Ratings info

How Glicko-2 ratings on this site are calculated, and how well they predict match results.

How a rating is calculated

Every player starts at 1500 with a wide uncertainty band. After each round, the rating moves toward the result: beating a higher-rated opponent gains more than beating a lower-rated one, and losing to someone lower costs more than losing to someone higher. How far it moves depends on how uncertain the system still is about you (and your opponent), so new players swing a lot and settle after about 16 games. A year away widens the band again.

That is Glicko-2, one rating period per round, kept separately for Juniors, Seniors, and Masters. Players are ranked once their rating deviation (RD) is 72 or lower.

1500 RD 100

First event

Unseen players start at 1500. A wide band means early results move the number a lot.

~16 games

Established

Around 16 rated games, RD drops to 72. The player is ranked and peaks start counting.

1 yr idle

Time off

A year away pushes an RD-60 player back toward provisional. The next event is treated as uncertain again.

— Log loss, all matches

— Log loss, both established

— Calibration slope, all

— Calibration slope, established

— Brier score, established

0.6931 Coin-flip log loss

Better than a coin flip

Distance from 0.6931. Established pairs in gold, all matches in green.

Match volume

Established pairs (both RD ≤ 72) versus every rated match.

Actual vs predicted score

Each match is scored with the expected score stored before the round, so this is out of sample. The dashed diagonal is perfectly calibrated. Marker size scales with how many matches fell in that bucket.

5-point buckets with at least 30 matches. Hover a point for the bucket.

Calibration residual

Actual minus predicted for established pairs. Bars above zero mean favorites underperformed the model (or underdogs overperformed). Gold is established; blue is all matches.

Favorite win rate by rating gap

Established pairs. The model’s expected favorite win rate versus what actually happened.

Gap	Matches	Won	Predicted	Error

By division

Log loss (lower is better) and calibration slope (1.00 is a perfect diagonal).

Division	All matches	Log loss	Slope	Established	Log loss	Slope

Juniors sit on the diagonal. Masters and Seniors are still slightly overconfident.