Data Cannot Swim: When Numbers Fail Before the Water and Human Emotion

I have spent 5 years hunting for anomalies in swimming data. I believe in data sequences longer than your emotions. But Kazan was the day I learned that a 99% probability can still die at the betting table. And that is why I am writing this article — not to praise the power of analytics, but to map the limits of it.

Hook: When the model collapses in the water

In July 2026, at the Australian National Swimming Championships in Brisbane, I sat in the analysis room with three screens full of data. My model — built from 14,000 historical races, integrating power indices, stroke efficiency, and pool-adjustment coefficients — predicted that 19-year-old swimmer Liam Carter would break the national record in the 200m freestyle with 87% probability. He had swum 1:45.32 in the heats, 0.8 seconds faster than his previous best. Every metric was green. His improvement slope over the past 18 months was 2.1% per quarter — an unprecedented trajectory for his age.

At exactly 19:42, the starting signal sounded. Carter dove in with a reaction time of 0.62 seconds — 0.04 seconds faster than his average. First 50m: 24.1 seconds, right on plan. 100m: 50.8, still within expectations. But at 150m, I saw something no spreadsheet would ever capture: Carter's left shoulder began to sit 3 cm lower than his right. His stroke rate dropped from 42 cycles per minute to 38. He finished in 1:47.89 — 2.57 seconds slower than predicted. Not because of a lack of fitness. Not because of wrong tactics. But because of a cramp in his intercostal muscle — something no algorithm in the world could have predicted.

Kazan was the day I learned that a 99% probability can still die at the betting table. But Brisbane was the day I learned that even an 87% probability can die in the water.

Context: Methodology and the limits of swimming data

Before diving into analysis, I need to clarify one thing: swimming data is not like football or basketball data. In football, you have xG, PPDA, distance covered — metrics that can be updated minute by minute. In swimming, you only have a few data points in a race: reaction time, split times, stroke rate, stroke length, and energy conversion efficiency. This creates a fundamental problem: the data sample is too small to draw statistically robust conclusions.

I built my prediction model on three pillars. First, historical data from major competitions — Olympics, World Championships, and national championships of Australia, the US, China, and European nations. Second, biometric data — height, arm span, muscle mass, body fat percentage — collected from public training centers. Third, performance data by phase — quarterly improvement rates, consistency between races, and recovery capacity after high-intensity training blocks.

But I also learned that data has its limits. Numbers have no gender, but the people who read them do. A 19-year-old athlete can have perfect biometrics, but if he just broke up with his girlfriend, failed his university entrance exam, and is facing pressure from his parents — none of those factors will ever appear in my spreadsheet. I do not believe in emotions. I believe in data sequences longer than your emotions. But I am also humble enough to admit that that data sequence is never complete.

In this article, I will analyze three typical cases — one from the Tokyo 2026 Olympics, one from the 2026 World Championships, and one from the 2026 Australian National Championships — to demonstrate that swimming data, no matter how powerful, still has blind spots that cannot be filled. I will not conclude that "data is useless" — that would betray my own profession. I will conclude that data is a tool, not a scripture. And the person using that tool must understand its limits.

Core: Three typical cases of data failure

Case 1: Tokyo 2026 Olympics — When psychological pressure erases physical advantage

At the Tokyo 2026 Olympics, I was hired by a major Brisbane betting company as an analytics expert for the men's 100m breaststroke. My model predicted that Adam Peaty — the world record holder with a time of 56.88 seconds — would win gold with 94% probability. My basis was solid: Peaty had been undefeated for 7 years, had an average time of 57.2 seconds in his last 12 races, and his power index was 12% higher than his closest rival. My model also accounted for pool conditions — the Tokyo Aquatics Centre was designed with an advanced wave-dampening system that minimized interference from adjacent lanes.

