Win Probability Added

Read each bar from the centre line. The pieces add up to the black tick, which is the character's win rate on the tier list minus 50%. The piece touching the centre line is the character's own; everything further out was the ranks, the players or the opponents. A long bar with a short first piece is a character that wins because of who plays it.

Bars that run past ±6 points are cut at the edge of the chart; the table below has their full figures. A bar that long comes from a character whose players are hard to tell apart from it, and its interval says how little it is known.

None of the 19 characters' WPA survives correction. Closest: Jacky at −3.8 points (p 0.015, q 0.283).

A win rate mixes three things: who the character's players were paired against, how good those players are, and the character. WPA predicts every match from the first two and gives the character whatever is left over.

1. Ranks. Each match starts from the win chance the two ranks give, read off the ladder's own rank-against-rank record (the Rank Impact report). A 35 facing a 31 is not an even match, whatever either of them picked.
2. Player skill. Each player gets one number: how much better or worse than their rank they play, fitted over every match they played with every character. A player who beats most people at their rank gets a positive number, whichever character they were on.
3. The character. Each character gets one number, fitted at the same time: how much more it wins than the ranks and the players predict. In points, that is its WPA: in a match that was even on ranks and players, picking it makes the chance 50% plus the WPA. Zero is the average of the nineteen characters, each counted once.
4. Opponents' characters. What is left of the bar is the characters it met. The baseline counts every character once and the most-picked characters sit below it, so everyone faces a slightly soft field; this piece is that.

The character's number is read most directly off people who play more than one character, the same person on two different characters, and off opponents, since every match has two characters in it. The table counts how many of each character's players also played another. A character whose players almost all play only it is harder to separate from them, and its interval is wider for it.

What it cannot say. Characters are chosen, not dealt, so this is an adjusted comparison and not proof of cause: a character that players pick up while they are improving, or bring out only in matchups they like, carries that into its number. Skill is one number per player for the whole period. And only matches with both players in the band count, as on every banded report.

The interval: players are dealt into 20 groups by a hash of their id, and the model is refitted 20 times, each time leaving out one group's players and all their matches. The spread of those refits is the error (a grouped jackknife). With 20 groups the p-value is read on a t distribution with 19 degrees of freedom (the 95% line is at t = 2.09), and q corrects the p-values for the characters tested together (Benjamini–Hochberg, 5%).