The Riftline score

What the score is

A performance score answers a question the scoreboard cannot: who played well, independently of who won. KDA cannot answer it, because a support who never dies and never kills has the same KDA as one who does nothing, and a jungler who takes every objective and dies doing it looks worse than one who farmed camps all game. The Riftline score measures seven things a player does, places each against players in the same role, and combines them with weights that are published on this page and never change without an audit.

It is a score for one game, not a rating of a player. Ten games of it say something about a player; one game says something about that game. A profile averages it over a role, and only once 10 scored games in that role stand behind the average, because below that one game moves the average by a full point.

The seven components

Each component becomes a percentile: where this game's value sits among every game Riftline holds in the same queue and the same role. Percentiles rather than standard scores, because damage and gold have long tails and one 40-minute stomp should not stretch the scale for everyone. Role-relative, so a support is never judged on farm and a top laner is never judged on vision, and so the ten scores in a lobby share one scale, which is what makes a placement mean anything.

Kill participation (Share of the team's kills you took part in)

Kills and assists as a share of the team’s kills. It says whether a player was where the fights were, which is why it is the biggest weight for a support and a jungler.

Damage share (Share of the team's damage to champions)

The player’s share of the team’s damage to champions. A share, so it is fair across game lengths, and unfair to a player whose team did nothing, which the next component corrects for.

Damage per gold (Damage to champions per 1,000 gold earned)

Damage to champions per 1,000 gold earned. The one number here that is the player’s own rather than a share of a team total: it asks what the gold became.

Economy (Gold earned per minute)

Gold per minute: farm, kills and objectives, all in one rate.

Survival (Share of the game spent alive)

The share of the game spent alive. Deaths are costed by time, not counted, so a death at minute 3 and a death at minute 35 weigh what they cost the team.

Objectives (Towers, plates and epic monsters taken)

Towers, plates and epic monsters the player took part in.

Vision (Vision score per minute)

Vision score per minute, which is the component that lets a support score a 9.

The weights, by role

The seven percentiles are combined with these weights and the result is put on a scale of 0 to 10. The weights are set by hand, per role, and published here. A fit to wins would choose different ones, and the audit below shows what it would choose and why it is not adopted.

Reading across a row: kill participation is a support's and a jungler's biggest weight, because being at the fights is most of their job; vision is a support's second; economy and survival are what a top laner is asked for. Reading down a column, a role's weights add up to one, so every role averages the same score across the corpus, which is the check that a support and a mid laner are measured on the same scale. This is version 2 of the weights.

Version 2 added damage per gold. Kill participation and damage share are both shares of a team total, so a player on a team that did little looked good on them for doing a little more. Damage per gold is the player's own. It took its weight from damage share, the component it overlaps, and on the same games it raised the score's AUC from 0.720 to 0.731 and every role's AUC with it. Spreading the weight across four components instead did worse, so that table was not shipped.

When there is no score

A lobby without ten players in lane roles gets no score: ARAM and Arena have no lanes, so a role-relative measure has nothing to measure against. A remake gets none. A queue and role with fewer than 200 games in the corpus gets none either, because a percentile against 40 games describes those 40 games, not the player. In each case the match row shows a dash and says why, never a guess. A score is computed the moment a game is stored, so a profile's newest games carry one at once.

Badges

Badges name what a score cannot: the tank who took a third of the team's damage and the enchanter who shielded thousands both look ordinary by KDA, which is exactly what the usual badges miss. Each badge is a rule, and the thresholds were swept against the corpus rather than picked.

How well it tracks wins

A score that measures play should tend to be higher on the winning team without being a restatement of the result. Every night the score is graded on every ranked solo game Riftline holds, per role, and the grade is published here whether it is flattering or not.

On 3.8k games and 38k scored players, winners averaged 5.68 and losers 4.28. A random winner outscored a random loser 72% of the time: that is the AUC, where 50% is a coin and 100% would mean the score simply reads the result. The lobby's top scorer was on the winning team 87.9% of the time and its bottom scorer on the losing team 83.9% of the time. The gap between those and 100% is the point: a great game on the losing side exists, and the score is meant to find it.

Each card shows the win rate by score tenth: a score that means anything climbs from left to right, and every role's does. The second table sets the published weights beside the weights a logistic fit of winning on the seven percentiles would choose.

The fitted weights are not adopted, and the reason is the most important sentence on this page. Winners take objectives, earn gold and stay alive partly because they are already winning, so a fit to wins rewards being on the winning team: it puts kill participation, damage share and vision at nothing and objectives, economy and survival at nearly everything. That would make the score a second copy of the result. The fit is a check on the hand-set weights, not a replacement for them.

What it is not

It is not a rating, an MMR or a rank, and Riftline does not turn it into one: Riot's rank is the only ranking here, and it is shown beside the score, never combined with it. It cannot see what a player's team asked of them, so a mid laner who roamed all game to save a losing bot lane scores lower than the numbers deserve. It knows nothing about the enemy's strength: a 9 against a weak lobby and a 9 against a strong one read the same. And it is one game. Read a profile's average, with its sample, before reading any single score.