusion.

I do not trust emotion. I trust a data series longer than your emotion. But a data series only grows long when the person collecting it is willing to record the places where they could not measure.

Signals for the next cycle

For the rest of this season, what I track is not the average of any indicator. I track the fill rate of the databases I use: what percentage of swims carry all four splits, what percentage carry reaction times, what percentage carry notes on pool conditions. A database with a 71% fill rate will produce a markedly worse model than one at 94%, even when both hold the same number of rows.

For readers of odds sheets, the signal sits elsewhere. When a ranking is published with the name of a recovering athlete missing entirely, that absence is data. When a meet does not publish split times, that too is data about the transparency of its organisers.

What I leave for the next cycle, not to be answered immediately: how strong is your model on the data you do not have?