How to read a stat percentile — and when to distrust it
Every unit page here says something like "11th percentile, rank 124 of 139". This is exactly what that means, exactly how it is calculated, and the four situations where it will mislead you.
Open any unit page on this site and you will find a section headed How this unit compares, containing sentences like:
Measured against 139 comparable cards, Goku (Youth) is below average for its cohort (5th percentile on average). Its strongest number is HP at 7125 — 11th percentile, rank 124 of 139.
It is the most-repeated construction on the site, so here is precisely what it claims. It is a statement about where a number sits in a distribution. It is not a tier rating, not a recommendation, and not an opinion about whether the unit is good.
How the cohort is chosen
A percentile is meaningless without saying "among what". The comparison group is built in three steps, taking the first that yields at least six members:
- Same rarity and same identity group. The identity group is whatever that game uses to slot a unit into a role — Class in Arknights, Path in Honkai: Star Rail, Element in Wuthering Waves, Attack type in Blue Archive, Type in Dokkan. So a 6★ Caster is measured against 6★ Casters.
- Same rarity only, if the first group is too small to say anything about a distribution.
- The whole roster, as a last resort.
The cohort in use is always named on the page — "vs 139 UR Super STR cards" — so you never have to guess which one you are looking at.
Rarity is always part of the cohort because rarities cap at different levels and cost different amounts to raise. A 4★ ranked against 6★s would sit at the bottom of every stat, which is true and completely uninformative. That reasoning is set out at more length in why max-level stats mislead.
How the number is calculated
For each stat, the cohort's values are sorted, and the unit's percentile is its midrank position:
percentile = (values strictly below + half the values exactly equal) ÷ cohort size
The "half the values exactly equal" term matters more than it looks. The naive definition — the fraction strictly below — reports 0th percentile for a unit sitting at the most common value in a heavily tied distribution. That reads as damning when the unit is merely typical. Midrank puts ties in the middle of their band, which is what a reader assumes the number means.
Rank is counted the other way: the number of cohort members strictly above, plus one. So "rank 124 of 139" means 123 cards have more of that stat. Median is the middle value of the sorted cohort, and the percentage beside it is the unit's gap to that median.
What gets left out on purpose
- Stats where a lower number is better — attack interval, deployment cost, redeploy time — are excluded entirely. The percentile wording assumes higher is better, and an operator with a fast 1.6-second interval was being labelled with its own weakest stat. Inverting them per stat, per game, was more likely to produce a confidently wrong sentence than a useful one.
- Stats with fewer than three distinct values in the cohort are skipped. "Block 1 · 0th percentile · rank 1 of 26" is noise: every operator in the class blocks one enemy.
- Cohorts under six members do not get a percentile at all. Two units is not a distribution.
The four situations where it will mislead you
1. It is computed on max-level stats
The roster value for a unit is its maximum. Growth curves cross, so a unit's percentile at the level you are actually playing it can differ substantially from the one on the page. If that matters for your decision, use the game's Stat Calculator at your real level rather than the percentile.
2. Stats are not the same thing as power
This is the big one. A percentile measures the numbers on the card. It cannot see a skill that doubles attack for eight seconds, a passive that ignores defence, a talent that only fires below half health, or the four-unit synergy that is the actual reason anyone builds the unit.
In games where the interesting design lives in skills rather than statlines — which is most of them — a unit can sit in the 20th percentile for raw ATK and be among the strongest things in the game. The percentile is not wrong about that; it is silent about it.
3. A narrow cohort makes ordinary units look extreme
A cohort of exactly six behaves very differently from one of 139. Move one place in a group of six and your percentile jumps by roughly seventeen points. Where the cohort is small, treat the rank ("2nd of 7") as the real information and the percentile as decoration.
4. The average across stats is a blunt summary
The one-line verdict — above average, middle-of-the-pack, below average — is the mean of the unit's percentiles across the stats it was ranked on, thresholded at 60 and 40. Averaging a 95th-percentile ATK with a 5th-percentile DEF produces "middle-of-the-pack", which is arithmetically fair and describes neither number. Read the per-stat rows; the summary exists to orient you, not to conclude for you.
What it is genuinely good for
- Spotting outliers. A unit in the 95th percentile on one stat and the 10th on another has a shape, and that shape is usually the design intent.
- Sanity-checking a recommendation. If a unit is described as a defensive wall and sits at the 15th percentile for DEF and HP within its own class, the defence is coming from a skill.
- Comparing like with like across a large roster, which is precisely where eyeballing a stat sheet stops working. Dokkan has thousands of cards; nobody holds that distribution in their head.
Used that way it is a good tool. Used as a tier list it will lie to you, and it was never meant to be one.