The problem Glicko addresses
A basic rating gives each player one number and treats it as equally reliable for everyone. That is awkward in practice. A player with 500 recent games has a rating that probably sits close to their true level. A player with five games, or one who has not played for a year, has a number that could be well off. Plain Elo ratings treat them the same, using one update size for both.
Mark Glickman designed the Glicko system in the 1990s to fix this by attaching a measure of confidence to each rating. A later version, Glicko-2, added a third quantity. The result is a rating that comes with a margin of error.
Rating and rating deviation
In Glicko, each player has two main numbers.
- The rating, the best current estimate of skill, on a scale similar to Elo’s.
- The rating deviation (RD), a measure of how uncertain that estimate is. Large means the system is unsure; small means it is fairly confident.
New players commonly start with a high RD, often 350 in the standard description, and a default rating around 1500. As a rough guide, you can think of the true skill as likely lying within about two RDs either side of the rating. For a 1500 rating with an RD of 350, that is a wide range, roughly 800 to 2200. After many games, an RD of 50 gives a range of roughly 1400 to 1600. The range is a rule of thumb, not a guarantee, but it conveys the idea: the system knows what it does not know.
How uncertainty changes the update
The central benefit is that the size of an update depends on the RD of the player and of the opponent.
- A player with a high RD moves a lot. Their rating is a weak estimate, so each result is treated as strong evidence and shifts it substantially. This lets new or returning players find their level in far fewer games than a fixed small K would allow.
- A player with a low RD moves a little. Their rating is already well supported, so a single result adds little.
- Opponent uncertainty matters too. Beating an opponent whose own rating is shaky counts for less than beating one whose rating is well established, because a shaky rating is poor evidence of how strong they really are.
Compare this with Elo, where a single K-factor is a compromise. Elo with a large K is responsive but noisy; with a small K it is stable but slow. Glicko effectively lets K shrink on its own as confidence grows, and grow again when confidence is lost.
Uncertainty grows when players are away
RD does not only fall. If a player does not play for a rating period, their RD increases, reflecting the idea that skill may have changed while they were away. Come back after a long break and the system is less sure of your number, so your first games can move it quickly. In a similar way, the rating of a quiet player becomes less certain even though their actual number has not changed.
This is a more honest handling of inactivity than leaving an old rating frozen. Whether a returning player is better or worse than before, a high RD lets the system correct itself quickly in either direction.
Glicko-2 and volatility
Glicko-2 adds a volatility value, which describes how erratic a player’s results are. A player whose performance swings widely gets a higher volatility, which keeps their RD from shrinking as fast. Someone who performs consistently has lower volatility, and their rating settles faster.
It is worth separating this from the day-to-day sense of rating volatility, meaning swings in a displayed rank. Glicko-2 volatility is a modelled property of a player, not a description of recent ups and downs, though the two are related.
What the confidence value is used for
RD is useful beyond updates.
- Matchmaking. A system can place players with high RD more cautiously or move them through levels quickly, so they are not stuck in mismatched games.
- Display. Some ladders show a conservative rating, such as the rating minus a multiple of the RD, so that a player must demonstrate skill to rank highly. This prevents a lucky newcomer from leaping to the top of a leaderboard on a few games. It also means two players with the same estimated skill can appear at different positions depending on how certain the system is.
- Placement. The process of using a short run of games to find a starting level is essentially a period of high uncertainty being reduced quickly, as in placement matches.
- Detecting unusual accounts. An account that wins far more than its rating predicts can be flagged, one of the ways systems respond to smurf accounts.
Limits of the approach
Glicko is not a cure-all.
- It still assumes a single skill number. Matchups, roles and form are not captured.
- It depends on rating periods. The original version works on batches of games within a period, which makes the details differ from continuous per-game updates, though newer variants adapt it.
- The confidence is model-based. An RD is only as reliable as the assumptions behind it. It is a guide to how much to trust a number, not a guarantee.
- Fast movement can feel odd. Players with high RD may see large swings, which can look like instability even when the system is working as intended.
Reading uncertainty as a player
The practical takeaway is that a rating without a margin hides information. Early in a season, after a break or in a new mode, your number is a rough estimate and may move a lot. As games accumulate, it should become steadier. If you want to judge whether you have really improved, wait until the uncertainty has shrunk and the sample is large enough that a few results do not decide the story.