Ranking Scores and Percentiles: A Calm Interpretation

Abstract

You saw a number. It was on a results sheet, a club league table, or the leaderboard on the erg's own website, and it told you where you placed. The number is the part that survives the week; the time, the conditions and the people who turned up are gone by Friday, and the rank stays in your head.

The instinct is to read it as a verdict on you. It is not one, for a specific reason rather than a consoling one: a rank is a position in a named list, and a percentile is that position rescaled onto a hundred-point axis. Both describe a group, not a person.

The context

Three different things get called "your ranking", and they are not the same measurement.

There is the place on a leaderboard. There is the percentile on a club results sheet, computed by whoever ran the race from the field that turned up. And there is the club league table, which counts metres logged. A rower told "you were 74th percentile" or "we came third in the league" has been handed one of these, and it is worth knowing which before doing anything with it.

Read it as a position, not a verdict. A percentile is a statement about the population it was computed against, and a rank orders without measuring.

Start with the filters, because on Concept2's own rankings they are the first thing you see. The event. The season. The age range. The weight class. The sex. The country. The state. Whether unverified results count. Which RowErg type the piece was set on. Whether the row is adaptive-classified. Ten choices, and every one changes the pool your result is placed in ([1] Concept2 2012). Set them differently and the same rower, on the same piece, occupies a different place. Nothing about the rowing changes; everything about the number does.

The core idea

A percentile describes a distribution, not a person

Miller's method paper on percentile ranks makes the unit of meaning explicit. Each observation is scored relative to a larger reference pool, and the histogram of ranks shows where a group tends to fall relative to that pool ([5] Miller 2021). The number only acquires meaning once the pool is named.

This is why a percentile in a club race, a percentile on a national list, and a percentile in a season's age-group standings are three different quantities rather than three views of one. The club race's population is the people who raced it. The national list's population is the people who logged a qualifying piece to it, under those filters, this season. The age-group standing's population is the same logbook population sliced by age band. Three pools, three numbers, one rower.

And be careful with the vocabulary, because it slides. Concept2 publishes a place, not a percentile. Anywhere you meet a percentile, somebody calculated it separately against a pool of results they defined, and the place on the board is only one of the things that pool can be measured against.

And check which way round it was counted. A 74th percentile usually means 74% of the field was behind you on the metric being counted, but sheets differ in what they percentile and in the direction they count. Read the number with its convention.

And a league table is not a percentile at all. Concept2's affiliation standings rank clubs by total metres logged, with a per-member column beside it and a filter that separates clubs of two to five members from clubs of fifty-one or more ([3] Concept2 2015). A small club can lead the per-member column that a large club leads on total metres, and neither column says anything about how fast anybody rows. It says how many metres a group logged.

A rank orders; it does not measure the gap

A rank is an ordinal quantity: categories in an order, where equal spacing between neighbours is an assumption, not a fact. The closest test of that assumption comes from a different field, medical exams, where examiners place candidates on a slider between a typical fail at 0 and a typical excellent at 1, and all four intervals between the grades came out significantly different from the equal-interval value of 0.25 ([4] McGown 2022). It is an analogy rather than a rowing study, and what transfers is the property, not the subject: an ordered category does not come with evenly spaced steps.

Concretely, in a list where the leader is five seconds clear of second place and the tail is packed twenty seconds deep, second place is a much bigger performance step than second from last. Both are one place. The gap between adjacent places is set by the shape of the distribution.

A percentile has the same property one level down, and it is worth being precise. The percentile scale itself is evenly spaced — the 60th and the 61st are one point apart, exactly as the 1st and the 2nd are. The performance behind those points is not. Take a typical field, bunched in the middle with the leaders a thin tail: the gap between the rower at the 60th percentile and the rower at the 61st is small, and the gap between the 1st and the 2nd is large. That is a property of the shape, not of the scale, and a different shape would look different.

The list is rebuilt, and your result moves inside it

Two things move underneath the number.

The season moves. Concept2's ranking season runs from May 1 to April 30, and on May 1 the previous twelve months of data is archived and a new season starts on a blank slate ([2] Concept2 2015). Only pieces from the current season are eligible to be ranked at all, so a piece rowed in June and a piece rowed the following April sit in the same season and compete against each other. On the first of May, the whole comparison set is replaced. A percentile is a snapshot of a pool rebuilt once a year.

Your own result moves too, because the measurement wobbles. Schabort's reliability work on this exact test — eight well-trained rowers, three 2,000 m trials at three-day intervals — found the coefficient of variation for mean power at 2.0%, a retest correlation of 0.96, and variability expressed as finishing time at about one third of the mean-power figure ([6] Schabort 1999). Hopkins's paper explains why the wording matters: typical error, expressed as a coefficient of variation, is the right way to say how much a test wobbles. He also sets a bar this study does not clear: reasonable precision for a reliability estimate needs roughly fifty participants and at least three trials, and the rowing study had eight participants and three trials ([7] Hopkins 2000). Read 2.0% as the best available estimate of the wobble rather than a settled constant.

