How to read a win rate

A win rate is not a measurement until you know how many matches went into it, and that second number changes what the first one means more than any argument about the brawler will. A rate computed from a few dozen matches and a rate computed from tens of thousands look identical on a page: same two digits, same decimal point, same confident little arrow next to them. They are not the same kind of claim. One of them is a measurement and the other is a rumour with a decimal point.

Half of the reason is the intuition everybody already has about coins. Flip one ten times, get six heads, and nobody concludes the coin is bent. Flip it ten thousand times and get six thousand heads and something is clearly going on. Wins and losses arrive with an enormous amount of noise attached — a teammate who disconnects, a map that suits you, an opponent having a bad evening — and the only thing that grinds noise down is volume.

Two samples of the same size are not worth the same

The other half is the part that gets missed, and it is why where a sample came from matters as much as how large it is. Imagine two brawlers with fifty recorded matches each. The first fifty come from fifty different players. The second fifty come from three. Any table would show the same count for both. The second one is worth a fraction of the first.

The arithmetic behind any confidence figure assumes each match is an independent draw, like a separate coin flip. Matches from the same player are nothing of the sort. They share that player's skill, their trophy range, the maps their rotation happened to be showing that week, the teammates they queue with, the time of day they play and whatever mood they were in. Ten matches from one account is closer to one loud opinion repeated ten times than to ten separate observations. The count says fifty; the amount of independent information in it is much smaller.

Fifty matches from fifty players is fifty weak opinions. Fifty matches from three players is three strong ones wearing a costume.

Stacked on top of that is a selection problem. A player with a long history on one brawler did not arrive there at random. They chose it, they kept choosing it, and they are better with it than the average person who picks it up for an evening. So when a sample concentrates into a handful of accounts, it stops describing the brawler and starts describing its specialists. That is a genuinely interesting thing to know. It is just not the thing the label on the number says it is.

What a rate has to carry before you trust it

• The number of decided matches behind it. Without a denominator on the page there is no measurement, only a figure with a decimal point in it.
• How many different players those matches came from. Fifty matches spread across fifty accounts and fifty matches from three are not the same evidence, and only the first resembles what the arithmetic assumes.
• The mode, and wherever it exists the map. A brawler that is excellent in a walled Gem Grab and useless on an open Bounty layout has one overall figure that describes neither.
• The trophy range. The same brawler can be punished at the top of the ladder and perfectly comfortable several thousand trophies below it.
• When it was measured, relative to the last balance change. A window that straddles a patch is measuring two different games and averaging them together.

A table worth reading also refuses to let a thin number behave like a thick one. If almost nothing is known about a brawler, the honest estimate is that it is roughly ordinary, and a list that lets three lucky matches sit at the top is reporting luck under a heading that says otherwise. So when an unfamiliar name appears at the head of a ranking, the first thing to look at is not the rate. It is the count beside it, and how quickly that name disappears again a week later.

What none of that fixes

Corrections deal with noise. They do not deal with a number pointed at a different question than the one you are asking. Figures gathered near the top of the ladder describe matches where everybody hits their shots and nobody wastes a Super. Several thousand trophies below that, a brawler being punished up there can be entirely fine, and a brawler being counted on can demand execution nobody in your matches is producing. Showdown is harder still: it reports where you finished rather than whether you won, so a win rate for it has to be manufactured out of placements before it can be printed at all — and a number that had to be invented before it could be shown is not one to plan around.

PiperA marksman is the clearest case of one figure hiding two answers: the same brawler is a different proposition on an open layout and a walled one.→

1. How many matches is this built on? If the page does not say, the number is not a measurement and should not be treated as one.
2. What is it being compared against? A rate means nothing on its own; it means something next to the same rate in the same mode, on the same map, in the same window.
3. Would I change my pick if it moved by a point in either direction? If not, the difference you are staring at is noise and you can stop staring at it.