The Data Void in F1: When Silence Gets Read as Safety

In March 2026, inside AC Milan's data room, I sat in front of a spreadsheet covering 20 Serie A matches from the 2026-17 season and saw something so absurd I had to read it three times. Milan's expected goals figure at the San Siro was 1.85. Away from home, it dropped to 1.02. Yet the team's actual goal count in those two situations was almost identical.

A side that creates nearly twice as many chances at home while scoring at the same rate on the road can be suffering from a psychological problem in attack, or from a data problem. I chose the second hypothesis and spent six days cross-checking video against raw data. The result: a sensor in the south-west corner of the stadium was lagging by 0.2 seconds. Every goalkeeper-initiated build-up was being time-stamped incorrectly, and that error propagated through the entire chain of metrics behind it.

The 14-page internal report I wrote afterwards was not about tactics. It was about measuring equipment. Head coach Vincenzo Montella used the findings to shift more ball circulation to the right flank; Milan won 5 of their last 8 matches and secured a Europa League place. But what I carried with me to today, at 57, is not that ticket. It is the awareness that a system can be wrong for months without making a single sound, because nobody is checking the thing that does the measuring.

Formula 1 today is a sport run by sensors. Each car carries hundreds of data channels: brake temperatures, tyre pressures, torque, battery state, steering angle, longitudinal and lateral acceleration, fuel flow. Outside the car sits another data layer — positioning systems and timing loops around the circuit, plus models that interpolate them into metrics such as expected lap time and tyre degradation rate.

The 2026 technical regulations push the electrical share of the power unit to roughly half of total output, bring active aerodynamics back, and force teams to manage energy differently lap by lap. At the same time, the cost cap and aerodynamic testing restrictions squeeze real track time. Less time on circuit means more decisions taken on simulation, on numerical models and on tracking data.

That is a reasonable economic trade-off. But it shifts risk in a way rarely discussed: when you withdraw one verification source and increase the weight of another, potential accuracy goes up, and the number of failure points that produce no alarm goes up by exactly the same proportion.

Three kinds of silent failure. The first is latency: the sensor still works, still transmits, but transmits late. At Milan, 0.2 seconds was enough to invert a conclusion about an entire attack. In a lap lasting more than 80 seconds, 0.2 seconds slows nobody down, but it is enough to record a braking phase earlier than it happened — and from there every calculation about braking distance, about the point of rotation, about the moment of throttle application drifts with it.

The second is calibration drift. A sensor can be perfectly punctual and still wrong. A tyre reading of 1.9 bar when the true value is 2.0 generates no system error at all; it merely generates a wrong conclusion repeated long enough to become doctrine inside the engineering office.

The third is transmission failure. The data exists on the car but never reaches the pit wall: a failed logging port, a broken extraction step, an empty file. This is the most dangerous kind, because the first two leave a wrong number for someone to doubt, while this one leaves a gap — and a gap produces no error message.

Why gaps get read as safety. In a technical briefing, people look up at the screen, see no red warning, and conclude the car is fine. But a screen with no red warning carries two entirely different meanings. The first is that the system is clean. The second is that the system has lost the ability to speak. Those two meanings look identical on the screen and identical in the minutes.

This is the most common reasoning error in any analysis room, and it does not come from carelessness. It comes from how we are trained: we are taught methods for reading data, and nobody teaches us to read the absence of data. Data only tells part of the story; the rest lies with those who know how to listen — and listening here includes hearing an abnormal silence.

Two data layers are not equals. On-car telemetry is sampled at high frequency, wired directly into the car's electrical system, and carries an internal time stamp. External tracking data comes from devices placed around the circuit, synchronised to each other through a common reference frame, and depends on every receiving station working. When the two layers disagree, the analytical community generally assumes the external layer is wrong. That assumption holds most of the time, but it causes people to overlook the opposite possibility.

At Milan in 2026, nobody doubted the equipment until somebody sat down to cross-check it. Every collapse has a precondition; few people are willing to look at it beforehand.

Confidence tagging. A simple practice almost no team writes down: attach a confidence tag to every number before it enters a decision. High when the figure has been cross-verified through at least two independent sources. Medium when it comes from a single source or is inferred from historical analogy. Low when it is directional only and insufficient to change a car setup.

Formula 1 does not use those words, but the industry has done this instinctively for decades. A chief engineer never assigns the same weight to a track measurement and to a simulation output. The problem is that instinct cannot be transferred through a data file. A confidence tag can.

State the measurement conditions. "The car is 0.3 seconds slower in sector 2" is a meaningless sentence without three things: which timing loop measured sector 2, what the track temperature was, and how many laps the tyre set had covered. Slower by 0.3 seconds than what, under which conditions, at what moment. Every tracking number belongs on the operating table, not on the altar.

What is not in the table. There is another data layer that never passes through a sensor: the human voice on the radio. The rhythm of an answer, the length of a pause before an engineer says "understood", the hesitation when confirming a setup change — these have no unit of measurement and appear in no data file. They are still data. They simply require a different instrument, and that instrument is experience.

A source is also a measuring device. In the driver market, the question of who said it matters as much as what was said. A journalist present in the paddock, an aggregator repeating a story, and an anonymous account do not share the same unit of measurement. Treating them as equal is a calibration error, not a moral one. And in a regular season, where each round is a small chapter, downgrading a rumour requires no counter-evidence — only the absence of supporting evidence.

In the 2026 power unit era, most driver decisions are energy decisions: when to deploy the full battery, when to recover, when to accept a time loss to gain track position. Those decisions rest almost entirely on information supplied by the pit wall. If that data layer is wrong, the driver has made no mistake — he is simply executing an instruction built on a wrong number. The history of this sport is full of such cases remembered as driver errors.

The cost cap consequence. When budgets are constrained, what usually gets cut is the final verification layer: one fewer analyst, one fewer video cross-check session, one fewer pre-season equipment inspection. Those cuts do not show up in the results sheet immediately, so they look cheap. In reality they merely move the cost into the future in the form of undetected error — a cost with no invoice, only consequences.

A review that finds no problems is usually a bad sign. If a team audits its entire data system and identifies no blind spot, the likely explanation is that they have not looked in the right place. The absence of a recorded risk is not the same as the absence of risk. A spotless empty dataset is not necessarily a healthy dataset; it may be a dataset that has stopped working.

The industry is moving in the opposite direction. Every season brings more channels, more sensors, more models. The marginal value of channel number 180 is close to zero. Its marginal cost is not: every new channel is a point that can lag, drift or fall silent, and one more line to read in a meeting that is already too long.

More troubling is the industry's defensive reflex. When a wrong conclusion is discovered, the default response is to add data — more channels, more sources, more models. It is rarely the right response, which would be to check whether the existing measuring equipment is still accurate. We tend to buy a new scale rather than recalibrate the one we already own.

And the final verification layer remains human. An engineer who spends six days cross-checking video against raw data cannot be replaced by an algorithm, because the algorithm itself is the suspect. The cost cap is not wrong. But cutting precisely there is a wrong decision, and it is the kind of mistake that never appears on a timing sheet.

The next competitive edge in Formula 1 probably lies not in collecting more data but in measurement discipline: publishing measurement conditions alongside the number, attaching confidence tags next to conclusions, and inspecting your own instruments before inspecting your rivals. Teams that manage this will not be faster over one lap, but they will be wrong less often across a season. And if your dataset vanished one morning, how long would it take you to notice?