NEO! VOD

Melee Tournament Tier List: data and method

Every number on the Melee Tournament Tier List and its character and matchup pages, as CSV, with how it is made. Free to reuse under CC BY-SA 3.0 with credit to NEO! VOD and Liquipedia.

tiers.csvmatchups.csvstages.csv

Ranges

RangeSinceThroughGamesSetsPlayersRankedsigma_tsigma_e
20262026-01-012026-11-1512,0253,4831,29290.050.2
20252025-01-012025-12-3117,7305,2261,61890.050.05
20242024-01-012024-12-2213,4073,6051,00190.050.05
20232023-01-012023-12-1710,8543,0061,00470.050.2
20222022-01-012022-12-3116,7104,9501,59370.050.1
20212021-01-012021-12-206,8371,92466460.050.05
20202020-01-012020-11-224,0941,12552660.050.05
20192019-01-012019-12-2817,4044,9471,40990.050.1
20182018-01-012018-12-2918,1535,3451,44380.050.2
20172017-01-012017-12-3021,5746,4911,744100.050.05
20162016-01-012016-12-1816,6435,2261,53390.050.05
20152015-01-012015-12-2611,0203,7471,20670.050.4
20142014-01-012014-12-276,2882,16267850.050.05
2009–20132009-01-012013-12-3012,5484,31488660.050.05
All time2002-01-012026-11-15186,62155,9437,993160.250.1

tiers.csv

One row per character per range, ranked characters first in tier order, then characters with too few games.

range
y2026 ... y2014 for one calendar year, y2009_13 for 2009 to 2013 together, all for every year together (all time)
since, through
first and last game date in the range
tier
S, A, B or C; empty if the character is not ranked in the range
in_equilibrium_share
share of the resamples in which the character is in the maximum-entropy Nash equilibrium
equilibrium_pick_rate
probability the equilibrium picks the character, averaged over resamples
win_rate_vs_equilibrium
game win rate against an opponent who picks by the equilibrium (bagged matrix)
games, players
games with the character in the range; players with at least 20 games on it

matchups.csv

One row per ordered pair of ranked characters per range.

character, opponent
both orders are listed, so each matchup appears twice
win_rate
probability the character wins a game against the opponent between equally skilled players (mean of the resampled matrices)
win_rate_lo90, win_rate_hi90
5th and 95th percentiles over the resamples
games
games between the two characters in the range
raw_wins
games the character won, unadjusted for player skill

stages.csv

One row per ordered pair of characters ranked since 2022 per legal stage, fitted on all games since 2022 with a recorded stage.

character, opponent
both orders are listed
stage
Battlefield, Yoshi's Story, Dream Land, Final Destination, Fountain of Dreams, Pokémon Stadium (Pokémon Stadium includes Frozen Pokémon Stadium)
win_rate
probability the character wins a game on the stage between equally skilled players (logit mean of the resamples)
win_rate_lo90, win_rate_hi90
5th and 95th percentiles over the resamples
games, raw_wins
games between the two characters on the stage, and the character's wins, unadjusted

Method

Data. Singles games from offline Melee tournament brackets on Liquipedia where both players' characters are recorded: the first are from 2002, and single periods start in 2009. All time fits every game in one model. Events Liquipedia lists as online are left out, so 2020 and 2021 cover only the few offline events held then. 2009 to 2013 are fitted together because each of those years has under 3,500 games. Each range is fitted on its own games only.

Model. The log-odds that player i on character c beats player j on character d is (skill of i) minus (skill of j) plus M[c, d]. A player's skill is the sum of an overall term, a term for the year and a term for the character they are on, plus a term for playing a character other than their main. The matchup term is M[c, d] = t[c] - t[d] + E[c, d], with E antisymmetric: t is overall character strength and E the part of a matchup not explained by it. All terms have normal priors centered on zero, and the fit is the posterior mode (L-BFGS-B). Because skill is fitted per player and per character, a character's numbers come from how the same players do across characters and opponents, not from who plays it.

Priors. The prior widths sigma_t (character strength) and sigma_e (matchup terms) are chosen by log loss on a held-out 10% of sets, using the one-standard-error rule: of the settings within one paired standard error of the best, the most shrunk one. Small samples are pulled towards even rather than over-fitted. The chosen values are in the table above.

Resampling. The fit is repeated on 200 resamples of the sets (whole sets, so the games of one set stay together). Matchup win rates are the mean of the resampled matrices on the logit scale; intervals are the 5th and 95th percentiles.

Tiers. For each resample, the matrix defines a two-player game of picking a character blind. Its maximum-entropy Nash equilibrium is the most spread-out mix of characters that no pick can beat on average. A character is S if it is in the equilibrium in at least 50% of resamples, A if in at least 10%, B if its median win rate against the equilibrium over the resamples is at least 43.8% (logit -0.25), and C otherwise. The matchup matrix and the win rate against the equilibrium shown on the pages use the mean of the resampled matrices. Characters need 500 games and 10 players with at least 20 games on them to be ranked. The tier list page orders characters by tier, then by share of resamples in the equilibrium.

Stages. Stage win rates come from the same model with three more terms, fitted on the 83,758 games since 2022 (of 89,609) whose stage is recorded and is one of the six legal stages; Frozen Pokémon Stadium is counted as Pokémon Stadium. The terms are how much better a character does on a stage than on its average stage, the part of a matchup specific to a stage, and one shared term for the player who chose the stage (the loser of the previous game; zero on game 1, where stages are struck). Their prior widths (0.05 and 0.05) are chosen on held-out sets with the one-standard-error rule, as above; the stage terms improve held-out log loss by 0.0014 per game (standard error 0.0003). The fitted term for the player who chose the stage is +0.027 logit; the set plans below leave it out, since it is the same on every stage. Intervals come from 200 resamples of the sets.

Strikes, bans and counterpicks. Each matchup page solves the whole set as a two-player game by backward induction, under the Platfight Melee ruleset (July 2026): game 1 on the last of the five starters (Battlefield, Yoshi's Story, Dream Land, Final Destination, Fountain of Dreams) left after 1-2-2-1 striking; later games on a stage picked by the loser of the previous game from all six (adding Pokémon Stadium), never one they already won on in the set; in best of 3 only, the winner of the previous game first bans one stage. Each player maximises their chance of winning the set, not just the next game, and both keep their characters for the whole set. The same plan is recomputed on each resample, and the page reports how often each choice comes out the same.

Limits. Character choice is not random: players counterpick on later games and pick characters they are good with. The per-character skill terms absorb the second. As a check on the first, the same fit on game 1s only (before any counterpick) puts the same three characters in S tier all time: Peach, Jigglypuff and Ice Climbers; lower tiers move more, so read B and C as rough. Single years have far fewer games, so most single-year matchups sit near 50%.

Cite

NEO! VOD, "Melee Tournament Tier List", Nov 15, 2026, https://neovod.com/tierlist/. Data from Liquipedia (CC BY-SA 3.0).