"They're 87th percentile from three." Great. Now what do you do with that on Saturday?
Percentiles have quietly become the language of basketball analytics, and for good reason. They rank a player or a team against everyone else in the league, so you instantly know whether a number is good, ordinary, or a problem. But a percentile handed to a bench with no context isn't intelligence. It's trivia. Worse, it can be a lie.
Here's how I use them, and how I make sure they help a coach instead of just impressing one.
The four rules
Always show the value and the percentile. "87th percentile from three" could be a player taking two attempts a game. The percentile tells you the rank, the raw number tells you whether it's real. A coach needs both: the rank to know it matters, the value to know how much. A percentile on its own is the most confident way to be wrong.
Mind the sample. Six games into a season, percentiles are noise wearing a suit. I won't put them in front of staff until there's enough of a season to mean something, and I set a minimum-minutes floor so a bench player's hot week doesn't read like a star. Every percentile that reaches a coach gets stamped with what it's drawn from: "vs league, 13 games." If you can't say what it's measured against, don't put it on the page.
Say which population it's against. This is the one people skip. A percentile is a rank inside a group, and the group is a choice somebody made. Against every player in the league, a starting centre's rebounding rate looks enormous. Against other centres, it might be ordinary. Neither ranking is wrong and they lead to opposite plans, so the population belongs next to the number every single time. Cross-league is worse again: 88th percentile in one competition and 88th percentile in another are not the same player, because the two distributions do not overlap.
Treat it as a flag, not a verdict. A poor defensive-rebounding percentile doesn't tell you why. Are they small? Do they leak out early for transition? Is one player dragging the whole number down? The percentile says "look here." It doesn't say "here's the answer." That's the coach's job, and it's the part no model does for you.
Where percentiles quietly mislead
The distribution is not flat. Percentiles are ranks, so the gap between 50th and 60th is usually tiny and the gap between 90th and 99th is enormous. Ten percentile points at the top of a distribution can be a different species of player. Ten in the middle is often nothing at all. A coach reading "they went from 55th to 70th" hears improvement. Frequently there isn't any.
A team percentile can be one player. Team-level ranks are averages, and an average hides who produced it. A bottom-third pick-and-roll defence might be five players who are all mediocre at it, or four who are fine and one who is a genuine hole. Those two teams get attacked in completely different ways: the first with volume, the second by hunting one matchup. The team percentile cannot distinguish them, so it should never be the last thing you look at.
Rate and volume ranks describe different players. A high percentile on efficiency and a low one on usage is a player who is very good at a small job. Present the efficiency rank alone and a coach prepares for a primary option who will not appear. Both ranks, side by side, or neither.
Small samples reward the wrong people. The tail of any distribution fills up with players who have not played much, because low sample means high variance means extreme ranks. If your leaderboard is topped by people you have never heard of, that is not a discovery. That is your minutes floor being too low.
What I put on a page
Value, rank, population, sample. Four things, one line, every time:
3PT: 38.1% on 5.4 attempts, 84th percentile among guards, 13 games.
That takes no more space than the version with just the rank and it cannot be misread. If any of the four is missing, the line does not go on the page, and I would rather send a scout with fewer numbers than one where a coach has to guess what a number is measured against.
The philosophy underneath it
Model for breadth, scout for judgement. Percentiles are the breadth: a fast, honest way to find where a team is strong and where it's soft, across more of the roster than you could ever watch properly. The judgement is yours. Which of those flags actually changes how you guard them, and which is just a number.
So when a teardown says an opponent's pick-and-roll defence sits in the bottom third of the league, that's not the headline. The headline is what you do about it: hammer the ball screen, early and often, because the evidence says it's where they break. The percentile pointed at the door. The coach decides whether to walk through it.