How Ratings Work

The basics: Glicko-2

Ratings run on Glicko-2 (Mark Glickman, 2013) — the same system lichess uses for its ladders. Every player has three numbers, not one: a rating (where the system thinks your skill is, starting at 1500), a rating deviation or RD(how confident it is in that rating — high when you're new or inactive, low once you've played a lot), and a volatility (how erratic your results have been, which lets the system react faster to a player whose skill is genuinely swinging). Only rating and RD show up on the site; volatility is internal bookkeeping.

A big rating gap between two players translates into a lopsided expected score, and a low-RD (confident) opponent counts for more evidence than a high-RD (uncertain) one — beating an established player moves you more than beating a first-timer at the same rating.

A tournament is one rating period

Glicko-2 updates in batches called rating periods — here, one tournament is one period. Standings are converted into match points(3 per win, 1 per tie, the same scoring Swiss tournaments run on) and grouped into bands of equal record — everyone on a 3-1 record is the same band, not ranked against each other by seed or tiebreaker. Every player is then scored one implied game against every other player in the field: a win against a lower band, a loss against a higher one, a draw against their own band. That whole batch of implied games is fed into the standard Glicko-2 update for each player, all computed against everyone's rating going into the event — nobody's mid-event results affect anyone else's update for that same event.

Evidence weighting: a Swiss event is its round count

An N-player field implies N−1 comparisons per player, but a Swiss tournament only actually plays each person about rounds games. Treating every implied comparison as full-strength evidence would let field size alone — not skill demonstrated — dominate the math, so every implied game is weighted by

w = min(1, rounds ÷ (N − 1))

using the tournament's recorded round count, or ⌈log₂ N⌉when it isn't recorded. A Swiss event is worth the games it actually played, spread evenly across the field comparison — big fields still move ratings more than small ones (there's simply more going on), just not in proportion to N−1.

Ranked board: gated by confidence, not attendance

The leaderboard shows two sections. Established players — those whose rating deviation has settled to RD ≤ 100 — get a rank number and a public rating. Everyone else is Unranked (shown as Pendingon their own profile): the system simply doesn't have enough evidence yet to stand behind a number for them. This is a direct readout of the model's own confidence, not a fixed events-played rule — a player who jumps into a huge, well-rounded event can settle in faster than one who plays a string of small pods, because RD tracks evidence, not attendance.

Full recompute, every time

Ratings are never patched incrementally. Every submission triggers a full chronological recompute of the entire history — one rating period per tournament, oldest first — so a late-reported tournament from last month slots into its correct place in time and every rating downstream of it updates. This also means corrections are painless: fix the data, and the ratings fix themselves.

Tiers

Tier bands are fit to the Established population's actual rating spread (see the leaderboard for current thresholds and the live distribution). Players above the RD ≤ 100 gate show as Pending regardless of rating.