But I missed a factor that data cannot measure: the pressure of defending a title under the scrutiny of the British media. Peaty had to face 47 interviews in the 10 days before the race. He slept an average of 5.2 hours per night — 1.8 hours less than usual. In the final, Peaty finished in 57.37 seconds — 0.49 seconds slower than my model predicted. He still won gold, but only 0.16 seconds ahead of Dutch rival Arno Kamminga — the narrowest margin of his career at a major competition.

Numbers have no gender, but the people who read them do. My model could not know that Peaty cried in the changing room before the race because of the pressure of an entire nation's expectations. My model could not know that he called his mother at 2 AM the night before the competition. All those factors — what I call "emotional data" — created a 0.49-second gap between prediction and reality.

The lesson from Tokyo: physical data can predict 90% of outcomes, but the remaining 10% — the part that decides between gold and silver — lies beyond the reach of any algorithm. I adjusted my model after Tokyo, adding a "expectation pressure" coefficient based on media coverage levels and each athlete's history of performing under pressure. But I know that coefficient is still a rough estimate — it can never replace truly understanding the human being.

Case 2: 2026 World Championships — When wrong tactics destroy data advantage

The 2026 World Championships in Fukuoka, Japan, was one of the most data-rich competitions in swimming history. Every race was recorded by 12 underwater cameras, collecting data on elbow angle, head position, and kick efficiency in milliseconds. I used this dataset to analyze the women's 200m butterfly — a race where I predicted Summer McIntosh of Canada would win with 78% probability.

My basis: McIntosh had an average time of 2:04.5 in her last 8 races, 1.2 seconds faster than her closest rival. She had the most consistent stroke rate in the top 10 — a standard deviation of only 0.8% between strokes, compared to the 2.3% average of her rivals. And she had a height advantage — 1.73m compared to the 1.68m average of the top 10 — allowing her a longer arm span, reducing the number of strokes needed per 50m.

But I missed a crucial tactical factor: McIntosh's coach decided to change tactics at the last minute. Instead of swimming with a "negative split" — first 50m 1.5 seconds slower than the last 50m — as my model predicted, they chose an "even split" strategy — maintaining a consistent pace throughout the race. Result: McIntosh finished in 2:05.8, only winning silver, 0.3 seconds behind American Regan Smith.

Interestingly, my data showed that McIntosh's "even split" tactic would produce a predicted time of 2:05.2 — 0.7 seconds slower than the "negative split" tactic. But I could not know that her coach decided to change tactics just 2 hours before the race, based on McIntosh's own feeling after the warm-up. That was a decision based on intuition — not data — and it destroyed the advantage my data predicted.

Kazan was the day I learned that a 99% probability can still die at the betting table. Fukuoka was the day I learned that even a 78% probability can die because of a human tactical decision. I cannot predict a coach's decision — not because I lack data, but because that decision does not follow any rule that can be modeled.

Case 3: 2026 Australian National Championships — When the human body betrays every prediction

Back to Liam Carter — the 19-year-old athlete I mentioned in the opening. Carter is a typical case of data failing before the complexity of the human body. My model predicted he would swim 1:45.3 in the 200m freestyle final — based on his 2.1% quarterly improvement trajectory, superior power index, and his 1:45.32 heat performance. But he finished in 1:47.89 — 2.57 seconds slower than predicted.

The cause: a cramp in his intercostal muscle — a muscle between the ribs, responsible for deep breathing and maintaining body position in the water. This cramp did not appear in any biometric test. It was not related to hydration levels, not related to sleep quality, not related to training intensity. It was simply a random event of the human body — an event with a probability of about 0.3% in any race, according to sports medicine data.

But here is the key point: that 0.3% probability is not a fixed number. It increases significantly when an athlete is in a period of increased training intensity — as Carter had been for 6 weeks before the competition. It increases when an athlete is sleep-deprived — Carter slept an average of 6.1 hours per night during competition week, 1.2 hours less than usual. And it increases when an athlete is under expectation pressure — Carter had been hailed by the Australian media as "the future of Australian swimming" after his heat performance.

