Every Leaderboard Contains Two Hidden Decisions

Kim Jin-Hyeon won 28 of 28 duels in 2024 J1 — a flawless 100%, and the top of any unfiltered duel table. He is a goalkeeper. The table is not lying; it is answering a question nobody meant to ask.

The sentence that causes the most damage

"The numbers don't lie" is the most dangerous sentence in data scouting, and it is dangerous precisely because it is true in a narrow sense and false in the sense people use it.

The underlying figures are what they are. But a leaderboard is not a figure — it is a construction, and it rests on two decisions that are almost never displayed alongside it. Who was allowed into the sample. And who each player is being measured against.

Get either wrong and the output is not a weak ranking. It is not a ranking at all.

Trap one — the sample, and the appeal of 100%

Sort 2024 J1 duel win rate with no conditions and the top of the table is a wall of perfect records. The outright leader is Cerezo Osaka goalkeeper Kim Jin-Hyeon, who won all 28 duels he contested. Beneath him sit players who won the single duel they entered, or both of their two.

Dribbling behaves identically. The unfiltered success-rate table is led by one attempt and one success, or nine of nine. Every one of those entries is arithmetically correct and none of them describes a player.

The correction is a minimum-attempts floor, and its effect is drastic.

The floors do two things at once. They remove the small-denominator entries, and they produce a stable baseline against which any individual figure can be read. Without the second, a number has no meaning even when the player is real: 45.0% dribble success is not interpretable until you know the cohort sits there.

This is why every leaderboard on this site states its threshold in the heading — "30+ shots", "200+ duels". A rate ranking that does not publish its floor should be treated as unread.

Trap two — the cohort, and the same number meaning opposite things

Fix the sample and a second problem is untouched: the players inside it are not doing the same job.

Inside that 74-player duel cohort, the split by position is severe. Defenders average 56.1%, midfielders 50.6%, forwards 42.2%. The reason is structural rather than about ability — defenders contest duels they are favoured to win, stopping progress in their own half; forwards contest duels they are not, with their back to goal in front of a set defence.

The consequence is that a single figure carries opposite meanings depending on who holds it. A 55% defender is unremarkable, marginally below his positional baseline. A 55% forward is roughly thirteen points clear of his — an outlier for the role.

Rank the two together and the table sorts by position with performance as a rounding error. This is treated at length in our duel win-rate analysis; the point here is narrower. A properly filtered sample with the wrong cohort is still useless, and the filtering step gives false confidence that the work has been done.

The three rules we run, and the one that costs most

Against these two traps the site applies three layers.

1. A minimum-attempts floor per metric. Stated in the heading of every leaderboard, and set per metric rather than globally.

2. Percentiles computed within position groups. Full-backs against full-backs, forwards against forwards, on the player map and everywhere else.

3. A minutes floor of roughly 900 — about ten matches. Anyone below it is removed from gems, the map and percentiles entirely. In the 2024 data, 240 players cleared it.

The third rule is the one worth being honest about, because it is the most expensive. A 900-minute floor removes rotation players and mid-season arrivals — and in a prospect-scouting context those are disproportionately the young players the screen exists to find. A 19-year-old with 700 excellent minutes is invisible to our percentiles.

We accept that cost because the alternative is worse: a gems list populated by small-sample noise is not merely less useful, it actively misdirects the scouting time it was built to allocate. But it is a trade, not a free improvement, and anyone using these rankings should know which players they cannot contain.

What the rules still do not fix

Floors and cohorts address who is compared to whom. They leave the underlying metric untouched.

A correctly filtered, correctly cohorted duel win rate still cannot separate ground from aerial contests, still cannot distinguish a duel won and possession retained from one won and immediately conceded, and still says nothing about how often a player was placed in duels at all. A properly built dribble leaderboard still cannot tell a successful dribble that advanced play from one that went sideways.

Clearing both traps produces a number that can be read. It does not produce a number that answers a recruitment question on its own — and the confidence a clean-looking leaderboard generates is itself something to guard against.