← Analysis

The bike is where the time is. The run is where the day is decided.

19 September 2026 · Alex Schouleur · 9 min

Last time I took a 2024 study of 16,611 professional results, which found the bike is the split most correlated with finish time, and said I didn't trust the question. Not the study — the question. On a long-course day the bike is half your race. A split that is half of the total will correlate with the total almost whatever happens. I called it a part–whole correlation and promised three things: replicate it on amateurs, correct for the artefact, and be honest about what the correction changes.

Here is all three, on 665,769 amateur finishers across 394 full-distance races between 2002 and 2022 — a dataset published openly by Beat Knechtle and colleagues alongside their own paper, which is why I can show you the numbers rather than describe them.

Finding 1 — the professional result replicates, and it survives the correction

The fix for a part–whole correlation is simple. Instead of correlating each discipline with the total, correlate it with the sum of the other two. The shared part disappears, and the question becomes the one an athlete actually asks: does a good bike go with a good swim and a good run?

I checked the method on synthetic data first — three legs drawn independently, no relationship between them. Against the total, each leg shows r² above 0.5, which is pure artefact. Against the other two, everything collapses below 0.05, as it should.

On 665,769 real finishers:

Look at what the correction does. Against the total, the run edges out the bike — 0.817 to 0.805 — which is exactly the artefact at work, because at full distance the run is long enough to compete with the bike on duration alone. Correct for it and the order flips: the bike is the most connected discipline, 0.474 against 0.437.

So the professional finding holds on amateurs, at twice the distance, and it holds better after the correction than before it. That is a stronger claim than the original study made, and it is the opposite of what usually happens when you control for an artefact.

In plain language: the bike tells you most about the athlete. Someone who rides well in an age-group field tends to swim and run well too. The swim tells you least — it is the most specialised skill of the three, and plenty of strong triathletes are ordinary swimmers.

Finding 2 — but the run is where finish times actually spread out

This is the one I had wrong before I looked.

There is an exact way to ask where the spread in finish times comes from. The variance of the total decomposes, with no approximation, into the covariance of each leg with the total. So each leg has a share of the time and a share of the variance, and you can set them side by side. If a leg's share of variance exceeds its share of time, it is doing more than its length to separate people.

The bike takes half the day and contributes 41 % of the spread. The run takes a bit over a third of the day and contributes 47 %. The swim sits below one: it contributes slightly less than its length.

Both of these are true at once, and they are not in conflict:

• The bike tells you who you are dealing with. Most connected discipline; a good rider is usually a good triathlete.
• The run tells you how the day went. It is where finish times fan out, well beyond what its duration justifies.

That second sentence matches what every triathlete knows in their legs and almost no predictor models properly. Bike times are compressed — by physics, by power limits, by the plain fact that there is a ceiling on how fast anyone rides 180 km. Run times are not. People crack on the run. The spread is the cracking.

Finding 3 — transitions are worth more than twice their length

In the same decomposition, T1 sits at ×2.0 and T2 at ×2.4 — together, 2 % of the day for 4.4 % of the spread. They are the only parts of the day that punch above their weight by more than 20 %.

Keep it in proportion: 1 % of your day producing 2 % of the spread is still small, and nobody wins a long-distance race in T1. But look at the actual minutes.

The median athlete spends fourteen minutes standing still. Between the quarter of the field that is quick in transition and the slowest 5 %, there are eighteen minutes — more than most of the gains people spend a winter chasing, and more than most people's entire aero budget.

Meanwhile, the pace calculators assume five minutes.

Finding 4 — the mistake I nearly published

I almost ran these numbers on all 403 races in the file. Then I looked at each race's median splits, and nine of them were impossible.

One had a median swim of 9.9 minutes. Another 16.4 minutes. Several sat between 29 and 52 minutes. For 3.8 km, in open water, in an amateur field, a median under 55 minutes cannot happen — the world's best cover the distance in about 45. One race had a median bike of 3.2 hours for 180 km, which would be 56 km/h.

These are not data errors. They are shortened races — a swim cut or cancelled for weather or water quality, a bike course shortened on the day. It happens, and the results are recorded as if the race were normal. Nine races out of 403, about 2 % of them, carrying 13,189 finishers.

Leaving them in doesn't wreck the analysis, but it isn't harmless either. It drags the swim's corrected r² from 0.295 down to 0.258 — an eighth of the effect — because a 10-minute swim in a full-distance race is noise that correlates with nothing. Every other number in the tables above moves by a hundredth or less.

I mention this for two reasons. The first is that it is the kind of thing that quietly makes analyses wrong, and I would rather show you the filter than hide it. The second is that if you are comparing your own race to a field average from any source, it is worth knowing that some of those races weren't the race they say they were.

What else is in there

Two figures worth keeping.

Which race you choose is worth about a sixth of everything. The race identity alone explains 16.3 % of the variation in finish times across the whole dataset. Median finish times by race run from 10.6 to 14.7 hours, with the middle 80 % between 11.5 and 13.7. That is a two-hour spread between a fast course in a good year and a hard one — for the same athlete population. It is also the single best argument for predicting against a specific course rather than against a distance.

The field itself: median finish 12.57 hours, with 80 % of finishers between 10.3 and 15.3.

What this does not say

It does not say the bike matters less than you thought. It says the bike is where the time is and the run is where the variation is — two different claims, both useful.

It does not tell you where your minutes are. These are field statistics and you are one person. A field that spreads out on the run says nothing about whether you will. The question that matters to you — where does an hour of training buy the most minutes on race day — is a question about your own history, and it is what the predictor is for.

And it cannot say anything about men versus women, or about age groups. The published file carries split times and a race identifier, and nothing else — no sex, no age group, no country, no year. The study behind it used those variables; the deposited data does not include them. I would rather tell you that than quietly imply a breakdown I cannot do.

Correction, 23 September 2026. The numbers in this piece were recomputed with the code published with the site (analysis/plos.py), whose row filter also requires the five splits to add up to the finish time within two minutes. The field is 665,769 finishers instead of 665,308, the swim's corrected r² is 0.295 instead of 0.300, the time shares now sum to 100 % (they were 99 % before), and the transition ratios are ×2.0 and ×2.4 instead of ×2.16 and ×2.72. No conclusion changes. The first version of this article and the rest of the site quoted two different finisher counts; they now quote one.