I could not predict that cramp. But I could — and should — have predicted that Carter's probability of an incident was higher than average, based on the risk factors I already knew. That is the biggest lesson from Brisbane: data is not just about predicting outcomes, but about assessing risk. And I failed to assess that risk.

Contrarian: Correlation is not causation — and data never tells the whole story

One of the most common mistakes in sports data analysis is confusing correlation with causation. I have seen countless analyses conclude that "athlete A swims faster because they have a longer arm span" — but that does not mean a longer arm span causes faster swimming. Perhaps athlete A has a longer arm span because they are taller, and being taller means having more muscle mass, and having more muscle mass means being able to generate more propulsive force. The true causal relationship lies somewhere much deeper than the surface of the numbers.

I developed a rule for myself: before concluding any causal relationship from data, I must write down my hypothesis before collecting the data. This forces me to clearly identify the causal mechanism I believe exists — before I can be influenced by what the data shows. This rule has saved me from countless false conclusions.

For example: in my analysis of Carter, I found a strong correlation between arm span and improvement rate — a correlation coefficient of 0.87. But when I wrote down my hypothesis before collecting the data, I realized I had no clear causal mechanism to explain why a longer arm span would cause faster improvement. Perhaps athletes with longer arm spans tend to be trained by better coaches — because they are considered to have higher potential. Perhaps they tend to train harder — because they receive more media attention. That correlation could reflect a host of confounding factors I cannot control.

Numbers have no gender, but the people who read them do. And the people who read them — whether analysts, coaches, or athletes — always carry their own biases and expectations. I have learned that the best way to combat those biases is to force myself to write down my hypothesis before looking at the data. It is not perfect — I can still be influenced by what I want to see — but it creates a useful barrier against self-deception.

Another blind spot of swimming data is its dependence on specific pool conditions. An athlete can swim 0.5 seconds faster in a pool with a better wave-dampening system, at a lower altitude, or at an ideal water temperature. These factors are not always recorded in the data — and when they are recorded, they are often not standardized across different competitions. I have seen athletes with excellent performances in one specific pool, but who fail in another — not because they are worse, but because the pool conditions differ.

Takeaway: Signals for the next round — and a bigger question

So, what do we learn from these cases? I believe there are three main lessons.

First, data is a tool, not a scripture. It can tell us probabilities, but it cannot tell us what will happen. A prediction model with 94% probability still has a 6% chance of being wrong — and in the world of swimming, that 6% can be the difference between a gold medal and finishing outside the top 3.

Second, emotional data — psychological pressure, mental state, personal motivation — is an indispensable part of the overall picture. We cannot measure it in milliseconds, but we can — and should — find ways to incorporate it into our analysis. I have started collecting data on each athlete's history of performing under pressure, media coverage levels before major competitions, and sleep quality during competition week. These data are not perfect, but they are much better than ignoring the human factor entirely.

Third, and most importantly, we need to be humble before the complexity of the human body. A cramp, a last-minute tactical decision, a nightmare before the race — all these factors can destroy our best predictions. That does not mean data is useless. It means we need to use data more intelligently — not just to predict outcomes, but to assess risk and prepare for unexpected situations.

Kazan was the day I learned that a 99% probability can still die at the betting table. Brisbane was the day I learned that even an 87% probability can die in the water. And I believe my biggest lesson is still to come — because each new race brings new variables, new challenges, and new limits of data.

I do not believe in emotions. I believe in data sequences longer than your emotions. But I also believe that that data sequence — no matter how long — will never tell the whole story. And the biggest question I am facing is not "how to improve my prediction model," but "how to accept that there are things that cannot be predicted."

That is a question I still do not have an answer to. But I know that the answer — when I find it — will not come from data. It will come from a deeper understanding of human beings, of emotions, and of the things that cannot be quantified. And perhaps, that is the most important lesson that swimming — and sport in general — can teach us about the limits of data.