When the Data Feed Goes Dark: Lessons From a 0.2-Second Sensor Lag at San Siro

The third monitor in the analysis room that evening showed nothing but a flat line. No ripple, no peak, no dip. A telemetry channel had gone silent in the middle of a data review, and what caught my attention was not the line itself but the fact that nobody in the room stood up to check it. Three people kept their heads down over their spreadsheets, kept typing, kept nodding to one another as though that flat line were a normal result.

I have sat in enough rooms like that to know one thing: data does not lie on its own, but a data feed will. When the feed goes quiet, people tend to fill the gap with story — a plausible hypothesis, a name, an explanation prepared in advance. A gap in the data does not create truth; it only clears a seat for guesswork.

This is not new to me. In 2026, while serving on the AC Milan coaching staff, I was asked to validate the motion dataset from twenty Serie A matches in the 2026-17 season. The job sounded dull: cross-check one file against another, see which figures matched and which did not. That is where I learned that most errors in sports analysis do not come from the algorithm. They come from the input.

The dataset showed AC Milan's home xG at San Siro was 1.85, while the away figure was 1.02. Nearly double. Read at face value, that is the story of a team that performs far better in front of its own crowd. But when I pulled the actual goals, the two values were level. A team creating nearly twice as much and scoring the same — either the attack was wasteful, or the data was broken. I chose to test the second option first.

Cross-referencing the video, I found the cause: the sensor mounted at the south-west corner of the stands was lagging by 0.2 seconds. Two-tenths of a second. Enough to timestamp every goalkeeper build-up incorrectly, enough to credit a long pass to the wrong receiver, enough to make an entire season look like a different story. I wrote a fourteen-page internal report recommending recalibration. Head coach Vincenzo Montella used the findings to shift more circulation to the right flank; the team won five of its last eight matches and qualified for the Europa League.

Every collapse has a precondition; few people bother to look at it beforehand. Data is no different. A 0.2-second drift does not ruin a training session. It smoulders, it runs through match after match, and only when the end-of-season table appears does anyone act surprised.

In a modern match, data does not come from a single source. It comes from dozens of measurement points: GPS units worn on players' backs, sensors inside the ball, multi-angle cameras, and at some stadiums, antenna arrays built around the stands. Each point has its own sampling rate, its own error, its own calibration routine. When they agree, the picture is sharp enough that people forget it was stitched together from scattered fragments. When one fragment is off, the whole picture tilts — but in a way nobody notices, because the rest still looks fine.

Every tracking number belongs on an operating table, not on an altar. PPDA, xG, transition counts, high-intensity distance — all of them are stories that have been compressed. Compression loses things. A low PPDA says a team presses hard, but it does not say where the press happens, with how many players, or at what point in the match. The same value placed in two different contexts describes two different teams.

There is a distinction few outside the industry notice: precision and accuracy. A sensor can repeat its reading perfectly, off by exactly 0.2 seconds every single time, and therefore look entirely trustworthy. Systematic error is more dangerous than random error because it produces no outlier to make anyone suspicious. It produces a pattern. And humans are very good at finding meaning in patterns born from faults.

An expected model, whether xG or any other predictive index, is only a way of asking a question. It answers the question its designer wanted to ask, using the data its collector managed to gather. Ask the wrong question, or ask the right question of a skewed dataset, and the answer still reads fluently. A model does not know how to doubt itself.

I once took thousands of taunts for a single short post. World Cup 2026, Germany against South Korea, minute 70: I wrote that Germany's defensive line was holding an average of 68 metres high, that their press had failed 17 times, and that South Korea had already produced 12 counter-attacks; if they did not drop the block, the goal would come from an aerial situation. In the 90th minute plus three, Kim Young-gwon scored exactly to that script. Many said I had turned emotion into arithmetic. But Gazzetta dello Sport reprinted my distorted-trapezoid diagram of the German back line, and that is what I remember.

The lesson lay elsewhere. The figure of 68 metres high helps nobody picture anything. But when I wrote that the gap between centre-back and goalkeeper was as wide as a vertical rectangle, that the back line had unzipped like a broken zipper all the way to the valve box, readers began to see the match. Data only tells part of the story; the rest lies in whether people know how to listen. A figure only lives once it is translated into spatial imagery.

On the pit lane, the thing I watch most closely is not the timing sheet. It is the pitch of the engineer's voice on the radio. When a driver reports the car sliding at turn four while the sensors say grip is still inside the safe window, the control room goes quiet for a beat longer than usual. That pause is data. The hesitation in the reply is data. The engineer lowering his voice to tell the driver everything is fine — that is data too, and it appears in no table.

The driver's body is also a sensor. It has its own error, its own fatigue, its own memory of an old impact. So when a driver's feel diverges from the numbers, the first move is not to pick a side but to check which one was calibrated most recently. This is why I put a note on measurement conditions at the top of every analysis, and why I never quote a figure that has not been cross-checked against at least two sources.

An empty grandstand is the same kind of problem. An empty grandstand does not kill the contest, but it takes away something the numbers cannot measure. I am not talking about vague emotion. I mean countable things: how many times the referee has to intervene, the delay in a defensive line's reaction after each call, whether radio traffic thickens or thins. The only way to reach that is through observed behaviour, not a spreadsheet.

Over a long season, error does not sit still. It accumulates. Each round adds a little drift to the larger picture, and by April people are analysing a team that never existed. That is why I distrust conclusions drawn from an uncleaned dataset, however well they match the general mood.

At a race, the same set of tyres can tell two different stories depending on where track temperature is sampled. A sensor in the shade of the grandstand and one in the middle of the track differ by a few degrees, and a few degrees is enough to flip the call on pitting a lap early or a lap late. Nobody is wrong. Two people are simply reading two different truths.

What troubles me most in this industry is not the wrong analysis. It is the confidence built on a foundation nobody inspected. Teams add analysts, buy software, build dashboards. Very few hire one more technician to calibrate the sensors. The lowest-paid person in the building is deciding the quality of the most expensive decisions. That is the execution blind spot, and it appears in nobody's report.

Meanwhile contracts are still signed on the strength of unverified tables. A midfielder with a beautiful progressive-passing figure arrives, and cannot be fitted into the system already running. A contract only looks good on paper until someone tries to fit it into a running system. When it does not fit, blame usually lands on the player; rarely does anyone go back and inspect the sensor that produced the original figure.

So I keep an old habit. Before each round I look at the calibration sheet before the results sheet. I read the technical notes before the predictions. It sounds slow, but it is far cheaper than writing a long analysis of a line that died at the start of the session.

The next round will bring more numbers, more charts, more conclusions delivered very smoothly. The question I carry into this weekend is simple: of all those figures, how many did somebody stand up and verify before they became a headline?