One thing worth being straight about, because it is easy to get backwards: Concept2 ranks pieces by the time they took ([1] Concept2 2012). Time is also the steadier of the two numbers, so the thing on the board is the least noisy one available. The watts on the monitor are the noisy ones, and the leaderboard does not rank on them.

The people in the list chose to be in it

This is the part most often left out. A logbook population is not a random sample of rowers. It is people who own a machine, chose to log the result, and wanted to appear on a board. Nothing in the percentile corrects for that, and a longer list does not make it fairer. Comparing a small club event's percentile with a national list's compares two different populations, and the larger list is not automatically the more honest one.

Age sharpens the point. In age-group competition the population inside a category is selected, not sampled. The relative age effect, a systematic review of age-group swimming, was present in more than half the participants analysed at 58.65%, stronger in the younger categories and around twice as pronounced in men as in women ([8] Lorenzo-Calvo 2021). That review is about youth categories cut by birth date rather than by decade, so its figures do not transfer to a decade-wide band. The transferable point is narrower and still worth having: an age band contains a gradient, not a level field. Even highly trained women masters athletes lose aerobic capacity at between 0.36 and 0.84 millilitres per kilogram per minute each year, which is 0.5% to 2.4% ([10] Vangsgaard 2026). Over the ten years between the bottom and the top of a decade band, that compounds: the 40 to 49 band is not the same pool at 40 as it is at 49.

The season moves the result as well as the pool. Fifteen collegiate male rowers followed across a 37-week season showed significant changes in their 2,000 m ergometer time between the off-season, pre-season and in-season phases ([9] Iguchi 2020). A piece logged mid-training-year and a piece logged at the end of it are not the same measurement, and a season-long leaderboard holds both.

The how-to

Six questions to put to any rank or percentile, in the order worth asking.

Which list is this? Find the filters before you look at the number, and write them down ([1] Concept2 2012). If you cannot name the filters, you cannot name the population, and the number is decoration.

What does one place actually cost you here? Look at the times either side of your position, not at the position. The gap to the rower above and below is the only thing the rank is really telling you.

How wide is the noise band? On this test the coefficient of variation for mean power was 2.0%, and variability expressed as finishing time was about one third of it ([6] Schabort 1999). Those figures describe how much a single result wobbles, and the second is the one that matters here, because the leaderboard ranks time. Read one result as carrying that much uncertainty, and repeat the piece under the same conditions before you read anything into a change.

Which season, and which machine? Only the current season's pieces are ranked, and RowErg, RowErg on Slides and Dynamic RowErg results sit in separate sections you can filter independently ([2] Concept2 2015). An archived season's times sit in that season's list, not this one.

How wide is your band? If you are at the older end of an age band, the people you are compared with are not in the same physiological place you were ten years ago.

Is it even a rank? Metres logged is not a percentile. If the number came from a league table, treat it as a count with a denominator and nothing more.

The trade-offs

What a rank is genuinely good for.

A percentile is a compact, portable way of comparing positions across distributions of different shapes ([5] Miller 2021). Used properly, against a named population with the filters written down, it is an honest summary of a relative position and a basis for a training priority. Relative to people like me, this is where I sit.

What it is not good for.

It is not a verdict on your fitness. It is a position on one list, on one day, against one pool. It is not automatically comparable across lists: the club race, the national board and the league table are three different questions with three different populations behind them, and only one was ever about you. Two percentiles can be compared when the metric, the convention and the pool are the same. And it is not a measure of the gap between places.

There is a cost here too. The moment a number becomes the target, the behaviours that distort it become available: choosing which piece to rank, which list to read, which race to enter. The ranking is an honour-based system, and only one piece of yours ever shows up in any given view ([1] Concept2 2012). Editing a ranked piece removes it, and a birthday that moves you into the next age category means choosing which piece you keep ([2] Concept2 2015). Those are not bugs. They are the system telling you that what is ranked is what you chose to submit, on a machine, that day.

What to do tomorrow

Write the filters next to the number. Which list, which season, which band, which machine. It takes a minute, and it is the difference between a measurement you can use and a number you can only react to.

Keep the time and use the rank for what it is good for. Your own time is the thing that trends. The percentile places you against other people. Mixing them up is what produces a bad week.

Repeat the piece under the same conditions before you read anything into a change. The wobble around a single result is not a rounding detail.

Stop comparing percentiles across lists, and stop borrowing a club race's percentile for a national board.