How the Tier List Is Calculated

We use only the weekly data (win rate, pick rate, ban rate) that Starward's official data site publishes for the 6000–8000 rating band. We don't estimate or backfill anything we can't get, such as card pick rate or match counts.

We use the 6000–8000 band because it's the only one where the pick-weighted average win rate comes out to exactly 50% (i.e., matches within the band are self-contained) — that's what makes win rate a well-defined probability of winning when played.

• Win rate — The probability that the character wins when played.
• Pick rate — The probability that a randomly chosen player's played character is this one (its share of all match slots).
• Ban rate — When in hand, the conditional probability of being banned. The total across all characters comes out to about 1400%, which confirms it's a conditional probability, not a share.

Win rate axis = win rate (shrinkage-adjusted) / Presence axis = pick rate ÷ (1 − ban rate)

• Win rate shrinkage — When a character's pick rate is low, its win rate tends to run high — partly statistical noise, partly because the few players who do pick it are specialists. We pull it back toward the theoretical 50% in proportion to how thin the data is, so a lucky streak alone doesn't put a character at the top.
• Ban-adjusted pick rate — Pick rate ÷ (1 − ban rate). A character banned in 70% of matches can only show up in the ones where it survived the ban phase, so its raw pick rate understates how strong it is. Dividing out the ban rate recovers how often it would be picked if it were never banned.
• Within-cost standardization — The win rate axis and presence axis differ hugely in scale, so we align their medians and spread within each cost before comparing them (a robust z-score). Every evaluation is always "relative standing within the same cost."

We split both the win rate axis and the presence axis into High/Mid/Low bands, then combine them into 5 tiers — S / A+ / A / A− / B. Characters in the Low win-rate band are capped at A− regardless of presence (an editorial choice, so a character that's actually losing isn't shown as mid-tier). Since it's not a fixed percentage like "the top 20% is always S," a week with no standout characters can leave S empty. The table groups the A tier (A+/A/A−) into a single row, distinguished by frame color and a badge in the card's top-left corner.

High: z ≥ +0.35 / Mid: in between / Low: z ≤ −0.35 (relative standing within cost, robust z-score)

Characters with a displayed pick rate of 0.1% or lower are marked with * because their win rate carries a lot of statistical noise. Shrinkage pulls their score back toward neutral, but their evaluation is still less reliable than other characters'.

• The band thresholds (±0.35) are an editorial choice. Characters near a boundary can easily shift to a neighboring tier from week to week. We no longer use ban rate itself as a scoring axis — ban only contributes through the corrected pick rate (pick rate divided back by 1 − ban rate).
• Ban rate includes a reputation component — "banned because it's famous" — which may cause some characters to be over-rated relative to their actual strength.
• We can't get data on team composition (which characters/costs are paired together), so we can't evaluate synergy.

The full method spec and statistical validation notes are in docs/tier-methodology.md and docs/data-observations.md in the repository.