Trofeo· The Trofeo team
Glicko-2 and the problem of draws
Chess breaks two of Elo's assumptions: draws are everywhere, and not every rating is equally trustworthy. Glicko-2 fixes both - here's how, without the math.
Elo was invented for chess, so it feels strange to say chess is where Elo struggles. But the game changed around it. Two things in particular trip up plain Elo at the board: draws are common - a serious chess result, not a rare tie - and rating trust varies wildly between the player who's logged 300 games this year and the one who just came back after a decade away.
Glicko-2 is Elo's successor, built to handle exactly these. It's the lineage behind the major online chess sites, and it's worth understanding what it adds - because it isn't a rewrite of Elo so much as Elo with a confidence dial bolted on.
The missing number: how sure are we?
Elo gives a player one number and treats it as gospel. Glicko-2 attaches a second number to it: a measure of reliability, usually called RD (rating deviation). RD answers the question Elo never asks - how confident are we that this rating is right?
A useful way to picture it: your true rating is somewhere inside a band, and RD is the width of that band.
- A new player has a wide band. The system has barely any evidence, so it keeps an open mind - and lets the rating move quickly.
- An active regular has a narrow band. Hundreds of recent games have pinned them down, so each new result nudges the rating only slightly.
- A returning player had a narrow band, but it widens again during the time away. The system reasons, sensibly, that it no longer knows whether you've improved, rusted, or stayed put.
This is the single biggest practical difference from Elo. Two players can both be rated 1500, but if one is a fresh sign-up and the other a steady club player, a single game means very different things for each. Glicko-2 tracks that; Elo can't.
Why the band matters in practice
The width of the band controls how fast the rating reacts:
| Player | Rating | Band (RD) | What one loss does |
|---|---|---|---|
| New sign-up | 1500 | Wide | Big drop - the system updates eagerly |
| Active regular | 1500 | Narrow | Small drop - it's already confident |
| Just back after a year | 1500 | Re-widened | Big move again, until they've re-established |
So the same scoreline moves a newcomer's rating sharply and a veteran's gently. That's not inconsistency - it's the system being honest about how much it has learned about each of them. New players reach a fair rating within a handful of games instead of grinding for a season, and one bad night can't wreck a well-established rating.
Draws, done properly
In chess a draw is a real outcome with real meaning, and who you drew against is the whole story. Glicko-2 treats a draw as a result worth half a win and rates it against expectation:
- Draw a stronger player and your rating ticks up. The system expected you to lose, so holding the draw is evidence you're better than it thought.
- Draw a weaker player and your rating ticks down. You were expected to win; not winning is mild evidence you're a touch overrated.
- Draw an equal and almost nothing happens - the result matched the forecast.
And because the update is scaled by the bands, a draw between two confident veterans barely moves anything, while a newcomer drawing a master jumps more - appropriately, since it's a louder surprise.
So when do you reach for it?
Glicko-2 is the pick when your game is 1v1, draws are part of the fabric, and the gap between trustworthy and untrustworthy ratings actually matters - which is to say, chess. (Trofeo runs Glicko-2 for chess by default, with draws as first-class results and the rating band feeding how fast ratings settle.) For clean win/loss 1v1 games like darts or pool, plain Elo is simpler and perfectly good; the extra machinery earns its keep specifically where draws and rating-trust are in play.
The honest caveats
- It still takes games to converge. A wide band narrows with evidence, not with time. A rating off six games is provisional no matter how precise the number looks.
- RD is not a skill stat. A low band means "we're confident," not "this player is good." A confidently-rated 1100 is still a 1100 - the band is about certainty, not strength.
- It doesn't see preparation or form. Glicko-2 knows your results, not that your opponent studied your favorite opening all weekend. Like every rating system, it models the long run, not tonight.
The one-sentence version
Glicko-2 keeps Elo's core bet - your rating is a promise about your results - and adds two things chess demands: a confidence band that lets new and rusty ratings move fast while settled ones hold steady, and draws that count for exactly what they're worth against the player across the board.
Run your chess ladder on it and the ratings stop pretending to know more than they do - which, paradoxically, is what makes them worth trusting.