In the summer of 2026, at the Foro Italico in Rome, organisers of the World Aquatics Championships had to reprint the record board almost every session. Across eight days, more than 40 world records fell, a figure unprecedented in modern swimming history. The anomaly lay not in the number, but in the conditions that produced it: polyurethane swimsuits, which FINA banned outright from 1 January 2026. In the men's 200m freestyle final, Paul Biedermann swam 1:42.00, breaking Michael Phelps's 1:42.96 set in Beijing 2026.

The technical context matters, because most online comparison tables skip it. The 2026–2026 polyurethane suits were airtight structures that lifted the body higher, cut drag and stabilised the stroke line. Their effect was unevenly distributed: the shorter the event, the smaller the effect; the longer the event, the greater the buoyancy advantage. Suits did not shift every record at the same speed.

FINA laid out a roadmap restricting suit thickness and buoyancy from 2026. The result was a noisy data band confined to 2026–2026, sitting in the middle of the continuous series any forecasting model needs. Every regression model on swimming performance since has had to handle a problem with no clean answer: those records are real, but the conditions that produced them no longer exist.

What interests me as a numbers person is the consequence at the data layer. Swimming analysis has three layers: final time, 50m splits, and derived metrics such as stroke rate and distance per cycle. Polyurethane hit the third layer hardest, the second indirectly, and left the faintest trace on the first. A model reading only final times sees nothing unusual. A model reading splits sees it immediately: swimmers lost less speed over the final 100m than they did in other seasons. That pattern does not repeat after 2026.

This is the error class that costs me the most time when building models: the data is not wrong, but the context that generated it is dead. Algorithms do not know that. They learn distributions, and the 2026–2026 distribution pushes the regression line higher than reality. When the 2026–2026 season returned to normal, models that had not been updated kept predicting in an overly optimistic direction, and they erred systematically rather than randomly.

The irony is that those same records created a reverse effect. According to World Aquatics data, Biedermann's Rome 2026 records have still not been erased from the world record board. They exist as a data island in a sea that has cooled. Anyone building a time-series ranking must decide: keep them in place, flag them, or drop them from the sample. No option is neutral. Each tells a different story about who swims fastest.

This is where the story touches my trade. Any swimming dataset has empty cells. A split missing due to equipment failure. An event without enough heats. A rising swimmer without enough samples for a stable average. Empty data is normal. How people handle empty cells is the problem.

Faced with a blank cell, the default reaction of most is to fill it with a story. That swimmer was holding back. She is hiding her cards. He just changed training programmes. Those lines sound reasonable, and that is exactly the danger. A blank filled with a hypothesis becomes a fact on the next read, then a premise on the third analysis. After a few cycles, nobody remembers it was ever blank. Numbers have no gender, but the people reading them do. And those readers have another habit: they fear white space more than they fear being wrong.

Records are usually treated as the most trustworthy indicator, when in reality they are the most misleading. A record is a single data point, measured on someone's best day, under one specific condition, in one specific psychological state. It does not represent average ability, still less future ability.

Amateur analysts default to personal bests as the comparison benchmark, because that is the easiest number to find. That is a convenient choice, not a correct one. When a swimmer goes three seconds slower than their personal best, the right question is not whether she is slowing down, but where those three seconds sit against her own standard deviation over the past 24 months. For most swimmers, that range is far wider than fans imagine.

I once watched a handicap-prediction model skew through almost an entire heat schedule because its whole training set sat inside the noisy band. Nobody cheated. A blank cell simply was not flagged, and a story was written into it.

Limits of the data: this article is about numbers, but the most important part sits outside them. I have no data on how a swimmer feels stepping onto the blocks after injury. I cannot measure the pressure of one Olympic berth inside a family. I cannot quantify what happens in someone's head when they know they are swimming faster than the world record pace at the final 50m. Those things are data too — data for which we simply lack the instruments.

What I keep after rebuilding the swimming dataset each time is not who is fastest. It is: which blank cell in my table is currently filled by a story, and if I remove that story, how long does my conclusion survive?