Every season, round three produces a league leader who won't be top ten by round twelve. The number wasn't wrong. The sample was, and the difference between those two things is most of analytics literacy.
A percentile is a comparison: this player against every other player, this team against the league. Comparisons are only as stable as the data underneath them, and early in a season nothing is stable. Three games of shooting is a weekend's mood. A guard who went 9-of-16 from deep across two rounds will sit at the top of the three-point table wearing a 56 percent that has almost no chance of surviving the winter. Put that in front of a bench as "elite shooter, 98th percentile" and you've scouted a coin flip.
Noise wearing a suit
The problem isn't that small-sample numbers are useless. It's that a percentile makes them look finished. "Fourth in the league in scoring" reads with the same authority in round three as in round eighteen; the rank hides how little sits behind it. Raw averages at least invite suspicion. A percentile arrives dressed for a board meeting.
The volatility runs in both directions, which is what makes it dangerous. Early-season percentiles don't just crown false stars; they bury real ones. A genuine scorer who opened against the league's two best defences can sit in the 30s looking like a role player. Act on either read and you've built a plan against a player who doesn't exist.
The discipline
Three rules keep percentiles honest, and none of them needs a model.
Set a floor before you look. Minimum games and minimum minutes before a percentile is allowed on the page. Where the floor sits is a judgement call by league, but having one isn't optional; without it, every hot bench week reads like a breakout.
Stamp every number with its sample. Not "she's 87th percentile," but "87th percentile, 5 games." The stamp does the reader's risk assessment for them. In my scouts every tendency carries a confidence grade that is really just a sample-size label: a read that has held across most of a season is worth building a game plan on; a read from a fortnight is a hypothesis you check at warm-up.
Expect the middle. Extreme early numbers move toward ordinary as games accumulate, because the extremes were mostly luck and luck doesn't repeat on schedule. The practical version: the smaller the sample, the more your projection should lean on what the player has been across her career and level, and the less on the last three box scores. The scorching start usually cools. The horror start usually corrects. Plan for the player, not the fortnight.
The question that does the work
All of this compresses into one habit. When a number surprises you, ask what it's drawn from before you ask what it means. Sample first, story second. Most bad analytics takes at this level aren't calculation errors; they're strong conclusions resting on weak denominators, delivered with a percentile's confidence.
The fix makes the numbers more useful, not less. A percentile with a floor under it, a stamp on it and a season behind it is one of the fastest ways to find where a team is strong and where it's soft. Treat the round-three version the way a scout treats any two-week read: a hypothesis, graded provisional, checked again when the games accumulate.